Loading...
1/*
2 * linux/mm/page_alloc.c
3 *
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
6 *
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
15 */
16
17#include <linux/stddef.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/interrupt.h>
21#include <linux/pagemap.h>
22#include <linux/jiffies.h>
23#include <linux/bootmem.h>
24#include <linux/memblock.h>
25#include <linux/compiler.h>
26#include <linux/kernel.h>
27#include <linux/kmemcheck.h>
28#include <linux/module.h>
29#include <linux/suspend.h>
30#include <linux/pagevec.h>
31#include <linux/blkdev.h>
32#include <linux/slab.h>
33#include <linux/ratelimit.h>
34#include <linux/oom.h>
35#include <linux/notifier.h>
36#include <linux/topology.h>
37#include <linux/sysctl.h>
38#include <linux/cpu.h>
39#include <linux/cpuset.h>
40#include <linux/memory_hotplug.h>
41#include <linux/nodemask.h>
42#include <linux/vmalloc.h>
43#include <linux/vmstat.h>
44#include <linux/mempolicy.h>
45#include <linux/stop_machine.h>
46#include <linux/sort.h>
47#include <linux/pfn.h>
48#include <linux/backing-dev.h>
49#include <linux/fault-inject.h>
50#include <linux/page-isolation.h>
51#include <linux/page_cgroup.h>
52#include <linux/debugobjects.h>
53#include <linux/kmemleak.h>
54#include <linux/compaction.h>
55#include <trace/events/kmem.h>
56#include <linux/ftrace_event.h>
57#include <linux/memcontrol.h>
58#include <linux/prefetch.h>
59#include <linux/mm_inline.h>
60#include <linux/migrate.h>
61#include <linux/page-debug-flags.h>
62#include <linux/hugetlb.h>
63#include <linux/sched/rt.h>
64
65#include <asm/sections.h>
66#include <asm/tlbflush.h>
67#include <asm/div64.h>
68#include "internal.h"
69
70/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71static DEFINE_MUTEX(pcp_batch_high_lock);
72
73#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74DEFINE_PER_CPU(int, numa_node);
75EXPORT_PER_CPU_SYMBOL(numa_node);
76#endif
77
78#ifdef CONFIG_HAVE_MEMORYLESS_NODES
79/*
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
84 */
85DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
86EXPORT_PER_CPU_SYMBOL(_numa_mem_);
87#endif
88
89/*
90 * Array of node states.
91 */
92nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
93 [N_POSSIBLE] = NODE_MASK_ALL,
94 [N_ONLINE] = { { [0] = 1UL } },
95#ifndef CONFIG_NUMA
96 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
97#ifdef CONFIG_HIGHMEM
98 [N_HIGH_MEMORY] = { { [0] = 1UL } },
99#endif
100#ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY] = { { [0] = 1UL } },
102#endif
103 [N_CPU] = { { [0] = 1UL } },
104#endif /* NUMA */
105};
106EXPORT_SYMBOL(node_states);
107
108/* Protect totalram_pages and zone->managed_pages */
109static DEFINE_SPINLOCK(managed_page_count_lock);
110
111unsigned long totalram_pages __read_mostly;
112unsigned long totalreserve_pages __read_mostly;
113/*
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
118 */
119unsigned long dirty_balance_reserve __read_mostly;
120
121int percpu_pagelist_fraction;
122gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
123
124#ifdef CONFIG_PM_SLEEP
125/*
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
132 */
133
134static gfp_t saved_gfp_mask;
135
136void pm_restore_gfp_mask(void)
137{
138 WARN_ON(!mutex_is_locked(&pm_mutex));
139 if (saved_gfp_mask) {
140 gfp_allowed_mask = saved_gfp_mask;
141 saved_gfp_mask = 0;
142 }
143}
144
145void pm_restrict_gfp_mask(void)
146{
147 WARN_ON(!mutex_is_locked(&pm_mutex));
148 WARN_ON(saved_gfp_mask);
149 saved_gfp_mask = gfp_allowed_mask;
150 gfp_allowed_mask &= ~GFP_IOFS;
151}
152
153bool pm_suspended_storage(void)
154{
155 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
156 return false;
157 return true;
158}
159#endif /* CONFIG_PM_SLEEP */
160
161#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162int pageblock_order __read_mostly;
163#endif
164
165static void __free_pages_ok(struct page *page, unsigned int order);
166
167/*
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
174 *
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
177 */
178int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
179#ifdef CONFIG_ZONE_DMA
180 256,
181#endif
182#ifdef CONFIG_ZONE_DMA32
183 256,
184#endif
185#ifdef CONFIG_HIGHMEM
186 32,
187#endif
188 32,
189};
190
191EXPORT_SYMBOL(totalram_pages);
192
193static char * const zone_names[MAX_NR_ZONES] = {
194#ifdef CONFIG_ZONE_DMA
195 "DMA",
196#endif
197#ifdef CONFIG_ZONE_DMA32
198 "DMA32",
199#endif
200 "Normal",
201#ifdef CONFIG_HIGHMEM
202 "HighMem",
203#endif
204 "Movable",
205};
206
207int min_free_kbytes = 1024;
208int user_min_free_kbytes = -1;
209
210static unsigned long __meminitdata nr_kernel_pages;
211static unsigned long __meminitdata nr_all_pages;
212static unsigned long __meminitdata dma_reserve;
213
214#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
216static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
217static unsigned long __initdata required_kernelcore;
218static unsigned long __initdata required_movablecore;
219static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
220
221/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
222int movable_zone;
223EXPORT_SYMBOL(movable_zone);
224#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
225
226#if MAX_NUMNODES > 1
227int nr_node_ids __read_mostly = MAX_NUMNODES;
228int nr_online_nodes __read_mostly = 1;
229EXPORT_SYMBOL(nr_node_ids);
230EXPORT_SYMBOL(nr_online_nodes);
231#endif
232
233int page_group_by_mobility_disabled __read_mostly;
234
235void set_pageblock_migratetype(struct page *page, int migratetype)
236{
237 if (unlikely(page_group_by_mobility_disabled &&
238 migratetype < MIGRATE_PCPTYPES))
239 migratetype = MIGRATE_UNMOVABLE;
240
241 set_pageblock_flags_group(page, (unsigned long)migratetype,
242 PB_migrate, PB_migrate_end);
243}
244
245bool oom_killer_disabled __read_mostly;
246
247#ifdef CONFIG_DEBUG_VM
248static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
249{
250 int ret = 0;
251 unsigned seq;
252 unsigned long pfn = page_to_pfn(page);
253 unsigned long sp, start_pfn;
254
255 do {
256 seq = zone_span_seqbegin(zone);
257 start_pfn = zone->zone_start_pfn;
258 sp = zone->spanned_pages;
259 if (!zone_spans_pfn(zone, pfn))
260 ret = 1;
261 } while (zone_span_seqretry(zone, seq));
262
263 if (ret)
264 pr_err("page %lu outside zone [ %lu - %lu ]\n",
265 pfn, start_pfn, start_pfn + sp);
266
267 return ret;
268}
269
270static int page_is_consistent(struct zone *zone, struct page *page)
271{
272 if (!pfn_valid_within(page_to_pfn(page)))
273 return 0;
274 if (zone != page_zone(page))
275 return 0;
276
277 return 1;
278}
279/*
280 * Temporary debugging check for pages not lying within a given zone.
281 */
282static int bad_range(struct zone *zone, struct page *page)
283{
284 if (page_outside_zone_boundaries(zone, page))
285 return 1;
286 if (!page_is_consistent(zone, page))
287 return 1;
288
289 return 0;
290}
291#else
292static inline int bad_range(struct zone *zone, struct page *page)
293{
294 return 0;
295}
296#endif
297
298static void bad_page(struct page *page, const char *reason,
299 unsigned long bad_flags)
300{
301 static unsigned long resume;
302 static unsigned long nr_shown;
303 static unsigned long nr_unshown;
304
305 /* Don't complain about poisoned pages */
306 if (PageHWPoison(page)) {
307 page_mapcount_reset(page); /* remove PageBuddy */
308 return;
309 }
310
311 /*
312 * Allow a burst of 60 reports, then keep quiet for that minute;
313 * or allow a steady drip of one report per second.
314 */
315 if (nr_shown == 60) {
316 if (time_before(jiffies, resume)) {
317 nr_unshown++;
318 goto out;
319 }
320 if (nr_unshown) {
321 printk(KERN_ALERT
322 "BUG: Bad page state: %lu messages suppressed\n",
323 nr_unshown);
324 nr_unshown = 0;
325 }
326 nr_shown = 0;
327 }
328 if (nr_shown++ == 0)
329 resume = jiffies + 60 * HZ;
330
331 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
332 current->comm, page_to_pfn(page));
333 dump_page_badflags(page, reason, bad_flags);
334
335 print_modules();
336 dump_stack();
337out:
338 /* Leave bad fields for debug, except PageBuddy could make trouble */
339 page_mapcount_reset(page); /* remove PageBuddy */
340 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
341}
342
343/*
344 * Higher-order pages are called "compound pages". They are structured thusly:
345 *
346 * The first PAGE_SIZE page is called the "head page".
347 *
348 * The remaining PAGE_SIZE pages are called "tail pages".
349 *
350 * All pages have PG_compound set. All tail pages have their ->first_page
351 * pointing at the head page.
352 *
353 * The first tail page's ->lru.next holds the address of the compound page's
354 * put_page() function. Its ->lru.prev holds the order of allocation.
355 * This usage means that zero-order pages may not be compound.
356 */
357
358static void free_compound_page(struct page *page)
359{
360 __free_pages_ok(page, compound_order(page));
361}
362
363void prep_compound_page(struct page *page, unsigned long order)
364{
365 int i;
366 int nr_pages = 1 << order;
367
368 set_compound_page_dtor(page, free_compound_page);
369 set_compound_order(page, order);
370 __SetPageHead(page);
371 for (i = 1; i < nr_pages; i++) {
372 struct page *p = page + i;
373 set_page_count(p, 0);
374 p->first_page = page;
375 /* Make sure p->first_page is always valid for PageTail() */
376 smp_wmb();
377 __SetPageTail(p);
378 }
379}
380
381/* update __split_huge_page_refcount if you change this function */
382static int destroy_compound_page(struct page *page, unsigned long order)
383{
384 int i;
385 int nr_pages = 1 << order;
386 int bad = 0;
387
388 if (unlikely(compound_order(page) != order)) {
389 bad_page(page, "wrong compound order", 0);
390 bad++;
391 }
392
393 __ClearPageHead(page);
394
395 for (i = 1; i < nr_pages; i++) {
396 struct page *p = page + i;
397
398 if (unlikely(!PageTail(p))) {
399 bad_page(page, "PageTail not set", 0);
400 bad++;
401 } else if (unlikely(p->first_page != page)) {
402 bad_page(page, "first_page not consistent", 0);
403 bad++;
404 }
405 __ClearPageTail(p);
406 }
407
408 return bad;
409}
410
411static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
412{
413 int i;
414
415 /*
416 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
417 * and __GFP_HIGHMEM from hard or soft interrupt context.
418 */
419 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
420 for (i = 0; i < (1 << order); i++)
421 clear_highpage(page + i);
422}
423
424#ifdef CONFIG_DEBUG_PAGEALLOC
425unsigned int _debug_guardpage_minorder;
426
427static int __init debug_guardpage_minorder_setup(char *buf)
428{
429 unsigned long res;
430
431 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
432 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
433 return 0;
434 }
435 _debug_guardpage_minorder = res;
436 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
437 return 0;
438}
439__setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
440
441static inline void set_page_guard_flag(struct page *page)
442{
443 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
444}
445
446static inline void clear_page_guard_flag(struct page *page)
447{
448 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
449}
450#else
451static inline void set_page_guard_flag(struct page *page) { }
452static inline void clear_page_guard_flag(struct page *page) { }
453#endif
454
455static inline void set_page_order(struct page *page, int order)
456{
457 set_page_private(page, order);
458 __SetPageBuddy(page);
459}
460
461static inline void rmv_page_order(struct page *page)
462{
463 __ClearPageBuddy(page);
464 set_page_private(page, 0);
465}
466
467/*
468 * Locate the struct page for both the matching buddy in our
469 * pair (buddy1) and the combined O(n+1) page they form (page).
470 *
471 * 1) Any buddy B1 will have an order O twin B2 which satisfies
472 * the following equation:
473 * B2 = B1 ^ (1 << O)
474 * For example, if the starting buddy (buddy2) is #8 its order
475 * 1 buddy is #10:
476 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
477 *
478 * 2) Any buddy B will have an order O+1 parent P which
479 * satisfies the following equation:
480 * P = B & ~(1 << O)
481 *
482 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
483 */
484static inline unsigned long
485__find_buddy_index(unsigned long page_idx, unsigned int order)
486{
487 return page_idx ^ (1 << order);
488}
489
490/*
491 * This function checks whether a page is free && is the buddy
492 * we can do coalesce a page and its buddy if
493 * (a) the buddy is not in a hole &&
494 * (b) the buddy is in the buddy system &&
495 * (c) a page and its buddy have the same order &&
496 * (d) a page and its buddy are in the same zone.
497 *
498 * For recording whether a page is in the buddy system, we set ->_mapcount
499 * PAGE_BUDDY_MAPCOUNT_VALUE.
500 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
501 * serialized by zone->lock.
502 *
503 * For recording page's order, we use page_private(page).
504 */
505static inline int page_is_buddy(struct page *page, struct page *buddy,
506 int order)
507{
508 if (!pfn_valid_within(page_to_pfn(buddy)))
509 return 0;
510
511 if (page_zone_id(page) != page_zone_id(buddy))
512 return 0;
513
514 if (page_is_guard(buddy) && page_order(buddy) == order) {
515 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
516 return 1;
517 }
518
519 if (PageBuddy(buddy) && page_order(buddy) == order) {
520 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
521 return 1;
522 }
523 return 0;
524}
525
526/*
527 * Freeing function for a buddy system allocator.
528 *
529 * The concept of a buddy system is to maintain direct-mapped table
530 * (containing bit values) for memory blocks of various "orders".
531 * The bottom level table contains the map for the smallest allocatable
532 * units of memory (here, pages), and each level above it describes
533 * pairs of units from the levels below, hence, "buddies".
534 * At a high level, all that happens here is marking the table entry
535 * at the bottom level available, and propagating the changes upward
536 * as necessary, plus some accounting needed to play nicely with other
537 * parts of the VM system.
538 * At each level, we keep a list of pages, which are heads of continuous
539 * free pages of length of (1 << order) and marked with _mapcount
540 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
541 * field.
542 * So when we are allocating or freeing one, we can derive the state of the
543 * other. That is, if we allocate a small block, and both were
544 * free, the remainder of the region must be split into blocks.
545 * If a block is freed, and its buddy is also free, then this
546 * triggers coalescing into a block of larger size.
547 *
548 * -- nyc
549 */
550
551static inline void __free_one_page(struct page *page,
552 struct zone *zone, unsigned int order,
553 int migratetype)
554{
555 unsigned long page_idx;
556 unsigned long combined_idx;
557 unsigned long uninitialized_var(buddy_idx);
558 struct page *buddy;
559
560 VM_BUG_ON(!zone_is_initialized(zone));
561
562 if (unlikely(PageCompound(page)))
563 if (unlikely(destroy_compound_page(page, order)))
564 return;
565
566 VM_BUG_ON(migratetype == -1);
567
568 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
569
570 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
571 VM_BUG_ON_PAGE(bad_range(zone, page), page);
572
573 while (order < MAX_ORDER-1) {
574 buddy_idx = __find_buddy_index(page_idx, order);
575 buddy = page + (buddy_idx - page_idx);
576 if (!page_is_buddy(page, buddy, order))
577 break;
578 /*
579 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
580 * merge with it and move up one order.
581 */
582 if (page_is_guard(buddy)) {
583 clear_page_guard_flag(buddy);
584 set_page_private(page, 0);
585 __mod_zone_freepage_state(zone, 1 << order,
586 migratetype);
587 } else {
588 list_del(&buddy->lru);
589 zone->free_area[order].nr_free--;
590 rmv_page_order(buddy);
591 }
592 combined_idx = buddy_idx & page_idx;
593 page = page + (combined_idx - page_idx);
594 page_idx = combined_idx;
595 order++;
596 }
597 set_page_order(page, order);
598
599 /*
600 * If this is not the largest possible page, check if the buddy
601 * of the next-highest order is free. If it is, it's possible
602 * that pages are being freed that will coalesce soon. In case,
603 * that is happening, add the free page to the tail of the list
604 * so it's less likely to be used soon and more likely to be merged
605 * as a higher order page
606 */
607 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
608 struct page *higher_page, *higher_buddy;
609 combined_idx = buddy_idx & page_idx;
610 higher_page = page + (combined_idx - page_idx);
611 buddy_idx = __find_buddy_index(combined_idx, order + 1);
612 higher_buddy = higher_page + (buddy_idx - combined_idx);
613 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
614 list_add_tail(&page->lru,
615 &zone->free_area[order].free_list[migratetype]);
616 goto out;
617 }
618 }
619
620 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
621out:
622 zone->free_area[order].nr_free++;
623}
624
625static inline int free_pages_check(struct page *page)
626{
627 const char *bad_reason = NULL;
628 unsigned long bad_flags = 0;
629
630 if (unlikely(page_mapcount(page)))
631 bad_reason = "nonzero mapcount";
632 if (unlikely(page->mapping != NULL))
633 bad_reason = "non-NULL mapping";
634 if (unlikely(atomic_read(&page->_count) != 0))
635 bad_reason = "nonzero _count";
636 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
637 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
638 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
639 }
640 if (unlikely(mem_cgroup_bad_page_check(page)))
641 bad_reason = "cgroup check failed";
642 if (unlikely(bad_reason)) {
643 bad_page(page, bad_reason, bad_flags);
644 return 1;
645 }
646 page_cpupid_reset_last(page);
647 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
648 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
649 return 0;
650}
651
652/*
653 * Frees a number of pages from the PCP lists
654 * Assumes all pages on list are in same zone, and of same order.
655 * count is the number of pages to free.
656 *
657 * If the zone was previously in an "all pages pinned" state then look to
658 * see if this freeing clears that state.
659 *
660 * And clear the zone's pages_scanned counter, to hold off the "all pages are
661 * pinned" detection logic.
662 */
663static void free_pcppages_bulk(struct zone *zone, int count,
664 struct per_cpu_pages *pcp)
665{
666 int migratetype = 0;
667 int batch_free = 0;
668 int to_free = count;
669
670 spin_lock(&zone->lock);
671 zone->pages_scanned = 0;
672
673 while (to_free) {
674 struct page *page;
675 struct list_head *list;
676
677 /*
678 * Remove pages from lists in a round-robin fashion. A
679 * batch_free count is maintained that is incremented when an
680 * empty list is encountered. This is so more pages are freed
681 * off fuller lists instead of spinning excessively around empty
682 * lists
683 */
684 do {
685 batch_free++;
686 if (++migratetype == MIGRATE_PCPTYPES)
687 migratetype = 0;
688 list = &pcp->lists[migratetype];
689 } while (list_empty(list));
690
691 /* This is the only non-empty list. Free them all. */
692 if (batch_free == MIGRATE_PCPTYPES)
693 batch_free = to_free;
694
695 do {
696 int mt; /* migratetype of the to-be-freed page */
697
698 page = list_entry(list->prev, struct page, lru);
699 /* must delete as __free_one_page list manipulates */
700 list_del(&page->lru);
701 mt = get_freepage_migratetype(page);
702 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
703 __free_one_page(page, zone, 0, mt);
704 trace_mm_page_pcpu_drain(page, 0, mt);
705 if (likely(!is_migrate_isolate_page(page))) {
706 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
707 if (is_migrate_cma(mt))
708 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
709 }
710 } while (--to_free && --batch_free && !list_empty(list));
711 }
712 spin_unlock(&zone->lock);
713}
714
715static void free_one_page(struct zone *zone, struct page *page, int order,
716 int migratetype)
717{
718 spin_lock(&zone->lock);
719 zone->pages_scanned = 0;
720
721 __free_one_page(page, zone, order, migratetype);
722 if (unlikely(!is_migrate_isolate(migratetype)))
723 __mod_zone_freepage_state(zone, 1 << order, migratetype);
724 spin_unlock(&zone->lock);
725}
726
727static bool free_pages_prepare(struct page *page, unsigned int order)
728{
729 int i;
730 int bad = 0;
731
732 trace_mm_page_free(page, order);
733 kmemcheck_free_shadow(page, order);
734
735 if (PageAnon(page))
736 page->mapping = NULL;
737 for (i = 0; i < (1 << order); i++)
738 bad += free_pages_check(page + i);
739 if (bad)
740 return false;
741
742 if (!PageHighMem(page)) {
743 debug_check_no_locks_freed(page_address(page),
744 PAGE_SIZE << order);
745 debug_check_no_obj_freed(page_address(page),
746 PAGE_SIZE << order);
747 }
748 arch_free_page(page, order);
749 kernel_map_pages(page, 1 << order, 0);
750
751 return true;
752}
753
754static void __free_pages_ok(struct page *page, unsigned int order)
755{
756 unsigned long flags;
757 int migratetype;
758
759 if (!free_pages_prepare(page, order))
760 return;
761
762 local_irq_save(flags);
763 __count_vm_events(PGFREE, 1 << order);
764 migratetype = get_pageblock_migratetype(page);
765 set_freepage_migratetype(page, migratetype);
766 free_one_page(page_zone(page), page, order, migratetype);
767 local_irq_restore(flags);
768}
769
770void __init __free_pages_bootmem(struct page *page, unsigned int order)
771{
772 unsigned int nr_pages = 1 << order;
773 struct page *p = page;
774 unsigned int loop;
775
776 prefetchw(p);
777 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
778 prefetchw(p + 1);
779 __ClearPageReserved(p);
780 set_page_count(p, 0);
781 }
782 __ClearPageReserved(p);
783 set_page_count(p, 0);
784
785 page_zone(page)->managed_pages += nr_pages;
786 set_page_refcounted(page);
787 __free_pages(page, order);
788}
789
790#ifdef CONFIG_CMA
791/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
792void __init init_cma_reserved_pageblock(struct page *page)
793{
794 unsigned i = pageblock_nr_pages;
795 struct page *p = page;
796
797 do {
798 __ClearPageReserved(p);
799 set_page_count(p, 0);
800 } while (++p, --i);
801
802 set_page_refcounted(page);
803 set_pageblock_migratetype(page, MIGRATE_CMA);
804 __free_pages(page, pageblock_order);
805 adjust_managed_page_count(page, pageblock_nr_pages);
806}
807#endif
808
809/*
810 * The order of subdivision here is critical for the IO subsystem.
811 * Please do not alter this order without good reasons and regression
812 * testing. Specifically, as large blocks of memory are subdivided,
813 * the order in which smaller blocks are delivered depends on the order
814 * they're subdivided in this function. This is the primary factor
815 * influencing the order in which pages are delivered to the IO
816 * subsystem according to empirical testing, and this is also justified
817 * by considering the behavior of a buddy system containing a single
818 * large block of memory acted on by a series of small allocations.
819 * This behavior is a critical factor in sglist merging's success.
820 *
821 * -- nyc
822 */
823static inline void expand(struct zone *zone, struct page *page,
824 int low, int high, struct free_area *area,
825 int migratetype)
826{
827 unsigned long size = 1 << high;
828
829 while (high > low) {
830 area--;
831 high--;
832 size >>= 1;
833 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
834
835#ifdef CONFIG_DEBUG_PAGEALLOC
836 if (high < debug_guardpage_minorder()) {
837 /*
838 * Mark as guard pages (or page), that will allow to
839 * merge back to allocator when buddy will be freed.
840 * Corresponding page table entries will not be touched,
841 * pages will stay not present in virtual address space
842 */
843 INIT_LIST_HEAD(&page[size].lru);
844 set_page_guard_flag(&page[size]);
845 set_page_private(&page[size], high);
846 /* Guard pages are not available for any usage */
847 __mod_zone_freepage_state(zone, -(1 << high),
848 migratetype);
849 continue;
850 }
851#endif
852 list_add(&page[size].lru, &area->free_list[migratetype]);
853 area->nr_free++;
854 set_page_order(&page[size], high);
855 }
856}
857
858/*
859 * This page is about to be returned from the page allocator
860 */
861static inline int check_new_page(struct page *page)
862{
863 const char *bad_reason = NULL;
864 unsigned long bad_flags = 0;
865
866 if (unlikely(page_mapcount(page)))
867 bad_reason = "nonzero mapcount";
868 if (unlikely(page->mapping != NULL))
869 bad_reason = "non-NULL mapping";
870 if (unlikely(atomic_read(&page->_count) != 0))
871 bad_reason = "nonzero _count";
872 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
873 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
874 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
875 }
876 if (unlikely(mem_cgroup_bad_page_check(page)))
877 bad_reason = "cgroup check failed";
878 if (unlikely(bad_reason)) {
879 bad_page(page, bad_reason, bad_flags);
880 return 1;
881 }
882 return 0;
883}
884
885static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
886{
887 int i;
888
889 for (i = 0; i < (1 << order); i++) {
890 struct page *p = page + i;
891 if (unlikely(check_new_page(p)))
892 return 1;
893 }
894
895 set_page_private(page, 0);
896 set_page_refcounted(page);
897
898 arch_alloc_page(page, order);
899 kernel_map_pages(page, 1 << order, 1);
900
901 if (gfp_flags & __GFP_ZERO)
902 prep_zero_page(page, order, gfp_flags);
903
904 if (order && (gfp_flags & __GFP_COMP))
905 prep_compound_page(page, order);
906
907 return 0;
908}
909
910/*
911 * Go through the free lists for the given migratetype and remove
912 * the smallest available page from the freelists
913 */
914static inline
915struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
916 int migratetype)
917{
918 unsigned int current_order;
919 struct free_area *area;
920 struct page *page;
921
922 /* Find a page of the appropriate size in the preferred list */
923 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
924 area = &(zone->free_area[current_order]);
925 if (list_empty(&area->free_list[migratetype]))
926 continue;
927
928 page = list_entry(area->free_list[migratetype].next,
929 struct page, lru);
930 list_del(&page->lru);
931 rmv_page_order(page);
932 area->nr_free--;
933 expand(zone, page, order, current_order, area, migratetype);
934 return page;
935 }
936
937 return NULL;
938}
939
940
941/*
942 * This array describes the order lists are fallen back to when
943 * the free lists for the desirable migrate type are depleted
944 */
945static int fallbacks[MIGRATE_TYPES][4] = {
946 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
947 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
948#ifdef CONFIG_CMA
949 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
950 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
951#else
952 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
953#endif
954 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
955#ifdef CONFIG_MEMORY_ISOLATION
956 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
957#endif
958};
959
960/*
961 * Move the free pages in a range to the free lists of the requested type.
962 * Note that start_page and end_pages are not aligned on a pageblock
963 * boundary. If alignment is required, use move_freepages_block()
964 */
965int move_freepages(struct zone *zone,
966 struct page *start_page, struct page *end_page,
967 int migratetype)
968{
969 struct page *page;
970 unsigned long order;
971 int pages_moved = 0;
972
973#ifndef CONFIG_HOLES_IN_ZONE
974 /*
975 * page_zone is not safe to call in this context when
976 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
977 * anyway as we check zone boundaries in move_freepages_block().
978 * Remove at a later date when no bug reports exist related to
979 * grouping pages by mobility
980 */
981 BUG_ON(page_zone(start_page) != page_zone(end_page));
982#endif
983
984 for (page = start_page; page <= end_page;) {
985 /* Make sure we are not inadvertently changing nodes */
986 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
987
988 if (!pfn_valid_within(page_to_pfn(page))) {
989 page++;
990 continue;
991 }
992
993 if (!PageBuddy(page)) {
994 page++;
995 continue;
996 }
997
998 order = page_order(page);
999 list_move(&page->lru,
1000 &zone->free_area[order].free_list[migratetype]);
1001 set_freepage_migratetype(page, migratetype);
1002 page += 1 << order;
1003 pages_moved += 1 << order;
1004 }
1005
1006 return pages_moved;
1007}
1008
1009int move_freepages_block(struct zone *zone, struct page *page,
1010 int migratetype)
1011{
1012 unsigned long start_pfn, end_pfn;
1013 struct page *start_page, *end_page;
1014
1015 start_pfn = page_to_pfn(page);
1016 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1017 start_page = pfn_to_page(start_pfn);
1018 end_page = start_page + pageblock_nr_pages - 1;
1019 end_pfn = start_pfn + pageblock_nr_pages - 1;
1020
1021 /* Do not cross zone boundaries */
1022 if (!zone_spans_pfn(zone, start_pfn))
1023 start_page = page;
1024 if (!zone_spans_pfn(zone, end_pfn))
1025 return 0;
1026
1027 return move_freepages(zone, start_page, end_page, migratetype);
1028}
1029
1030static void change_pageblock_range(struct page *pageblock_page,
1031 int start_order, int migratetype)
1032{
1033 int nr_pageblocks = 1 << (start_order - pageblock_order);
1034
1035 while (nr_pageblocks--) {
1036 set_pageblock_migratetype(pageblock_page, migratetype);
1037 pageblock_page += pageblock_nr_pages;
1038 }
1039}
1040
1041/*
1042 * If breaking a large block of pages, move all free pages to the preferred
1043 * allocation list. If falling back for a reclaimable kernel allocation, be
1044 * more aggressive about taking ownership of free pages.
1045 *
1046 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1047 * nor move CMA pages to different free lists. We don't want unmovable pages
1048 * to be allocated from MIGRATE_CMA areas.
1049 *
1050 * Returns the new migratetype of the pageblock (or the same old migratetype
1051 * if it was unchanged).
1052 */
1053static int try_to_steal_freepages(struct zone *zone, struct page *page,
1054 int start_type, int fallback_type)
1055{
1056 int current_order = page_order(page);
1057
1058 /*
1059 * When borrowing from MIGRATE_CMA, we need to release the excess
1060 * buddy pages to CMA itself.
1061 */
1062 if (is_migrate_cma(fallback_type))
1063 return fallback_type;
1064
1065 /* Take ownership for orders >= pageblock_order */
1066 if (current_order >= pageblock_order) {
1067 change_pageblock_range(page, current_order, start_type);
1068 return start_type;
1069 }
1070
1071 if (current_order >= pageblock_order / 2 ||
1072 start_type == MIGRATE_RECLAIMABLE ||
1073 page_group_by_mobility_disabled) {
1074 int pages;
1075
1076 pages = move_freepages_block(zone, page, start_type);
1077
1078 /* Claim the whole block if over half of it is free */
1079 if (pages >= (1 << (pageblock_order-1)) ||
1080 page_group_by_mobility_disabled) {
1081
1082 set_pageblock_migratetype(page, start_type);
1083 return start_type;
1084 }
1085
1086 }
1087
1088 return fallback_type;
1089}
1090
1091/* Remove an element from the buddy allocator from the fallback list */
1092static inline struct page *
1093__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1094{
1095 struct free_area *area;
1096 int current_order;
1097 struct page *page;
1098 int migratetype, new_type, i;
1099
1100 /* Find the largest possible block of pages in the other list */
1101 for (current_order = MAX_ORDER-1; current_order >= order;
1102 --current_order) {
1103 for (i = 0;; i++) {
1104 migratetype = fallbacks[start_migratetype][i];
1105
1106 /* MIGRATE_RESERVE handled later if necessary */
1107 if (migratetype == MIGRATE_RESERVE)
1108 break;
1109
1110 area = &(zone->free_area[current_order]);
1111 if (list_empty(&area->free_list[migratetype]))
1112 continue;
1113
1114 page = list_entry(area->free_list[migratetype].next,
1115 struct page, lru);
1116 area->nr_free--;
1117
1118 new_type = try_to_steal_freepages(zone, page,
1119 start_migratetype,
1120 migratetype);
1121
1122 /* Remove the page from the freelists */
1123 list_del(&page->lru);
1124 rmv_page_order(page);
1125
1126 expand(zone, page, order, current_order, area,
1127 new_type);
1128
1129 trace_mm_page_alloc_extfrag(page, order, current_order,
1130 start_migratetype, migratetype, new_type);
1131
1132 return page;
1133 }
1134 }
1135
1136 return NULL;
1137}
1138
1139/*
1140 * Do the hard work of removing an element from the buddy allocator.
1141 * Call me with the zone->lock already held.
1142 */
1143static struct page *__rmqueue(struct zone *zone, unsigned int order,
1144 int migratetype)
1145{
1146 struct page *page;
1147
1148retry_reserve:
1149 page = __rmqueue_smallest(zone, order, migratetype);
1150
1151 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1152 page = __rmqueue_fallback(zone, order, migratetype);
1153
1154 /*
1155 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1156 * is used because __rmqueue_smallest is an inline function
1157 * and we want just one call site
1158 */
1159 if (!page) {
1160 migratetype = MIGRATE_RESERVE;
1161 goto retry_reserve;
1162 }
1163 }
1164
1165 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1166 return page;
1167}
1168
1169/*
1170 * Obtain a specified number of elements from the buddy allocator, all under
1171 * a single hold of the lock, for efficiency. Add them to the supplied list.
1172 * Returns the number of new pages which were placed at *list.
1173 */
1174static int rmqueue_bulk(struct zone *zone, unsigned int order,
1175 unsigned long count, struct list_head *list,
1176 int migratetype, int cold)
1177{
1178 int mt = migratetype, i;
1179
1180 spin_lock(&zone->lock);
1181 for (i = 0; i < count; ++i) {
1182 struct page *page = __rmqueue(zone, order, migratetype);
1183 if (unlikely(page == NULL))
1184 break;
1185
1186 /*
1187 * Split buddy pages returned by expand() are received here
1188 * in physical page order. The page is added to the callers and
1189 * list and the list head then moves forward. From the callers
1190 * perspective, the linked list is ordered by page number in
1191 * some conditions. This is useful for IO devices that can
1192 * merge IO requests if the physical pages are ordered
1193 * properly.
1194 */
1195 if (likely(cold == 0))
1196 list_add(&page->lru, list);
1197 else
1198 list_add_tail(&page->lru, list);
1199 if (IS_ENABLED(CONFIG_CMA)) {
1200 mt = get_pageblock_migratetype(page);
1201 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1202 mt = migratetype;
1203 }
1204 set_freepage_migratetype(page, mt);
1205 list = &page->lru;
1206 if (is_migrate_cma(mt))
1207 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1208 -(1 << order));
1209 }
1210 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1211 spin_unlock(&zone->lock);
1212 return i;
1213}
1214
1215#ifdef CONFIG_NUMA
1216/*
1217 * Called from the vmstat counter updater to drain pagesets of this
1218 * currently executing processor on remote nodes after they have
1219 * expired.
1220 *
1221 * Note that this function must be called with the thread pinned to
1222 * a single processor.
1223 */
1224void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1225{
1226 unsigned long flags;
1227 int to_drain;
1228 unsigned long batch;
1229
1230 local_irq_save(flags);
1231 batch = ACCESS_ONCE(pcp->batch);
1232 if (pcp->count >= batch)
1233 to_drain = batch;
1234 else
1235 to_drain = pcp->count;
1236 if (to_drain > 0) {
1237 free_pcppages_bulk(zone, to_drain, pcp);
1238 pcp->count -= to_drain;
1239 }
1240 local_irq_restore(flags);
1241}
1242#endif
1243
1244/*
1245 * Drain pages of the indicated processor.
1246 *
1247 * The processor must either be the current processor and the
1248 * thread pinned to the current processor or a processor that
1249 * is not online.
1250 */
1251static void drain_pages(unsigned int cpu)
1252{
1253 unsigned long flags;
1254 struct zone *zone;
1255
1256 for_each_populated_zone(zone) {
1257 struct per_cpu_pageset *pset;
1258 struct per_cpu_pages *pcp;
1259
1260 local_irq_save(flags);
1261 pset = per_cpu_ptr(zone->pageset, cpu);
1262
1263 pcp = &pset->pcp;
1264 if (pcp->count) {
1265 free_pcppages_bulk(zone, pcp->count, pcp);
1266 pcp->count = 0;
1267 }
1268 local_irq_restore(flags);
1269 }
1270}
1271
1272/*
1273 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1274 */
1275void drain_local_pages(void *arg)
1276{
1277 drain_pages(smp_processor_id());
1278}
1279
1280/*
1281 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1282 *
1283 * Note that this code is protected against sending an IPI to an offline
1284 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1285 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1286 * nothing keeps CPUs from showing up after we populated the cpumask and
1287 * before the call to on_each_cpu_mask().
1288 */
1289void drain_all_pages(void)
1290{
1291 int cpu;
1292 struct per_cpu_pageset *pcp;
1293 struct zone *zone;
1294
1295 /*
1296 * Allocate in the BSS so we wont require allocation in
1297 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1298 */
1299 static cpumask_t cpus_with_pcps;
1300
1301 /*
1302 * We don't care about racing with CPU hotplug event
1303 * as offline notification will cause the notified
1304 * cpu to drain that CPU pcps and on_each_cpu_mask
1305 * disables preemption as part of its processing
1306 */
1307 for_each_online_cpu(cpu) {
1308 bool has_pcps = false;
1309 for_each_populated_zone(zone) {
1310 pcp = per_cpu_ptr(zone->pageset, cpu);
1311 if (pcp->pcp.count) {
1312 has_pcps = true;
1313 break;
1314 }
1315 }
1316 if (has_pcps)
1317 cpumask_set_cpu(cpu, &cpus_with_pcps);
1318 else
1319 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1320 }
1321 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1322}
1323
1324#ifdef CONFIG_HIBERNATION
1325
1326void mark_free_pages(struct zone *zone)
1327{
1328 unsigned long pfn, max_zone_pfn;
1329 unsigned long flags;
1330 int order, t;
1331 struct list_head *curr;
1332
1333 if (zone_is_empty(zone))
1334 return;
1335
1336 spin_lock_irqsave(&zone->lock, flags);
1337
1338 max_zone_pfn = zone_end_pfn(zone);
1339 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1340 if (pfn_valid(pfn)) {
1341 struct page *page = pfn_to_page(pfn);
1342
1343 if (!swsusp_page_is_forbidden(page))
1344 swsusp_unset_page_free(page);
1345 }
1346
1347 for_each_migratetype_order(order, t) {
1348 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1349 unsigned long i;
1350
1351 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1352 for (i = 0; i < (1UL << order); i++)
1353 swsusp_set_page_free(pfn_to_page(pfn + i));
1354 }
1355 }
1356 spin_unlock_irqrestore(&zone->lock, flags);
1357}
1358#endif /* CONFIG_PM */
1359
1360/*
1361 * Free a 0-order page
1362 * cold == 1 ? free a cold page : free a hot page
1363 */
1364void free_hot_cold_page(struct page *page, int cold)
1365{
1366 struct zone *zone = page_zone(page);
1367 struct per_cpu_pages *pcp;
1368 unsigned long flags;
1369 int migratetype;
1370
1371 if (!free_pages_prepare(page, 0))
1372 return;
1373
1374 migratetype = get_pageblock_migratetype(page);
1375 set_freepage_migratetype(page, migratetype);
1376 local_irq_save(flags);
1377 __count_vm_event(PGFREE);
1378
1379 /*
1380 * We only track unmovable, reclaimable and movable on pcp lists.
1381 * Free ISOLATE pages back to the allocator because they are being
1382 * offlined but treat RESERVE as movable pages so we can get those
1383 * areas back if necessary. Otherwise, we may have to free
1384 * excessively into the page allocator
1385 */
1386 if (migratetype >= MIGRATE_PCPTYPES) {
1387 if (unlikely(is_migrate_isolate(migratetype))) {
1388 free_one_page(zone, page, 0, migratetype);
1389 goto out;
1390 }
1391 migratetype = MIGRATE_MOVABLE;
1392 }
1393
1394 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1395 if (cold)
1396 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1397 else
1398 list_add(&page->lru, &pcp->lists[migratetype]);
1399 pcp->count++;
1400 if (pcp->count >= pcp->high) {
1401 unsigned long batch = ACCESS_ONCE(pcp->batch);
1402 free_pcppages_bulk(zone, batch, pcp);
1403 pcp->count -= batch;
1404 }
1405
1406out:
1407 local_irq_restore(flags);
1408}
1409
1410/*
1411 * Free a list of 0-order pages
1412 */
1413void free_hot_cold_page_list(struct list_head *list, int cold)
1414{
1415 struct page *page, *next;
1416
1417 list_for_each_entry_safe(page, next, list, lru) {
1418 trace_mm_page_free_batched(page, cold);
1419 free_hot_cold_page(page, cold);
1420 }
1421}
1422
1423/*
1424 * split_page takes a non-compound higher-order page, and splits it into
1425 * n (1<<order) sub-pages: page[0..n]
1426 * Each sub-page must be freed individually.
1427 *
1428 * Note: this is probably too low level an operation for use in drivers.
1429 * Please consult with lkml before using this in your driver.
1430 */
1431void split_page(struct page *page, unsigned int order)
1432{
1433 int i;
1434
1435 VM_BUG_ON_PAGE(PageCompound(page), page);
1436 VM_BUG_ON_PAGE(!page_count(page), page);
1437
1438#ifdef CONFIG_KMEMCHECK
1439 /*
1440 * Split shadow pages too, because free(page[0]) would
1441 * otherwise free the whole shadow.
1442 */
1443 if (kmemcheck_page_is_tracked(page))
1444 split_page(virt_to_page(page[0].shadow), order);
1445#endif
1446
1447 for (i = 1; i < (1 << order); i++)
1448 set_page_refcounted(page + i);
1449}
1450EXPORT_SYMBOL_GPL(split_page);
1451
1452static int __isolate_free_page(struct page *page, unsigned int order)
1453{
1454 unsigned long watermark;
1455 struct zone *zone;
1456 int mt;
1457
1458 BUG_ON(!PageBuddy(page));
1459
1460 zone = page_zone(page);
1461 mt = get_pageblock_migratetype(page);
1462
1463 if (!is_migrate_isolate(mt)) {
1464 /* Obey watermarks as if the page was being allocated */
1465 watermark = low_wmark_pages(zone) + (1 << order);
1466 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1467 return 0;
1468
1469 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1470 }
1471
1472 /* Remove page from free list */
1473 list_del(&page->lru);
1474 zone->free_area[order].nr_free--;
1475 rmv_page_order(page);
1476
1477 /* Set the pageblock if the isolated page is at least a pageblock */
1478 if (order >= pageblock_order - 1) {
1479 struct page *endpage = page + (1 << order) - 1;
1480 for (; page < endpage; page += pageblock_nr_pages) {
1481 int mt = get_pageblock_migratetype(page);
1482 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1483 set_pageblock_migratetype(page,
1484 MIGRATE_MOVABLE);
1485 }
1486 }
1487
1488 return 1UL << order;
1489}
1490
1491/*
1492 * Similar to split_page except the page is already free. As this is only
1493 * being used for migration, the migratetype of the block also changes.
1494 * As this is called with interrupts disabled, the caller is responsible
1495 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1496 * are enabled.
1497 *
1498 * Note: this is probably too low level an operation for use in drivers.
1499 * Please consult with lkml before using this in your driver.
1500 */
1501int split_free_page(struct page *page)
1502{
1503 unsigned int order;
1504 int nr_pages;
1505
1506 order = page_order(page);
1507
1508 nr_pages = __isolate_free_page(page, order);
1509 if (!nr_pages)
1510 return 0;
1511
1512 /* Split into individual pages */
1513 set_page_refcounted(page);
1514 split_page(page, order);
1515 return nr_pages;
1516}
1517
1518/*
1519 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1520 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1521 * or two.
1522 */
1523static inline
1524struct page *buffered_rmqueue(struct zone *preferred_zone,
1525 struct zone *zone, int order, gfp_t gfp_flags,
1526 int migratetype)
1527{
1528 unsigned long flags;
1529 struct page *page;
1530 int cold = !!(gfp_flags & __GFP_COLD);
1531
1532again:
1533 if (likely(order == 0)) {
1534 struct per_cpu_pages *pcp;
1535 struct list_head *list;
1536
1537 local_irq_save(flags);
1538 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1539 list = &pcp->lists[migratetype];
1540 if (list_empty(list)) {
1541 pcp->count += rmqueue_bulk(zone, 0,
1542 pcp->batch, list,
1543 migratetype, cold);
1544 if (unlikely(list_empty(list)))
1545 goto failed;
1546 }
1547
1548 if (cold)
1549 page = list_entry(list->prev, struct page, lru);
1550 else
1551 page = list_entry(list->next, struct page, lru);
1552
1553 list_del(&page->lru);
1554 pcp->count--;
1555 } else {
1556 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1557 /*
1558 * __GFP_NOFAIL is not to be used in new code.
1559 *
1560 * All __GFP_NOFAIL callers should be fixed so that they
1561 * properly detect and handle allocation failures.
1562 *
1563 * We most definitely don't want callers attempting to
1564 * allocate greater than order-1 page units with
1565 * __GFP_NOFAIL.
1566 */
1567 WARN_ON_ONCE(order > 1);
1568 }
1569 spin_lock_irqsave(&zone->lock, flags);
1570 page = __rmqueue(zone, order, migratetype);
1571 spin_unlock(&zone->lock);
1572 if (!page)
1573 goto failed;
1574 __mod_zone_freepage_state(zone, -(1 << order),
1575 get_pageblock_migratetype(page));
1576 }
1577
1578 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1579
1580 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1581 zone_statistics(preferred_zone, zone, gfp_flags);
1582 local_irq_restore(flags);
1583
1584 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1585 if (prep_new_page(page, order, gfp_flags))
1586 goto again;
1587 return page;
1588
1589failed:
1590 local_irq_restore(flags);
1591 return NULL;
1592}
1593
1594#ifdef CONFIG_FAIL_PAGE_ALLOC
1595
1596static struct {
1597 struct fault_attr attr;
1598
1599 u32 ignore_gfp_highmem;
1600 u32 ignore_gfp_wait;
1601 u32 min_order;
1602} fail_page_alloc = {
1603 .attr = FAULT_ATTR_INITIALIZER,
1604 .ignore_gfp_wait = 1,
1605 .ignore_gfp_highmem = 1,
1606 .min_order = 1,
1607};
1608
1609static int __init setup_fail_page_alloc(char *str)
1610{
1611 return setup_fault_attr(&fail_page_alloc.attr, str);
1612}
1613__setup("fail_page_alloc=", setup_fail_page_alloc);
1614
1615static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1616{
1617 if (order < fail_page_alloc.min_order)
1618 return false;
1619 if (gfp_mask & __GFP_NOFAIL)
1620 return false;
1621 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1622 return false;
1623 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1624 return false;
1625
1626 return should_fail(&fail_page_alloc.attr, 1 << order);
1627}
1628
1629#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1630
1631static int __init fail_page_alloc_debugfs(void)
1632{
1633 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1634 struct dentry *dir;
1635
1636 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1637 &fail_page_alloc.attr);
1638 if (IS_ERR(dir))
1639 return PTR_ERR(dir);
1640
1641 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1642 &fail_page_alloc.ignore_gfp_wait))
1643 goto fail;
1644 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1645 &fail_page_alloc.ignore_gfp_highmem))
1646 goto fail;
1647 if (!debugfs_create_u32("min-order", mode, dir,
1648 &fail_page_alloc.min_order))
1649 goto fail;
1650
1651 return 0;
1652fail:
1653 debugfs_remove_recursive(dir);
1654
1655 return -ENOMEM;
1656}
1657
1658late_initcall(fail_page_alloc_debugfs);
1659
1660#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1661
1662#else /* CONFIG_FAIL_PAGE_ALLOC */
1663
1664static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1665{
1666 return false;
1667}
1668
1669#endif /* CONFIG_FAIL_PAGE_ALLOC */
1670
1671/*
1672 * Return true if free pages are above 'mark'. This takes into account the order
1673 * of the allocation.
1674 */
1675static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1676 int classzone_idx, int alloc_flags, long free_pages)
1677{
1678 /* free_pages my go negative - that's OK */
1679 long min = mark;
1680 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1681 int o;
1682 long free_cma = 0;
1683
1684 free_pages -= (1 << order) - 1;
1685 if (alloc_flags & ALLOC_HIGH)
1686 min -= min / 2;
1687 if (alloc_flags & ALLOC_HARDER)
1688 min -= min / 4;
1689#ifdef CONFIG_CMA
1690 /* If allocation can't use CMA areas don't use free CMA pages */
1691 if (!(alloc_flags & ALLOC_CMA))
1692 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1693#endif
1694
1695 if (free_pages - free_cma <= min + lowmem_reserve)
1696 return false;
1697 for (o = 0; o < order; o++) {
1698 /* At the next order, this order's pages become unavailable */
1699 free_pages -= z->free_area[o].nr_free << o;
1700
1701 /* Require fewer higher order pages to be free */
1702 min >>= 1;
1703
1704 if (free_pages <= min)
1705 return false;
1706 }
1707 return true;
1708}
1709
1710bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1711 int classzone_idx, int alloc_flags)
1712{
1713 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1714 zone_page_state(z, NR_FREE_PAGES));
1715}
1716
1717bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1718 int classzone_idx, int alloc_flags)
1719{
1720 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1721
1722 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1723 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1724
1725 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1726 free_pages);
1727}
1728
1729#ifdef CONFIG_NUMA
1730/*
1731 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1732 * skip over zones that are not allowed by the cpuset, or that have
1733 * been recently (in last second) found to be nearly full. See further
1734 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1735 * that have to skip over a lot of full or unallowed zones.
1736 *
1737 * If the zonelist cache is present in the passed zonelist, then
1738 * returns a pointer to the allowed node mask (either the current
1739 * tasks mems_allowed, or node_states[N_MEMORY].)
1740 *
1741 * If the zonelist cache is not available for this zonelist, does
1742 * nothing and returns NULL.
1743 *
1744 * If the fullzones BITMAP in the zonelist cache is stale (more than
1745 * a second since last zap'd) then we zap it out (clear its bits.)
1746 *
1747 * We hold off even calling zlc_setup, until after we've checked the
1748 * first zone in the zonelist, on the theory that most allocations will
1749 * be satisfied from that first zone, so best to examine that zone as
1750 * quickly as we can.
1751 */
1752static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1753{
1754 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1755 nodemask_t *allowednodes; /* zonelist_cache approximation */
1756
1757 zlc = zonelist->zlcache_ptr;
1758 if (!zlc)
1759 return NULL;
1760
1761 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1762 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1763 zlc->last_full_zap = jiffies;
1764 }
1765
1766 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1767 &cpuset_current_mems_allowed :
1768 &node_states[N_MEMORY];
1769 return allowednodes;
1770}
1771
1772/*
1773 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1774 * if it is worth looking at further for free memory:
1775 * 1) Check that the zone isn't thought to be full (doesn't have its
1776 * bit set in the zonelist_cache fullzones BITMAP).
1777 * 2) Check that the zones node (obtained from the zonelist_cache
1778 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1779 * Return true (non-zero) if zone is worth looking at further, or
1780 * else return false (zero) if it is not.
1781 *
1782 * This check -ignores- the distinction between various watermarks,
1783 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1784 * found to be full for any variation of these watermarks, it will
1785 * be considered full for up to one second by all requests, unless
1786 * we are so low on memory on all allowed nodes that we are forced
1787 * into the second scan of the zonelist.
1788 *
1789 * In the second scan we ignore this zonelist cache and exactly
1790 * apply the watermarks to all zones, even it is slower to do so.
1791 * We are low on memory in the second scan, and should leave no stone
1792 * unturned looking for a free page.
1793 */
1794static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1795 nodemask_t *allowednodes)
1796{
1797 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1798 int i; /* index of *z in zonelist zones */
1799 int n; /* node that zone *z is on */
1800
1801 zlc = zonelist->zlcache_ptr;
1802 if (!zlc)
1803 return 1;
1804
1805 i = z - zonelist->_zonerefs;
1806 n = zlc->z_to_n[i];
1807
1808 /* This zone is worth trying if it is allowed but not full */
1809 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1810}
1811
1812/*
1813 * Given 'z' scanning a zonelist, set the corresponding bit in
1814 * zlc->fullzones, so that subsequent attempts to allocate a page
1815 * from that zone don't waste time re-examining it.
1816 */
1817static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1818{
1819 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1820 int i; /* index of *z in zonelist zones */
1821
1822 zlc = zonelist->zlcache_ptr;
1823 if (!zlc)
1824 return;
1825
1826 i = z - zonelist->_zonerefs;
1827
1828 set_bit(i, zlc->fullzones);
1829}
1830
1831/*
1832 * clear all zones full, called after direct reclaim makes progress so that
1833 * a zone that was recently full is not skipped over for up to a second
1834 */
1835static void zlc_clear_zones_full(struct zonelist *zonelist)
1836{
1837 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1838
1839 zlc = zonelist->zlcache_ptr;
1840 if (!zlc)
1841 return;
1842
1843 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1844}
1845
1846static bool zone_local(struct zone *local_zone, struct zone *zone)
1847{
1848 return local_zone->node == zone->node;
1849}
1850
1851static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1852{
1853 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1854}
1855
1856static void __paginginit init_zone_allows_reclaim(int nid)
1857{
1858 int i;
1859
1860 for_each_node_state(i, N_MEMORY)
1861 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1862 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1863 else
1864 zone_reclaim_mode = 1;
1865}
1866
1867#else /* CONFIG_NUMA */
1868
1869static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1870{
1871 return NULL;
1872}
1873
1874static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1875 nodemask_t *allowednodes)
1876{
1877 return 1;
1878}
1879
1880static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1881{
1882}
1883
1884static void zlc_clear_zones_full(struct zonelist *zonelist)
1885{
1886}
1887
1888static bool zone_local(struct zone *local_zone, struct zone *zone)
1889{
1890 return true;
1891}
1892
1893static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1894{
1895 return true;
1896}
1897
1898static inline void init_zone_allows_reclaim(int nid)
1899{
1900}
1901#endif /* CONFIG_NUMA */
1902
1903/*
1904 * get_page_from_freelist goes through the zonelist trying to allocate
1905 * a page.
1906 */
1907static struct page *
1908get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1909 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1910 struct zone *preferred_zone, int migratetype)
1911{
1912 struct zoneref *z;
1913 struct page *page = NULL;
1914 int classzone_idx;
1915 struct zone *zone;
1916 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1917 int zlc_active = 0; /* set if using zonelist_cache */
1918 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1919
1920 classzone_idx = zone_idx(preferred_zone);
1921zonelist_scan:
1922 /*
1923 * Scan zonelist, looking for a zone with enough free.
1924 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1925 */
1926 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1927 high_zoneidx, nodemask) {
1928 unsigned long mark;
1929
1930 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1931 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1932 continue;
1933 if ((alloc_flags & ALLOC_CPUSET) &&
1934 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1935 continue;
1936 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1937 if (unlikely(alloc_flags & ALLOC_NO_WATERMARKS))
1938 goto try_this_zone;
1939 /*
1940 * Distribute pages in proportion to the individual
1941 * zone size to ensure fair page aging. The zone a
1942 * page was allocated in should have no effect on the
1943 * time the page has in memory before being reclaimed.
1944 */
1945 if (alloc_flags & ALLOC_FAIR) {
1946 if (!zone_local(preferred_zone, zone))
1947 continue;
1948 if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1949 continue;
1950 }
1951 /*
1952 * When allocating a page cache page for writing, we
1953 * want to get it from a zone that is within its dirty
1954 * limit, such that no single zone holds more than its
1955 * proportional share of globally allowed dirty pages.
1956 * The dirty limits take into account the zone's
1957 * lowmem reserves and high watermark so that kswapd
1958 * should be able to balance it without having to
1959 * write pages from its LRU list.
1960 *
1961 * This may look like it could increase pressure on
1962 * lower zones by failing allocations in higher zones
1963 * before they are full. But the pages that do spill
1964 * over are limited as the lower zones are protected
1965 * by this very same mechanism. It should not become
1966 * a practical burden to them.
1967 *
1968 * XXX: For now, allow allocations to potentially
1969 * exceed the per-zone dirty limit in the slowpath
1970 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1971 * which is important when on a NUMA setup the allowed
1972 * zones are together not big enough to reach the
1973 * global limit. The proper fix for these situations
1974 * will require awareness of zones in the
1975 * dirty-throttling and the flusher threads.
1976 */
1977 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1978 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1979 goto this_zone_full;
1980
1981 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1982 if (!zone_watermark_ok(zone, order, mark,
1983 classzone_idx, alloc_flags)) {
1984 int ret;
1985
1986 if (IS_ENABLED(CONFIG_NUMA) &&
1987 !did_zlc_setup && nr_online_nodes > 1) {
1988 /*
1989 * we do zlc_setup if there are multiple nodes
1990 * and before considering the first zone allowed
1991 * by the cpuset.
1992 */
1993 allowednodes = zlc_setup(zonelist, alloc_flags);
1994 zlc_active = 1;
1995 did_zlc_setup = 1;
1996 }
1997
1998 if (zone_reclaim_mode == 0 ||
1999 !zone_allows_reclaim(preferred_zone, zone))
2000 goto this_zone_full;
2001
2002 /*
2003 * As we may have just activated ZLC, check if the first
2004 * eligible zone has failed zone_reclaim recently.
2005 */
2006 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2007 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2008 continue;
2009
2010 ret = zone_reclaim(zone, gfp_mask, order);
2011 switch (ret) {
2012 case ZONE_RECLAIM_NOSCAN:
2013 /* did not scan */
2014 continue;
2015 case ZONE_RECLAIM_FULL:
2016 /* scanned but unreclaimable */
2017 continue;
2018 default:
2019 /* did we reclaim enough */
2020 if (zone_watermark_ok(zone, order, mark,
2021 classzone_idx, alloc_flags))
2022 goto try_this_zone;
2023
2024 /*
2025 * Failed to reclaim enough to meet watermark.
2026 * Only mark the zone full if checking the min
2027 * watermark or if we failed to reclaim just
2028 * 1<<order pages or else the page allocator
2029 * fastpath will prematurely mark zones full
2030 * when the watermark is between the low and
2031 * min watermarks.
2032 */
2033 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2034 ret == ZONE_RECLAIM_SOME)
2035 goto this_zone_full;
2036
2037 continue;
2038 }
2039 }
2040
2041try_this_zone:
2042 page = buffered_rmqueue(preferred_zone, zone, order,
2043 gfp_mask, migratetype);
2044 if (page)
2045 break;
2046this_zone_full:
2047 if (IS_ENABLED(CONFIG_NUMA))
2048 zlc_mark_zone_full(zonelist, z);
2049 }
2050
2051 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2052 /* Disable zlc cache for second zonelist scan */
2053 zlc_active = 0;
2054 goto zonelist_scan;
2055 }
2056
2057 if (page)
2058 /*
2059 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2060 * necessary to allocate the page. The expectation is
2061 * that the caller is taking steps that will free more
2062 * memory. The caller should avoid the page being used
2063 * for !PFMEMALLOC purposes.
2064 */
2065 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2066
2067 return page;
2068}
2069
2070/*
2071 * Large machines with many possible nodes should not always dump per-node
2072 * meminfo in irq context.
2073 */
2074static inline bool should_suppress_show_mem(void)
2075{
2076 bool ret = false;
2077
2078#if NODES_SHIFT > 8
2079 ret = in_interrupt();
2080#endif
2081 return ret;
2082}
2083
2084static DEFINE_RATELIMIT_STATE(nopage_rs,
2085 DEFAULT_RATELIMIT_INTERVAL,
2086 DEFAULT_RATELIMIT_BURST);
2087
2088void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2089{
2090 unsigned int filter = SHOW_MEM_FILTER_NODES;
2091
2092 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2093 debug_guardpage_minorder() > 0)
2094 return;
2095
2096 /*
2097 * This documents exceptions given to allocations in certain
2098 * contexts that are allowed to allocate outside current's set
2099 * of allowed nodes.
2100 */
2101 if (!(gfp_mask & __GFP_NOMEMALLOC))
2102 if (test_thread_flag(TIF_MEMDIE) ||
2103 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2104 filter &= ~SHOW_MEM_FILTER_NODES;
2105 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2106 filter &= ~SHOW_MEM_FILTER_NODES;
2107
2108 if (fmt) {
2109 struct va_format vaf;
2110 va_list args;
2111
2112 va_start(args, fmt);
2113
2114 vaf.fmt = fmt;
2115 vaf.va = &args;
2116
2117 pr_warn("%pV", &vaf);
2118
2119 va_end(args);
2120 }
2121
2122 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2123 current->comm, order, gfp_mask);
2124
2125 dump_stack();
2126 if (!should_suppress_show_mem())
2127 show_mem(filter);
2128}
2129
2130static inline int
2131should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2132 unsigned long did_some_progress,
2133 unsigned long pages_reclaimed)
2134{
2135 /* Do not loop if specifically requested */
2136 if (gfp_mask & __GFP_NORETRY)
2137 return 0;
2138
2139 /* Always retry if specifically requested */
2140 if (gfp_mask & __GFP_NOFAIL)
2141 return 1;
2142
2143 /*
2144 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2145 * making forward progress without invoking OOM. Suspend also disables
2146 * storage devices so kswapd will not help. Bail if we are suspending.
2147 */
2148 if (!did_some_progress && pm_suspended_storage())
2149 return 0;
2150
2151 /*
2152 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2153 * means __GFP_NOFAIL, but that may not be true in other
2154 * implementations.
2155 */
2156 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2157 return 1;
2158
2159 /*
2160 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2161 * specified, then we retry until we no longer reclaim any pages
2162 * (above), or we've reclaimed an order of pages at least as
2163 * large as the allocation's order. In both cases, if the
2164 * allocation still fails, we stop retrying.
2165 */
2166 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2167 return 1;
2168
2169 return 0;
2170}
2171
2172static inline struct page *
2173__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2174 struct zonelist *zonelist, enum zone_type high_zoneidx,
2175 nodemask_t *nodemask, struct zone *preferred_zone,
2176 int migratetype)
2177{
2178 struct page *page;
2179
2180 /* Acquire the OOM killer lock for the zones in zonelist */
2181 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2182 schedule_timeout_uninterruptible(1);
2183 return NULL;
2184 }
2185
2186 /*
2187 * Go through the zonelist yet one more time, keep very high watermark
2188 * here, this is only to catch a parallel oom killing, we must fail if
2189 * we're still under heavy pressure.
2190 */
2191 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2192 order, zonelist, high_zoneidx,
2193 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2194 preferred_zone, migratetype);
2195 if (page)
2196 goto out;
2197
2198 if (!(gfp_mask & __GFP_NOFAIL)) {
2199 /* The OOM killer will not help higher order allocs */
2200 if (order > PAGE_ALLOC_COSTLY_ORDER)
2201 goto out;
2202 /* The OOM killer does not needlessly kill tasks for lowmem */
2203 if (high_zoneidx < ZONE_NORMAL)
2204 goto out;
2205 /*
2206 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2207 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2208 * The caller should handle page allocation failure by itself if
2209 * it specifies __GFP_THISNODE.
2210 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2211 */
2212 if (gfp_mask & __GFP_THISNODE)
2213 goto out;
2214 }
2215 /* Exhausted what can be done so it's blamo time */
2216 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2217
2218out:
2219 clear_zonelist_oom(zonelist, gfp_mask);
2220 return page;
2221}
2222
2223#ifdef CONFIG_COMPACTION
2224/* Try memory compaction for high-order allocations before reclaim */
2225static struct page *
2226__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2227 struct zonelist *zonelist, enum zone_type high_zoneidx,
2228 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2229 int migratetype, bool sync_migration,
2230 bool *contended_compaction, bool *deferred_compaction,
2231 unsigned long *did_some_progress)
2232{
2233 if (!order)
2234 return NULL;
2235
2236 if (compaction_deferred(preferred_zone, order)) {
2237 *deferred_compaction = true;
2238 return NULL;
2239 }
2240
2241 current->flags |= PF_MEMALLOC;
2242 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2243 nodemask, sync_migration,
2244 contended_compaction);
2245 current->flags &= ~PF_MEMALLOC;
2246
2247 if (*did_some_progress != COMPACT_SKIPPED) {
2248 struct page *page;
2249
2250 /* Page migration frees to the PCP lists but we want merging */
2251 drain_pages(get_cpu());
2252 put_cpu();
2253
2254 page = get_page_from_freelist(gfp_mask, nodemask,
2255 order, zonelist, high_zoneidx,
2256 alloc_flags & ~ALLOC_NO_WATERMARKS,
2257 preferred_zone, migratetype);
2258 if (page) {
2259 preferred_zone->compact_blockskip_flush = false;
2260 compaction_defer_reset(preferred_zone, order, true);
2261 count_vm_event(COMPACTSUCCESS);
2262 return page;
2263 }
2264
2265 /*
2266 * It's bad if compaction run occurs and fails.
2267 * The most likely reason is that pages exist,
2268 * but not enough to satisfy watermarks.
2269 */
2270 count_vm_event(COMPACTFAIL);
2271
2272 /*
2273 * As async compaction considers a subset of pageblocks, only
2274 * defer if the failure was a sync compaction failure.
2275 */
2276 if (sync_migration)
2277 defer_compaction(preferred_zone, order);
2278
2279 cond_resched();
2280 }
2281
2282 return NULL;
2283}
2284#else
2285static inline struct page *
2286__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2287 struct zonelist *zonelist, enum zone_type high_zoneidx,
2288 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2289 int migratetype, bool sync_migration,
2290 bool *contended_compaction, bool *deferred_compaction,
2291 unsigned long *did_some_progress)
2292{
2293 return NULL;
2294}
2295#endif /* CONFIG_COMPACTION */
2296
2297/* Perform direct synchronous page reclaim */
2298static int
2299__perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2300 nodemask_t *nodemask)
2301{
2302 struct reclaim_state reclaim_state;
2303 int progress;
2304
2305 cond_resched();
2306
2307 /* We now go into synchronous reclaim */
2308 cpuset_memory_pressure_bump();
2309 current->flags |= PF_MEMALLOC;
2310 lockdep_set_current_reclaim_state(gfp_mask);
2311 reclaim_state.reclaimed_slab = 0;
2312 current->reclaim_state = &reclaim_state;
2313
2314 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2315
2316 current->reclaim_state = NULL;
2317 lockdep_clear_current_reclaim_state();
2318 current->flags &= ~PF_MEMALLOC;
2319
2320 cond_resched();
2321
2322 return progress;
2323}
2324
2325/* The really slow allocator path where we enter direct reclaim */
2326static inline struct page *
2327__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2328 struct zonelist *zonelist, enum zone_type high_zoneidx,
2329 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2330 int migratetype, unsigned long *did_some_progress)
2331{
2332 struct page *page = NULL;
2333 bool drained = false;
2334
2335 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2336 nodemask);
2337 if (unlikely(!(*did_some_progress)))
2338 return NULL;
2339
2340 /* After successful reclaim, reconsider all zones for allocation */
2341 if (IS_ENABLED(CONFIG_NUMA))
2342 zlc_clear_zones_full(zonelist);
2343
2344retry:
2345 page = get_page_from_freelist(gfp_mask, nodemask, order,
2346 zonelist, high_zoneidx,
2347 alloc_flags & ~ALLOC_NO_WATERMARKS,
2348 preferred_zone, migratetype);
2349
2350 /*
2351 * If an allocation failed after direct reclaim, it could be because
2352 * pages are pinned on the per-cpu lists. Drain them and try again
2353 */
2354 if (!page && !drained) {
2355 drain_all_pages();
2356 drained = true;
2357 goto retry;
2358 }
2359
2360 return page;
2361}
2362
2363/*
2364 * This is called in the allocator slow-path if the allocation request is of
2365 * sufficient urgency to ignore watermarks and take other desperate measures
2366 */
2367static inline struct page *
2368__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2369 struct zonelist *zonelist, enum zone_type high_zoneidx,
2370 nodemask_t *nodemask, struct zone *preferred_zone,
2371 int migratetype)
2372{
2373 struct page *page;
2374
2375 do {
2376 page = get_page_from_freelist(gfp_mask, nodemask, order,
2377 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2378 preferred_zone, migratetype);
2379
2380 if (!page && gfp_mask & __GFP_NOFAIL)
2381 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2382 } while (!page && (gfp_mask & __GFP_NOFAIL));
2383
2384 return page;
2385}
2386
2387static void reset_alloc_batches(struct zonelist *zonelist,
2388 enum zone_type high_zoneidx,
2389 struct zone *preferred_zone)
2390{
2391 struct zoneref *z;
2392 struct zone *zone;
2393
2394 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2395 /*
2396 * Only reset the batches of zones that were actually
2397 * considered in the fairness pass, we don't want to
2398 * trash fairness information for zones that are not
2399 * actually part of this zonelist's round-robin cycle.
2400 */
2401 if (!zone_local(preferred_zone, zone))
2402 continue;
2403 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2404 high_wmark_pages(zone) - low_wmark_pages(zone) -
2405 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2406 }
2407}
2408
2409static void wake_all_kswapds(unsigned int order,
2410 struct zonelist *zonelist,
2411 enum zone_type high_zoneidx,
2412 struct zone *preferred_zone)
2413{
2414 struct zoneref *z;
2415 struct zone *zone;
2416
2417 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2418 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2419}
2420
2421static inline int
2422gfp_to_alloc_flags(gfp_t gfp_mask)
2423{
2424 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2425 const gfp_t wait = gfp_mask & __GFP_WAIT;
2426
2427 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2428 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2429
2430 /*
2431 * The caller may dip into page reserves a bit more if the caller
2432 * cannot run direct reclaim, or if the caller has realtime scheduling
2433 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2434 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2435 */
2436 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2437
2438 if (!wait) {
2439 /*
2440 * Not worth trying to allocate harder for
2441 * __GFP_NOMEMALLOC even if it can't schedule.
2442 */
2443 if (!(gfp_mask & __GFP_NOMEMALLOC))
2444 alloc_flags |= ALLOC_HARDER;
2445 /*
2446 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2447 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2448 */
2449 alloc_flags &= ~ALLOC_CPUSET;
2450 } else if (unlikely(rt_task(current)) && !in_interrupt())
2451 alloc_flags |= ALLOC_HARDER;
2452
2453 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2454 if (gfp_mask & __GFP_MEMALLOC)
2455 alloc_flags |= ALLOC_NO_WATERMARKS;
2456 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2457 alloc_flags |= ALLOC_NO_WATERMARKS;
2458 else if (!in_interrupt() &&
2459 ((current->flags & PF_MEMALLOC) ||
2460 unlikely(test_thread_flag(TIF_MEMDIE))))
2461 alloc_flags |= ALLOC_NO_WATERMARKS;
2462 }
2463#ifdef CONFIG_CMA
2464 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2465 alloc_flags |= ALLOC_CMA;
2466#endif
2467 return alloc_flags;
2468}
2469
2470bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2471{
2472 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2473}
2474
2475static inline struct page *
2476__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2477 struct zonelist *zonelist, enum zone_type high_zoneidx,
2478 nodemask_t *nodemask, struct zone *preferred_zone,
2479 int migratetype)
2480{
2481 const gfp_t wait = gfp_mask & __GFP_WAIT;
2482 struct page *page = NULL;
2483 int alloc_flags;
2484 unsigned long pages_reclaimed = 0;
2485 unsigned long did_some_progress;
2486 bool sync_migration = false;
2487 bool deferred_compaction = false;
2488 bool contended_compaction = false;
2489
2490 /*
2491 * In the slowpath, we sanity check order to avoid ever trying to
2492 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2493 * be using allocators in order of preference for an area that is
2494 * too large.
2495 */
2496 if (order >= MAX_ORDER) {
2497 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2498 return NULL;
2499 }
2500
2501 /*
2502 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2503 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2504 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2505 * using a larger set of nodes after it has established that the
2506 * allowed per node queues are empty and that nodes are
2507 * over allocated.
2508 */
2509 if (IS_ENABLED(CONFIG_NUMA) &&
2510 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2511 goto nopage;
2512
2513restart:
2514 if (!(gfp_mask & __GFP_NO_KSWAPD))
2515 wake_all_kswapds(order, zonelist, high_zoneidx, preferred_zone);
2516
2517 /*
2518 * OK, we're below the kswapd watermark and have kicked background
2519 * reclaim. Now things get more complex, so set up alloc_flags according
2520 * to how we want to proceed.
2521 */
2522 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2523
2524 /*
2525 * Find the true preferred zone if the allocation is unconstrained by
2526 * cpusets.
2527 */
2528 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2529 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2530 &preferred_zone);
2531
2532rebalance:
2533 /* This is the last chance, in general, before the goto nopage. */
2534 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2535 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2536 preferred_zone, migratetype);
2537 if (page)
2538 goto got_pg;
2539
2540 /* Allocate without watermarks if the context allows */
2541 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2542 /*
2543 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2544 * the allocation is high priority and these type of
2545 * allocations are system rather than user orientated
2546 */
2547 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2548
2549 page = __alloc_pages_high_priority(gfp_mask, order,
2550 zonelist, high_zoneidx, nodemask,
2551 preferred_zone, migratetype);
2552 if (page) {
2553 goto got_pg;
2554 }
2555 }
2556
2557 /* Atomic allocations - we can't balance anything */
2558 if (!wait) {
2559 /*
2560 * All existing users of the deprecated __GFP_NOFAIL are
2561 * blockable, so warn of any new users that actually allow this
2562 * type of allocation to fail.
2563 */
2564 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2565 goto nopage;
2566 }
2567
2568 /* Avoid recursion of direct reclaim */
2569 if (current->flags & PF_MEMALLOC)
2570 goto nopage;
2571
2572 /* Avoid allocations with no watermarks from looping endlessly */
2573 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2574 goto nopage;
2575
2576 /*
2577 * Try direct compaction. The first pass is asynchronous. Subsequent
2578 * attempts after direct reclaim are synchronous
2579 */
2580 page = __alloc_pages_direct_compact(gfp_mask, order,
2581 zonelist, high_zoneidx,
2582 nodemask,
2583 alloc_flags, preferred_zone,
2584 migratetype, sync_migration,
2585 &contended_compaction,
2586 &deferred_compaction,
2587 &did_some_progress);
2588 if (page)
2589 goto got_pg;
2590 sync_migration = true;
2591
2592 /*
2593 * If compaction is deferred for high-order allocations, it is because
2594 * sync compaction recently failed. In this is the case and the caller
2595 * requested a movable allocation that does not heavily disrupt the
2596 * system then fail the allocation instead of entering direct reclaim.
2597 */
2598 if ((deferred_compaction || contended_compaction) &&
2599 (gfp_mask & __GFP_NO_KSWAPD))
2600 goto nopage;
2601
2602 /* Try direct reclaim and then allocating */
2603 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2604 zonelist, high_zoneidx,
2605 nodemask,
2606 alloc_flags, preferred_zone,
2607 migratetype, &did_some_progress);
2608 if (page)
2609 goto got_pg;
2610
2611 /*
2612 * If we failed to make any progress reclaiming, then we are
2613 * running out of options and have to consider going OOM
2614 */
2615 if (!did_some_progress) {
2616 if (oom_gfp_allowed(gfp_mask)) {
2617 if (oom_killer_disabled)
2618 goto nopage;
2619 /* Coredumps can quickly deplete all memory reserves */
2620 if ((current->flags & PF_DUMPCORE) &&
2621 !(gfp_mask & __GFP_NOFAIL))
2622 goto nopage;
2623 page = __alloc_pages_may_oom(gfp_mask, order,
2624 zonelist, high_zoneidx,
2625 nodemask, preferred_zone,
2626 migratetype);
2627 if (page)
2628 goto got_pg;
2629
2630 if (!(gfp_mask & __GFP_NOFAIL)) {
2631 /*
2632 * The oom killer is not called for high-order
2633 * allocations that may fail, so if no progress
2634 * is being made, there are no other options and
2635 * retrying is unlikely to help.
2636 */
2637 if (order > PAGE_ALLOC_COSTLY_ORDER)
2638 goto nopage;
2639 /*
2640 * The oom killer is not called for lowmem
2641 * allocations to prevent needlessly killing
2642 * innocent tasks.
2643 */
2644 if (high_zoneidx < ZONE_NORMAL)
2645 goto nopage;
2646 }
2647
2648 goto restart;
2649 }
2650 }
2651
2652 /* Check if we should retry the allocation */
2653 pages_reclaimed += did_some_progress;
2654 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2655 pages_reclaimed)) {
2656 /* Wait for some write requests to complete then retry */
2657 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2658 goto rebalance;
2659 } else {
2660 /*
2661 * High-order allocations do not necessarily loop after
2662 * direct reclaim and reclaim/compaction depends on compaction
2663 * being called after reclaim so call directly if necessary
2664 */
2665 page = __alloc_pages_direct_compact(gfp_mask, order,
2666 zonelist, high_zoneidx,
2667 nodemask,
2668 alloc_flags, preferred_zone,
2669 migratetype, sync_migration,
2670 &contended_compaction,
2671 &deferred_compaction,
2672 &did_some_progress);
2673 if (page)
2674 goto got_pg;
2675 }
2676
2677nopage:
2678 warn_alloc_failed(gfp_mask, order, NULL);
2679 return page;
2680got_pg:
2681 if (kmemcheck_enabled)
2682 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2683
2684 return page;
2685}
2686
2687/*
2688 * This is the 'heart' of the zoned buddy allocator.
2689 */
2690struct page *
2691__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2692 struct zonelist *zonelist, nodemask_t *nodemask)
2693{
2694 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2695 struct zone *preferred_zone;
2696 struct page *page = NULL;
2697 int migratetype = allocflags_to_migratetype(gfp_mask);
2698 unsigned int cpuset_mems_cookie;
2699 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2700 struct mem_cgroup *memcg = NULL;
2701
2702 gfp_mask &= gfp_allowed_mask;
2703
2704 lockdep_trace_alloc(gfp_mask);
2705
2706 might_sleep_if(gfp_mask & __GFP_WAIT);
2707
2708 if (should_fail_alloc_page(gfp_mask, order))
2709 return NULL;
2710
2711 /*
2712 * Check the zones suitable for the gfp_mask contain at least one
2713 * valid zone. It's possible to have an empty zonelist as a result
2714 * of GFP_THISNODE and a memoryless node
2715 */
2716 if (unlikely(!zonelist->_zonerefs->zone))
2717 return NULL;
2718
2719 /*
2720 * Will only have any effect when __GFP_KMEMCG is set. This is
2721 * verified in the (always inline) callee
2722 */
2723 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2724 return NULL;
2725
2726retry_cpuset:
2727 cpuset_mems_cookie = read_mems_allowed_begin();
2728
2729 /* The preferred zone is used for statistics later */
2730 first_zones_zonelist(zonelist, high_zoneidx,
2731 nodemask ? : &cpuset_current_mems_allowed,
2732 &preferred_zone);
2733 if (!preferred_zone)
2734 goto out;
2735
2736#ifdef CONFIG_CMA
2737 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2738 alloc_flags |= ALLOC_CMA;
2739#endif
2740retry:
2741 /* First allocation attempt */
2742 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2743 zonelist, high_zoneidx, alloc_flags,
2744 preferred_zone, migratetype);
2745 if (unlikely(!page)) {
2746 /*
2747 * The first pass makes sure allocations are spread
2748 * fairly within the local node. However, the local
2749 * node might have free pages left after the fairness
2750 * batches are exhausted, and remote zones haven't
2751 * even been considered yet. Try once more without
2752 * fairness, and include remote zones now, before
2753 * entering the slowpath and waking kswapd: prefer
2754 * spilling to a remote zone over swapping locally.
2755 */
2756 if (alloc_flags & ALLOC_FAIR) {
2757 reset_alloc_batches(zonelist, high_zoneidx,
2758 preferred_zone);
2759 alloc_flags &= ~ALLOC_FAIR;
2760 goto retry;
2761 }
2762 /*
2763 * Runtime PM, block IO and its error handling path
2764 * can deadlock because I/O on the device might not
2765 * complete.
2766 */
2767 gfp_mask = memalloc_noio_flags(gfp_mask);
2768 page = __alloc_pages_slowpath(gfp_mask, order,
2769 zonelist, high_zoneidx, nodemask,
2770 preferred_zone, migratetype);
2771 }
2772
2773 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2774
2775out:
2776 /*
2777 * When updating a task's mems_allowed, it is possible to race with
2778 * parallel threads in such a way that an allocation can fail while
2779 * the mask is being updated. If a page allocation is about to fail,
2780 * check if the cpuset changed during allocation and if so, retry.
2781 */
2782 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2783 goto retry_cpuset;
2784
2785 memcg_kmem_commit_charge(page, memcg, order);
2786
2787 return page;
2788}
2789EXPORT_SYMBOL(__alloc_pages_nodemask);
2790
2791/*
2792 * Common helper functions.
2793 */
2794unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2795{
2796 struct page *page;
2797
2798 /*
2799 * __get_free_pages() returns a 32-bit address, which cannot represent
2800 * a highmem page
2801 */
2802 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2803
2804 page = alloc_pages(gfp_mask, order);
2805 if (!page)
2806 return 0;
2807 return (unsigned long) page_address(page);
2808}
2809EXPORT_SYMBOL(__get_free_pages);
2810
2811unsigned long get_zeroed_page(gfp_t gfp_mask)
2812{
2813 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2814}
2815EXPORT_SYMBOL(get_zeroed_page);
2816
2817void __free_pages(struct page *page, unsigned int order)
2818{
2819 if (put_page_testzero(page)) {
2820 if (order == 0)
2821 free_hot_cold_page(page, 0);
2822 else
2823 __free_pages_ok(page, order);
2824 }
2825}
2826
2827EXPORT_SYMBOL(__free_pages);
2828
2829void free_pages(unsigned long addr, unsigned int order)
2830{
2831 if (addr != 0) {
2832 VM_BUG_ON(!virt_addr_valid((void *)addr));
2833 __free_pages(virt_to_page((void *)addr), order);
2834 }
2835}
2836
2837EXPORT_SYMBOL(free_pages);
2838
2839/*
2840 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2841 * pages allocated with __GFP_KMEMCG.
2842 *
2843 * Those pages are accounted to a particular memcg, embedded in the
2844 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2845 * for that information only to find out that it is NULL for users who have no
2846 * interest in that whatsoever, we provide these functions.
2847 *
2848 * The caller knows better which flags it relies on.
2849 */
2850void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2851{
2852 memcg_kmem_uncharge_pages(page, order);
2853 __free_pages(page, order);
2854}
2855
2856void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2857{
2858 if (addr != 0) {
2859 VM_BUG_ON(!virt_addr_valid((void *)addr));
2860 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2861 }
2862}
2863
2864static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2865{
2866 if (addr) {
2867 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2868 unsigned long used = addr + PAGE_ALIGN(size);
2869
2870 split_page(virt_to_page((void *)addr), order);
2871 while (used < alloc_end) {
2872 free_page(used);
2873 used += PAGE_SIZE;
2874 }
2875 }
2876 return (void *)addr;
2877}
2878
2879/**
2880 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2881 * @size: the number of bytes to allocate
2882 * @gfp_mask: GFP flags for the allocation
2883 *
2884 * This function is similar to alloc_pages(), except that it allocates the
2885 * minimum number of pages to satisfy the request. alloc_pages() can only
2886 * allocate memory in power-of-two pages.
2887 *
2888 * This function is also limited by MAX_ORDER.
2889 *
2890 * Memory allocated by this function must be released by free_pages_exact().
2891 */
2892void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2893{
2894 unsigned int order = get_order(size);
2895 unsigned long addr;
2896
2897 addr = __get_free_pages(gfp_mask, order);
2898 return make_alloc_exact(addr, order, size);
2899}
2900EXPORT_SYMBOL(alloc_pages_exact);
2901
2902/**
2903 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2904 * pages on a node.
2905 * @nid: the preferred node ID where memory should be allocated
2906 * @size: the number of bytes to allocate
2907 * @gfp_mask: GFP flags for the allocation
2908 *
2909 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2910 * back.
2911 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2912 * but is not exact.
2913 */
2914void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2915{
2916 unsigned order = get_order(size);
2917 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2918 if (!p)
2919 return NULL;
2920 return make_alloc_exact((unsigned long)page_address(p), order, size);
2921}
2922EXPORT_SYMBOL(alloc_pages_exact_nid);
2923
2924/**
2925 * free_pages_exact - release memory allocated via alloc_pages_exact()
2926 * @virt: the value returned by alloc_pages_exact.
2927 * @size: size of allocation, same value as passed to alloc_pages_exact().
2928 *
2929 * Release the memory allocated by a previous call to alloc_pages_exact.
2930 */
2931void free_pages_exact(void *virt, size_t size)
2932{
2933 unsigned long addr = (unsigned long)virt;
2934 unsigned long end = addr + PAGE_ALIGN(size);
2935
2936 while (addr < end) {
2937 free_page(addr);
2938 addr += PAGE_SIZE;
2939 }
2940}
2941EXPORT_SYMBOL(free_pages_exact);
2942
2943/**
2944 * nr_free_zone_pages - count number of pages beyond high watermark
2945 * @offset: The zone index of the highest zone
2946 *
2947 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2948 * high watermark within all zones at or below a given zone index. For each
2949 * zone, the number of pages is calculated as:
2950 * managed_pages - high_pages
2951 */
2952static unsigned long nr_free_zone_pages(int offset)
2953{
2954 struct zoneref *z;
2955 struct zone *zone;
2956
2957 /* Just pick one node, since fallback list is circular */
2958 unsigned long sum = 0;
2959
2960 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2961
2962 for_each_zone_zonelist(zone, z, zonelist, offset) {
2963 unsigned long size = zone->managed_pages;
2964 unsigned long high = high_wmark_pages(zone);
2965 if (size > high)
2966 sum += size - high;
2967 }
2968
2969 return sum;
2970}
2971
2972/**
2973 * nr_free_buffer_pages - count number of pages beyond high watermark
2974 *
2975 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2976 * watermark within ZONE_DMA and ZONE_NORMAL.
2977 */
2978unsigned long nr_free_buffer_pages(void)
2979{
2980 return nr_free_zone_pages(gfp_zone(GFP_USER));
2981}
2982EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2983
2984/**
2985 * nr_free_pagecache_pages - count number of pages beyond high watermark
2986 *
2987 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2988 * high watermark within all zones.
2989 */
2990unsigned long nr_free_pagecache_pages(void)
2991{
2992 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2993}
2994
2995static inline void show_node(struct zone *zone)
2996{
2997 if (IS_ENABLED(CONFIG_NUMA))
2998 printk("Node %d ", zone_to_nid(zone));
2999}
3000
3001void si_meminfo(struct sysinfo *val)
3002{
3003 val->totalram = totalram_pages;
3004 val->sharedram = 0;
3005 val->freeram = global_page_state(NR_FREE_PAGES);
3006 val->bufferram = nr_blockdev_pages();
3007 val->totalhigh = totalhigh_pages;
3008 val->freehigh = nr_free_highpages();
3009 val->mem_unit = PAGE_SIZE;
3010}
3011
3012EXPORT_SYMBOL(si_meminfo);
3013
3014#ifdef CONFIG_NUMA
3015void si_meminfo_node(struct sysinfo *val, int nid)
3016{
3017 int zone_type; /* needs to be signed */
3018 unsigned long managed_pages = 0;
3019 pg_data_t *pgdat = NODE_DATA(nid);
3020
3021 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3022 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3023 val->totalram = managed_pages;
3024 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3025#ifdef CONFIG_HIGHMEM
3026 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3027 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3028 NR_FREE_PAGES);
3029#else
3030 val->totalhigh = 0;
3031 val->freehigh = 0;
3032#endif
3033 val->mem_unit = PAGE_SIZE;
3034}
3035#endif
3036
3037/*
3038 * Determine whether the node should be displayed or not, depending on whether
3039 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3040 */
3041bool skip_free_areas_node(unsigned int flags, int nid)
3042{
3043 bool ret = false;
3044 unsigned int cpuset_mems_cookie;
3045
3046 if (!(flags & SHOW_MEM_FILTER_NODES))
3047 goto out;
3048
3049 do {
3050 cpuset_mems_cookie = read_mems_allowed_begin();
3051 ret = !node_isset(nid, cpuset_current_mems_allowed);
3052 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3053out:
3054 return ret;
3055}
3056
3057#define K(x) ((x) << (PAGE_SHIFT-10))
3058
3059static void show_migration_types(unsigned char type)
3060{
3061 static const char types[MIGRATE_TYPES] = {
3062 [MIGRATE_UNMOVABLE] = 'U',
3063 [MIGRATE_RECLAIMABLE] = 'E',
3064 [MIGRATE_MOVABLE] = 'M',
3065 [MIGRATE_RESERVE] = 'R',
3066#ifdef CONFIG_CMA
3067 [MIGRATE_CMA] = 'C',
3068#endif
3069#ifdef CONFIG_MEMORY_ISOLATION
3070 [MIGRATE_ISOLATE] = 'I',
3071#endif
3072 };
3073 char tmp[MIGRATE_TYPES + 1];
3074 char *p = tmp;
3075 int i;
3076
3077 for (i = 0; i < MIGRATE_TYPES; i++) {
3078 if (type & (1 << i))
3079 *p++ = types[i];
3080 }
3081
3082 *p = '\0';
3083 printk("(%s) ", tmp);
3084}
3085
3086/*
3087 * Show free area list (used inside shift_scroll-lock stuff)
3088 * We also calculate the percentage fragmentation. We do this by counting the
3089 * memory on each free list with the exception of the first item on the list.
3090 * Suppresses nodes that are not allowed by current's cpuset if
3091 * SHOW_MEM_FILTER_NODES is passed.
3092 */
3093void show_free_areas(unsigned int filter)
3094{
3095 int cpu;
3096 struct zone *zone;
3097
3098 for_each_populated_zone(zone) {
3099 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3100 continue;
3101 show_node(zone);
3102 printk("%s per-cpu:\n", zone->name);
3103
3104 for_each_online_cpu(cpu) {
3105 struct per_cpu_pageset *pageset;
3106
3107 pageset = per_cpu_ptr(zone->pageset, cpu);
3108
3109 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3110 cpu, pageset->pcp.high,
3111 pageset->pcp.batch, pageset->pcp.count);
3112 }
3113 }
3114
3115 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3116 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3117 " unevictable:%lu"
3118 " dirty:%lu writeback:%lu unstable:%lu\n"
3119 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3120 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3121 " free_cma:%lu\n",
3122 global_page_state(NR_ACTIVE_ANON),
3123 global_page_state(NR_INACTIVE_ANON),
3124 global_page_state(NR_ISOLATED_ANON),
3125 global_page_state(NR_ACTIVE_FILE),
3126 global_page_state(NR_INACTIVE_FILE),
3127 global_page_state(NR_ISOLATED_FILE),
3128 global_page_state(NR_UNEVICTABLE),
3129 global_page_state(NR_FILE_DIRTY),
3130 global_page_state(NR_WRITEBACK),
3131 global_page_state(NR_UNSTABLE_NFS),
3132 global_page_state(NR_FREE_PAGES),
3133 global_page_state(NR_SLAB_RECLAIMABLE),
3134 global_page_state(NR_SLAB_UNRECLAIMABLE),
3135 global_page_state(NR_FILE_MAPPED),
3136 global_page_state(NR_SHMEM),
3137 global_page_state(NR_PAGETABLE),
3138 global_page_state(NR_BOUNCE),
3139 global_page_state(NR_FREE_CMA_PAGES));
3140
3141 for_each_populated_zone(zone) {
3142 int i;
3143
3144 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3145 continue;
3146 show_node(zone);
3147 printk("%s"
3148 " free:%lukB"
3149 " min:%lukB"
3150 " low:%lukB"
3151 " high:%lukB"
3152 " active_anon:%lukB"
3153 " inactive_anon:%lukB"
3154 " active_file:%lukB"
3155 " inactive_file:%lukB"
3156 " unevictable:%lukB"
3157 " isolated(anon):%lukB"
3158 " isolated(file):%lukB"
3159 " present:%lukB"
3160 " managed:%lukB"
3161 " mlocked:%lukB"
3162 " dirty:%lukB"
3163 " writeback:%lukB"
3164 " mapped:%lukB"
3165 " shmem:%lukB"
3166 " slab_reclaimable:%lukB"
3167 " slab_unreclaimable:%lukB"
3168 " kernel_stack:%lukB"
3169 " pagetables:%lukB"
3170 " unstable:%lukB"
3171 " bounce:%lukB"
3172 " free_cma:%lukB"
3173 " writeback_tmp:%lukB"
3174 " pages_scanned:%lu"
3175 " all_unreclaimable? %s"
3176 "\n",
3177 zone->name,
3178 K(zone_page_state(zone, NR_FREE_PAGES)),
3179 K(min_wmark_pages(zone)),
3180 K(low_wmark_pages(zone)),
3181 K(high_wmark_pages(zone)),
3182 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3183 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3184 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3185 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3186 K(zone_page_state(zone, NR_UNEVICTABLE)),
3187 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3188 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3189 K(zone->present_pages),
3190 K(zone->managed_pages),
3191 K(zone_page_state(zone, NR_MLOCK)),
3192 K(zone_page_state(zone, NR_FILE_DIRTY)),
3193 K(zone_page_state(zone, NR_WRITEBACK)),
3194 K(zone_page_state(zone, NR_FILE_MAPPED)),
3195 K(zone_page_state(zone, NR_SHMEM)),
3196 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3197 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3198 zone_page_state(zone, NR_KERNEL_STACK) *
3199 THREAD_SIZE / 1024,
3200 K(zone_page_state(zone, NR_PAGETABLE)),
3201 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3202 K(zone_page_state(zone, NR_BOUNCE)),
3203 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3204 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3205 zone->pages_scanned,
3206 (!zone_reclaimable(zone) ? "yes" : "no")
3207 );
3208 printk("lowmem_reserve[]:");
3209 for (i = 0; i < MAX_NR_ZONES; i++)
3210 printk(" %lu", zone->lowmem_reserve[i]);
3211 printk("\n");
3212 }
3213
3214 for_each_populated_zone(zone) {
3215 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3216 unsigned char types[MAX_ORDER];
3217
3218 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3219 continue;
3220 show_node(zone);
3221 printk("%s: ", zone->name);
3222
3223 spin_lock_irqsave(&zone->lock, flags);
3224 for (order = 0; order < MAX_ORDER; order++) {
3225 struct free_area *area = &zone->free_area[order];
3226 int type;
3227
3228 nr[order] = area->nr_free;
3229 total += nr[order] << order;
3230
3231 types[order] = 0;
3232 for (type = 0; type < MIGRATE_TYPES; type++) {
3233 if (!list_empty(&area->free_list[type]))
3234 types[order] |= 1 << type;
3235 }
3236 }
3237 spin_unlock_irqrestore(&zone->lock, flags);
3238 for (order = 0; order < MAX_ORDER; order++) {
3239 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3240 if (nr[order])
3241 show_migration_types(types[order]);
3242 }
3243 printk("= %lukB\n", K(total));
3244 }
3245
3246 hugetlb_show_meminfo();
3247
3248 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3249
3250 show_swap_cache_info();
3251}
3252
3253static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3254{
3255 zoneref->zone = zone;
3256 zoneref->zone_idx = zone_idx(zone);
3257}
3258
3259/*
3260 * Builds allocation fallback zone lists.
3261 *
3262 * Add all populated zones of a node to the zonelist.
3263 */
3264static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3265 int nr_zones)
3266{
3267 struct zone *zone;
3268 enum zone_type zone_type = MAX_NR_ZONES;
3269
3270 do {
3271 zone_type--;
3272 zone = pgdat->node_zones + zone_type;
3273 if (populated_zone(zone)) {
3274 zoneref_set_zone(zone,
3275 &zonelist->_zonerefs[nr_zones++]);
3276 check_highest_zone(zone_type);
3277 }
3278 } while (zone_type);
3279
3280 return nr_zones;
3281}
3282
3283
3284/*
3285 * zonelist_order:
3286 * 0 = automatic detection of better ordering.
3287 * 1 = order by ([node] distance, -zonetype)
3288 * 2 = order by (-zonetype, [node] distance)
3289 *
3290 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3291 * the same zonelist. So only NUMA can configure this param.
3292 */
3293#define ZONELIST_ORDER_DEFAULT 0
3294#define ZONELIST_ORDER_NODE 1
3295#define ZONELIST_ORDER_ZONE 2
3296
3297/* zonelist order in the kernel.
3298 * set_zonelist_order() will set this to NODE or ZONE.
3299 */
3300static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3301static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3302
3303
3304#ifdef CONFIG_NUMA
3305/* The value user specified ....changed by config */
3306static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3307/* string for sysctl */
3308#define NUMA_ZONELIST_ORDER_LEN 16
3309char numa_zonelist_order[16] = "default";
3310
3311/*
3312 * interface for configure zonelist ordering.
3313 * command line option "numa_zonelist_order"
3314 * = "[dD]efault - default, automatic configuration.
3315 * = "[nN]ode - order by node locality, then by zone within node
3316 * = "[zZ]one - order by zone, then by locality within zone
3317 */
3318
3319static int __parse_numa_zonelist_order(char *s)
3320{
3321 if (*s == 'd' || *s == 'D') {
3322 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3323 } else if (*s == 'n' || *s == 'N') {
3324 user_zonelist_order = ZONELIST_ORDER_NODE;
3325 } else if (*s == 'z' || *s == 'Z') {
3326 user_zonelist_order = ZONELIST_ORDER_ZONE;
3327 } else {
3328 printk(KERN_WARNING
3329 "Ignoring invalid numa_zonelist_order value: "
3330 "%s\n", s);
3331 return -EINVAL;
3332 }
3333 return 0;
3334}
3335
3336static __init int setup_numa_zonelist_order(char *s)
3337{
3338 int ret;
3339
3340 if (!s)
3341 return 0;
3342
3343 ret = __parse_numa_zonelist_order(s);
3344 if (ret == 0)
3345 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3346
3347 return ret;
3348}
3349early_param("numa_zonelist_order", setup_numa_zonelist_order);
3350
3351/*
3352 * sysctl handler for numa_zonelist_order
3353 */
3354int numa_zonelist_order_handler(ctl_table *table, int write,
3355 void __user *buffer, size_t *length,
3356 loff_t *ppos)
3357{
3358 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3359 int ret;
3360 static DEFINE_MUTEX(zl_order_mutex);
3361
3362 mutex_lock(&zl_order_mutex);
3363 if (write) {
3364 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3365 ret = -EINVAL;
3366 goto out;
3367 }
3368 strcpy(saved_string, (char *)table->data);
3369 }
3370 ret = proc_dostring(table, write, buffer, length, ppos);
3371 if (ret)
3372 goto out;
3373 if (write) {
3374 int oldval = user_zonelist_order;
3375
3376 ret = __parse_numa_zonelist_order((char *)table->data);
3377 if (ret) {
3378 /*
3379 * bogus value. restore saved string
3380 */
3381 strncpy((char *)table->data, saved_string,
3382 NUMA_ZONELIST_ORDER_LEN);
3383 user_zonelist_order = oldval;
3384 } else if (oldval != user_zonelist_order) {
3385 mutex_lock(&zonelists_mutex);
3386 build_all_zonelists(NULL, NULL);
3387 mutex_unlock(&zonelists_mutex);
3388 }
3389 }
3390out:
3391 mutex_unlock(&zl_order_mutex);
3392 return ret;
3393}
3394
3395
3396#define MAX_NODE_LOAD (nr_online_nodes)
3397static int node_load[MAX_NUMNODES];
3398
3399/**
3400 * find_next_best_node - find the next node that should appear in a given node's fallback list
3401 * @node: node whose fallback list we're appending
3402 * @used_node_mask: nodemask_t of already used nodes
3403 *
3404 * We use a number of factors to determine which is the next node that should
3405 * appear on a given node's fallback list. The node should not have appeared
3406 * already in @node's fallback list, and it should be the next closest node
3407 * according to the distance array (which contains arbitrary distance values
3408 * from each node to each node in the system), and should also prefer nodes
3409 * with no CPUs, since presumably they'll have very little allocation pressure
3410 * on them otherwise.
3411 * It returns -1 if no node is found.
3412 */
3413static int find_next_best_node(int node, nodemask_t *used_node_mask)
3414{
3415 int n, val;
3416 int min_val = INT_MAX;
3417 int best_node = NUMA_NO_NODE;
3418 const struct cpumask *tmp = cpumask_of_node(0);
3419
3420 /* Use the local node if we haven't already */
3421 if (!node_isset(node, *used_node_mask)) {
3422 node_set(node, *used_node_mask);
3423 return node;
3424 }
3425
3426 for_each_node_state(n, N_MEMORY) {
3427
3428 /* Don't want a node to appear more than once */
3429 if (node_isset(n, *used_node_mask))
3430 continue;
3431
3432 /* Use the distance array to find the distance */
3433 val = node_distance(node, n);
3434
3435 /* Penalize nodes under us ("prefer the next node") */
3436 val += (n < node);
3437
3438 /* Give preference to headless and unused nodes */
3439 tmp = cpumask_of_node(n);
3440 if (!cpumask_empty(tmp))
3441 val += PENALTY_FOR_NODE_WITH_CPUS;
3442
3443 /* Slight preference for less loaded node */
3444 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3445 val += node_load[n];
3446
3447 if (val < min_val) {
3448 min_val = val;
3449 best_node = n;
3450 }
3451 }
3452
3453 if (best_node >= 0)
3454 node_set(best_node, *used_node_mask);
3455
3456 return best_node;
3457}
3458
3459
3460/*
3461 * Build zonelists ordered by node and zones within node.
3462 * This results in maximum locality--normal zone overflows into local
3463 * DMA zone, if any--but risks exhausting DMA zone.
3464 */
3465static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3466{
3467 int j;
3468 struct zonelist *zonelist;
3469
3470 zonelist = &pgdat->node_zonelists[0];
3471 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3472 ;
3473 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3474 zonelist->_zonerefs[j].zone = NULL;
3475 zonelist->_zonerefs[j].zone_idx = 0;
3476}
3477
3478/*
3479 * Build gfp_thisnode zonelists
3480 */
3481static void build_thisnode_zonelists(pg_data_t *pgdat)
3482{
3483 int j;
3484 struct zonelist *zonelist;
3485
3486 zonelist = &pgdat->node_zonelists[1];
3487 j = build_zonelists_node(pgdat, zonelist, 0);
3488 zonelist->_zonerefs[j].zone = NULL;
3489 zonelist->_zonerefs[j].zone_idx = 0;
3490}
3491
3492/*
3493 * Build zonelists ordered by zone and nodes within zones.
3494 * This results in conserving DMA zone[s] until all Normal memory is
3495 * exhausted, but results in overflowing to remote node while memory
3496 * may still exist in local DMA zone.
3497 */
3498static int node_order[MAX_NUMNODES];
3499
3500static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3501{
3502 int pos, j, node;
3503 int zone_type; /* needs to be signed */
3504 struct zone *z;
3505 struct zonelist *zonelist;
3506
3507 zonelist = &pgdat->node_zonelists[0];
3508 pos = 0;
3509 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3510 for (j = 0; j < nr_nodes; j++) {
3511 node = node_order[j];
3512 z = &NODE_DATA(node)->node_zones[zone_type];
3513 if (populated_zone(z)) {
3514 zoneref_set_zone(z,
3515 &zonelist->_zonerefs[pos++]);
3516 check_highest_zone(zone_type);
3517 }
3518 }
3519 }
3520 zonelist->_zonerefs[pos].zone = NULL;
3521 zonelist->_zonerefs[pos].zone_idx = 0;
3522}
3523
3524static int default_zonelist_order(void)
3525{
3526 int nid, zone_type;
3527 unsigned long low_kmem_size, total_size;
3528 struct zone *z;
3529 int average_size;
3530 /*
3531 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3532 * If they are really small and used heavily, the system can fall
3533 * into OOM very easily.
3534 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3535 */
3536 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3537 low_kmem_size = 0;
3538 total_size = 0;
3539 for_each_online_node(nid) {
3540 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3541 z = &NODE_DATA(nid)->node_zones[zone_type];
3542 if (populated_zone(z)) {
3543 if (zone_type < ZONE_NORMAL)
3544 low_kmem_size += z->managed_pages;
3545 total_size += z->managed_pages;
3546 } else if (zone_type == ZONE_NORMAL) {
3547 /*
3548 * If any node has only lowmem, then node order
3549 * is preferred to allow kernel allocations
3550 * locally; otherwise, they can easily infringe
3551 * on other nodes when there is an abundance of
3552 * lowmem available to allocate from.
3553 */
3554 return ZONELIST_ORDER_NODE;
3555 }
3556 }
3557 }
3558 if (!low_kmem_size || /* there are no DMA area. */
3559 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3560 return ZONELIST_ORDER_NODE;
3561 /*
3562 * look into each node's config.
3563 * If there is a node whose DMA/DMA32 memory is very big area on
3564 * local memory, NODE_ORDER may be suitable.
3565 */
3566 average_size = total_size /
3567 (nodes_weight(node_states[N_MEMORY]) + 1);
3568 for_each_online_node(nid) {
3569 low_kmem_size = 0;
3570 total_size = 0;
3571 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3572 z = &NODE_DATA(nid)->node_zones[zone_type];
3573 if (populated_zone(z)) {
3574 if (zone_type < ZONE_NORMAL)
3575 low_kmem_size += z->present_pages;
3576 total_size += z->present_pages;
3577 }
3578 }
3579 if (low_kmem_size &&
3580 total_size > average_size && /* ignore small node */
3581 low_kmem_size > total_size * 70/100)
3582 return ZONELIST_ORDER_NODE;
3583 }
3584 return ZONELIST_ORDER_ZONE;
3585}
3586
3587static void set_zonelist_order(void)
3588{
3589 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3590 current_zonelist_order = default_zonelist_order();
3591 else
3592 current_zonelist_order = user_zonelist_order;
3593}
3594
3595static void build_zonelists(pg_data_t *pgdat)
3596{
3597 int j, node, load;
3598 enum zone_type i;
3599 nodemask_t used_mask;
3600 int local_node, prev_node;
3601 struct zonelist *zonelist;
3602 int order = current_zonelist_order;
3603
3604 /* initialize zonelists */
3605 for (i = 0; i < MAX_ZONELISTS; i++) {
3606 zonelist = pgdat->node_zonelists + i;
3607 zonelist->_zonerefs[0].zone = NULL;
3608 zonelist->_zonerefs[0].zone_idx = 0;
3609 }
3610
3611 /* NUMA-aware ordering of nodes */
3612 local_node = pgdat->node_id;
3613 load = nr_online_nodes;
3614 prev_node = local_node;
3615 nodes_clear(used_mask);
3616
3617 memset(node_order, 0, sizeof(node_order));
3618 j = 0;
3619
3620 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3621 /*
3622 * We don't want to pressure a particular node.
3623 * So adding penalty to the first node in same
3624 * distance group to make it round-robin.
3625 */
3626 if (node_distance(local_node, node) !=
3627 node_distance(local_node, prev_node))
3628 node_load[node] = load;
3629
3630 prev_node = node;
3631 load--;
3632 if (order == ZONELIST_ORDER_NODE)
3633 build_zonelists_in_node_order(pgdat, node);
3634 else
3635 node_order[j++] = node; /* remember order */
3636 }
3637
3638 if (order == ZONELIST_ORDER_ZONE) {
3639 /* calculate node order -- i.e., DMA last! */
3640 build_zonelists_in_zone_order(pgdat, j);
3641 }
3642
3643 build_thisnode_zonelists(pgdat);
3644}
3645
3646/* Construct the zonelist performance cache - see further mmzone.h */
3647static void build_zonelist_cache(pg_data_t *pgdat)
3648{
3649 struct zonelist *zonelist;
3650 struct zonelist_cache *zlc;
3651 struct zoneref *z;
3652
3653 zonelist = &pgdat->node_zonelists[0];
3654 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3655 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3656 for (z = zonelist->_zonerefs; z->zone; z++)
3657 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3658}
3659
3660#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3661/*
3662 * Return node id of node used for "local" allocations.
3663 * I.e., first node id of first zone in arg node's generic zonelist.
3664 * Used for initializing percpu 'numa_mem', which is used primarily
3665 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3666 */
3667int local_memory_node(int node)
3668{
3669 struct zone *zone;
3670
3671 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3672 gfp_zone(GFP_KERNEL),
3673 NULL,
3674 &zone);
3675 return zone->node;
3676}
3677#endif
3678
3679#else /* CONFIG_NUMA */
3680
3681static void set_zonelist_order(void)
3682{
3683 current_zonelist_order = ZONELIST_ORDER_ZONE;
3684}
3685
3686static void build_zonelists(pg_data_t *pgdat)
3687{
3688 int node, local_node;
3689 enum zone_type j;
3690 struct zonelist *zonelist;
3691
3692 local_node = pgdat->node_id;
3693
3694 zonelist = &pgdat->node_zonelists[0];
3695 j = build_zonelists_node(pgdat, zonelist, 0);
3696
3697 /*
3698 * Now we build the zonelist so that it contains the zones
3699 * of all the other nodes.
3700 * We don't want to pressure a particular node, so when
3701 * building the zones for node N, we make sure that the
3702 * zones coming right after the local ones are those from
3703 * node N+1 (modulo N)
3704 */
3705 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3706 if (!node_online(node))
3707 continue;
3708 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3709 }
3710 for (node = 0; node < local_node; node++) {
3711 if (!node_online(node))
3712 continue;
3713 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3714 }
3715
3716 zonelist->_zonerefs[j].zone = NULL;
3717 zonelist->_zonerefs[j].zone_idx = 0;
3718}
3719
3720/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3721static void build_zonelist_cache(pg_data_t *pgdat)
3722{
3723 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3724}
3725
3726#endif /* CONFIG_NUMA */
3727
3728/*
3729 * Boot pageset table. One per cpu which is going to be used for all
3730 * zones and all nodes. The parameters will be set in such a way
3731 * that an item put on a list will immediately be handed over to
3732 * the buddy list. This is safe since pageset manipulation is done
3733 * with interrupts disabled.
3734 *
3735 * The boot_pagesets must be kept even after bootup is complete for
3736 * unused processors and/or zones. They do play a role for bootstrapping
3737 * hotplugged processors.
3738 *
3739 * zoneinfo_show() and maybe other functions do
3740 * not check if the processor is online before following the pageset pointer.
3741 * Other parts of the kernel may not check if the zone is available.
3742 */
3743static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3744static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3745static void setup_zone_pageset(struct zone *zone);
3746
3747/*
3748 * Global mutex to protect against size modification of zonelists
3749 * as well as to serialize pageset setup for the new populated zone.
3750 */
3751DEFINE_MUTEX(zonelists_mutex);
3752
3753/* return values int ....just for stop_machine() */
3754static int __build_all_zonelists(void *data)
3755{
3756 int nid;
3757 int cpu;
3758 pg_data_t *self = data;
3759
3760#ifdef CONFIG_NUMA
3761 memset(node_load, 0, sizeof(node_load));
3762#endif
3763
3764 if (self && !node_online(self->node_id)) {
3765 build_zonelists(self);
3766 build_zonelist_cache(self);
3767 }
3768
3769 for_each_online_node(nid) {
3770 pg_data_t *pgdat = NODE_DATA(nid);
3771
3772 build_zonelists(pgdat);
3773 build_zonelist_cache(pgdat);
3774 }
3775
3776 /*
3777 * Initialize the boot_pagesets that are going to be used
3778 * for bootstrapping processors. The real pagesets for
3779 * each zone will be allocated later when the per cpu
3780 * allocator is available.
3781 *
3782 * boot_pagesets are used also for bootstrapping offline
3783 * cpus if the system is already booted because the pagesets
3784 * are needed to initialize allocators on a specific cpu too.
3785 * F.e. the percpu allocator needs the page allocator which
3786 * needs the percpu allocator in order to allocate its pagesets
3787 * (a chicken-egg dilemma).
3788 */
3789 for_each_possible_cpu(cpu) {
3790 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3791
3792#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3793 /*
3794 * We now know the "local memory node" for each node--
3795 * i.e., the node of the first zone in the generic zonelist.
3796 * Set up numa_mem percpu variable for on-line cpus. During
3797 * boot, only the boot cpu should be on-line; we'll init the
3798 * secondary cpus' numa_mem as they come on-line. During
3799 * node/memory hotplug, we'll fixup all on-line cpus.
3800 */
3801 if (cpu_online(cpu))
3802 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3803#endif
3804 }
3805
3806 return 0;
3807}
3808
3809/*
3810 * Called with zonelists_mutex held always
3811 * unless system_state == SYSTEM_BOOTING.
3812 */
3813void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3814{
3815 set_zonelist_order();
3816
3817 if (system_state == SYSTEM_BOOTING) {
3818 __build_all_zonelists(NULL);
3819 mminit_verify_zonelist();
3820 cpuset_init_current_mems_allowed();
3821 } else {
3822#ifdef CONFIG_MEMORY_HOTPLUG
3823 if (zone)
3824 setup_zone_pageset(zone);
3825#endif
3826 /* we have to stop all cpus to guarantee there is no user
3827 of zonelist */
3828 stop_machine(__build_all_zonelists, pgdat, NULL);
3829 /* cpuset refresh routine should be here */
3830 }
3831 vm_total_pages = nr_free_pagecache_pages();
3832 /*
3833 * Disable grouping by mobility if the number of pages in the
3834 * system is too low to allow the mechanism to work. It would be
3835 * more accurate, but expensive to check per-zone. This check is
3836 * made on memory-hotadd so a system can start with mobility
3837 * disabled and enable it later
3838 */
3839 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3840 page_group_by_mobility_disabled = 1;
3841 else
3842 page_group_by_mobility_disabled = 0;
3843
3844 printk("Built %i zonelists in %s order, mobility grouping %s. "
3845 "Total pages: %ld\n",
3846 nr_online_nodes,
3847 zonelist_order_name[current_zonelist_order],
3848 page_group_by_mobility_disabled ? "off" : "on",
3849 vm_total_pages);
3850#ifdef CONFIG_NUMA
3851 printk("Policy zone: %s\n", zone_names[policy_zone]);
3852#endif
3853}
3854
3855/*
3856 * Helper functions to size the waitqueue hash table.
3857 * Essentially these want to choose hash table sizes sufficiently
3858 * large so that collisions trying to wait on pages are rare.
3859 * But in fact, the number of active page waitqueues on typical
3860 * systems is ridiculously low, less than 200. So this is even
3861 * conservative, even though it seems large.
3862 *
3863 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3864 * waitqueues, i.e. the size of the waitq table given the number of pages.
3865 */
3866#define PAGES_PER_WAITQUEUE 256
3867
3868#ifndef CONFIG_MEMORY_HOTPLUG
3869static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3870{
3871 unsigned long size = 1;
3872
3873 pages /= PAGES_PER_WAITQUEUE;
3874
3875 while (size < pages)
3876 size <<= 1;
3877
3878 /*
3879 * Once we have dozens or even hundreds of threads sleeping
3880 * on IO we've got bigger problems than wait queue collision.
3881 * Limit the size of the wait table to a reasonable size.
3882 */
3883 size = min(size, 4096UL);
3884
3885 return max(size, 4UL);
3886}
3887#else
3888/*
3889 * A zone's size might be changed by hot-add, so it is not possible to determine
3890 * a suitable size for its wait_table. So we use the maximum size now.
3891 *
3892 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3893 *
3894 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3895 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3896 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3897 *
3898 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3899 * or more by the traditional way. (See above). It equals:
3900 *
3901 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3902 * ia64(16K page size) : = ( 8G + 4M)byte.
3903 * powerpc (64K page size) : = (32G +16M)byte.
3904 */
3905static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3906{
3907 return 4096UL;
3908}
3909#endif
3910
3911/*
3912 * This is an integer logarithm so that shifts can be used later
3913 * to extract the more random high bits from the multiplicative
3914 * hash function before the remainder is taken.
3915 */
3916static inline unsigned long wait_table_bits(unsigned long size)
3917{
3918 return ffz(~size);
3919}
3920
3921/*
3922 * Check if a pageblock contains reserved pages
3923 */
3924static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3925{
3926 unsigned long pfn;
3927
3928 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3929 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3930 return 1;
3931 }
3932 return 0;
3933}
3934
3935/*
3936 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3937 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3938 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3939 * higher will lead to a bigger reserve which will get freed as contiguous
3940 * blocks as reclaim kicks in
3941 */
3942static void setup_zone_migrate_reserve(struct zone *zone)
3943{
3944 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3945 struct page *page;
3946 unsigned long block_migratetype;
3947 int reserve;
3948 int old_reserve;
3949
3950 /*
3951 * Get the start pfn, end pfn and the number of blocks to reserve
3952 * We have to be careful to be aligned to pageblock_nr_pages to
3953 * make sure that we always check pfn_valid for the first page in
3954 * the block.
3955 */
3956 start_pfn = zone->zone_start_pfn;
3957 end_pfn = zone_end_pfn(zone);
3958 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3959 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3960 pageblock_order;
3961
3962 /*
3963 * Reserve blocks are generally in place to help high-order atomic
3964 * allocations that are short-lived. A min_free_kbytes value that
3965 * would result in more than 2 reserve blocks for atomic allocations
3966 * is assumed to be in place to help anti-fragmentation for the
3967 * future allocation of hugepages at runtime.
3968 */
3969 reserve = min(2, reserve);
3970 old_reserve = zone->nr_migrate_reserve_block;
3971
3972 /* When memory hot-add, we almost always need to do nothing */
3973 if (reserve == old_reserve)
3974 return;
3975 zone->nr_migrate_reserve_block = reserve;
3976
3977 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3978 if (!pfn_valid(pfn))
3979 continue;
3980 page = pfn_to_page(pfn);
3981
3982 /* Watch out for overlapping nodes */
3983 if (page_to_nid(page) != zone_to_nid(zone))
3984 continue;
3985
3986 block_migratetype = get_pageblock_migratetype(page);
3987
3988 /* Only test what is necessary when the reserves are not met */
3989 if (reserve > 0) {
3990 /*
3991 * Blocks with reserved pages will never free, skip
3992 * them.
3993 */
3994 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3995 if (pageblock_is_reserved(pfn, block_end_pfn))
3996 continue;
3997
3998 /* If this block is reserved, account for it */
3999 if (block_migratetype == MIGRATE_RESERVE) {
4000 reserve--;
4001 continue;
4002 }
4003
4004 /* Suitable for reserving if this block is movable */
4005 if (block_migratetype == MIGRATE_MOVABLE) {
4006 set_pageblock_migratetype(page,
4007 MIGRATE_RESERVE);
4008 move_freepages_block(zone, page,
4009 MIGRATE_RESERVE);
4010 reserve--;
4011 continue;
4012 }
4013 } else if (!old_reserve) {
4014 /*
4015 * At boot time we don't need to scan the whole zone
4016 * for turning off MIGRATE_RESERVE.
4017 */
4018 break;
4019 }
4020
4021 /*
4022 * If the reserve is met and this is a previous reserved block,
4023 * take it back
4024 */
4025 if (block_migratetype == MIGRATE_RESERVE) {
4026 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4027 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4028 }
4029 }
4030}
4031
4032/*
4033 * Initially all pages are reserved - free ones are freed
4034 * up by free_all_bootmem() once the early boot process is
4035 * done. Non-atomic initialization, single-pass.
4036 */
4037void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4038 unsigned long start_pfn, enum memmap_context context)
4039{
4040 struct page *page;
4041 unsigned long end_pfn = start_pfn + size;
4042 unsigned long pfn;
4043 struct zone *z;
4044
4045 if (highest_memmap_pfn < end_pfn - 1)
4046 highest_memmap_pfn = end_pfn - 1;
4047
4048 z = &NODE_DATA(nid)->node_zones[zone];
4049 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4050 /*
4051 * There can be holes in boot-time mem_map[]s
4052 * handed to this function. They do not
4053 * exist on hotplugged memory.
4054 */
4055 if (context == MEMMAP_EARLY) {
4056 if (!early_pfn_valid(pfn))
4057 continue;
4058 if (!early_pfn_in_nid(pfn, nid))
4059 continue;
4060 }
4061 page = pfn_to_page(pfn);
4062 set_page_links(page, zone, nid, pfn);
4063 mminit_verify_page_links(page, zone, nid, pfn);
4064 init_page_count(page);
4065 page_mapcount_reset(page);
4066 page_cpupid_reset_last(page);
4067 SetPageReserved(page);
4068 /*
4069 * Mark the block movable so that blocks are reserved for
4070 * movable at startup. This will force kernel allocations
4071 * to reserve their blocks rather than leaking throughout
4072 * the address space during boot when many long-lived
4073 * kernel allocations are made. Later some blocks near
4074 * the start are marked MIGRATE_RESERVE by
4075 * setup_zone_migrate_reserve()
4076 *
4077 * bitmap is created for zone's valid pfn range. but memmap
4078 * can be created for invalid pages (for alignment)
4079 * check here not to call set_pageblock_migratetype() against
4080 * pfn out of zone.
4081 */
4082 if ((z->zone_start_pfn <= pfn)
4083 && (pfn < zone_end_pfn(z))
4084 && !(pfn & (pageblock_nr_pages - 1)))
4085 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4086
4087 INIT_LIST_HEAD(&page->lru);
4088#ifdef WANT_PAGE_VIRTUAL
4089 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4090 if (!is_highmem_idx(zone))
4091 set_page_address(page, __va(pfn << PAGE_SHIFT));
4092#endif
4093 }
4094}
4095
4096static void __meminit zone_init_free_lists(struct zone *zone)
4097{
4098 int order, t;
4099 for_each_migratetype_order(order, t) {
4100 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4101 zone->free_area[order].nr_free = 0;
4102 }
4103}
4104
4105#ifndef __HAVE_ARCH_MEMMAP_INIT
4106#define memmap_init(size, nid, zone, start_pfn) \
4107 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4108#endif
4109
4110static int __meminit zone_batchsize(struct zone *zone)
4111{
4112#ifdef CONFIG_MMU
4113 int batch;
4114
4115 /*
4116 * The per-cpu-pages pools are set to around 1000th of the
4117 * size of the zone. But no more than 1/2 of a meg.
4118 *
4119 * OK, so we don't know how big the cache is. So guess.
4120 */
4121 batch = zone->managed_pages / 1024;
4122 if (batch * PAGE_SIZE > 512 * 1024)
4123 batch = (512 * 1024) / PAGE_SIZE;
4124 batch /= 4; /* We effectively *= 4 below */
4125 if (batch < 1)
4126 batch = 1;
4127
4128 /*
4129 * Clamp the batch to a 2^n - 1 value. Having a power
4130 * of 2 value was found to be more likely to have
4131 * suboptimal cache aliasing properties in some cases.
4132 *
4133 * For example if 2 tasks are alternately allocating
4134 * batches of pages, one task can end up with a lot
4135 * of pages of one half of the possible page colors
4136 * and the other with pages of the other colors.
4137 */
4138 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4139
4140 return batch;
4141
4142#else
4143 /* The deferral and batching of frees should be suppressed under NOMMU
4144 * conditions.
4145 *
4146 * The problem is that NOMMU needs to be able to allocate large chunks
4147 * of contiguous memory as there's no hardware page translation to
4148 * assemble apparent contiguous memory from discontiguous pages.
4149 *
4150 * Queueing large contiguous runs of pages for batching, however,
4151 * causes the pages to actually be freed in smaller chunks. As there
4152 * can be a significant delay between the individual batches being
4153 * recycled, this leads to the once large chunks of space being
4154 * fragmented and becoming unavailable for high-order allocations.
4155 */
4156 return 0;
4157#endif
4158}
4159
4160/*
4161 * pcp->high and pcp->batch values are related and dependent on one another:
4162 * ->batch must never be higher then ->high.
4163 * The following function updates them in a safe manner without read side
4164 * locking.
4165 *
4166 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4167 * those fields changing asynchronously (acording the the above rule).
4168 *
4169 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4170 * outside of boot time (or some other assurance that no concurrent updaters
4171 * exist).
4172 */
4173static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4174 unsigned long batch)
4175{
4176 /* start with a fail safe value for batch */
4177 pcp->batch = 1;
4178 smp_wmb();
4179
4180 /* Update high, then batch, in order */
4181 pcp->high = high;
4182 smp_wmb();
4183
4184 pcp->batch = batch;
4185}
4186
4187/* a companion to pageset_set_high() */
4188static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4189{
4190 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4191}
4192
4193static void pageset_init(struct per_cpu_pageset *p)
4194{
4195 struct per_cpu_pages *pcp;
4196 int migratetype;
4197
4198 memset(p, 0, sizeof(*p));
4199
4200 pcp = &p->pcp;
4201 pcp->count = 0;
4202 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4203 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4204}
4205
4206static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4207{
4208 pageset_init(p);
4209 pageset_set_batch(p, batch);
4210}
4211
4212/*
4213 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4214 * to the value high for the pageset p.
4215 */
4216static void pageset_set_high(struct per_cpu_pageset *p,
4217 unsigned long high)
4218{
4219 unsigned long batch = max(1UL, high / 4);
4220 if ((high / 4) > (PAGE_SHIFT * 8))
4221 batch = PAGE_SHIFT * 8;
4222
4223 pageset_update(&p->pcp, high, batch);
4224}
4225
4226static void __meminit pageset_set_high_and_batch(struct zone *zone,
4227 struct per_cpu_pageset *pcp)
4228{
4229 if (percpu_pagelist_fraction)
4230 pageset_set_high(pcp,
4231 (zone->managed_pages /
4232 percpu_pagelist_fraction));
4233 else
4234 pageset_set_batch(pcp, zone_batchsize(zone));
4235}
4236
4237static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4238{
4239 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4240
4241 pageset_init(pcp);
4242 pageset_set_high_and_batch(zone, pcp);
4243}
4244
4245static void __meminit setup_zone_pageset(struct zone *zone)
4246{
4247 int cpu;
4248 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4249 for_each_possible_cpu(cpu)
4250 zone_pageset_init(zone, cpu);
4251}
4252
4253/*
4254 * Allocate per cpu pagesets and initialize them.
4255 * Before this call only boot pagesets were available.
4256 */
4257void __init setup_per_cpu_pageset(void)
4258{
4259 struct zone *zone;
4260
4261 for_each_populated_zone(zone)
4262 setup_zone_pageset(zone);
4263}
4264
4265static noinline __init_refok
4266int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4267{
4268 int i;
4269 size_t alloc_size;
4270
4271 /*
4272 * The per-page waitqueue mechanism uses hashed waitqueues
4273 * per zone.
4274 */
4275 zone->wait_table_hash_nr_entries =
4276 wait_table_hash_nr_entries(zone_size_pages);
4277 zone->wait_table_bits =
4278 wait_table_bits(zone->wait_table_hash_nr_entries);
4279 alloc_size = zone->wait_table_hash_nr_entries
4280 * sizeof(wait_queue_head_t);
4281
4282 if (!slab_is_available()) {
4283 zone->wait_table = (wait_queue_head_t *)
4284 memblock_virt_alloc_node_nopanic(
4285 alloc_size, zone->zone_pgdat->node_id);
4286 } else {
4287 /*
4288 * This case means that a zone whose size was 0 gets new memory
4289 * via memory hot-add.
4290 * But it may be the case that a new node was hot-added. In
4291 * this case vmalloc() will not be able to use this new node's
4292 * memory - this wait_table must be initialized to use this new
4293 * node itself as well.
4294 * To use this new node's memory, further consideration will be
4295 * necessary.
4296 */
4297 zone->wait_table = vmalloc(alloc_size);
4298 }
4299 if (!zone->wait_table)
4300 return -ENOMEM;
4301
4302 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4303 init_waitqueue_head(zone->wait_table + i);
4304
4305 return 0;
4306}
4307
4308static __meminit void zone_pcp_init(struct zone *zone)
4309{
4310 /*
4311 * per cpu subsystem is not up at this point. The following code
4312 * relies on the ability of the linker to provide the
4313 * offset of a (static) per cpu variable into the per cpu area.
4314 */
4315 zone->pageset = &boot_pageset;
4316
4317 if (populated_zone(zone))
4318 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4319 zone->name, zone->present_pages,
4320 zone_batchsize(zone));
4321}
4322
4323int __meminit init_currently_empty_zone(struct zone *zone,
4324 unsigned long zone_start_pfn,
4325 unsigned long size,
4326 enum memmap_context context)
4327{
4328 struct pglist_data *pgdat = zone->zone_pgdat;
4329 int ret;
4330 ret = zone_wait_table_init(zone, size);
4331 if (ret)
4332 return ret;
4333 pgdat->nr_zones = zone_idx(zone) + 1;
4334
4335 zone->zone_start_pfn = zone_start_pfn;
4336
4337 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4338 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4339 pgdat->node_id,
4340 (unsigned long)zone_idx(zone),
4341 zone_start_pfn, (zone_start_pfn + size));
4342
4343 zone_init_free_lists(zone);
4344
4345 return 0;
4346}
4347
4348#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4349#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4350/*
4351 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4352 * Architectures may implement their own version but if add_active_range()
4353 * was used and there are no special requirements, this is a convenient
4354 * alternative
4355 */
4356int __meminit __early_pfn_to_nid(unsigned long pfn)
4357{
4358 unsigned long start_pfn, end_pfn;
4359 int nid;
4360 /*
4361 * NOTE: The following SMP-unsafe globals are only used early in boot
4362 * when the kernel is running single-threaded.
4363 */
4364 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4365 static int __meminitdata last_nid;
4366
4367 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4368 return last_nid;
4369
4370 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4371 if (nid != -1) {
4372 last_start_pfn = start_pfn;
4373 last_end_pfn = end_pfn;
4374 last_nid = nid;
4375 }
4376
4377 return nid;
4378}
4379#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4380
4381int __meminit early_pfn_to_nid(unsigned long pfn)
4382{
4383 int nid;
4384
4385 nid = __early_pfn_to_nid(pfn);
4386 if (nid >= 0)
4387 return nid;
4388 /* just returns 0 */
4389 return 0;
4390}
4391
4392#ifdef CONFIG_NODES_SPAN_OTHER_NODES
4393bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4394{
4395 int nid;
4396
4397 nid = __early_pfn_to_nid(pfn);
4398 if (nid >= 0 && nid != node)
4399 return false;
4400 return true;
4401}
4402#endif
4403
4404/**
4405 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4406 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4407 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4408 *
4409 * If an architecture guarantees that all ranges registered with
4410 * add_active_ranges() contain no holes and may be freed, this
4411 * this function may be used instead of calling memblock_free_early_nid()
4412 * manually.
4413 */
4414void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4415{
4416 unsigned long start_pfn, end_pfn;
4417 int i, this_nid;
4418
4419 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4420 start_pfn = min(start_pfn, max_low_pfn);
4421 end_pfn = min(end_pfn, max_low_pfn);
4422
4423 if (start_pfn < end_pfn)
4424 memblock_free_early_nid(PFN_PHYS(start_pfn),
4425 (end_pfn - start_pfn) << PAGE_SHIFT,
4426 this_nid);
4427 }
4428}
4429
4430/**
4431 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4432 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4433 *
4434 * If an architecture guarantees that all ranges registered with
4435 * add_active_ranges() contain no holes and may be freed, this
4436 * function may be used instead of calling memory_present() manually.
4437 */
4438void __init sparse_memory_present_with_active_regions(int nid)
4439{
4440 unsigned long start_pfn, end_pfn;
4441 int i, this_nid;
4442
4443 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4444 memory_present(this_nid, start_pfn, end_pfn);
4445}
4446
4447/**
4448 * get_pfn_range_for_nid - Return the start and end page frames for a node
4449 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4450 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4451 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4452 *
4453 * It returns the start and end page frame of a node based on information
4454 * provided by an arch calling add_active_range(). If called for a node
4455 * with no available memory, a warning is printed and the start and end
4456 * PFNs will be 0.
4457 */
4458void __meminit get_pfn_range_for_nid(unsigned int nid,
4459 unsigned long *start_pfn, unsigned long *end_pfn)
4460{
4461 unsigned long this_start_pfn, this_end_pfn;
4462 int i;
4463
4464 *start_pfn = -1UL;
4465 *end_pfn = 0;
4466
4467 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4468 *start_pfn = min(*start_pfn, this_start_pfn);
4469 *end_pfn = max(*end_pfn, this_end_pfn);
4470 }
4471
4472 if (*start_pfn == -1UL)
4473 *start_pfn = 0;
4474}
4475
4476/*
4477 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4478 * assumption is made that zones within a node are ordered in monotonic
4479 * increasing memory addresses so that the "highest" populated zone is used
4480 */
4481static void __init find_usable_zone_for_movable(void)
4482{
4483 int zone_index;
4484 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4485 if (zone_index == ZONE_MOVABLE)
4486 continue;
4487
4488 if (arch_zone_highest_possible_pfn[zone_index] >
4489 arch_zone_lowest_possible_pfn[zone_index])
4490 break;
4491 }
4492
4493 VM_BUG_ON(zone_index == -1);
4494 movable_zone = zone_index;
4495}
4496
4497/*
4498 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4499 * because it is sized independent of architecture. Unlike the other zones,
4500 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4501 * in each node depending on the size of each node and how evenly kernelcore
4502 * is distributed. This helper function adjusts the zone ranges
4503 * provided by the architecture for a given node by using the end of the
4504 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4505 * zones within a node are in order of monotonic increases memory addresses
4506 */
4507static void __meminit adjust_zone_range_for_zone_movable(int nid,
4508 unsigned long zone_type,
4509 unsigned long node_start_pfn,
4510 unsigned long node_end_pfn,
4511 unsigned long *zone_start_pfn,
4512 unsigned long *zone_end_pfn)
4513{
4514 /* Only adjust if ZONE_MOVABLE is on this node */
4515 if (zone_movable_pfn[nid]) {
4516 /* Size ZONE_MOVABLE */
4517 if (zone_type == ZONE_MOVABLE) {
4518 *zone_start_pfn = zone_movable_pfn[nid];
4519 *zone_end_pfn = min(node_end_pfn,
4520 arch_zone_highest_possible_pfn[movable_zone]);
4521
4522 /* Adjust for ZONE_MOVABLE starting within this range */
4523 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4524 *zone_end_pfn > zone_movable_pfn[nid]) {
4525 *zone_end_pfn = zone_movable_pfn[nid];
4526
4527 /* Check if this whole range is within ZONE_MOVABLE */
4528 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4529 *zone_start_pfn = *zone_end_pfn;
4530 }
4531}
4532
4533/*
4534 * Return the number of pages a zone spans in a node, including holes
4535 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4536 */
4537static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4538 unsigned long zone_type,
4539 unsigned long node_start_pfn,
4540 unsigned long node_end_pfn,
4541 unsigned long *ignored)
4542{
4543 unsigned long zone_start_pfn, zone_end_pfn;
4544
4545 /* Get the start and end of the zone */
4546 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4547 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4548 adjust_zone_range_for_zone_movable(nid, zone_type,
4549 node_start_pfn, node_end_pfn,
4550 &zone_start_pfn, &zone_end_pfn);
4551
4552 /* Check that this node has pages within the zone's required range */
4553 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4554 return 0;
4555
4556 /* Move the zone boundaries inside the node if necessary */
4557 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4558 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4559
4560 /* Return the spanned pages */
4561 return zone_end_pfn - zone_start_pfn;
4562}
4563
4564/*
4565 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4566 * then all holes in the requested range will be accounted for.
4567 */
4568unsigned long __meminit __absent_pages_in_range(int nid,
4569 unsigned long range_start_pfn,
4570 unsigned long range_end_pfn)
4571{
4572 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4573 unsigned long start_pfn, end_pfn;
4574 int i;
4575
4576 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4577 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4578 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4579 nr_absent -= end_pfn - start_pfn;
4580 }
4581 return nr_absent;
4582}
4583
4584/**
4585 * absent_pages_in_range - Return number of page frames in holes within a range
4586 * @start_pfn: The start PFN to start searching for holes
4587 * @end_pfn: The end PFN to stop searching for holes
4588 *
4589 * It returns the number of pages frames in memory holes within a range.
4590 */
4591unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4592 unsigned long end_pfn)
4593{
4594 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4595}
4596
4597/* Return the number of page frames in holes in a zone on a node */
4598static unsigned long __meminit zone_absent_pages_in_node(int nid,
4599 unsigned long zone_type,
4600 unsigned long node_start_pfn,
4601 unsigned long node_end_pfn,
4602 unsigned long *ignored)
4603{
4604 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4605 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4606 unsigned long zone_start_pfn, zone_end_pfn;
4607
4608 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4609 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4610
4611 adjust_zone_range_for_zone_movable(nid, zone_type,
4612 node_start_pfn, node_end_pfn,
4613 &zone_start_pfn, &zone_end_pfn);
4614 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4615}
4616
4617#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4618static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4619 unsigned long zone_type,
4620 unsigned long node_start_pfn,
4621 unsigned long node_end_pfn,
4622 unsigned long *zones_size)
4623{
4624 return zones_size[zone_type];
4625}
4626
4627static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4628 unsigned long zone_type,
4629 unsigned long node_start_pfn,
4630 unsigned long node_end_pfn,
4631 unsigned long *zholes_size)
4632{
4633 if (!zholes_size)
4634 return 0;
4635
4636 return zholes_size[zone_type];
4637}
4638
4639#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4640
4641static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4642 unsigned long node_start_pfn,
4643 unsigned long node_end_pfn,
4644 unsigned long *zones_size,
4645 unsigned long *zholes_size)
4646{
4647 unsigned long realtotalpages, totalpages = 0;
4648 enum zone_type i;
4649
4650 for (i = 0; i < MAX_NR_ZONES; i++)
4651 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4652 node_start_pfn,
4653 node_end_pfn,
4654 zones_size);
4655 pgdat->node_spanned_pages = totalpages;
4656
4657 realtotalpages = totalpages;
4658 for (i = 0; i < MAX_NR_ZONES; i++)
4659 realtotalpages -=
4660 zone_absent_pages_in_node(pgdat->node_id, i,
4661 node_start_pfn, node_end_pfn,
4662 zholes_size);
4663 pgdat->node_present_pages = realtotalpages;
4664 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4665 realtotalpages);
4666}
4667
4668#ifndef CONFIG_SPARSEMEM
4669/*
4670 * Calculate the size of the zone->blockflags rounded to an unsigned long
4671 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4672 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4673 * round what is now in bits to nearest long in bits, then return it in
4674 * bytes.
4675 */
4676static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4677{
4678 unsigned long usemapsize;
4679
4680 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4681 usemapsize = roundup(zonesize, pageblock_nr_pages);
4682 usemapsize = usemapsize >> pageblock_order;
4683 usemapsize *= NR_PAGEBLOCK_BITS;
4684 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4685
4686 return usemapsize / 8;
4687}
4688
4689static void __init setup_usemap(struct pglist_data *pgdat,
4690 struct zone *zone,
4691 unsigned long zone_start_pfn,
4692 unsigned long zonesize)
4693{
4694 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4695 zone->pageblock_flags = NULL;
4696 if (usemapsize)
4697 zone->pageblock_flags =
4698 memblock_virt_alloc_node_nopanic(usemapsize,
4699 pgdat->node_id);
4700}
4701#else
4702static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4703 unsigned long zone_start_pfn, unsigned long zonesize) {}
4704#endif /* CONFIG_SPARSEMEM */
4705
4706#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4707
4708/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4709void __paginginit set_pageblock_order(void)
4710{
4711 unsigned int order;
4712
4713 /* Check that pageblock_nr_pages has not already been setup */
4714 if (pageblock_order)
4715 return;
4716
4717 if (HPAGE_SHIFT > PAGE_SHIFT)
4718 order = HUGETLB_PAGE_ORDER;
4719 else
4720 order = MAX_ORDER - 1;
4721
4722 /*
4723 * Assume the largest contiguous order of interest is a huge page.
4724 * This value may be variable depending on boot parameters on IA64 and
4725 * powerpc.
4726 */
4727 pageblock_order = order;
4728}
4729#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4730
4731/*
4732 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4733 * is unused as pageblock_order is set at compile-time. See
4734 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4735 * the kernel config
4736 */
4737void __paginginit set_pageblock_order(void)
4738{
4739}
4740
4741#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4742
4743static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4744 unsigned long present_pages)
4745{
4746 unsigned long pages = spanned_pages;
4747
4748 /*
4749 * Provide a more accurate estimation if there are holes within
4750 * the zone and SPARSEMEM is in use. If there are holes within the
4751 * zone, each populated memory region may cost us one or two extra
4752 * memmap pages due to alignment because memmap pages for each
4753 * populated regions may not naturally algined on page boundary.
4754 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4755 */
4756 if (spanned_pages > present_pages + (present_pages >> 4) &&
4757 IS_ENABLED(CONFIG_SPARSEMEM))
4758 pages = present_pages;
4759
4760 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4761}
4762
4763/*
4764 * Set up the zone data structures:
4765 * - mark all pages reserved
4766 * - mark all memory queues empty
4767 * - clear the memory bitmaps
4768 *
4769 * NOTE: pgdat should get zeroed by caller.
4770 */
4771static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4772 unsigned long node_start_pfn, unsigned long node_end_pfn,
4773 unsigned long *zones_size, unsigned long *zholes_size)
4774{
4775 enum zone_type j;
4776 int nid = pgdat->node_id;
4777 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4778 int ret;
4779
4780 pgdat_resize_init(pgdat);
4781#ifdef CONFIG_NUMA_BALANCING
4782 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4783 pgdat->numabalancing_migrate_nr_pages = 0;
4784 pgdat->numabalancing_migrate_next_window = jiffies;
4785#endif
4786 init_waitqueue_head(&pgdat->kswapd_wait);
4787 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4788 pgdat_page_cgroup_init(pgdat);
4789
4790 for (j = 0; j < MAX_NR_ZONES; j++) {
4791 struct zone *zone = pgdat->node_zones + j;
4792 unsigned long size, realsize, freesize, memmap_pages;
4793
4794 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4795 node_end_pfn, zones_size);
4796 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4797 node_start_pfn,
4798 node_end_pfn,
4799 zholes_size);
4800
4801 /*
4802 * Adjust freesize so that it accounts for how much memory
4803 * is used by this zone for memmap. This affects the watermark
4804 * and per-cpu initialisations
4805 */
4806 memmap_pages = calc_memmap_size(size, realsize);
4807 if (freesize >= memmap_pages) {
4808 freesize -= memmap_pages;
4809 if (memmap_pages)
4810 printk(KERN_DEBUG
4811 " %s zone: %lu pages used for memmap\n",
4812 zone_names[j], memmap_pages);
4813 } else
4814 printk(KERN_WARNING
4815 " %s zone: %lu pages exceeds freesize %lu\n",
4816 zone_names[j], memmap_pages, freesize);
4817
4818 /* Account for reserved pages */
4819 if (j == 0 && freesize > dma_reserve) {
4820 freesize -= dma_reserve;
4821 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4822 zone_names[0], dma_reserve);
4823 }
4824
4825 if (!is_highmem_idx(j))
4826 nr_kernel_pages += freesize;
4827 /* Charge for highmem memmap if there are enough kernel pages */
4828 else if (nr_kernel_pages > memmap_pages * 2)
4829 nr_kernel_pages -= memmap_pages;
4830 nr_all_pages += freesize;
4831
4832 zone->spanned_pages = size;
4833 zone->present_pages = realsize;
4834 /*
4835 * Set an approximate value for lowmem here, it will be adjusted
4836 * when the bootmem allocator frees pages into the buddy system.
4837 * And all highmem pages will be managed by the buddy system.
4838 */
4839 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4840#ifdef CONFIG_NUMA
4841 zone->node = nid;
4842 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4843 / 100;
4844 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4845#endif
4846 zone->name = zone_names[j];
4847 spin_lock_init(&zone->lock);
4848 spin_lock_init(&zone->lru_lock);
4849 zone_seqlock_init(zone);
4850 zone->zone_pgdat = pgdat;
4851 zone_pcp_init(zone);
4852
4853 /* For bootup, initialized properly in watermark setup */
4854 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4855
4856 lruvec_init(&zone->lruvec);
4857 if (!size)
4858 continue;
4859
4860 set_pageblock_order();
4861 setup_usemap(pgdat, zone, zone_start_pfn, size);
4862 ret = init_currently_empty_zone(zone, zone_start_pfn,
4863 size, MEMMAP_EARLY);
4864 BUG_ON(ret);
4865 memmap_init(size, nid, j, zone_start_pfn);
4866 zone_start_pfn += size;
4867 }
4868}
4869
4870static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4871{
4872 /* Skip empty nodes */
4873 if (!pgdat->node_spanned_pages)
4874 return;
4875
4876#ifdef CONFIG_FLAT_NODE_MEM_MAP
4877 /* ia64 gets its own node_mem_map, before this, without bootmem */
4878 if (!pgdat->node_mem_map) {
4879 unsigned long size, start, end;
4880 struct page *map;
4881
4882 /*
4883 * The zone's endpoints aren't required to be MAX_ORDER
4884 * aligned but the node_mem_map endpoints must be in order
4885 * for the buddy allocator to function correctly.
4886 */
4887 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4888 end = pgdat_end_pfn(pgdat);
4889 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4890 size = (end - start) * sizeof(struct page);
4891 map = alloc_remap(pgdat->node_id, size);
4892 if (!map)
4893 map = memblock_virt_alloc_node_nopanic(size,
4894 pgdat->node_id);
4895 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4896 }
4897#ifndef CONFIG_NEED_MULTIPLE_NODES
4898 /*
4899 * With no DISCONTIG, the global mem_map is just set as node 0's
4900 */
4901 if (pgdat == NODE_DATA(0)) {
4902 mem_map = NODE_DATA(0)->node_mem_map;
4903#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4904 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4905 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4906#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4907 }
4908#endif
4909#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4910}
4911
4912void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4913 unsigned long node_start_pfn, unsigned long *zholes_size)
4914{
4915 pg_data_t *pgdat = NODE_DATA(nid);
4916 unsigned long start_pfn = 0;
4917 unsigned long end_pfn = 0;
4918
4919 /* pg_data_t should be reset to zero when it's allocated */
4920 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4921
4922 pgdat->node_id = nid;
4923 pgdat->node_start_pfn = node_start_pfn;
4924 if (node_state(nid, N_MEMORY))
4925 init_zone_allows_reclaim(nid);
4926#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4927 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4928#endif
4929 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4930 zones_size, zholes_size);
4931
4932 alloc_node_mem_map(pgdat);
4933#ifdef CONFIG_FLAT_NODE_MEM_MAP
4934 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4935 nid, (unsigned long)pgdat,
4936 (unsigned long)pgdat->node_mem_map);
4937#endif
4938
4939 free_area_init_core(pgdat, start_pfn, end_pfn,
4940 zones_size, zholes_size);
4941}
4942
4943#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4944
4945#if MAX_NUMNODES > 1
4946/*
4947 * Figure out the number of possible node ids.
4948 */
4949void __init setup_nr_node_ids(void)
4950{
4951 unsigned int node;
4952 unsigned int highest = 0;
4953
4954 for_each_node_mask(node, node_possible_map)
4955 highest = node;
4956 nr_node_ids = highest + 1;
4957}
4958#endif
4959
4960/**
4961 * node_map_pfn_alignment - determine the maximum internode alignment
4962 *
4963 * This function should be called after node map is populated and sorted.
4964 * It calculates the maximum power of two alignment which can distinguish
4965 * all the nodes.
4966 *
4967 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4968 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4969 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4970 * shifted, 1GiB is enough and this function will indicate so.
4971 *
4972 * This is used to test whether pfn -> nid mapping of the chosen memory
4973 * model has fine enough granularity to avoid incorrect mapping for the
4974 * populated node map.
4975 *
4976 * Returns the determined alignment in pfn's. 0 if there is no alignment
4977 * requirement (single node).
4978 */
4979unsigned long __init node_map_pfn_alignment(void)
4980{
4981 unsigned long accl_mask = 0, last_end = 0;
4982 unsigned long start, end, mask;
4983 int last_nid = -1;
4984 int i, nid;
4985
4986 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4987 if (!start || last_nid < 0 || last_nid == nid) {
4988 last_nid = nid;
4989 last_end = end;
4990 continue;
4991 }
4992
4993 /*
4994 * Start with a mask granular enough to pin-point to the
4995 * start pfn and tick off bits one-by-one until it becomes
4996 * too coarse to separate the current node from the last.
4997 */
4998 mask = ~((1 << __ffs(start)) - 1);
4999 while (mask && last_end <= (start & (mask << 1)))
5000 mask <<= 1;
5001
5002 /* accumulate all internode masks */
5003 accl_mask |= mask;
5004 }
5005
5006 /* convert mask to number of pages */
5007 return ~accl_mask + 1;
5008}
5009
5010/* Find the lowest pfn for a node */
5011static unsigned long __init find_min_pfn_for_node(int nid)
5012{
5013 unsigned long min_pfn = ULONG_MAX;
5014 unsigned long start_pfn;
5015 int i;
5016
5017 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5018 min_pfn = min(min_pfn, start_pfn);
5019
5020 if (min_pfn == ULONG_MAX) {
5021 printk(KERN_WARNING
5022 "Could not find start_pfn for node %d\n", nid);
5023 return 0;
5024 }
5025
5026 return min_pfn;
5027}
5028
5029/**
5030 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5031 *
5032 * It returns the minimum PFN based on information provided via
5033 * add_active_range().
5034 */
5035unsigned long __init find_min_pfn_with_active_regions(void)
5036{
5037 return find_min_pfn_for_node(MAX_NUMNODES);
5038}
5039
5040/*
5041 * early_calculate_totalpages()
5042 * Sum pages in active regions for movable zone.
5043 * Populate N_MEMORY for calculating usable_nodes.
5044 */
5045static unsigned long __init early_calculate_totalpages(void)
5046{
5047 unsigned long totalpages = 0;
5048 unsigned long start_pfn, end_pfn;
5049 int i, nid;
5050
5051 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5052 unsigned long pages = end_pfn - start_pfn;
5053
5054 totalpages += pages;
5055 if (pages)
5056 node_set_state(nid, N_MEMORY);
5057 }
5058 return totalpages;
5059}
5060
5061/*
5062 * Find the PFN the Movable zone begins in each node. Kernel memory
5063 * is spread evenly between nodes as long as the nodes have enough
5064 * memory. When they don't, some nodes will have more kernelcore than
5065 * others
5066 */
5067static void __init find_zone_movable_pfns_for_nodes(void)
5068{
5069 int i, nid;
5070 unsigned long usable_startpfn;
5071 unsigned long kernelcore_node, kernelcore_remaining;
5072 /* save the state before borrow the nodemask */
5073 nodemask_t saved_node_state = node_states[N_MEMORY];
5074 unsigned long totalpages = early_calculate_totalpages();
5075 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5076 struct memblock_region *r;
5077
5078 /* Need to find movable_zone earlier when movable_node is specified. */
5079 find_usable_zone_for_movable();
5080
5081 /*
5082 * If movable_node is specified, ignore kernelcore and movablecore
5083 * options.
5084 */
5085 if (movable_node_is_enabled()) {
5086 for_each_memblock(memory, r) {
5087 if (!memblock_is_hotpluggable(r))
5088 continue;
5089
5090 nid = r->nid;
5091
5092 usable_startpfn = PFN_DOWN(r->base);
5093 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5094 min(usable_startpfn, zone_movable_pfn[nid]) :
5095 usable_startpfn;
5096 }
5097
5098 goto out2;
5099 }
5100
5101 /*
5102 * If movablecore=nn[KMG] was specified, calculate what size of
5103 * kernelcore that corresponds so that memory usable for
5104 * any allocation type is evenly spread. If both kernelcore
5105 * and movablecore are specified, then the value of kernelcore
5106 * will be used for required_kernelcore if it's greater than
5107 * what movablecore would have allowed.
5108 */
5109 if (required_movablecore) {
5110 unsigned long corepages;
5111
5112 /*
5113 * Round-up so that ZONE_MOVABLE is at least as large as what
5114 * was requested by the user
5115 */
5116 required_movablecore =
5117 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5118 corepages = totalpages - required_movablecore;
5119
5120 required_kernelcore = max(required_kernelcore, corepages);
5121 }
5122
5123 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5124 if (!required_kernelcore)
5125 goto out;
5126
5127 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5128 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5129
5130restart:
5131 /* Spread kernelcore memory as evenly as possible throughout nodes */
5132 kernelcore_node = required_kernelcore / usable_nodes;
5133 for_each_node_state(nid, N_MEMORY) {
5134 unsigned long start_pfn, end_pfn;
5135
5136 /*
5137 * Recalculate kernelcore_node if the division per node
5138 * now exceeds what is necessary to satisfy the requested
5139 * amount of memory for the kernel
5140 */
5141 if (required_kernelcore < kernelcore_node)
5142 kernelcore_node = required_kernelcore / usable_nodes;
5143
5144 /*
5145 * As the map is walked, we track how much memory is usable
5146 * by the kernel using kernelcore_remaining. When it is
5147 * 0, the rest of the node is usable by ZONE_MOVABLE
5148 */
5149 kernelcore_remaining = kernelcore_node;
5150
5151 /* Go through each range of PFNs within this node */
5152 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5153 unsigned long size_pages;
5154
5155 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5156 if (start_pfn >= end_pfn)
5157 continue;
5158
5159 /* Account for what is only usable for kernelcore */
5160 if (start_pfn < usable_startpfn) {
5161 unsigned long kernel_pages;
5162 kernel_pages = min(end_pfn, usable_startpfn)
5163 - start_pfn;
5164
5165 kernelcore_remaining -= min(kernel_pages,
5166 kernelcore_remaining);
5167 required_kernelcore -= min(kernel_pages,
5168 required_kernelcore);
5169
5170 /* Continue if range is now fully accounted */
5171 if (end_pfn <= usable_startpfn) {
5172
5173 /*
5174 * Push zone_movable_pfn to the end so
5175 * that if we have to rebalance
5176 * kernelcore across nodes, we will
5177 * not double account here
5178 */
5179 zone_movable_pfn[nid] = end_pfn;
5180 continue;
5181 }
5182 start_pfn = usable_startpfn;
5183 }
5184
5185 /*
5186 * The usable PFN range for ZONE_MOVABLE is from
5187 * start_pfn->end_pfn. Calculate size_pages as the
5188 * number of pages used as kernelcore
5189 */
5190 size_pages = end_pfn - start_pfn;
5191 if (size_pages > kernelcore_remaining)
5192 size_pages = kernelcore_remaining;
5193 zone_movable_pfn[nid] = start_pfn + size_pages;
5194
5195 /*
5196 * Some kernelcore has been met, update counts and
5197 * break if the kernelcore for this node has been
5198 * satisfied
5199 */
5200 required_kernelcore -= min(required_kernelcore,
5201 size_pages);
5202 kernelcore_remaining -= size_pages;
5203 if (!kernelcore_remaining)
5204 break;
5205 }
5206 }
5207
5208 /*
5209 * If there is still required_kernelcore, we do another pass with one
5210 * less node in the count. This will push zone_movable_pfn[nid] further
5211 * along on the nodes that still have memory until kernelcore is
5212 * satisfied
5213 */
5214 usable_nodes--;
5215 if (usable_nodes && required_kernelcore > usable_nodes)
5216 goto restart;
5217
5218out2:
5219 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5220 for (nid = 0; nid < MAX_NUMNODES; nid++)
5221 zone_movable_pfn[nid] =
5222 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5223
5224out:
5225 /* restore the node_state */
5226 node_states[N_MEMORY] = saved_node_state;
5227}
5228
5229/* Any regular or high memory on that node ? */
5230static void check_for_memory(pg_data_t *pgdat, int nid)
5231{
5232 enum zone_type zone_type;
5233
5234 if (N_MEMORY == N_NORMAL_MEMORY)
5235 return;
5236
5237 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5238 struct zone *zone = &pgdat->node_zones[zone_type];
5239 if (populated_zone(zone)) {
5240 node_set_state(nid, N_HIGH_MEMORY);
5241 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5242 zone_type <= ZONE_NORMAL)
5243 node_set_state(nid, N_NORMAL_MEMORY);
5244 break;
5245 }
5246 }
5247}
5248
5249/**
5250 * free_area_init_nodes - Initialise all pg_data_t and zone data
5251 * @max_zone_pfn: an array of max PFNs for each zone
5252 *
5253 * This will call free_area_init_node() for each active node in the system.
5254 * Using the page ranges provided by add_active_range(), the size of each
5255 * zone in each node and their holes is calculated. If the maximum PFN
5256 * between two adjacent zones match, it is assumed that the zone is empty.
5257 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5258 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5259 * starts where the previous one ended. For example, ZONE_DMA32 starts
5260 * at arch_max_dma_pfn.
5261 */
5262void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5263{
5264 unsigned long start_pfn, end_pfn;
5265 int i, nid;
5266
5267 /* Record where the zone boundaries are */
5268 memset(arch_zone_lowest_possible_pfn, 0,
5269 sizeof(arch_zone_lowest_possible_pfn));
5270 memset(arch_zone_highest_possible_pfn, 0,
5271 sizeof(arch_zone_highest_possible_pfn));
5272 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5273 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5274 for (i = 1; i < MAX_NR_ZONES; i++) {
5275 if (i == ZONE_MOVABLE)
5276 continue;
5277 arch_zone_lowest_possible_pfn[i] =
5278 arch_zone_highest_possible_pfn[i-1];
5279 arch_zone_highest_possible_pfn[i] =
5280 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5281 }
5282 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5283 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5284
5285 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5286 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5287 find_zone_movable_pfns_for_nodes();
5288
5289 /* Print out the zone ranges */
5290 printk("Zone ranges:\n");
5291 for (i = 0; i < MAX_NR_ZONES; i++) {
5292 if (i == ZONE_MOVABLE)
5293 continue;
5294 printk(KERN_CONT " %-8s ", zone_names[i]);
5295 if (arch_zone_lowest_possible_pfn[i] ==
5296 arch_zone_highest_possible_pfn[i])
5297 printk(KERN_CONT "empty\n");
5298 else
5299 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5300 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5301 (arch_zone_highest_possible_pfn[i]
5302 << PAGE_SHIFT) - 1);
5303 }
5304
5305 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5306 printk("Movable zone start for each node\n");
5307 for (i = 0; i < MAX_NUMNODES; i++) {
5308 if (zone_movable_pfn[i])
5309 printk(" Node %d: %#010lx\n", i,
5310 zone_movable_pfn[i] << PAGE_SHIFT);
5311 }
5312
5313 /* Print out the early node map */
5314 printk("Early memory node ranges\n");
5315 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5316 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5317 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5318
5319 /* Initialise every node */
5320 mminit_verify_pageflags_layout();
5321 setup_nr_node_ids();
5322 for_each_online_node(nid) {
5323 pg_data_t *pgdat = NODE_DATA(nid);
5324 free_area_init_node(nid, NULL,
5325 find_min_pfn_for_node(nid), NULL);
5326
5327 /* Any memory on that node */
5328 if (pgdat->node_present_pages)
5329 node_set_state(nid, N_MEMORY);
5330 check_for_memory(pgdat, nid);
5331 }
5332}
5333
5334static int __init cmdline_parse_core(char *p, unsigned long *core)
5335{
5336 unsigned long long coremem;
5337 if (!p)
5338 return -EINVAL;
5339
5340 coremem = memparse(p, &p);
5341 *core = coremem >> PAGE_SHIFT;
5342
5343 /* Paranoid check that UL is enough for the coremem value */
5344 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5345
5346 return 0;
5347}
5348
5349/*
5350 * kernelcore=size sets the amount of memory for use for allocations that
5351 * cannot be reclaimed or migrated.
5352 */
5353static int __init cmdline_parse_kernelcore(char *p)
5354{
5355 return cmdline_parse_core(p, &required_kernelcore);
5356}
5357
5358/*
5359 * movablecore=size sets the amount of memory for use for allocations that
5360 * can be reclaimed or migrated.
5361 */
5362static int __init cmdline_parse_movablecore(char *p)
5363{
5364 return cmdline_parse_core(p, &required_movablecore);
5365}
5366
5367early_param("kernelcore", cmdline_parse_kernelcore);
5368early_param("movablecore", cmdline_parse_movablecore);
5369
5370#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5371
5372void adjust_managed_page_count(struct page *page, long count)
5373{
5374 spin_lock(&managed_page_count_lock);
5375 page_zone(page)->managed_pages += count;
5376 totalram_pages += count;
5377#ifdef CONFIG_HIGHMEM
5378 if (PageHighMem(page))
5379 totalhigh_pages += count;
5380#endif
5381 spin_unlock(&managed_page_count_lock);
5382}
5383EXPORT_SYMBOL(adjust_managed_page_count);
5384
5385unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5386{
5387 void *pos;
5388 unsigned long pages = 0;
5389
5390 start = (void *)PAGE_ALIGN((unsigned long)start);
5391 end = (void *)((unsigned long)end & PAGE_MASK);
5392 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5393 if ((unsigned int)poison <= 0xFF)
5394 memset(pos, poison, PAGE_SIZE);
5395 free_reserved_page(virt_to_page(pos));
5396 }
5397
5398 if (pages && s)
5399 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5400 s, pages << (PAGE_SHIFT - 10), start, end);
5401
5402 return pages;
5403}
5404EXPORT_SYMBOL(free_reserved_area);
5405
5406#ifdef CONFIG_HIGHMEM
5407void free_highmem_page(struct page *page)
5408{
5409 __free_reserved_page(page);
5410 totalram_pages++;
5411 page_zone(page)->managed_pages++;
5412 totalhigh_pages++;
5413}
5414#endif
5415
5416
5417void __init mem_init_print_info(const char *str)
5418{
5419 unsigned long physpages, codesize, datasize, rosize, bss_size;
5420 unsigned long init_code_size, init_data_size;
5421
5422 physpages = get_num_physpages();
5423 codesize = _etext - _stext;
5424 datasize = _edata - _sdata;
5425 rosize = __end_rodata - __start_rodata;
5426 bss_size = __bss_stop - __bss_start;
5427 init_data_size = __init_end - __init_begin;
5428 init_code_size = _einittext - _sinittext;
5429
5430 /*
5431 * Detect special cases and adjust section sizes accordingly:
5432 * 1) .init.* may be embedded into .data sections
5433 * 2) .init.text.* may be out of [__init_begin, __init_end],
5434 * please refer to arch/tile/kernel/vmlinux.lds.S.
5435 * 3) .rodata.* may be embedded into .text or .data sections.
5436 */
5437#define adj_init_size(start, end, size, pos, adj) \
5438 do { \
5439 if (start <= pos && pos < end && size > adj) \
5440 size -= adj; \
5441 } while (0)
5442
5443 adj_init_size(__init_begin, __init_end, init_data_size,
5444 _sinittext, init_code_size);
5445 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5446 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5447 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5448 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5449
5450#undef adj_init_size
5451
5452 printk("Memory: %luK/%luK available "
5453 "(%luK kernel code, %luK rwdata, %luK rodata, "
5454 "%luK init, %luK bss, %luK reserved"
5455#ifdef CONFIG_HIGHMEM
5456 ", %luK highmem"
5457#endif
5458 "%s%s)\n",
5459 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5460 codesize >> 10, datasize >> 10, rosize >> 10,
5461 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5462 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5463#ifdef CONFIG_HIGHMEM
5464 totalhigh_pages << (PAGE_SHIFT-10),
5465#endif
5466 str ? ", " : "", str ? str : "");
5467}
5468
5469/**
5470 * set_dma_reserve - set the specified number of pages reserved in the first zone
5471 * @new_dma_reserve: The number of pages to mark reserved
5472 *
5473 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5474 * In the DMA zone, a significant percentage may be consumed by kernel image
5475 * and other unfreeable allocations which can skew the watermarks badly. This
5476 * function may optionally be used to account for unfreeable pages in the
5477 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5478 * smaller per-cpu batchsize.
5479 */
5480void __init set_dma_reserve(unsigned long new_dma_reserve)
5481{
5482 dma_reserve = new_dma_reserve;
5483}
5484
5485void __init free_area_init(unsigned long *zones_size)
5486{
5487 free_area_init_node(0, zones_size,
5488 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5489}
5490
5491static int page_alloc_cpu_notify(struct notifier_block *self,
5492 unsigned long action, void *hcpu)
5493{
5494 int cpu = (unsigned long)hcpu;
5495
5496 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5497 lru_add_drain_cpu(cpu);
5498 drain_pages(cpu);
5499
5500 /*
5501 * Spill the event counters of the dead processor
5502 * into the current processors event counters.
5503 * This artificially elevates the count of the current
5504 * processor.
5505 */
5506 vm_events_fold_cpu(cpu);
5507
5508 /*
5509 * Zero the differential counters of the dead processor
5510 * so that the vm statistics are consistent.
5511 *
5512 * This is only okay since the processor is dead and cannot
5513 * race with what we are doing.
5514 */
5515 cpu_vm_stats_fold(cpu);
5516 }
5517 return NOTIFY_OK;
5518}
5519
5520void __init page_alloc_init(void)
5521{
5522 hotcpu_notifier(page_alloc_cpu_notify, 0);
5523}
5524
5525/*
5526 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5527 * or min_free_kbytes changes.
5528 */
5529static void calculate_totalreserve_pages(void)
5530{
5531 struct pglist_data *pgdat;
5532 unsigned long reserve_pages = 0;
5533 enum zone_type i, j;
5534
5535 for_each_online_pgdat(pgdat) {
5536 for (i = 0; i < MAX_NR_ZONES; i++) {
5537 struct zone *zone = pgdat->node_zones + i;
5538 unsigned long max = 0;
5539
5540 /* Find valid and maximum lowmem_reserve in the zone */
5541 for (j = i; j < MAX_NR_ZONES; j++) {
5542 if (zone->lowmem_reserve[j] > max)
5543 max = zone->lowmem_reserve[j];
5544 }
5545
5546 /* we treat the high watermark as reserved pages. */
5547 max += high_wmark_pages(zone);
5548
5549 if (max > zone->managed_pages)
5550 max = zone->managed_pages;
5551 reserve_pages += max;
5552 /*
5553 * Lowmem reserves are not available to
5554 * GFP_HIGHUSER page cache allocations and
5555 * kswapd tries to balance zones to their high
5556 * watermark. As a result, neither should be
5557 * regarded as dirtyable memory, to prevent a
5558 * situation where reclaim has to clean pages
5559 * in order to balance the zones.
5560 */
5561 zone->dirty_balance_reserve = max;
5562 }
5563 }
5564 dirty_balance_reserve = reserve_pages;
5565 totalreserve_pages = reserve_pages;
5566}
5567
5568/*
5569 * setup_per_zone_lowmem_reserve - called whenever
5570 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5571 * has a correct pages reserved value, so an adequate number of
5572 * pages are left in the zone after a successful __alloc_pages().
5573 */
5574static void setup_per_zone_lowmem_reserve(void)
5575{
5576 struct pglist_data *pgdat;
5577 enum zone_type j, idx;
5578
5579 for_each_online_pgdat(pgdat) {
5580 for (j = 0; j < MAX_NR_ZONES; j++) {
5581 struct zone *zone = pgdat->node_zones + j;
5582 unsigned long managed_pages = zone->managed_pages;
5583
5584 zone->lowmem_reserve[j] = 0;
5585
5586 idx = j;
5587 while (idx) {
5588 struct zone *lower_zone;
5589
5590 idx--;
5591
5592 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5593 sysctl_lowmem_reserve_ratio[idx] = 1;
5594
5595 lower_zone = pgdat->node_zones + idx;
5596 lower_zone->lowmem_reserve[j] = managed_pages /
5597 sysctl_lowmem_reserve_ratio[idx];
5598 managed_pages += lower_zone->managed_pages;
5599 }
5600 }
5601 }
5602
5603 /* update totalreserve_pages */
5604 calculate_totalreserve_pages();
5605}
5606
5607static void __setup_per_zone_wmarks(void)
5608{
5609 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5610 unsigned long lowmem_pages = 0;
5611 struct zone *zone;
5612 unsigned long flags;
5613
5614 /* Calculate total number of !ZONE_HIGHMEM pages */
5615 for_each_zone(zone) {
5616 if (!is_highmem(zone))
5617 lowmem_pages += zone->managed_pages;
5618 }
5619
5620 for_each_zone(zone) {
5621 u64 tmp;
5622
5623 spin_lock_irqsave(&zone->lock, flags);
5624 tmp = (u64)pages_min * zone->managed_pages;
5625 do_div(tmp, lowmem_pages);
5626 if (is_highmem(zone)) {
5627 /*
5628 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5629 * need highmem pages, so cap pages_min to a small
5630 * value here.
5631 *
5632 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5633 * deltas controls asynch page reclaim, and so should
5634 * not be capped for highmem.
5635 */
5636 unsigned long min_pages;
5637
5638 min_pages = zone->managed_pages / 1024;
5639 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5640 zone->watermark[WMARK_MIN] = min_pages;
5641 } else {
5642 /*
5643 * If it's a lowmem zone, reserve a number of pages
5644 * proportionate to the zone's size.
5645 */
5646 zone->watermark[WMARK_MIN] = tmp;
5647 }
5648
5649 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5650 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5651
5652 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5653 high_wmark_pages(zone) -
5654 low_wmark_pages(zone) -
5655 zone_page_state(zone, NR_ALLOC_BATCH));
5656
5657 setup_zone_migrate_reserve(zone);
5658 spin_unlock_irqrestore(&zone->lock, flags);
5659 }
5660
5661 /* update totalreserve_pages */
5662 calculate_totalreserve_pages();
5663}
5664
5665/**
5666 * setup_per_zone_wmarks - called when min_free_kbytes changes
5667 * or when memory is hot-{added|removed}
5668 *
5669 * Ensures that the watermark[min,low,high] values for each zone are set
5670 * correctly with respect to min_free_kbytes.
5671 */
5672void setup_per_zone_wmarks(void)
5673{
5674 mutex_lock(&zonelists_mutex);
5675 __setup_per_zone_wmarks();
5676 mutex_unlock(&zonelists_mutex);
5677}
5678
5679/*
5680 * The inactive anon list should be small enough that the VM never has to
5681 * do too much work, but large enough that each inactive page has a chance
5682 * to be referenced again before it is swapped out.
5683 *
5684 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5685 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5686 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5687 * the anonymous pages are kept on the inactive list.
5688 *
5689 * total target max
5690 * memory ratio inactive anon
5691 * -------------------------------------
5692 * 10MB 1 5MB
5693 * 100MB 1 50MB
5694 * 1GB 3 250MB
5695 * 10GB 10 0.9GB
5696 * 100GB 31 3GB
5697 * 1TB 101 10GB
5698 * 10TB 320 32GB
5699 */
5700static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5701{
5702 unsigned int gb, ratio;
5703
5704 /* Zone size in gigabytes */
5705 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5706 if (gb)
5707 ratio = int_sqrt(10 * gb);
5708 else
5709 ratio = 1;
5710
5711 zone->inactive_ratio = ratio;
5712}
5713
5714static void __meminit setup_per_zone_inactive_ratio(void)
5715{
5716 struct zone *zone;
5717
5718 for_each_zone(zone)
5719 calculate_zone_inactive_ratio(zone);
5720}
5721
5722/*
5723 * Initialise min_free_kbytes.
5724 *
5725 * For small machines we want it small (128k min). For large machines
5726 * we want it large (64MB max). But it is not linear, because network
5727 * bandwidth does not increase linearly with machine size. We use
5728 *
5729 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5730 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5731 *
5732 * which yields
5733 *
5734 * 16MB: 512k
5735 * 32MB: 724k
5736 * 64MB: 1024k
5737 * 128MB: 1448k
5738 * 256MB: 2048k
5739 * 512MB: 2896k
5740 * 1024MB: 4096k
5741 * 2048MB: 5792k
5742 * 4096MB: 8192k
5743 * 8192MB: 11584k
5744 * 16384MB: 16384k
5745 */
5746int __meminit init_per_zone_wmark_min(void)
5747{
5748 unsigned long lowmem_kbytes;
5749 int new_min_free_kbytes;
5750
5751 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5752 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5753
5754 if (new_min_free_kbytes > user_min_free_kbytes) {
5755 min_free_kbytes = new_min_free_kbytes;
5756 if (min_free_kbytes < 128)
5757 min_free_kbytes = 128;
5758 if (min_free_kbytes > 65536)
5759 min_free_kbytes = 65536;
5760 } else {
5761 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5762 new_min_free_kbytes, user_min_free_kbytes);
5763 }
5764 setup_per_zone_wmarks();
5765 refresh_zone_stat_thresholds();
5766 setup_per_zone_lowmem_reserve();
5767 setup_per_zone_inactive_ratio();
5768 return 0;
5769}
5770module_init(init_per_zone_wmark_min)
5771
5772/*
5773 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5774 * that we can call two helper functions whenever min_free_kbytes
5775 * changes.
5776 */
5777int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5778 void __user *buffer, size_t *length, loff_t *ppos)
5779{
5780 int rc;
5781
5782 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5783 if (rc)
5784 return rc;
5785
5786 if (write) {
5787 user_min_free_kbytes = min_free_kbytes;
5788 setup_per_zone_wmarks();
5789 }
5790 return 0;
5791}
5792
5793#ifdef CONFIG_NUMA
5794int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5795 void __user *buffer, size_t *length, loff_t *ppos)
5796{
5797 struct zone *zone;
5798 int rc;
5799
5800 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5801 if (rc)
5802 return rc;
5803
5804 for_each_zone(zone)
5805 zone->min_unmapped_pages = (zone->managed_pages *
5806 sysctl_min_unmapped_ratio) / 100;
5807 return 0;
5808}
5809
5810int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5811 void __user *buffer, size_t *length, loff_t *ppos)
5812{
5813 struct zone *zone;
5814 int rc;
5815
5816 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5817 if (rc)
5818 return rc;
5819
5820 for_each_zone(zone)
5821 zone->min_slab_pages = (zone->managed_pages *
5822 sysctl_min_slab_ratio) / 100;
5823 return 0;
5824}
5825#endif
5826
5827/*
5828 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5829 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5830 * whenever sysctl_lowmem_reserve_ratio changes.
5831 *
5832 * The reserve ratio obviously has absolutely no relation with the
5833 * minimum watermarks. The lowmem reserve ratio can only make sense
5834 * if in function of the boot time zone sizes.
5835 */
5836int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5837 void __user *buffer, size_t *length, loff_t *ppos)
5838{
5839 proc_dointvec_minmax(table, write, buffer, length, ppos);
5840 setup_per_zone_lowmem_reserve();
5841 return 0;
5842}
5843
5844/*
5845 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5846 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5847 * pagelist can have before it gets flushed back to buddy allocator.
5848 */
5849int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5850 void __user *buffer, size_t *length, loff_t *ppos)
5851{
5852 struct zone *zone;
5853 unsigned int cpu;
5854 int ret;
5855
5856 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5857 if (!write || (ret < 0))
5858 return ret;
5859
5860 mutex_lock(&pcp_batch_high_lock);
5861 for_each_populated_zone(zone) {
5862 unsigned long high;
5863 high = zone->managed_pages / percpu_pagelist_fraction;
5864 for_each_possible_cpu(cpu)
5865 pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5866 high);
5867 }
5868 mutex_unlock(&pcp_batch_high_lock);
5869 return 0;
5870}
5871
5872int hashdist = HASHDIST_DEFAULT;
5873
5874#ifdef CONFIG_NUMA
5875static int __init set_hashdist(char *str)
5876{
5877 if (!str)
5878 return 0;
5879 hashdist = simple_strtoul(str, &str, 0);
5880 return 1;
5881}
5882__setup("hashdist=", set_hashdist);
5883#endif
5884
5885/*
5886 * allocate a large system hash table from bootmem
5887 * - it is assumed that the hash table must contain an exact power-of-2
5888 * quantity of entries
5889 * - limit is the number of hash buckets, not the total allocation size
5890 */
5891void *__init alloc_large_system_hash(const char *tablename,
5892 unsigned long bucketsize,
5893 unsigned long numentries,
5894 int scale,
5895 int flags,
5896 unsigned int *_hash_shift,
5897 unsigned int *_hash_mask,
5898 unsigned long low_limit,
5899 unsigned long high_limit)
5900{
5901 unsigned long long max = high_limit;
5902 unsigned long log2qty, size;
5903 void *table = NULL;
5904
5905 /* allow the kernel cmdline to have a say */
5906 if (!numentries) {
5907 /* round applicable memory size up to nearest megabyte */
5908 numentries = nr_kernel_pages;
5909
5910 /* It isn't necessary when PAGE_SIZE >= 1MB */
5911 if (PAGE_SHIFT < 20)
5912 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5913
5914 /* limit to 1 bucket per 2^scale bytes of low memory */
5915 if (scale > PAGE_SHIFT)
5916 numentries >>= (scale - PAGE_SHIFT);
5917 else
5918 numentries <<= (PAGE_SHIFT - scale);
5919
5920 /* Make sure we've got at least a 0-order allocation.. */
5921 if (unlikely(flags & HASH_SMALL)) {
5922 /* Makes no sense without HASH_EARLY */
5923 WARN_ON(!(flags & HASH_EARLY));
5924 if (!(numentries >> *_hash_shift)) {
5925 numentries = 1UL << *_hash_shift;
5926 BUG_ON(!numentries);
5927 }
5928 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5929 numentries = PAGE_SIZE / bucketsize;
5930 }
5931 numentries = roundup_pow_of_two(numentries);
5932
5933 /* limit allocation size to 1/16 total memory by default */
5934 if (max == 0) {
5935 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5936 do_div(max, bucketsize);
5937 }
5938 max = min(max, 0x80000000ULL);
5939
5940 if (numentries < low_limit)
5941 numentries = low_limit;
5942 if (numentries > max)
5943 numentries = max;
5944
5945 log2qty = ilog2(numentries);
5946
5947 do {
5948 size = bucketsize << log2qty;
5949 if (flags & HASH_EARLY)
5950 table = memblock_virt_alloc_nopanic(size, 0);
5951 else if (hashdist)
5952 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5953 else {
5954 /*
5955 * If bucketsize is not a power-of-two, we may free
5956 * some pages at the end of hash table which
5957 * alloc_pages_exact() automatically does
5958 */
5959 if (get_order(size) < MAX_ORDER) {
5960 table = alloc_pages_exact(size, GFP_ATOMIC);
5961 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5962 }
5963 }
5964 } while (!table && size > PAGE_SIZE && --log2qty);
5965
5966 if (!table)
5967 panic("Failed to allocate %s hash table\n", tablename);
5968
5969 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5970 tablename,
5971 (1UL << log2qty),
5972 ilog2(size) - PAGE_SHIFT,
5973 size);
5974
5975 if (_hash_shift)
5976 *_hash_shift = log2qty;
5977 if (_hash_mask)
5978 *_hash_mask = (1 << log2qty) - 1;
5979
5980 return table;
5981}
5982
5983/* Return a pointer to the bitmap storing bits affecting a block of pages */
5984static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5985 unsigned long pfn)
5986{
5987#ifdef CONFIG_SPARSEMEM
5988 return __pfn_to_section(pfn)->pageblock_flags;
5989#else
5990 return zone->pageblock_flags;
5991#endif /* CONFIG_SPARSEMEM */
5992}
5993
5994static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5995{
5996#ifdef CONFIG_SPARSEMEM
5997 pfn &= (PAGES_PER_SECTION-1);
5998 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5999#else
6000 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6001 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6002#endif /* CONFIG_SPARSEMEM */
6003}
6004
6005/**
6006 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
6007 * @page: The page within the block of interest
6008 * @start_bitidx: The first bit of interest to retrieve
6009 * @end_bitidx: The last bit of interest
6010 * returns pageblock_bits flags
6011 */
6012unsigned long get_pageblock_flags_group(struct page *page,
6013 int start_bitidx, int end_bitidx)
6014{
6015 struct zone *zone;
6016 unsigned long *bitmap;
6017 unsigned long pfn, bitidx;
6018 unsigned long flags = 0;
6019 unsigned long value = 1;
6020
6021 zone = page_zone(page);
6022 pfn = page_to_pfn(page);
6023 bitmap = get_pageblock_bitmap(zone, pfn);
6024 bitidx = pfn_to_bitidx(zone, pfn);
6025
6026 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6027 if (test_bit(bitidx + start_bitidx, bitmap))
6028 flags |= value;
6029
6030 return flags;
6031}
6032
6033/**
6034 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
6035 * @page: The page within the block of interest
6036 * @start_bitidx: The first bit of interest
6037 * @end_bitidx: The last bit of interest
6038 * @flags: The flags to set
6039 */
6040void set_pageblock_flags_group(struct page *page, unsigned long flags,
6041 int start_bitidx, int end_bitidx)
6042{
6043 struct zone *zone;
6044 unsigned long *bitmap;
6045 unsigned long pfn, bitidx;
6046 unsigned long value = 1;
6047
6048 zone = page_zone(page);
6049 pfn = page_to_pfn(page);
6050 bitmap = get_pageblock_bitmap(zone, pfn);
6051 bitidx = pfn_to_bitidx(zone, pfn);
6052 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6053
6054 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6055 if (flags & value)
6056 __set_bit(bitidx + start_bitidx, bitmap);
6057 else
6058 __clear_bit(bitidx + start_bitidx, bitmap);
6059}
6060
6061/*
6062 * This function checks whether pageblock includes unmovable pages or not.
6063 * If @count is not zero, it is okay to include less @count unmovable pages
6064 *
6065 * PageLRU check without isolation or lru_lock could race so that
6066 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6067 * expect this function should be exact.
6068 */
6069bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6070 bool skip_hwpoisoned_pages)
6071{
6072 unsigned long pfn, iter, found;
6073 int mt;
6074
6075 /*
6076 * For avoiding noise data, lru_add_drain_all() should be called
6077 * If ZONE_MOVABLE, the zone never contains unmovable pages
6078 */
6079 if (zone_idx(zone) == ZONE_MOVABLE)
6080 return false;
6081 mt = get_pageblock_migratetype(page);
6082 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6083 return false;
6084
6085 pfn = page_to_pfn(page);
6086 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6087 unsigned long check = pfn + iter;
6088
6089 if (!pfn_valid_within(check))
6090 continue;
6091
6092 page = pfn_to_page(check);
6093
6094 /*
6095 * Hugepages are not in LRU lists, but they're movable.
6096 * We need not scan over tail pages bacause we don't
6097 * handle each tail page individually in migration.
6098 */
6099 if (PageHuge(page)) {
6100 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6101 continue;
6102 }
6103
6104 /*
6105 * We can't use page_count without pin a page
6106 * because another CPU can free compound page.
6107 * This check already skips compound tails of THP
6108 * because their page->_count is zero at all time.
6109 */
6110 if (!atomic_read(&page->_count)) {
6111 if (PageBuddy(page))
6112 iter += (1 << page_order(page)) - 1;
6113 continue;
6114 }
6115
6116 /*
6117 * The HWPoisoned page may be not in buddy system, and
6118 * page_count() is not 0.
6119 */
6120 if (skip_hwpoisoned_pages && PageHWPoison(page))
6121 continue;
6122
6123 if (!PageLRU(page))
6124 found++;
6125 /*
6126 * If there are RECLAIMABLE pages, we need to check it.
6127 * But now, memory offline itself doesn't call shrink_slab()
6128 * and it still to be fixed.
6129 */
6130 /*
6131 * If the page is not RAM, page_count()should be 0.
6132 * we don't need more check. This is an _used_ not-movable page.
6133 *
6134 * The problematic thing here is PG_reserved pages. PG_reserved
6135 * is set to both of a memory hole page and a _used_ kernel
6136 * page at boot.
6137 */
6138 if (found > count)
6139 return true;
6140 }
6141 return false;
6142}
6143
6144bool is_pageblock_removable_nolock(struct page *page)
6145{
6146 struct zone *zone;
6147 unsigned long pfn;
6148
6149 /*
6150 * We have to be careful here because we are iterating over memory
6151 * sections which are not zone aware so we might end up outside of
6152 * the zone but still within the section.
6153 * We have to take care about the node as well. If the node is offline
6154 * its NODE_DATA will be NULL - see page_zone.
6155 */
6156 if (!node_online(page_to_nid(page)))
6157 return false;
6158
6159 zone = page_zone(page);
6160 pfn = page_to_pfn(page);
6161 if (!zone_spans_pfn(zone, pfn))
6162 return false;
6163
6164 return !has_unmovable_pages(zone, page, 0, true);
6165}
6166
6167#ifdef CONFIG_CMA
6168
6169static unsigned long pfn_max_align_down(unsigned long pfn)
6170{
6171 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6172 pageblock_nr_pages) - 1);
6173}
6174
6175static unsigned long pfn_max_align_up(unsigned long pfn)
6176{
6177 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6178 pageblock_nr_pages));
6179}
6180
6181/* [start, end) must belong to a single zone. */
6182static int __alloc_contig_migrate_range(struct compact_control *cc,
6183 unsigned long start, unsigned long end)
6184{
6185 /* This function is based on compact_zone() from compaction.c. */
6186 unsigned long nr_reclaimed;
6187 unsigned long pfn = start;
6188 unsigned int tries = 0;
6189 int ret = 0;
6190
6191 migrate_prep();
6192
6193 while (pfn < end || !list_empty(&cc->migratepages)) {
6194 if (fatal_signal_pending(current)) {
6195 ret = -EINTR;
6196 break;
6197 }
6198
6199 if (list_empty(&cc->migratepages)) {
6200 cc->nr_migratepages = 0;
6201 pfn = isolate_migratepages_range(cc->zone, cc,
6202 pfn, end, true);
6203 if (!pfn) {
6204 ret = -EINTR;
6205 break;
6206 }
6207 tries = 0;
6208 } else if (++tries == 5) {
6209 ret = ret < 0 ? ret : -EBUSY;
6210 break;
6211 }
6212
6213 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6214 &cc->migratepages);
6215 cc->nr_migratepages -= nr_reclaimed;
6216
6217 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6218 0, MIGRATE_SYNC, MR_CMA);
6219 }
6220 if (ret < 0) {
6221 putback_movable_pages(&cc->migratepages);
6222 return ret;
6223 }
6224 return 0;
6225}
6226
6227/**
6228 * alloc_contig_range() -- tries to allocate given range of pages
6229 * @start: start PFN to allocate
6230 * @end: one-past-the-last PFN to allocate
6231 * @migratetype: migratetype of the underlaying pageblocks (either
6232 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6233 * in range must have the same migratetype and it must
6234 * be either of the two.
6235 *
6236 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6237 * aligned, however it's the caller's responsibility to guarantee that
6238 * we are the only thread that changes migrate type of pageblocks the
6239 * pages fall in.
6240 *
6241 * The PFN range must belong to a single zone.
6242 *
6243 * Returns zero on success or negative error code. On success all
6244 * pages which PFN is in [start, end) are allocated for the caller and
6245 * need to be freed with free_contig_range().
6246 */
6247int alloc_contig_range(unsigned long start, unsigned long end,
6248 unsigned migratetype)
6249{
6250 unsigned long outer_start, outer_end;
6251 int ret = 0, order;
6252
6253 struct compact_control cc = {
6254 .nr_migratepages = 0,
6255 .order = -1,
6256 .zone = page_zone(pfn_to_page(start)),
6257 .sync = true,
6258 .ignore_skip_hint = true,
6259 };
6260 INIT_LIST_HEAD(&cc.migratepages);
6261
6262 /*
6263 * What we do here is we mark all pageblocks in range as
6264 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6265 * have different sizes, and due to the way page allocator
6266 * work, we align the range to biggest of the two pages so
6267 * that page allocator won't try to merge buddies from
6268 * different pageblocks and change MIGRATE_ISOLATE to some
6269 * other migration type.
6270 *
6271 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6272 * migrate the pages from an unaligned range (ie. pages that
6273 * we are interested in). This will put all the pages in
6274 * range back to page allocator as MIGRATE_ISOLATE.
6275 *
6276 * When this is done, we take the pages in range from page
6277 * allocator removing them from the buddy system. This way
6278 * page allocator will never consider using them.
6279 *
6280 * This lets us mark the pageblocks back as
6281 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6282 * aligned range but not in the unaligned, original range are
6283 * put back to page allocator so that buddy can use them.
6284 */
6285
6286 ret = start_isolate_page_range(pfn_max_align_down(start),
6287 pfn_max_align_up(end), migratetype,
6288 false);
6289 if (ret)
6290 return ret;
6291
6292 ret = __alloc_contig_migrate_range(&cc, start, end);
6293 if (ret)
6294 goto done;
6295
6296 /*
6297 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6298 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6299 * more, all pages in [start, end) are free in page allocator.
6300 * What we are going to do is to allocate all pages from
6301 * [start, end) (that is remove them from page allocator).
6302 *
6303 * The only problem is that pages at the beginning and at the
6304 * end of interesting range may be not aligned with pages that
6305 * page allocator holds, ie. they can be part of higher order
6306 * pages. Because of this, we reserve the bigger range and
6307 * once this is done free the pages we are not interested in.
6308 *
6309 * We don't have to hold zone->lock here because the pages are
6310 * isolated thus they won't get removed from buddy.
6311 */
6312
6313 lru_add_drain_all();
6314 drain_all_pages();
6315
6316 order = 0;
6317 outer_start = start;
6318 while (!PageBuddy(pfn_to_page(outer_start))) {
6319 if (++order >= MAX_ORDER) {
6320 ret = -EBUSY;
6321 goto done;
6322 }
6323 outer_start &= ~0UL << order;
6324 }
6325
6326 /* Make sure the range is really isolated. */
6327 if (test_pages_isolated(outer_start, end, false)) {
6328 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6329 outer_start, end);
6330 ret = -EBUSY;
6331 goto done;
6332 }
6333
6334
6335 /* Grab isolated pages from freelists. */
6336 outer_end = isolate_freepages_range(&cc, outer_start, end);
6337 if (!outer_end) {
6338 ret = -EBUSY;
6339 goto done;
6340 }
6341
6342 /* Free head and tail (if any) */
6343 if (start != outer_start)
6344 free_contig_range(outer_start, start - outer_start);
6345 if (end != outer_end)
6346 free_contig_range(end, outer_end - end);
6347
6348done:
6349 undo_isolate_page_range(pfn_max_align_down(start),
6350 pfn_max_align_up(end), migratetype);
6351 return ret;
6352}
6353
6354void free_contig_range(unsigned long pfn, unsigned nr_pages)
6355{
6356 unsigned int count = 0;
6357
6358 for (; nr_pages--; pfn++) {
6359 struct page *page = pfn_to_page(pfn);
6360
6361 count += page_count(page) != 1;
6362 __free_page(page);
6363 }
6364 WARN(count != 0, "%d pages are still in use!\n", count);
6365}
6366#endif
6367
6368#ifdef CONFIG_MEMORY_HOTPLUG
6369/*
6370 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6371 * page high values need to be recalulated.
6372 */
6373void __meminit zone_pcp_update(struct zone *zone)
6374{
6375 unsigned cpu;
6376 mutex_lock(&pcp_batch_high_lock);
6377 for_each_possible_cpu(cpu)
6378 pageset_set_high_and_batch(zone,
6379 per_cpu_ptr(zone->pageset, cpu));
6380 mutex_unlock(&pcp_batch_high_lock);
6381}
6382#endif
6383
6384void zone_pcp_reset(struct zone *zone)
6385{
6386 unsigned long flags;
6387 int cpu;
6388 struct per_cpu_pageset *pset;
6389
6390 /* avoid races with drain_pages() */
6391 local_irq_save(flags);
6392 if (zone->pageset != &boot_pageset) {
6393 for_each_online_cpu(cpu) {
6394 pset = per_cpu_ptr(zone->pageset, cpu);
6395 drain_zonestat(zone, pset);
6396 }
6397 free_percpu(zone->pageset);
6398 zone->pageset = &boot_pageset;
6399 }
6400 local_irq_restore(flags);
6401}
6402
6403#ifdef CONFIG_MEMORY_HOTREMOVE
6404/*
6405 * All pages in the range must be isolated before calling this.
6406 */
6407void
6408__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6409{
6410 struct page *page;
6411 struct zone *zone;
6412 int order, i;
6413 unsigned long pfn;
6414 unsigned long flags;
6415 /* find the first valid pfn */
6416 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6417 if (pfn_valid(pfn))
6418 break;
6419 if (pfn == end_pfn)
6420 return;
6421 zone = page_zone(pfn_to_page(pfn));
6422 spin_lock_irqsave(&zone->lock, flags);
6423 pfn = start_pfn;
6424 while (pfn < end_pfn) {
6425 if (!pfn_valid(pfn)) {
6426 pfn++;
6427 continue;
6428 }
6429 page = pfn_to_page(pfn);
6430 /*
6431 * The HWPoisoned page may be not in buddy system, and
6432 * page_count() is not 0.
6433 */
6434 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6435 pfn++;
6436 SetPageReserved(page);
6437 continue;
6438 }
6439
6440 BUG_ON(page_count(page));
6441 BUG_ON(!PageBuddy(page));
6442 order = page_order(page);
6443#ifdef CONFIG_DEBUG_VM
6444 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6445 pfn, 1 << order, end_pfn);
6446#endif
6447 list_del(&page->lru);
6448 rmv_page_order(page);
6449 zone->free_area[order].nr_free--;
6450 for (i = 0; i < (1 << order); i++)
6451 SetPageReserved((page+i));
6452 pfn += (1 << order);
6453 }
6454 spin_unlock_irqrestore(&zone->lock, flags);
6455}
6456#endif
6457
6458#ifdef CONFIG_MEMORY_FAILURE
6459bool is_free_buddy_page(struct page *page)
6460{
6461 struct zone *zone = page_zone(page);
6462 unsigned long pfn = page_to_pfn(page);
6463 unsigned long flags;
6464 int order;
6465
6466 spin_lock_irqsave(&zone->lock, flags);
6467 for (order = 0; order < MAX_ORDER; order++) {
6468 struct page *page_head = page - (pfn & ((1 << order) - 1));
6469
6470 if (PageBuddy(page_head) && page_order(page_head) >= order)
6471 break;
6472 }
6473 spin_unlock_irqrestore(&zone->lock, flags);
6474
6475 return order < MAX_ORDER;
6476}
6477#endif
6478
6479static const struct trace_print_flags pageflag_names[] = {
6480 {1UL << PG_locked, "locked" },
6481 {1UL << PG_error, "error" },
6482 {1UL << PG_referenced, "referenced" },
6483 {1UL << PG_uptodate, "uptodate" },
6484 {1UL << PG_dirty, "dirty" },
6485 {1UL << PG_lru, "lru" },
6486 {1UL << PG_active, "active" },
6487 {1UL << PG_slab, "slab" },
6488 {1UL << PG_owner_priv_1, "owner_priv_1" },
6489 {1UL << PG_arch_1, "arch_1" },
6490 {1UL << PG_reserved, "reserved" },
6491 {1UL << PG_private, "private" },
6492 {1UL << PG_private_2, "private_2" },
6493 {1UL << PG_writeback, "writeback" },
6494#ifdef CONFIG_PAGEFLAGS_EXTENDED
6495 {1UL << PG_head, "head" },
6496 {1UL << PG_tail, "tail" },
6497#else
6498 {1UL << PG_compound, "compound" },
6499#endif
6500 {1UL << PG_swapcache, "swapcache" },
6501 {1UL << PG_mappedtodisk, "mappedtodisk" },
6502 {1UL << PG_reclaim, "reclaim" },
6503 {1UL << PG_swapbacked, "swapbacked" },
6504 {1UL << PG_unevictable, "unevictable" },
6505#ifdef CONFIG_MMU
6506 {1UL << PG_mlocked, "mlocked" },
6507#endif
6508#ifdef CONFIG_ARCH_USES_PG_UNCACHED
6509 {1UL << PG_uncached, "uncached" },
6510#endif
6511#ifdef CONFIG_MEMORY_FAILURE
6512 {1UL << PG_hwpoison, "hwpoison" },
6513#endif
6514#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6515 {1UL << PG_compound_lock, "compound_lock" },
6516#endif
6517};
6518
6519static void dump_page_flags(unsigned long flags)
6520{
6521 const char *delim = "";
6522 unsigned long mask;
6523 int i;
6524
6525 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6526
6527 printk(KERN_ALERT "page flags: %#lx(", flags);
6528
6529 /* remove zone id */
6530 flags &= (1UL << NR_PAGEFLAGS) - 1;
6531
6532 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6533
6534 mask = pageflag_names[i].mask;
6535 if ((flags & mask) != mask)
6536 continue;
6537
6538 flags &= ~mask;
6539 printk("%s%s", delim, pageflag_names[i].name);
6540 delim = "|";
6541 }
6542
6543 /* check for left over flags */
6544 if (flags)
6545 printk("%s%#lx", delim, flags);
6546
6547 printk(")\n");
6548}
6549
6550void dump_page_badflags(struct page *page, const char *reason,
6551 unsigned long badflags)
6552{
6553 printk(KERN_ALERT
6554 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6555 page, atomic_read(&page->_count), page_mapcount(page),
6556 page->mapping, page->index);
6557 dump_page_flags(page->flags);
6558 if (reason)
6559 pr_alert("page dumped because: %s\n", reason);
6560 if (page->flags & badflags) {
6561 pr_alert("bad because of flags:\n");
6562 dump_page_flags(page->flags & badflags);
6563 }
6564 mem_cgroup_print_bad_page(page);
6565}
6566
6567void dump_page(struct page *page, const char *reason)
6568{
6569 dump_page_badflags(page, reason, 0);
6570}
6571EXPORT_SYMBOL(dump_page);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/page_alloc.c
4 *
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
7 *
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
16 */
17
18#include <linux/stddef.h>
19#include <linux/mm.h>
20#include <linux/highmem.h>
21#include <linux/swap.h>
22#include <linux/interrupt.h>
23#include <linux/pagemap.h>
24#include <linux/jiffies.h>
25#include <linux/memblock.h>
26#include <linux/compiler.h>
27#include <linux/kernel.h>
28#include <linux/kasan.h>
29#include <linux/module.h>
30#include <linux/suspend.h>
31#include <linux/pagevec.h>
32#include <linux/blkdev.h>
33#include <linux/slab.h>
34#include <linux/ratelimit.h>
35#include <linux/oom.h>
36#include <linux/topology.h>
37#include <linux/sysctl.h>
38#include <linux/cpu.h>
39#include <linux/cpuset.h>
40#include <linux/memory_hotplug.h>
41#include <linux/nodemask.h>
42#include <linux/vmalloc.h>
43#include <linux/vmstat.h>
44#include <linux/mempolicy.h>
45#include <linux/memremap.h>
46#include <linux/stop_machine.h>
47#include <linux/random.h>
48#include <linux/sort.h>
49#include <linux/pfn.h>
50#include <linux/backing-dev.h>
51#include <linux/fault-inject.h>
52#include <linux/page-isolation.h>
53#include <linux/debugobjects.h>
54#include <linux/kmemleak.h>
55#include <linux/compaction.h>
56#include <trace/events/kmem.h>
57#include <trace/events/oom.h>
58#include <linux/prefetch.h>
59#include <linux/mm_inline.h>
60#include <linux/migrate.h>
61#include <linux/hugetlb.h>
62#include <linux/sched/rt.h>
63#include <linux/sched/mm.h>
64#include <linux/page_owner.h>
65#include <linux/kthread.h>
66#include <linux/memcontrol.h>
67#include <linux/ftrace.h>
68#include <linux/lockdep.h>
69#include <linux/nmi.h>
70#include <linux/psi.h>
71
72#include <asm/sections.h>
73#include <asm/tlbflush.h>
74#include <asm/div64.h>
75#include "internal.h"
76#include "shuffle.h"
77
78/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
79static DEFINE_MUTEX(pcp_batch_high_lock);
80#define MIN_PERCPU_PAGELIST_FRACTION (8)
81
82#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
83DEFINE_PER_CPU(int, numa_node);
84EXPORT_PER_CPU_SYMBOL(numa_node);
85#endif
86
87DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
88
89#ifdef CONFIG_HAVE_MEMORYLESS_NODES
90/*
91 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
92 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
93 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
94 * defined in <linux/topology.h>.
95 */
96DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
97EXPORT_PER_CPU_SYMBOL(_numa_mem_);
98int _node_numa_mem_[MAX_NUMNODES];
99#endif
100
101/* work_structs for global per-cpu drains */
102struct pcpu_drain {
103 struct zone *zone;
104 struct work_struct work;
105};
106DEFINE_MUTEX(pcpu_drain_mutex);
107DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
108
109#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
110volatile unsigned long latent_entropy __latent_entropy;
111EXPORT_SYMBOL(latent_entropy);
112#endif
113
114/*
115 * Array of node states.
116 */
117nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
118 [N_POSSIBLE] = NODE_MASK_ALL,
119 [N_ONLINE] = { { [0] = 1UL } },
120#ifndef CONFIG_NUMA
121 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
122#ifdef CONFIG_HIGHMEM
123 [N_HIGH_MEMORY] = { { [0] = 1UL } },
124#endif
125 [N_MEMORY] = { { [0] = 1UL } },
126 [N_CPU] = { { [0] = 1UL } },
127#endif /* NUMA */
128};
129EXPORT_SYMBOL(node_states);
130
131atomic_long_t _totalram_pages __read_mostly;
132EXPORT_SYMBOL(_totalram_pages);
133unsigned long totalreserve_pages __read_mostly;
134unsigned long totalcma_pages __read_mostly;
135
136int percpu_pagelist_fraction;
137gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
138#ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
139DEFINE_STATIC_KEY_TRUE(init_on_alloc);
140#else
141DEFINE_STATIC_KEY_FALSE(init_on_alloc);
142#endif
143EXPORT_SYMBOL(init_on_alloc);
144
145#ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
146DEFINE_STATIC_KEY_TRUE(init_on_free);
147#else
148DEFINE_STATIC_KEY_FALSE(init_on_free);
149#endif
150EXPORT_SYMBOL(init_on_free);
151
152static int __init early_init_on_alloc(char *buf)
153{
154 int ret;
155 bool bool_result;
156
157 if (!buf)
158 return -EINVAL;
159 ret = kstrtobool(buf, &bool_result);
160 if (bool_result && page_poisoning_enabled())
161 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
162 if (bool_result)
163 static_branch_enable(&init_on_alloc);
164 else
165 static_branch_disable(&init_on_alloc);
166 return ret;
167}
168early_param("init_on_alloc", early_init_on_alloc);
169
170static int __init early_init_on_free(char *buf)
171{
172 int ret;
173 bool bool_result;
174
175 if (!buf)
176 return -EINVAL;
177 ret = kstrtobool(buf, &bool_result);
178 if (bool_result && page_poisoning_enabled())
179 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
180 if (bool_result)
181 static_branch_enable(&init_on_free);
182 else
183 static_branch_disable(&init_on_free);
184 return ret;
185}
186early_param("init_on_free", early_init_on_free);
187
188/*
189 * A cached value of the page's pageblock's migratetype, used when the page is
190 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
191 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
192 * Also the migratetype set in the page does not necessarily match the pcplist
193 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
194 * other index - this ensures that it will be put on the correct CMA freelist.
195 */
196static inline int get_pcppage_migratetype(struct page *page)
197{
198 return page->index;
199}
200
201static inline void set_pcppage_migratetype(struct page *page, int migratetype)
202{
203 page->index = migratetype;
204}
205
206#ifdef CONFIG_PM_SLEEP
207/*
208 * The following functions are used by the suspend/hibernate code to temporarily
209 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
210 * while devices are suspended. To avoid races with the suspend/hibernate code,
211 * they should always be called with system_transition_mutex held
212 * (gfp_allowed_mask also should only be modified with system_transition_mutex
213 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
214 * with that modification).
215 */
216
217static gfp_t saved_gfp_mask;
218
219void pm_restore_gfp_mask(void)
220{
221 WARN_ON(!mutex_is_locked(&system_transition_mutex));
222 if (saved_gfp_mask) {
223 gfp_allowed_mask = saved_gfp_mask;
224 saved_gfp_mask = 0;
225 }
226}
227
228void pm_restrict_gfp_mask(void)
229{
230 WARN_ON(!mutex_is_locked(&system_transition_mutex));
231 WARN_ON(saved_gfp_mask);
232 saved_gfp_mask = gfp_allowed_mask;
233 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
234}
235
236bool pm_suspended_storage(void)
237{
238 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
239 return false;
240 return true;
241}
242#endif /* CONFIG_PM_SLEEP */
243
244#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
245unsigned int pageblock_order __read_mostly;
246#endif
247
248static void __free_pages_ok(struct page *page, unsigned int order);
249
250/*
251 * results with 256, 32 in the lowmem_reserve sysctl:
252 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
253 * 1G machine -> (16M dma, 784M normal, 224M high)
254 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
255 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
256 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
257 *
258 * TBD: should special case ZONE_DMA32 machines here - in those we normally
259 * don't need any ZONE_NORMAL reservation
260 */
261int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
262#ifdef CONFIG_ZONE_DMA
263 [ZONE_DMA] = 256,
264#endif
265#ifdef CONFIG_ZONE_DMA32
266 [ZONE_DMA32] = 256,
267#endif
268 [ZONE_NORMAL] = 32,
269#ifdef CONFIG_HIGHMEM
270 [ZONE_HIGHMEM] = 0,
271#endif
272 [ZONE_MOVABLE] = 0,
273};
274
275static char * const zone_names[MAX_NR_ZONES] = {
276#ifdef CONFIG_ZONE_DMA
277 "DMA",
278#endif
279#ifdef CONFIG_ZONE_DMA32
280 "DMA32",
281#endif
282 "Normal",
283#ifdef CONFIG_HIGHMEM
284 "HighMem",
285#endif
286 "Movable",
287#ifdef CONFIG_ZONE_DEVICE
288 "Device",
289#endif
290};
291
292const char * const migratetype_names[MIGRATE_TYPES] = {
293 "Unmovable",
294 "Movable",
295 "Reclaimable",
296 "HighAtomic",
297#ifdef CONFIG_CMA
298 "CMA",
299#endif
300#ifdef CONFIG_MEMORY_ISOLATION
301 "Isolate",
302#endif
303};
304
305compound_page_dtor * const compound_page_dtors[] = {
306 NULL,
307 free_compound_page,
308#ifdef CONFIG_HUGETLB_PAGE
309 free_huge_page,
310#endif
311#ifdef CONFIG_TRANSPARENT_HUGEPAGE
312 free_transhuge_page,
313#endif
314};
315
316int min_free_kbytes = 1024;
317int user_min_free_kbytes = -1;
318#ifdef CONFIG_DISCONTIGMEM
319/*
320 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
321 * are not on separate NUMA nodes. Functionally this works but with
322 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
323 * quite small. By default, do not boost watermarks on discontigmem as in
324 * many cases very high-order allocations like THP are likely to be
325 * unsupported and the premature reclaim offsets the advantage of long-term
326 * fragmentation avoidance.
327 */
328int watermark_boost_factor __read_mostly;
329#else
330int watermark_boost_factor __read_mostly = 15000;
331#endif
332int watermark_scale_factor = 10;
333
334static unsigned long nr_kernel_pages __initdata;
335static unsigned long nr_all_pages __initdata;
336static unsigned long dma_reserve __initdata;
337
338#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
339static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
340static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
341static unsigned long required_kernelcore __initdata;
342static unsigned long required_kernelcore_percent __initdata;
343static unsigned long required_movablecore __initdata;
344static unsigned long required_movablecore_percent __initdata;
345static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
346static bool mirrored_kernelcore __meminitdata;
347
348/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
349int movable_zone;
350EXPORT_SYMBOL(movable_zone);
351#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
352
353#if MAX_NUMNODES > 1
354unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
355unsigned int nr_online_nodes __read_mostly = 1;
356EXPORT_SYMBOL(nr_node_ids);
357EXPORT_SYMBOL(nr_online_nodes);
358#endif
359
360int page_group_by_mobility_disabled __read_mostly;
361
362#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
363/*
364 * During boot we initialize deferred pages on-demand, as needed, but once
365 * page_alloc_init_late() has finished, the deferred pages are all initialized,
366 * and we can permanently disable that path.
367 */
368static DEFINE_STATIC_KEY_TRUE(deferred_pages);
369
370/*
371 * Calling kasan_free_pages() only after deferred memory initialization
372 * has completed. Poisoning pages during deferred memory init will greatly
373 * lengthen the process and cause problem in large memory systems as the
374 * deferred pages initialization is done with interrupt disabled.
375 *
376 * Assuming that there will be no reference to those newly initialized
377 * pages before they are ever allocated, this should have no effect on
378 * KASAN memory tracking as the poison will be properly inserted at page
379 * allocation time. The only corner case is when pages are allocated by
380 * on-demand allocation and then freed again before the deferred pages
381 * initialization is done, but this is not likely to happen.
382 */
383static inline void kasan_free_nondeferred_pages(struct page *page, int order)
384{
385 if (!static_branch_unlikely(&deferred_pages))
386 kasan_free_pages(page, order);
387}
388
389/* Returns true if the struct page for the pfn is uninitialised */
390static inline bool __meminit early_page_uninitialised(unsigned long pfn)
391{
392 int nid = early_pfn_to_nid(pfn);
393
394 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
395 return true;
396
397 return false;
398}
399
400/*
401 * Returns true when the remaining initialisation should be deferred until
402 * later in the boot cycle when it can be parallelised.
403 */
404static bool __meminit
405defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
406{
407 static unsigned long prev_end_pfn, nr_initialised;
408
409 /*
410 * prev_end_pfn static that contains the end of previous zone
411 * No need to protect because called very early in boot before smp_init.
412 */
413 if (prev_end_pfn != end_pfn) {
414 prev_end_pfn = end_pfn;
415 nr_initialised = 0;
416 }
417
418 /* Always populate low zones for address-constrained allocations */
419 if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
420 return false;
421
422 /*
423 * We start only with one section of pages, more pages are added as
424 * needed until the rest of deferred pages are initialized.
425 */
426 nr_initialised++;
427 if ((nr_initialised > PAGES_PER_SECTION) &&
428 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
429 NODE_DATA(nid)->first_deferred_pfn = pfn;
430 return true;
431 }
432 return false;
433}
434#else
435#define kasan_free_nondeferred_pages(p, o) kasan_free_pages(p, o)
436
437static inline bool early_page_uninitialised(unsigned long pfn)
438{
439 return false;
440}
441
442static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
443{
444 return false;
445}
446#endif
447
448/* Return a pointer to the bitmap storing bits affecting a block of pages */
449static inline unsigned long *get_pageblock_bitmap(struct page *page,
450 unsigned long pfn)
451{
452#ifdef CONFIG_SPARSEMEM
453 return section_to_usemap(__pfn_to_section(pfn));
454#else
455 return page_zone(page)->pageblock_flags;
456#endif /* CONFIG_SPARSEMEM */
457}
458
459static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
460{
461#ifdef CONFIG_SPARSEMEM
462 pfn &= (PAGES_PER_SECTION-1);
463 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
464#else
465 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
466 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
467#endif /* CONFIG_SPARSEMEM */
468}
469
470/**
471 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
472 * @page: The page within the block of interest
473 * @pfn: The target page frame number
474 * @end_bitidx: The last bit of interest to retrieve
475 * @mask: mask of bits that the caller is interested in
476 *
477 * Return: pageblock_bits flags
478 */
479static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
480 unsigned long pfn,
481 unsigned long end_bitidx,
482 unsigned long mask)
483{
484 unsigned long *bitmap;
485 unsigned long bitidx, word_bitidx;
486 unsigned long word;
487
488 bitmap = get_pageblock_bitmap(page, pfn);
489 bitidx = pfn_to_bitidx(page, pfn);
490 word_bitidx = bitidx / BITS_PER_LONG;
491 bitidx &= (BITS_PER_LONG-1);
492
493 word = bitmap[word_bitidx];
494 bitidx += end_bitidx;
495 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
496}
497
498unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
499 unsigned long end_bitidx,
500 unsigned long mask)
501{
502 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
503}
504
505static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
506{
507 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
508}
509
510/**
511 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
512 * @page: The page within the block of interest
513 * @flags: The flags to set
514 * @pfn: The target page frame number
515 * @end_bitidx: The last bit of interest
516 * @mask: mask of bits that the caller is interested in
517 */
518void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
519 unsigned long pfn,
520 unsigned long end_bitidx,
521 unsigned long mask)
522{
523 unsigned long *bitmap;
524 unsigned long bitidx, word_bitidx;
525 unsigned long old_word, word;
526
527 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
528 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
529
530 bitmap = get_pageblock_bitmap(page, pfn);
531 bitidx = pfn_to_bitidx(page, pfn);
532 word_bitidx = bitidx / BITS_PER_LONG;
533 bitidx &= (BITS_PER_LONG-1);
534
535 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
536
537 bitidx += end_bitidx;
538 mask <<= (BITS_PER_LONG - bitidx - 1);
539 flags <<= (BITS_PER_LONG - bitidx - 1);
540
541 word = READ_ONCE(bitmap[word_bitidx]);
542 for (;;) {
543 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
544 if (word == old_word)
545 break;
546 word = old_word;
547 }
548}
549
550void set_pageblock_migratetype(struct page *page, int migratetype)
551{
552 if (unlikely(page_group_by_mobility_disabled &&
553 migratetype < MIGRATE_PCPTYPES))
554 migratetype = MIGRATE_UNMOVABLE;
555
556 set_pageblock_flags_group(page, (unsigned long)migratetype,
557 PB_migrate, PB_migrate_end);
558}
559
560#ifdef CONFIG_DEBUG_VM
561static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
562{
563 int ret = 0;
564 unsigned seq;
565 unsigned long pfn = page_to_pfn(page);
566 unsigned long sp, start_pfn;
567
568 do {
569 seq = zone_span_seqbegin(zone);
570 start_pfn = zone->zone_start_pfn;
571 sp = zone->spanned_pages;
572 if (!zone_spans_pfn(zone, pfn))
573 ret = 1;
574 } while (zone_span_seqretry(zone, seq));
575
576 if (ret)
577 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
578 pfn, zone_to_nid(zone), zone->name,
579 start_pfn, start_pfn + sp);
580
581 return ret;
582}
583
584static int page_is_consistent(struct zone *zone, struct page *page)
585{
586 if (!pfn_valid_within(page_to_pfn(page)))
587 return 0;
588 if (zone != page_zone(page))
589 return 0;
590
591 return 1;
592}
593/*
594 * Temporary debugging check for pages not lying within a given zone.
595 */
596static int __maybe_unused bad_range(struct zone *zone, struct page *page)
597{
598 if (page_outside_zone_boundaries(zone, page))
599 return 1;
600 if (!page_is_consistent(zone, page))
601 return 1;
602
603 return 0;
604}
605#else
606static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
607{
608 return 0;
609}
610#endif
611
612static void bad_page(struct page *page, const char *reason,
613 unsigned long bad_flags)
614{
615 static unsigned long resume;
616 static unsigned long nr_shown;
617 static unsigned long nr_unshown;
618
619 /*
620 * Allow a burst of 60 reports, then keep quiet for that minute;
621 * or allow a steady drip of one report per second.
622 */
623 if (nr_shown == 60) {
624 if (time_before(jiffies, resume)) {
625 nr_unshown++;
626 goto out;
627 }
628 if (nr_unshown) {
629 pr_alert(
630 "BUG: Bad page state: %lu messages suppressed\n",
631 nr_unshown);
632 nr_unshown = 0;
633 }
634 nr_shown = 0;
635 }
636 if (nr_shown++ == 0)
637 resume = jiffies + 60 * HZ;
638
639 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
640 current->comm, page_to_pfn(page));
641 __dump_page(page, reason);
642 bad_flags &= page->flags;
643 if (bad_flags)
644 pr_alert("bad because of flags: %#lx(%pGp)\n",
645 bad_flags, &bad_flags);
646 dump_page_owner(page);
647
648 print_modules();
649 dump_stack();
650out:
651 /* Leave bad fields for debug, except PageBuddy could make trouble */
652 page_mapcount_reset(page); /* remove PageBuddy */
653 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
654}
655
656/*
657 * Higher-order pages are called "compound pages". They are structured thusly:
658 *
659 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
660 *
661 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
662 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
663 *
664 * The first tail page's ->compound_dtor holds the offset in array of compound
665 * page destructors. See compound_page_dtors.
666 *
667 * The first tail page's ->compound_order holds the order of allocation.
668 * This usage means that zero-order pages may not be compound.
669 */
670
671void free_compound_page(struct page *page)
672{
673 mem_cgroup_uncharge(page);
674 __free_pages_ok(page, compound_order(page));
675}
676
677void prep_compound_page(struct page *page, unsigned int order)
678{
679 int i;
680 int nr_pages = 1 << order;
681
682 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
683 set_compound_order(page, order);
684 __SetPageHead(page);
685 for (i = 1; i < nr_pages; i++) {
686 struct page *p = page + i;
687 set_page_count(p, 0);
688 p->mapping = TAIL_MAPPING;
689 set_compound_head(p, page);
690 }
691 atomic_set(compound_mapcount_ptr(page), -1);
692}
693
694#ifdef CONFIG_DEBUG_PAGEALLOC
695unsigned int _debug_guardpage_minorder;
696
697#ifdef CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT
698DEFINE_STATIC_KEY_TRUE(_debug_pagealloc_enabled);
699#else
700DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
701#endif
702EXPORT_SYMBOL(_debug_pagealloc_enabled);
703
704DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
705
706static int __init early_debug_pagealloc(char *buf)
707{
708 bool enable = false;
709
710 if (kstrtobool(buf, &enable))
711 return -EINVAL;
712
713 if (enable)
714 static_branch_enable(&_debug_pagealloc_enabled);
715
716 return 0;
717}
718early_param("debug_pagealloc", early_debug_pagealloc);
719
720static void init_debug_guardpage(void)
721{
722 if (!debug_pagealloc_enabled())
723 return;
724
725 if (!debug_guardpage_minorder())
726 return;
727
728 static_branch_enable(&_debug_guardpage_enabled);
729}
730
731static int __init debug_guardpage_minorder_setup(char *buf)
732{
733 unsigned long res;
734
735 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
736 pr_err("Bad debug_guardpage_minorder value\n");
737 return 0;
738 }
739 _debug_guardpage_minorder = res;
740 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
741 return 0;
742}
743early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
744
745static inline bool set_page_guard(struct zone *zone, struct page *page,
746 unsigned int order, int migratetype)
747{
748 if (!debug_guardpage_enabled())
749 return false;
750
751 if (order >= debug_guardpage_minorder())
752 return false;
753
754 __SetPageGuard(page);
755 INIT_LIST_HEAD(&page->lru);
756 set_page_private(page, order);
757 /* Guard pages are not available for any usage */
758 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
759
760 return true;
761}
762
763static inline void clear_page_guard(struct zone *zone, struct page *page,
764 unsigned int order, int migratetype)
765{
766 if (!debug_guardpage_enabled())
767 return;
768
769 __ClearPageGuard(page);
770
771 set_page_private(page, 0);
772 if (!is_migrate_isolate(migratetype))
773 __mod_zone_freepage_state(zone, (1 << order), migratetype);
774}
775#else
776static inline bool set_page_guard(struct zone *zone, struct page *page,
777 unsigned int order, int migratetype) { return false; }
778static inline void clear_page_guard(struct zone *zone, struct page *page,
779 unsigned int order, int migratetype) {}
780#endif
781
782static inline void set_page_order(struct page *page, unsigned int order)
783{
784 set_page_private(page, order);
785 __SetPageBuddy(page);
786}
787
788/*
789 * This function checks whether a page is free && is the buddy
790 * we can coalesce a page and its buddy if
791 * (a) the buddy is not in a hole (check before calling!) &&
792 * (b) the buddy is in the buddy system &&
793 * (c) a page and its buddy have the same order &&
794 * (d) a page and its buddy are in the same zone.
795 *
796 * For recording whether a page is in the buddy system, we set PageBuddy.
797 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
798 *
799 * For recording page's order, we use page_private(page).
800 */
801static inline int page_is_buddy(struct page *page, struct page *buddy,
802 unsigned int order)
803{
804 if (page_is_guard(buddy) && page_order(buddy) == order) {
805 if (page_zone_id(page) != page_zone_id(buddy))
806 return 0;
807
808 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
809
810 return 1;
811 }
812
813 if (PageBuddy(buddy) && page_order(buddy) == order) {
814 /*
815 * zone check is done late to avoid uselessly
816 * calculating zone/node ids for pages that could
817 * never merge.
818 */
819 if (page_zone_id(page) != page_zone_id(buddy))
820 return 0;
821
822 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
823
824 return 1;
825 }
826 return 0;
827}
828
829#ifdef CONFIG_COMPACTION
830static inline struct capture_control *task_capc(struct zone *zone)
831{
832 struct capture_control *capc = current->capture_control;
833
834 return capc &&
835 !(current->flags & PF_KTHREAD) &&
836 !capc->page &&
837 capc->cc->zone == zone &&
838 capc->cc->direct_compaction ? capc : NULL;
839}
840
841static inline bool
842compaction_capture(struct capture_control *capc, struct page *page,
843 int order, int migratetype)
844{
845 if (!capc || order != capc->cc->order)
846 return false;
847
848 /* Do not accidentally pollute CMA or isolated regions*/
849 if (is_migrate_cma(migratetype) ||
850 is_migrate_isolate(migratetype))
851 return false;
852
853 /*
854 * Do not let lower order allocations polluate a movable pageblock.
855 * This might let an unmovable request use a reclaimable pageblock
856 * and vice-versa but no more than normal fallback logic which can
857 * have trouble finding a high-order free page.
858 */
859 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
860 return false;
861
862 capc->page = page;
863 return true;
864}
865
866#else
867static inline struct capture_control *task_capc(struct zone *zone)
868{
869 return NULL;
870}
871
872static inline bool
873compaction_capture(struct capture_control *capc, struct page *page,
874 int order, int migratetype)
875{
876 return false;
877}
878#endif /* CONFIG_COMPACTION */
879
880/*
881 * Freeing function for a buddy system allocator.
882 *
883 * The concept of a buddy system is to maintain direct-mapped table
884 * (containing bit values) for memory blocks of various "orders".
885 * The bottom level table contains the map for the smallest allocatable
886 * units of memory (here, pages), and each level above it describes
887 * pairs of units from the levels below, hence, "buddies".
888 * At a high level, all that happens here is marking the table entry
889 * at the bottom level available, and propagating the changes upward
890 * as necessary, plus some accounting needed to play nicely with other
891 * parts of the VM system.
892 * At each level, we keep a list of pages, which are heads of continuous
893 * free pages of length of (1 << order) and marked with PageBuddy.
894 * Page's order is recorded in page_private(page) field.
895 * So when we are allocating or freeing one, we can derive the state of the
896 * other. That is, if we allocate a small block, and both were
897 * free, the remainder of the region must be split into blocks.
898 * If a block is freed, and its buddy is also free, then this
899 * triggers coalescing into a block of larger size.
900 *
901 * -- nyc
902 */
903
904static inline void __free_one_page(struct page *page,
905 unsigned long pfn,
906 struct zone *zone, unsigned int order,
907 int migratetype)
908{
909 unsigned long combined_pfn;
910 unsigned long uninitialized_var(buddy_pfn);
911 struct page *buddy;
912 unsigned int max_order;
913 struct capture_control *capc = task_capc(zone);
914
915 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
916
917 VM_BUG_ON(!zone_is_initialized(zone));
918 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
919
920 VM_BUG_ON(migratetype == -1);
921 if (likely(!is_migrate_isolate(migratetype)))
922 __mod_zone_freepage_state(zone, 1 << order, migratetype);
923
924 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
925 VM_BUG_ON_PAGE(bad_range(zone, page), page);
926
927continue_merging:
928 while (order < max_order - 1) {
929 if (compaction_capture(capc, page, order, migratetype)) {
930 __mod_zone_freepage_state(zone, -(1 << order),
931 migratetype);
932 return;
933 }
934 buddy_pfn = __find_buddy_pfn(pfn, order);
935 buddy = page + (buddy_pfn - pfn);
936
937 if (!pfn_valid_within(buddy_pfn))
938 goto done_merging;
939 if (!page_is_buddy(page, buddy, order))
940 goto done_merging;
941 /*
942 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
943 * merge with it and move up one order.
944 */
945 if (page_is_guard(buddy))
946 clear_page_guard(zone, buddy, order, migratetype);
947 else
948 del_page_from_free_area(buddy, &zone->free_area[order]);
949 combined_pfn = buddy_pfn & pfn;
950 page = page + (combined_pfn - pfn);
951 pfn = combined_pfn;
952 order++;
953 }
954 if (max_order < MAX_ORDER) {
955 /* If we are here, it means order is >= pageblock_order.
956 * We want to prevent merge between freepages on isolate
957 * pageblock and normal pageblock. Without this, pageblock
958 * isolation could cause incorrect freepage or CMA accounting.
959 *
960 * We don't want to hit this code for the more frequent
961 * low-order merging.
962 */
963 if (unlikely(has_isolate_pageblock(zone))) {
964 int buddy_mt;
965
966 buddy_pfn = __find_buddy_pfn(pfn, order);
967 buddy = page + (buddy_pfn - pfn);
968 buddy_mt = get_pageblock_migratetype(buddy);
969
970 if (migratetype != buddy_mt
971 && (is_migrate_isolate(migratetype) ||
972 is_migrate_isolate(buddy_mt)))
973 goto done_merging;
974 }
975 max_order++;
976 goto continue_merging;
977 }
978
979done_merging:
980 set_page_order(page, order);
981
982 /*
983 * If this is not the largest possible page, check if the buddy
984 * of the next-highest order is free. If it is, it's possible
985 * that pages are being freed that will coalesce soon. In case,
986 * that is happening, add the free page to the tail of the list
987 * so it's less likely to be used soon and more likely to be merged
988 * as a higher order page
989 */
990 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)
991 && !is_shuffle_order(order)) {
992 struct page *higher_page, *higher_buddy;
993 combined_pfn = buddy_pfn & pfn;
994 higher_page = page + (combined_pfn - pfn);
995 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
996 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
997 if (pfn_valid_within(buddy_pfn) &&
998 page_is_buddy(higher_page, higher_buddy, order + 1)) {
999 add_to_free_area_tail(page, &zone->free_area[order],
1000 migratetype);
1001 return;
1002 }
1003 }
1004
1005 if (is_shuffle_order(order))
1006 add_to_free_area_random(page, &zone->free_area[order],
1007 migratetype);
1008 else
1009 add_to_free_area(page, &zone->free_area[order], migratetype);
1010
1011}
1012
1013/*
1014 * A bad page could be due to a number of fields. Instead of multiple branches,
1015 * try and check multiple fields with one check. The caller must do a detailed
1016 * check if necessary.
1017 */
1018static inline bool page_expected_state(struct page *page,
1019 unsigned long check_flags)
1020{
1021 if (unlikely(atomic_read(&page->_mapcount) != -1))
1022 return false;
1023
1024 if (unlikely((unsigned long)page->mapping |
1025 page_ref_count(page) |
1026#ifdef CONFIG_MEMCG
1027 (unsigned long)page->mem_cgroup |
1028#endif
1029 (page->flags & check_flags)))
1030 return false;
1031
1032 return true;
1033}
1034
1035static void free_pages_check_bad(struct page *page)
1036{
1037 const char *bad_reason;
1038 unsigned long bad_flags;
1039
1040 bad_reason = NULL;
1041 bad_flags = 0;
1042
1043 if (unlikely(atomic_read(&page->_mapcount) != -1))
1044 bad_reason = "nonzero mapcount";
1045 if (unlikely(page->mapping != NULL))
1046 bad_reason = "non-NULL mapping";
1047 if (unlikely(page_ref_count(page) != 0))
1048 bad_reason = "nonzero _refcount";
1049 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
1050 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1051 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
1052 }
1053#ifdef CONFIG_MEMCG
1054 if (unlikely(page->mem_cgroup))
1055 bad_reason = "page still charged to cgroup";
1056#endif
1057 bad_page(page, bad_reason, bad_flags);
1058}
1059
1060static inline int free_pages_check(struct page *page)
1061{
1062 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1063 return 0;
1064
1065 /* Something has gone sideways, find it */
1066 free_pages_check_bad(page);
1067 return 1;
1068}
1069
1070static int free_tail_pages_check(struct page *head_page, struct page *page)
1071{
1072 int ret = 1;
1073
1074 /*
1075 * We rely page->lru.next never has bit 0 set, unless the page
1076 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1077 */
1078 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1079
1080 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1081 ret = 0;
1082 goto out;
1083 }
1084 switch (page - head_page) {
1085 case 1:
1086 /* the first tail page: ->mapping may be compound_mapcount() */
1087 if (unlikely(compound_mapcount(page))) {
1088 bad_page(page, "nonzero compound_mapcount", 0);
1089 goto out;
1090 }
1091 break;
1092 case 2:
1093 /*
1094 * the second tail page: ->mapping is
1095 * deferred_list.next -- ignore value.
1096 */
1097 break;
1098 default:
1099 if (page->mapping != TAIL_MAPPING) {
1100 bad_page(page, "corrupted mapping in tail page", 0);
1101 goto out;
1102 }
1103 break;
1104 }
1105 if (unlikely(!PageTail(page))) {
1106 bad_page(page, "PageTail not set", 0);
1107 goto out;
1108 }
1109 if (unlikely(compound_head(page) != head_page)) {
1110 bad_page(page, "compound_head not consistent", 0);
1111 goto out;
1112 }
1113 ret = 0;
1114out:
1115 page->mapping = NULL;
1116 clear_compound_head(page);
1117 return ret;
1118}
1119
1120static void kernel_init_free_pages(struct page *page, int numpages)
1121{
1122 int i;
1123
1124 for (i = 0; i < numpages; i++)
1125 clear_highpage(page + i);
1126}
1127
1128static __always_inline bool free_pages_prepare(struct page *page,
1129 unsigned int order, bool check_free)
1130{
1131 int bad = 0;
1132
1133 VM_BUG_ON_PAGE(PageTail(page), page);
1134
1135 trace_mm_page_free(page, order);
1136
1137 /*
1138 * Check tail pages before head page information is cleared to
1139 * avoid checking PageCompound for order-0 pages.
1140 */
1141 if (unlikely(order)) {
1142 bool compound = PageCompound(page);
1143 int i;
1144
1145 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1146
1147 if (compound)
1148 ClearPageDoubleMap(page);
1149 for (i = 1; i < (1 << order); i++) {
1150 if (compound)
1151 bad += free_tail_pages_check(page, page + i);
1152 if (unlikely(free_pages_check(page + i))) {
1153 bad++;
1154 continue;
1155 }
1156 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1157 }
1158 }
1159 if (PageMappingFlags(page))
1160 page->mapping = NULL;
1161 if (memcg_kmem_enabled() && PageKmemcg(page))
1162 __memcg_kmem_uncharge(page, order);
1163 if (check_free)
1164 bad += free_pages_check(page);
1165 if (bad)
1166 return false;
1167
1168 page_cpupid_reset_last(page);
1169 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1170 reset_page_owner(page, order);
1171
1172 if (!PageHighMem(page)) {
1173 debug_check_no_locks_freed(page_address(page),
1174 PAGE_SIZE << order);
1175 debug_check_no_obj_freed(page_address(page),
1176 PAGE_SIZE << order);
1177 }
1178 if (want_init_on_free())
1179 kernel_init_free_pages(page, 1 << order);
1180
1181 kernel_poison_pages(page, 1 << order, 0);
1182 /*
1183 * arch_free_page() can make the page's contents inaccessible. s390
1184 * does this. So nothing which can access the page's contents should
1185 * happen after this.
1186 */
1187 arch_free_page(page, order);
1188
1189 if (debug_pagealloc_enabled())
1190 kernel_map_pages(page, 1 << order, 0);
1191
1192 kasan_free_nondeferred_pages(page, order);
1193
1194 return true;
1195}
1196
1197#ifdef CONFIG_DEBUG_VM
1198/*
1199 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1200 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1201 * moved from pcp lists to free lists.
1202 */
1203static bool free_pcp_prepare(struct page *page)
1204{
1205 return free_pages_prepare(page, 0, true);
1206}
1207
1208static bool bulkfree_pcp_prepare(struct page *page)
1209{
1210 if (debug_pagealloc_enabled())
1211 return free_pages_check(page);
1212 else
1213 return false;
1214}
1215#else
1216/*
1217 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1218 * moving from pcp lists to free list in order to reduce overhead. With
1219 * debug_pagealloc enabled, they are checked also immediately when being freed
1220 * to the pcp lists.
1221 */
1222static bool free_pcp_prepare(struct page *page)
1223{
1224 if (debug_pagealloc_enabled())
1225 return free_pages_prepare(page, 0, true);
1226 else
1227 return free_pages_prepare(page, 0, false);
1228}
1229
1230static bool bulkfree_pcp_prepare(struct page *page)
1231{
1232 return free_pages_check(page);
1233}
1234#endif /* CONFIG_DEBUG_VM */
1235
1236static inline void prefetch_buddy(struct page *page)
1237{
1238 unsigned long pfn = page_to_pfn(page);
1239 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1240 struct page *buddy = page + (buddy_pfn - pfn);
1241
1242 prefetch(buddy);
1243}
1244
1245/*
1246 * Frees a number of pages from the PCP lists
1247 * Assumes all pages on list are in same zone, and of same order.
1248 * count is the number of pages to free.
1249 *
1250 * If the zone was previously in an "all pages pinned" state then look to
1251 * see if this freeing clears that state.
1252 *
1253 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1254 * pinned" detection logic.
1255 */
1256static void free_pcppages_bulk(struct zone *zone, int count,
1257 struct per_cpu_pages *pcp)
1258{
1259 int migratetype = 0;
1260 int batch_free = 0;
1261 int prefetch_nr = 0;
1262 bool isolated_pageblocks;
1263 struct page *page, *tmp;
1264 LIST_HEAD(head);
1265
1266 while (count) {
1267 struct list_head *list;
1268
1269 /*
1270 * Remove pages from lists in a round-robin fashion. A
1271 * batch_free count is maintained that is incremented when an
1272 * empty list is encountered. This is so more pages are freed
1273 * off fuller lists instead of spinning excessively around empty
1274 * lists
1275 */
1276 do {
1277 batch_free++;
1278 if (++migratetype == MIGRATE_PCPTYPES)
1279 migratetype = 0;
1280 list = &pcp->lists[migratetype];
1281 } while (list_empty(list));
1282
1283 /* This is the only non-empty list. Free them all. */
1284 if (batch_free == MIGRATE_PCPTYPES)
1285 batch_free = count;
1286
1287 do {
1288 page = list_last_entry(list, struct page, lru);
1289 /* must delete to avoid corrupting pcp list */
1290 list_del(&page->lru);
1291 pcp->count--;
1292
1293 if (bulkfree_pcp_prepare(page))
1294 continue;
1295
1296 list_add_tail(&page->lru, &head);
1297
1298 /*
1299 * We are going to put the page back to the global
1300 * pool, prefetch its buddy to speed up later access
1301 * under zone->lock. It is believed the overhead of
1302 * an additional test and calculating buddy_pfn here
1303 * can be offset by reduced memory latency later. To
1304 * avoid excessive prefetching due to large count, only
1305 * prefetch buddy for the first pcp->batch nr of pages.
1306 */
1307 if (prefetch_nr++ < pcp->batch)
1308 prefetch_buddy(page);
1309 } while (--count && --batch_free && !list_empty(list));
1310 }
1311
1312 spin_lock(&zone->lock);
1313 isolated_pageblocks = has_isolate_pageblock(zone);
1314
1315 /*
1316 * Use safe version since after __free_one_page(),
1317 * page->lru.next will not point to original list.
1318 */
1319 list_for_each_entry_safe(page, tmp, &head, lru) {
1320 int mt = get_pcppage_migratetype(page);
1321 /* MIGRATE_ISOLATE page should not go to pcplists */
1322 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1323 /* Pageblock could have been isolated meanwhile */
1324 if (unlikely(isolated_pageblocks))
1325 mt = get_pageblock_migratetype(page);
1326
1327 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1328 trace_mm_page_pcpu_drain(page, 0, mt);
1329 }
1330 spin_unlock(&zone->lock);
1331}
1332
1333static void free_one_page(struct zone *zone,
1334 struct page *page, unsigned long pfn,
1335 unsigned int order,
1336 int migratetype)
1337{
1338 spin_lock(&zone->lock);
1339 if (unlikely(has_isolate_pageblock(zone) ||
1340 is_migrate_isolate(migratetype))) {
1341 migratetype = get_pfnblock_migratetype(page, pfn);
1342 }
1343 __free_one_page(page, pfn, zone, order, migratetype);
1344 spin_unlock(&zone->lock);
1345}
1346
1347static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1348 unsigned long zone, int nid)
1349{
1350 mm_zero_struct_page(page);
1351 set_page_links(page, zone, nid, pfn);
1352 init_page_count(page);
1353 page_mapcount_reset(page);
1354 page_cpupid_reset_last(page);
1355 page_kasan_tag_reset(page);
1356
1357 INIT_LIST_HEAD(&page->lru);
1358#ifdef WANT_PAGE_VIRTUAL
1359 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1360 if (!is_highmem_idx(zone))
1361 set_page_address(page, __va(pfn << PAGE_SHIFT));
1362#endif
1363}
1364
1365#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1366static void __meminit init_reserved_page(unsigned long pfn)
1367{
1368 pg_data_t *pgdat;
1369 int nid, zid;
1370
1371 if (!early_page_uninitialised(pfn))
1372 return;
1373
1374 nid = early_pfn_to_nid(pfn);
1375 pgdat = NODE_DATA(nid);
1376
1377 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1378 struct zone *zone = &pgdat->node_zones[zid];
1379
1380 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1381 break;
1382 }
1383 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1384}
1385#else
1386static inline void init_reserved_page(unsigned long pfn)
1387{
1388}
1389#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1390
1391/*
1392 * Initialised pages do not have PageReserved set. This function is
1393 * called for each range allocated by the bootmem allocator and
1394 * marks the pages PageReserved. The remaining valid pages are later
1395 * sent to the buddy page allocator.
1396 */
1397void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1398{
1399 unsigned long start_pfn = PFN_DOWN(start);
1400 unsigned long end_pfn = PFN_UP(end);
1401
1402 for (; start_pfn < end_pfn; start_pfn++) {
1403 if (pfn_valid(start_pfn)) {
1404 struct page *page = pfn_to_page(start_pfn);
1405
1406 init_reserved_page(start_pfn);
1407
1408 /* Avoid false-positive PageTail() */
1409 INIT_LIST_HEAD(&page->lru);
1410
1411 /*
1412 * no need for atomic set_bit because the struct
1413 * page is not visible yet so nobody should
1414 * access it yet.
1415 */
1416 __SetPageReserved(page);
1417 }
1418 }
1419}
1420
1421static void __free_pages_ok(struct page *page, unsigned int order)
1422{
1423 unsigned long flags;
1424 int migratetype;
1425 unsigned long pfn = page_to_pfn(page);
1426
1427 if (!free_pages_prepare(page, order, true))
1428 return;
1429
1430 migratetype = get_pfnblock_migratetype(page, pfn);
1431 local_irq_save(flags);
1432 __count_vm_events(PGFREE, 1 << order);
1433 free_one_page(page_zone(page), page, pfn, order, migratetype);
1434 local_irq_restore(flags);
1435}
1436
1437void __free_pages_core(struct page *page, unsigned int order)
1438{
1439 unsigned int nr_pages = 1 << order;
1440 struct page *p = page;
1441 unsigned int loop;
1442
1443 prefetchw(p);
1444 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1445 prefetchw(p + 1);
1446 __ClearPageReserved(p);
1447 set_page_count(p, 0);
1448 }
1449 __ClearPageReserved(p);
1450 set_page_count(p, 0);
1451
1452 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1453 set_page_refcounted(page);
1454 __free_pages(page, order);
1455}
1456
1457#if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1458 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1459
1460static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1461
1462int __meminit early_pfn_to_nid(unsigned long pfn)
1463{
1464 static DEFINE_SPINLOCK(early_pfn_lock);
1465 int nid;
1466
1467 spin_lock(&early_pfn_lock);
1468 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1469 if (nid < 0)
1470 nid = first_online_node;
1471 spin_unlock(&early_pfn_lock);
1472
1473 return nid;
1474}
1475#endif
1476
1477#ifdef CONFIG_NODES_SPAN_OTHER_NODES
1478/* Only safe to use early in boot when initialisation is single-threaded */
1479static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1480{
1481 int nid;
1482
1483 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1484 if (nid >= 0 && nid != node)
1485 return false;
1486 return true;
1487}
1488
1489#else
1490static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1491{
1492 return true;
1493}
1494#endif
1495
1496
1497void __init memblock_free_pages(struct page *page, unsigned long pfn,
1498 unsigned int order)
1499{
1500 if (early_page_uninitialised(pfn))
1501 return;
1502 __free_pages_core(page, order);
1503}
1504
1505/*
1506 * Check that the whole (or subset of) a pageblock given by the interval of
1507 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1508 * with the migration of free compaction scanner. The scanners then need to
1509 * use only pfn_valid_within() check for arches that allow holes within
1510 * pageblocks.
1511 *
1512 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1513 *
1514 * It's possible on some configurations to have a setup like node0 node1 node0
1515 * i.e. it's possible that all pages within a zones range of pages do not
1516 * belong to a single zone. We assume that a border between node0 and node1
1517 * can occur within a single pageblock, but not a node0 node1 node0
1518 * interleaving within a single pageblock. It is therefore sufficient to check
1519 * the first and last page of a pageblock and avoid checking each individual
1520 * page in a pageblock.
1521 */
1522struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1523 unsigned long end_pfn, struct zone *zone)
1524{
1525 struct page *start_page;
1526 struct page *end_page;
1527
1528 /* end_pfn is one past the range we are checking */
1529 end_pfn--;
1530
1531 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1532 return NULL;
1533
1534 start_page = pfn_to_online_page(start_pfn);
1535 if (!start_page)
1536 return NULL;
1537
1538 if (page_zone(start_page) != zone)
1539 return NULL;
1540
1541 end_page = pfn_to_page(end_pfn);
1542
1543 /* This gives a shorter code than deriving page_zone(end_page) */
1544 if (page_zone_id(start_page) != page_zone_id(end_page))
1545 return NULL;
1546
1547 return start_page;
1548}
1549
1550void set_zone_contiguous(struct zone *zone)
1551{
1552 unsigned long block_start_pfn = zone->zone_start_pfn;
1553 unsigned long block_end_pfn;
1554
1555 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1556 for (; block_start_pfn < zone_end_pfn(zone);
1557 block_start_pfn = block_end_pfn,
1558 block_end_pfn += pageblock_nr_pages) {
1559
1560 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1561
1562 if (!__pageblock_pfn_to_page(block_start_pfn,
1563 block_end_pfn, zone))
1564 return;
1565 }
1566
1567 /* We confirm that there is no hole */
1568 zone->contiguous = true;
1569}
1570
1571void clear_zone_contiguous(struct zone *zone)
1572{
1573 zone->contiguous = false;
1574}
1575
1576#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1577static void __init deferred_free_range(unsigned long pfn,
1578 unsigned long nr_pages)
1579{
1580 struct page *page;
1581 unsigned long i;
1582
1583 if (!nr_pages)
1584 return;
1585
1586 page = pfn_to_page(pfn);
1587
1588 /* Free a large naturally-aligned chunk if possible */
1589 if (nr_pages == pageblock_nr_pages &&
1590 (pfn & (pageblock_nr_pages - 1)) == 0) {
1591 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1592 __free_pages_core(page, pageblock_order);
1593 return;
1594 }
1595
1596 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1597 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1598 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1599 __free_pages_core(page, 0);
1600 }
1601}
1602
1603/* Completion tracking for deferred_init_memmap() threads */
1604static atomic_t pgdat_init_n_undone __initdata;
1605static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1606
1607static inline void __init pgdat_init_report_one_done(void)
1608{
1609 if (atomic_dec_and_test(&pgdat_init_n_undone))
1610 complete(&pgdat_init_all_done_comp);
1611}
1612
1613/*
1614 * Returns true if page needs to be initialized or freed to buddy allocator.
1615 *
1616 * First we check if pfn is valid on architectures where it is possible to have
1617 * holes within pageblock_nr_pages. On systems where it is not possible, this
1618 * function is optimized out.
1619 *
1620 * Then, we check if a current large page is valid by only checking the validity
1621 * of the head pfn.
1622 */
1623static inline bool __init deferred_pfn_valid(unsigned long pfn)
1624{
1625 if (!pfn_valid_within(pfn))
1626 return false;
1627 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1628 return false;
1629 return true;
1630}
1631
1632/*
1633 * Free pages to buddy allocator. Try to free aligned pages in
1634 * pageblock_nr_pages sizes.
1635 */
1636static void __init deferred_free_pages(unsigned long pfn,
1637 unsigned long end_pfn)
1638{
1639 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1640 unsigned long nr_free = 0;
1641
1642 for (; pfn < end_pfn; pfn++) {
1643 if (!deferred_pfn_valid(pfn)) {
1644 deferred_free_range(pfn - nr_free, nr_free);
1645 nr_free = 0;
1646 } else if (!(pfn & nr_pgmask)) {
1647 deferred_free_range(pfn - nr_free, nr_free);
1648 nr_free = 1;
1649 touch_nmi_watchdog();
1650 } else {
1651 nr_free++;
1652 }
1653 }
1654 /* Free the last block of pages to allocator */
1655 deferred_free_range(pfn - nr_free, nr_free);
1656}
1657
1658/*
1659 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1660 * by performing it only once every pageblock_nr_pages.
1661 * Return number of pages initialized.
1662 */
1663static unsigned long __init deferred_init_pages(struct zone *zone,
1664 unsigned long pfn,
1665 unsigned long end_pfn)
1666{
1667 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1668 int nid = zone_to_nid(zone);
1669 unsigned long nr_pages = 0;
1670 int zid = zone_idx(zone);
1671 struct page *page = NULL;
1672
1673 for (; pfn < end_pfn; pfn++) {
1674 if (!deferred_pfn_valid(pfn)) {
1675 page = NULL;
1676 continue;
1677 } else if (!page || !(pfn & nr_pgmask)) {
1678 page = pfn_to_page(pfn);
1679 touch_nmi_watchdog();
1680 } else {
1681 page++;
1682 }
1683 __init_single_page(page, pfn, zid, nid);
1684 nr_pages++;
1685 }
1686 return (nr_pages);
1687}
1688
1689/*
1690 * This function is meant to pre-load the iterator for the zone init.
1691 * Specifically it walks through the ranges until we are caught up to the
1692 * first_init_pfn value and exits there. If we never encounter the value we
1693 * return false indicating there are no valid ranges left.
1694 */
1695static bool __init
1696deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1697 unsigned long *spfn, unsigned long *epfn,
1698 unsigned long first_init_pfn)
1699{
1700 u64 j;
1701
1702 /*
1703 * Start out by walking through the ranges in this zone that have
1704 * already been initialized. We don't need to do anything with them
1705 * so we just need to flush them out of the system.
1706 */
1707 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1708 if (*epfn <= first_init_pfn)
1709 continue;
1710 if (*spfn < first_init_pfn)
1711 *spfn = first_init_pfn;
1712 *i = j;
1713 return true;
1714 }
1715
1716 return false;
1717}
1718
1719/*
1720 * Initialize and free pages. We do it in two loops: first we initialize
1721 * struct page, then free to buddy allocator, because while we are
1722 * freeing pages we can access pages that are ahead (computing buddy
1723 * page in __free_one_page()).
1724 *
1725 * In order to try and keep some memory in the cache we have the loop
1726 * broken along max page order boundaries. This way we will not cause
1727 * any issues with the buddy page computation.
1728 */
1729static unsigned long __init
1730deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1731 unsigned long *end_pfn)
1732{
1733 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1734 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1735 unsigned long nr_pages = 0;
1736 u64 j = *i;
1737
1738 /* First we loop through and initialize the page values */
1739 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1740 unsigned long t;
1741
1742 if (mo_pfn <= *start_pfn)
1743 break;
1744
1745 t = min(mo_pfn, *end_pfn);
1746 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1747
1748 if (mo_pfn < *end_pfn) {
1749 *start_pfn = mo_pfn;
1750 break;
1751 }
1752 }
1753
1754 /* Reset values and now loop through freeing pages as needed */
1755 swap(j, *i);
1756
1757 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1758 unsigned long t;
1759
1760 if (mo_pfn <= spfn)
1761 break;
1762
1763 t = min(mo_pfn, epfn);
1764 deferred_free_pages(spfn, t);
1765
1766 if (mo_pfn <= epfn)
1767 break;
1768 }
1769
1770 return nr_pages;
1771}
1772
1773/* Initialise remaining memory on a node */
1774static int __init deferred_init_memmap(void *data)
1775{
1776 pg_data_t *pgdat = data;
1777 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1778 unsigned long spfn = 0, epfn = 0, nr_pages = 0;
1779 unsigned long first_init_pfn, flags;
1780 unsigned long start = jiffies;
1781 struct zone *zone;
1782 int zid;
1783 u64 i;
1784
1785 /* Bind memory initialisation thread to a local node if possible */
1786 if (!cpumask_empty(cpumask))
1787 set_cpus_allowed_ptr(current, cpumask);
1788
1789 pgdat_resize_lock(pgdat, &flags);
1790 first_init_pfn = pgdat->first_deferred_pfn;
1791 if (first_init_pfn == ULONG_MAX) {
1792 pgdat_resize_unlock(pgdat, &flags);
1793 pgdat_init_report_one_done();
1794 return 0;
1795 }
1796
1797 /* Sanity check boundaries */
1798 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1799 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1800 pgdat->first_deferred_pfn = ULONG_MAX;
1801
1802 /* Only the highest zone is deferred so find it */
1803 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1804 zone = pgdat->node_zones + zid;
1805 if (first_init_pfn < zone_end_pfn(zone))
1806 break;
1807 }
1808
1809 /* If the zone is empty somebody else may have cleared out the zone */
1810 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1811 first_init_pfn))
1812 goto zone_empty;
1813
1814 /*
1815 * Initialize and free pages in MAX_ORDER sized increments so
1816 * that we can avoid introducing any issues with the buddy
1817 * allocator.
1818 */
1819 while (spfn < epfn)
1820 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1821zone_empty:
1822 pgdat_resize_unlock(pgdat, &flags);
1823
1824 /* Sanity check that the next zone really is unpopulated */
1825 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1826
1827 pr_info("node %d initialised, %lu pages in %ums\n",
1828 pgdat->node_id, nr_pages, jiffies_to_msecs(jiffies - start));
1829
1830 pgdat_init_report_one_done();
1831 return 0;
1832}
1833
1834/*
1835 * If this zone has deferred pages, try to grow it by initializing enough
1836 * deferred pages to satisfy the allocation specified by order, rounded up to
1837 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1838 * of SECTION_SIZE bytes by initializing struct pages in increments of
1839 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1840 *
1841 * Return true when zone was grown, otherwise return false. We return true even
1842 * when we grow less than requested, to let the caller decide if there are
1843 * enough pages to satisfy the allocation.
1844 *
1845 * Note: We use noinline because this function is needed only during boot, and
1846 * it is called from a __ref function _deferred_grow_zone. This way we are
1847 * making sure that it is not inlined into permanent text section.
1848 */
1849static noinline bool __init
1850deferred_grow_zone(struct zone *zone, unsigned int order)
1851{
1852 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1853 pg_data_t *pgdat = zone->zone_pgdat;
1854 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1855 unsigned long spfn, epfn, flags;
1856 unsigned long nr_pages = 0;
1857 u64 i;
1858
1859 /* Only the last zone may have deferred pages */
1860 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1861 return false;
1862
1863 pgdat_resize_lock(pgdat, &flags);
1864
1865 /*
1866 * If deferred pages have been initialized while we were waiting for
1867 * the lock, return true, as the zone was grown. The caller will retry
1868 * this zone. We won't return to this function since the caller also
1869 * has this static branch.
1870 */
1871 if (!static_branch_unlikely(&deferred_pages)) {
1872 pgdat_resize_unlock(pgdat, &flags);
1873 return true;
1874 }
1875
1876 /*
1877 * If someone grew this zone while we were waiting for spinlock, return
1878 * true, as there might be enough pages already.
1879 */
1880 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1881 pgdat_resize_unlock(pgdat, &flags);
1882 return true;
1883 }
1884
1885 /* If the zone is empty somebody else may have cleared out the zone */
1886 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1887 first_deferred_pfn)) {
1888 pgdat->first_deferred_pfn = ULONG_MAX;
1889 pgdat_resize_unlock(pgdat, &flags);
1890 /* Retry only once. */
1891 return first_deferred_pfn != ULONG_MAX;
1892 }
1893
1894 /*
1895 * Initialize and free pages in MAX_ORDER sized increments so
1896 * that we can avoid introducing any issues with the buddy
1897 * allocator.
1898 */
1899 while (spfn < epfn) {
1900 /* update our first deferred PFN for this section */
1901 first_deferred_pfn = spfn;
1902
1903 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1904
1905 /* We should only stop along section boundaries */
1906 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1907 continue;
1908
1909 /* If our quota has been met we can stop here */
1910 if (nr_pages >= nr_pages_needed)
1911 break;
1912 }
1913
1914 pgdat->first_deferred_pfn = spfn;
1915 pgdat_resize_unlock(pgdat, &flags);
1916
1917 return nr_pages > 0;
1918}
1919
1920/*
1921 * deferred_grow_zone() is __init, but it is called from
1922 * get_page_from_freelist() during early boot until deferred_pages permanently
1923 * disables this call. This is why we have refdata wrapper to avoid warning,
1924 * and to ensure that the function body gets unloaded.
1925 */
1926static bool __ref
1927_deferred_grow_zone(struct zone *zone, unsigned int order)
1928{
1929 return deferred_grow_zone(zone, order);
1930}
1931
1932#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1933
1934void __init page_alloc_init_late(void)
1935{
1936 struct zone *zone;
1937 int nid;
1938
1939#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1940
1941 /* There will be num_node_state(N_MEMORY) threads */
1942 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1943 for_each_node_state(nid, N_MEMORY) {
1944 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1945 }
1946
1947 /* Block until all are initialised */
1948 wait_for_completion(&pgdat_init_all_done_comp);
1949
1950 /*
1951 * The number of managed pages has changed due to the initialisation
1952 * so the pcpu batch and high limits needs to be updated or the limits
1953 * will be artificially small.
1954 */
1955 for_each_populated_zone(zone)
1956 zone_pcp_update(zone);
1957
1958 /*
1959 * We initialized the rest of the deferred pages. Permanently disable
1960 * on-demand struct page initialization.
1961 */
1962 static_branch_disable(&deferred_pages);
1963
1964 /* Reinit limits that are based on free pages after the kernel is up */
1965 files_maxfiles_init();
1966#endif
1967
1968 /* Discard memblock private memory */
1969 memblock_discard();
1970
1971 for_each_node_state(nid, N_MEMORY)
1972 shuffle_free_memory(NODE_DATA(nid));
1973
1974 for_each_populated_zone(zone)
1975 set_zone_contiguous(zone);
1976
1977#ifdef CONFIG_DEBUG_PAGEALLOC
1978 init_debug_guardpage();
1979#endif
1980}
1981
1982#ifdef CONFIG_CMA
1983/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1984void __init init_cma_reserved_pageblock(struct page *page)
1985{
1986 unsigned i = pageblock_nr_pages;
1987 struct page *p = page;
1988
1989 do {
1990 __ClearPageReserved(p);
1991 set_page_count(p, 0);
1992 } while (++p, --i);
1993
1994 set_pageblock_migratetype(page, MIGRATE_CMA);
1995
1996 if (pageblock_order >= MAX_ORDER) {
1997 i = pageblock_nr_pages;
1998 p = page;
1999 do {
2000 set_page_refcounted(p);
2001 __free_pages(p, MAX_ORDER - 1);
2002 p += MAX_ORDER_NR_PAGES;
2003 } while (i -= MAX_ORDER_NR_PAGES);
2004 } else {
2005 set_page_refcounted(page);
2006 __free_pages(page, pageblock_order);
2007 }
2008
2009 adjust_managed_page_count(page, pageblock_nr_pages);
2010}
2011#endif
2012
2013/*
2014 * The order of subdivision here is critical for the IO subsystem.
2015 * Please do not alter this order without good reasons and regression
2016 * testing. Specifically, as large blocks of memory are subdivided,
2017 * the order in which smaller blocks are delivered depends on the order
2018 * they're subdivided in this function. This is the primary factor
2019 * influencing the order in which pages are delivered to the IO
2020 * subsystem according to empirical testing, and this is also justified
2021 * by considering the behavior of a buddy system containing a single
2022 * large block of memory acted on by a series of small allocations.
2023 * This behavior is a critical factor in sglist merging's success.
2024 *
2025 * -- nyc
2026 */
2027static inline void expand(struct zone *zone, struct page *page,
2028 int low, int high, struct free_area *area,
2029 int migratetype)
2030{
2031 unsigned long size = 1 << high;
2032
2033 while (high > low) {
2034 area--;
2035 high--;
2036 size >>= 1;
2037 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2038
2039 /*
2040 * Mark as guard pages (or page), that will allow to
2041 * merge back to allocator when buddy will be freed.
2042 * Corresponding page table entries will not be touched,
2043 * pages will stay not present in virtual address space
2044 */
2045 if (set_page_guard(zone, &page[size], high, migratetype))
2046 continue;
2047
2048 add_to_free_area(&page[size], area, migratetype);
2049 set_page_order(&page[size], high);
2050 }
2051}
2052
2053static void check_new_page_bad(struct page *page)
2054{
2055 const char *bad_reason = NULL;
2056 unsigned long bad_flags = 0;
2057
2058 if (unlikely(atomic_read(&page->_mapcount) != -1))
2059 bad_reason = "nonzero mapcount";
2060 if (unlikely(page->mapping != NULL))
2061 bad_reason = "non-NULL mapping";
2062 if (unlikely(page_ref_count(page) != 0))
2063 bad_reason = "nonzero _refcount";
2064 if (unlikely(page->flags & __PG_HWPOISON)) {
2065 bad_reason = "HWPoisoned (hardware-corrupted)";
2066 bad_flags = __PG_HWPOISON;
2067 /* Don't complain about hwpoisoned pages */
2068 page_mapcount_reset(page); /* remove PageBuddy */
2069 return;
2070 }
2071 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
2072 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
2073 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
2074 }
2075#ifdef CONFIG_MEMCG
2076 if (unlikely(page->mem_cgroup))
2077 bad_reason = "page still charged to cgroup";
2078#endif
2079 bad_page(page, bad_reason, bad_flags);
2080}
2081
2082/*
2083 * This page is about to be returned from the page allocator
2084 */
2085static inline int check_new_page(struct page *page)
2086{
2087 if (likely(page_expected_state(page,
2088 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2089 return 0;
2090
2091 check_new_page_bad(page);
2092 return 1;
2093}
2094
2095static inline bool free_pages_prezeroed(void)
2096{
2097 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2098 page_poisoning_enabled()) || want_init_on_free();
2099}
2100
2101#ifdef CONFIG_DEBUG_VM
2102/*
2103 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2104 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2105 * also checked when pcp lists are refilled from the free lists.
2106 */
2107static inline bool check_pcp_refill(struct page *page)
2108{
2109 if (debug_pagealloc_enabled())
2110 return check_new_page(page);
2111 else
2112 return false;
2113}
2114
2115static inline bool check_new_pcp(struct page *page)
2116{
2117 return check_new_page(page);
2118}
2119#else
2120/*
2121 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2122 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2123 * enabled, they are also checked when being allocated from the pcp lists.
2124 */
2125static inline bool check_pcp_refill(struct page *page)
2126{
2127 return check_new_page(page);
2128}
2129static inline bool check_new_pcp(struct page *page)
2130{
2131 if (debug_pagealloc_enabled())
2132 return check_new_page(page);
2133 else
2134 return false;
2135}
2136#endif /* CONFIG_DEBUG_VM */
2137
2138static bool check_new_pages(struct page *page, unsigned int order)
2139{
2140 int i;
2141 for (i = 0; i < (1 << order); i++) {
2142 struct page *p = page + i;
2143
2144 if (unlikely(check_new_page(p)))
2145 return true;
2146 }
2147
2148 return false;
2149}
2150
2151inline void post_alloc_hook(struct page *page, unsigned int order,
2152 gfp_t gfp_flags)
2153{
2154 set_page_private(page, 0);
2155 set_page_refcounted(page);
2156
2157 arch_alloc_page(page, order);
2158 if (debug_pagealloc_enabled())
2159 kernel_map_pages(page, 1 << order, 1);
2160 kasan_alloc_pages(page, order);
2161 kernel_poison_pages(page, 1 << order, 1);
2162 set_page_owner(page, order, gfp_flags);
2163}
2164
2165static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2166 unsigned int alloc_flags)
2167{
2168 post_alloc_hook(page, order, gfp_flags);
2169
2170 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2171 kernel_init_free_pages(page, 1 << order);
2172
2173 if (order && (gfp_flags & __GFP_COMP))
2174 prep_compound_page(page, order);
2175
2176 /*
2177 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2178 * allocate the page. The expectation is that the caller is taking
2179 * steps that will free more memory. The caller should avoid the page
2180 * being used for !PFMEMALLOC purposes.
2181 */
2182 if (alloc_flags & ALLOC_NO_WATERMARKS)
2183 set_page_pfmemalloc(page);
2184 else
2185 clear_page_pfmemalloc(page);
2186}
2187
2188/*
2189 * Go through the free lists for the given migratetype and remove
2190 * the smallest available page from the freelists
2191 */
2192static __always_inline
2193struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2194 int migratetype)
2195{
2196 unsigned int current_order;
2197 struct free_area *area;
2198 struct page *page;
2199
2200 /* Find a page of the appropriate size in the preferred list */
2201 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2202 area = &(zone->free_area[current_order]);
2203 page = get_page_from_free_area(area, migratetype);
2204 if (!page)
2205 continue;
2206 del_page_from_free_area(page, area);
2207 expand(zone, page, order, current_order, area, migratetype);
2208 set_pcppage_migratetype(page, migratetype);
2209 return page;
2210 }
2211
2212 return NULL;
2213}
2214
2215
2216/*
2217 * This array describes the order lists are fallen back to when
2218 * the free lists for the desirable migrate type are depleted
2219 */
2220static int fallbacks[MIGRATE_TYPES][4] = {
2221 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2222 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2223 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2224#ifdef CONFIG_CMA
2225 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2226#endif
2227#ifdef CONFIG_MEMORY_ISOLATION
2228 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2229#endif
2230};
2231
2232#ifdef CONFIG_CMA
2233static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2234 unsigned int order)
2235{
2236 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2237}
2238#else
2239static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2240 unsigned int order) { return NULL; }
2241#endif
2242
2243/*
2244 * Move the free pages in a range to the free lists of the requested type.
2245 * Note that start_page and end_pages are not aligned on a pageblock
2246 * boundary. If alignment is required, use move_freepages_block()
2247 */
2248static int move_freepages(struct zone *zone,
2249 struct page *start_page, struct page *end_page,
2250 int migratetype, int *num_movable)
2251{
2252 struct page *page;
2253 unsigned int order;
2254 int pages_moved = 0;
2255
2256 for (page = start_page; page <= end_page;) {
2257 if (!pfn_valid_within(page_to_pfn(page))) {
2258 page++;
2259 continue;
2260 }
2261
2262 if (!PageBuddy(page)) {
2263 /*
2264 * We assume that pages that could be isolated for
2265 * migration are movable. But we don't actually try
2266 * isolating, as that would be expensive.
2267 */
2268 if (num_movable &&
2269 (PageLRU(page) || __PageMovable(page)))
2270 (*num_movable)++;
2271
2272 page++;
2273 continue;
2274 }
2275
2276 /* Make sure we are not inadvertently changing nodes */
2277 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2278 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2279
2280 order = page_order(page);
2281 move_to_free_area(page, &zone->free_area[order], migratetype);
2282 page += 1 << order;
2283 pages_moved += 1 << order;
2284 }
2285
2286 return pages_moved;
2287}
2288
2289int move_freepages_block(struct zone *zone, struct page *page,
2290 int migratetype, int *num_movable)
2291{
2292 unsigned long start_pfn, end_pfn;
2293 struct page *start_page, *end_page;
2294
2295 if (num_movable)
2296 *num_movable = 0;
2297
2298 start_pfn = page_to_pfn(page);
2299 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2300 start_page = pfn_to_page(start_pfn);
2301 end_page = start_page + pageblock_nr_pages - 1;
2302 end_pfn = start_pfn + pageblock_nr_pages - 1;
2303
2304 /* Do not cross zone boundaries */
2305 if (!zone_spans_pfn(zone, start_pfn))
2306 start_page = page;
2307 if (!zone_spans_pfn(zone, end_pfn))
2308 return 0;
2309
2310 return move_freepages(zone, start_page, end_page, migratetype,
2311 num_movable);
2312}
2313
2314static void change_pageblock_range(struct page *pageblock_page,
2315 int start_order, int migratetype)
2316{
2317 int nr_pageblocks = 1 << (start_order - pageblock_order);
2318
2319 while (nr_pageblocks--) {
2320 set_pageblock_migratetype(pageblock_page, migratetype);
2321 pageblock_page += pageblock_nr_pages;
2322 }
2323}
2324
2325/*
2326 * When we are falling back to another migratetype during allocation, try to
2327 * steal extra free pages from the same pageblocks to satisfy further
2328 * allocations, instead of polluting multiple pageblocks.
2329 *
2330 * If we are stealing a relatively large buddy page, it is likely there will
2331 * be more free pages in the pageblock, so try to steal them all. For
2332 * reclaimable and unmovable allocations, we steal regardless of page size,
2333 * as fragmentation caused by those allocations polluting movable pageblocks
2334 * is worse than movable allocations stealing from unmovable and reclaimable
2335 * pageblocks.
2336 */
2337static bool can_steal_fallback(unsigned int order, int start_mt)
2338{
2339 /*
2340 * Leaving this order check is intended, although there is
2341 * relaxed order check in next check. The reason is that
2342 * we can actually steal whole pageblock if this condition met,
2343 * but, below check doesn't guarantee it and that is just heuristic
2344 * so could be changed anytime.
2345 */
2346 if (order >= pageblock_order)
2347 return true;
2348
2349 if (order >= pageblock_order / 2 ||
2350 start_mt == MIGRATE_RECLAIMABLE ||
2351 start_mt == MIGRATE_UNMOVABLE ||
2352 page_group_by_mobility_disabled)
2353 return true;
2354
2355 return false;
2356}
2357
2358static inline void boost_watermark(struct zone *zone)
2359{
2360 unsigned long max_boost;
2361
2362 if (!watermark_boost_factor)
2363 return;
2364
2365 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2366 watermark_boost_factor, 10000);
2367
2368 /*
2369 * high watermark may be uninitialised if fragmentation occurs
2370 * very early in boot so do not boost. We do not fall
2371 * through and boost by pageblock_nr_pages as failing
2372 * allocations that early means that reclaim is not going
2373 * to help and it may even be impossible to reclaim the
2374 * boosted watermark resulting in a hang.
2375 */
2376 if (!max_boost)
2377 return;
2378
2379 max_boost = max(pageblock_nr_pages, max_boost);
2380
2381 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2382 max_boost);
2383}
2384
2385/*
2386 * This function implements actual steal behaviour. If order is large enough,
2387 * we can steal whole pageblock. If not, we first move freepages in this
2388 * pageblock to our migratetype and determine how many already-allocated pages
2389 * are there in the pageblock with a compatible migratetype. If at least half
2390 * of pages are free or compatible, we can change migratetype of the pageblock
2391 * itself, so pages freed in the future will be put on the correct free list.
2392 */
2393static void steal_suitable_fallback(struct zone *zone, struct page *page,
2394 unsigned int alloc_flags, int start_type, bool whole_block)
2395{
2396 unsigned int current_order = page_order(page);
2397 struct free_area *area;
2398 int free_pages, movable_pages, alike_pages;
2399 int old_block_type;
2400
2401 old_block_type = get_pageblock_migratetype(page);
2402
2403 /*
2404 * This can happen due to races and we want to prevent broken
2405 * highatomic accounting.
2406 */
2407 if (is_migrate_highatomic(old_block_type))
2408 goto single_page;
2409
2410 /* Take ownership for orders >= pageblock_order */
2411 if (current_order >= pageblock_order) {
2412 change_pageblock_range(page, current_order, start_type);
2413 goto single_page;
2414 }
2415
2416 /*
2417 * Boost watermarks to increase reclaim pressure to reduce the
2418 * likelihood of future fallbacks. Wake kswapd now as the node
2419 * may be balanced overall and kswapd will not wake naturally.
2420 */
2421 boost_watermark(zone);
2422 if (alloc_flags & ALLOC_KSWAPD)
2423 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2424
2425 /* We are not allowed to try stealing from the whole block */
2426 if (!whole_block)
2427 goto single_page;
2428
2429 free_pages = move_freepages_block(zone, page, start_type,
2430 &movable_pages);
2431 /*
2432 * Determine how many pages are compatible with our allocation.
2433 * For movable allocation, it's the number of movable pages which
2434 * we just obtained. For other types it's a bit more tricky.
2435 */
2436 if (start_type == MIGRATE_MOVABLE) {
2437 alike_pages = movable_pages;
2438 } else {
2439 /*
2440 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2441 * to MOVABLE pageblock, consider all non-movable pages as
2442 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2443 * vice versa, be conservative since we can't distinguish the
2444 * exact migratetype of non-movable pages.
2445 */
2446 if (old_block_type == MIGRATE_MOVABLE)
2447 alike_pages = pageblock_nr_pages
2448 - (free_pages + movable_pages);
2449 else
2450 alike_pages = 0;
2451 }
2452
2453 /* moving whole block can fail due to zone boundary conditions */
2454 if (!free_pages)
2455 goto single_page;
2456
2457 /*
2458 * If a sufficient number of pages in the block are either free or of
2459 * comparable migratability as our allocation, claim the whole block.
2460 */
2461 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2462 page_group_by_mobility_disabled)
2463 set_pageblock_migratetype(page, start_type);
2464
2465 return;
2466
2467single_page:
2468 area = &zone->free_area[current_order];
2469 move_to_free_area(page, area, start_type);
2470}
2471
2472/*
2473 * Check whether there is a suitable fallback freepage with requested order.
2474 * If only_stealable is true, this function returns fallback_mt only if
2475 * we can steal other freepages all together. This would help to reduce
2476 * fragmentation due to mixed migratetype pages in one pageblock.
2477 */
2478int find_suitable_fallback(struct free_area *area, unsigned int order,
2479 int migratetype, bool only_stealable, bool *can_steal)
2480{
2481 int i;
2482 int fallback_mt;
2483
2484 if (area->nr_free == 0)
2485 return -1;
2486
2487 *can_steal = false;
2488 for (i = 0;; i++) {
2489 fallback_mt = fallbacks[migratetype][i];
2490 if (fallback_mt == MIGRATE_TYPES)
2491 break;
2492
2493 if (free_area_empty(area, fallback_mt))
2494 continue;
2495
2496 if (can_steal_fallback(order, migratetype))
2497 *can_steal = true;
2498
2499 if (!only_stealable)
2500 return fallback_mt;
2501
2502 if (*can_steal)
2503 return fallback_mt;
2504 }
2505
2506 return -1;
2507}
2508
2509/*
2510 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2511 * there are no empty page blocks that contain a page with a suitable order
2512 */
2513static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2514 unsigned int alloc_order)
2515{
2516 int mt;
2517 unsigned long max_managed, flags;
2518
2519 /*
2520 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2521 * Check is race-prone but harmless.
2522 */
2523 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2524 if (zone->nr_reserved_highatomic >= max_managed)
2525 return;
2526
2527 spin_lock_irqsave(&zone->lock, flags);
2528
2529 /* Recheck the nr_reserved_highatomic limit under the lock */
2530 if (zone->nr_reserved_highatomic >= max_managed)
2531 goto out_unlock;
2532
2533 /* Yoink! */
2534 mt = get_pageblock_migratetype(page);
2535 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2536 && !is_migrate_cma(mt)) {
2537 zone->nr_reserved_highatomic += pageblock_nr_pages;
2538 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2539 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2540 }
2541
2542out_unlock:
2543 spin_unlock_irqrestore(&zone->lock, flags);
2544}
2545
2546/*
2547 * Used when an allocation is about to fail under memory pressure. This
2548 * potentially hurts the reliability of high-order allocations when under
2549 * intense memory pressure but failed atomic allocations should be easier
2550 * to recover from than an OOM.
2551 *
2552 * If @force is true, try to unreserve a pageblock even though highatomic
2553 * pageblock is exhausted.
2554 */
2555static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2556 bool force)
2557{
2558 struct zonelist *zonelist = ac->zonelist;
2559 unsigned long flags;
2560 struct zoneref *z;
2561 struct zone *zone;
2562 struct page *page;
2563 int order;
2564 bool ret;
2565
2566 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2567 ac->nodemask) {
2568 /*
2569 * Preserve at least one pageblock unless memory pressure
2570 * is really high.
2571 */
2572 if (!force && zone->nr_reserved_highatomic <=
2573 pageblock_nr_pages)
2574 continue;
2575
2576 spin_lock_irqsave(&zone->lock, flags);
2577 for (order = 0; order < MAX_ORDER; order++) {
2578 struct free_area *area = &(zone->free_area[order]);
2579
2580 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2581 if (!page)
2582 continue;
2583
2584 /*
2585 * In page freeing path, migratetype change is racy so
2586 * we can counter several free pages in a pageblock
2587 * in this loop althoug we changed the pageblock type
2588 * from highatomic to ac->migratetype. So we should
2589 * adjust the count once.
2590 */
2591 if (is_migrate_highatomic_page(page)) {
2592 /*
2593 * It should never happen but changes to
2594 * locking could inadvertently allow a per-cpu
2595 * drain to add pages to MIGRATE_HIGHATOMIC
2596 * while unreserving so be safe and watch for
2597 * underflows.
2598 */
2599 zone->nr_reserved_highatomic -= min(
2600 pageblock_nr_pages,
2601 zone->nr_reserved_highatomic);
2602 }
2603
2604 /*
2605 * Convert to ac->migratetype and avoid the normal
2606 * pageblock stealing heuristics. Minimally, the caller
2607 * is doing the work and needs the pages. More
2608 * importantly, if the block was always converted to
2609 * MIGRATE_UNMOVABLE or another type then the number
2610 * of pageblocks that cannot be completely freed
2611 * may increase.
2612 */
2613 set_pageblock_migratetype(page, ac->migratetype);
2614 ret = move_freepages_block(zone, page, ac->migratetype,
2615 NULL);
2616 if (ret) {
2617 spin_unlock_irqrestore(&zone->lock, flags);
2618 return ret;
2619 }
2620 }
2621 spin_unlock_irqrestore(&zone->lock, flags);
2622 }
2623
2624 return false;
2625}
2626
2627/*
2628 * Try finding a free buddy page on the fallback list and put it on the free
2629 * list of requested migratetype, possibly along with other pages from the same
2630 * block, depending on fragmentation avoidance heuristics. Returns true if
2631 * fallback was found so that __rmqueue_smallest() can grab it.
2632 *
2633 * The use of signed ints for order and current_order is a deliberate
2634 * deviation from the rest of this file, to make the for loop
2635 * condition simpler.
2636 */
2637static __always_inline bool
2638__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2639 unsigned int alloc_flags)
2640{
2641 struct free_area *area;
2642 int current_order;
2643 int min_order = order;
2644 struct page *page;
2645 int fallback_mt;
2646 bool can_steal;
2647
2648 /*
2649 * Do not steal pages from freelists belonging to other pageblocks
2650 * i.e. orders < pageblock_order. If there are no local zones free,
2651 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2652 */
2653 if (alloc_flags & ALLOC_NOFRAGMENT)
2654 min_order = pageblock_order;
2655
2656 /*
2657 * Find the largest available free page in the other list. This roughly
2658 * approximates finding the pageblock with the most free pages, which
2659 * would be too costly to do exactly.
2660 */
2661 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2662 --current_order) {
2663 area = &(zone->free_area[current_order]);
2664 fallback_mt = find_suitable_fallback(area, current_order,
2665 start_migratetype, false, &can_steal);
2666 if (fallback_mt == -1)
2667 continue;
2668
2669 /*
2670 * We cannot steal all free pages from the pageblock and the
2671 * requested migratetype is movable. In that case it's better to
2672 * steal and split the smallest available page instead of the
2673 * largest available page, because even if the next movable
2674 * allocation falls back into a different pageblock than this
2675 * one, it won't cause permanent fragmentation.
2676 */
2677 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2678 && current_order > order)
2679 goto find_smallest;
2680
2681 goto do_steal;
2682 }
2683
2684 return false;
2685
2686find_smallest:
2687 for (current_order = order; current_order < MAX_ORDER;
2688 current_order++) {
2689 area = &(zone->free_area[current_order]);
2690 fallback_mt = find_suitable_fallback(area, current_order,
2691 start_migratetype, false, &can_steal);
2692 if (fallback_mt != -1)
2693 break;
2694 }
2695
2696 /*
2697 * This should not happen - we already found a suitable fallback
2698 * when looking for the largest page.
2699 */
2700 VM_BUG_ON(current_order == MAX_ORDER);
2701
2702do_steal:
2703 page = get_page_from_free_area(area, fallback_mt);
2704
2705 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2706 can_steal);
2707
2708 trace_mm_page_alloc_extfrag(page, order, current_order,
2709 start_migratetype, fallback_mt);
2710
2711 return true;
2712
2713}
2714
2715/*
2716 * Do the hard work of removing an element from the buddy allocator.
2717 * Call me with the zone->lock already held.
2718 */
2719static __always_inline struct page *
2720__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2721 unsigned int alloc_flags)
2722{
2723 struct page *page;
2724
2725retry:
2726 page = __rmqueue_smallest(zone, order, migratetype);
2727 if (unlikely(!page)) {
2728 if (migratetype == MIGRATE_MOVABLE)
2729 page = __rmqueue_cma_fallback(zone, order);
2730
2731 if (!page && __rmqueue_fallback(zone, order, migratetype,
2732 alloc_flags))
2733 goto retry;
2734 }
2735
2736 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2737 return page;
2738}
2739
2740/*
2741 * Obtain a specified number of elements from the buddy allocator, all under
2742 * a single hold of the lock, for efficiency. Add them to the supplied list.
2743 * Returns the number of new pages which were placed at *list.
2744 */
2745static int rmqueue_bulk(struct zone *zone, unsigned int order,
2746 unsigned long count, struct list_head *list,
2747 int migratetype, unsigned int alloc_flags)
2748{
2749 int i, alloced = 0;
2750
2751 spin_lock(&zone->lock);
2752 for (i = 0; i < count; ++i) {
2753 struct page *page = __rmqueue(zone, order, migratetype,
2754 alloc_flags);
2755 if (unlikely(page == NULL))
2756 break;
2757
2758 if (unlikely(check_pcp_refill(page)))
2759 continue;
2760
2761 /*
2762 * Split buddy pages returned by expand() are received here in
2763 * physical page order. The page is added to the tail of
2764 * caller's list. From the callers perspective, the linked list
2765 * is ordered by page number under some conditions. This is
2766 * useful for IO devices that can forward direction from the
2767 * head, thus also in the physical page order. This is useful
2768 * for IO devices that can merge IO requests if the physical
2769 * pages are ordered properly.
2770 */
2771 list_add_tail(&page->lru, list);
2772 alloced++;
2773 if (is_migrate_cma(get_pcppage_migratetype(page)))
2774 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2775 -(1 << order));
2776 }
2777
2778 /*
2779 * i pages were removed from the buddy list even if some leak due
2780 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2781 * on i. Do not confuse with 'alloced' which is the number of
2782 * pages added to the pcp list.
2783 */
2784 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2785 spin_unlock(&zone->lock);
2786 return alloced;
2787}
2788
2789#ifdef CONFIG_NUMA
2790/*
2791 * Called from the vmstat counter updater to drain pagesets of this
2792 * currently executing processor on remote nodes after they have
2793 * expired.
2794 *
2795 * Note that this function must be called with the thread pinned to
2796 * a single processor.
2797 */
2798void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2799{
2800 unsigned long flags;
2801 int to_drain, batch;
2802
2803 local_irq_save(flags);
2804 batch = READ_ONCE(pcp->batch);
2805 to_drain = min(pcp->count, batch);
2806 if (to_drain > 0)
2807 free_pcppages_bulk(zone, to_drain, pcp);
2808 local_irq_restore(flags);
2809}
2810#endif
2811
2812/*
2813 * Drain pcplists of the indicated processor and zone.
2814 *
2815 * The processor must either be the current processor and the
2816 * thread pinned to the current processor or a processor that
2817 * is not online.
2818 */
2819static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2820{
2821 unsigned long flags;
2822 struct per_cpu_pageset *pset;
2823 struct per_cpu_pages *pcp;
2824
2825 local_irq_save(flags);
2826 pset = per_cpu_ptr(zone->pageset, cpu);
2827
2828 pcp = &pset->pcp;
2829 if (pcp->count)
2830 free_pcppages_bulk(zone, pcp->count, pcp);
2831 local_irq_restore(flags);
2832}
2833
2834/*
2835 * Drain pcplists of all zones on the indicated processor.
2836 *
2837 * The processor must either be the current processor and the
2838 * thread pinned to the current processor or a processor that
2839 * is not online.
2840 */
2841static void drain_pages(unsigned int cpu)
2842{
2843 struct zone *zone;
2844
2845 for_each_populated_zone(zone) {
2846 drain_pages_zone(cpu, zone);
2847 }
2848}
2849
2850/*
2851 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2852 *
2853 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2854 * the single zone's pages.
2855 */
2856void drain_local_pages(struct zone *zone)
2857{
2858 int cpu = smp_processor_id();
2859
2860 if (zone)
2861 drain_pages_zone(cpu, zone);
2862 else
2863 drain_pages(cpu);
2864}
2865
2866static void drain_local_pages_wq(struct work_struct *work)
2867{
2868 struct pcpu_drain *drain;
2869
2870 drain = container_of(work, struct pcpu_drain, work);
2871
2872 /*
2873 * drain_all_pages doesn't use proper cpu hotplug protection so
2874 * we can race with cpu offline when the WQ can move this from
2875 * a cpu pinned worker to an unbound one. We can operate on a different
2876 * cpu which is allright but we also have to make sure to not move to
2877 * a different one.
2878 */
2879 preempt_disable();
2880 drain_local_pages(drain->zone);
2881 preempt_enable();
2882}
2883
2884/*
2885 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2886 *
2887 * When zone parameter is non-NULL, spill just the single zone's pages.
2888 *
2889 * Note that this can be extremely slow as the draining happens in a workqueue.
2890 */
2891void drain_all_pages(struct zone *zone)
2892{
2893 int cpu;
2894
2895 /*
2896 * Allocate in the BSS so we wont require allocation in
2897 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2898 */
2899 static cpumask_t cpus_with_pcps;
2900
2901 /*
2902 * Make sure nobody triggers this path before mm_percpu_wq is fully
2903 * initialized.
2904 */
2905 if (WARN_ON_ONCE(!mm_percpu_wq))
2906 return;
2907
2908 /*
2909 * Do not drain if one is already in progress unless it's specific to
2910 * a zone. Such callers are primarily CMA and memory hotplug and need
2911 * the drain to be complete when the call returns.
2912 */
2913 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2914 if (!zone)
2915 return;
2916 mutex_lock(&pcpu_drain_mutex);
2917 }
2918
2919 /*
2920 * We don't care about racing with CPU hotplug event
2921 * as offline notification will cause the notified
2922 * cpu to drain that CPU pcps and on_each_cpu_mask
2923 * disables preemption as part of its processing
2924 */
2925 for_each_online_cpu(cpu) {
2926 struct per_cpu_pageset *pcp;
2927 struct zone *z;
2928 bool has_pcps = false;
2929
2930 if (zone) {
2931 pcp = per_cpu_ptr(zone->pageset, cpu);
2932 if (pcp->pcp.count)
2933 has_pcps = true;
2934 } else {
2935 for_each_populated_zone(z) {
2936 pcp = per_cpu_ptr(z->pageset, cpu);
2937 if (pcp->pcp.count) {
2938 has_pcps = true;
2939 break;
2940 }
2941 }
2942 }
2943
2944 if (has_pcps)
2945 cpumask_set_cpu(cpu, &cpus_with_pcps);
2946 else
2947 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2948 }
2949
2950 for_each_cpu(cpu, &cpus_with_pcps) {
2951 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
2952
2953 drain->zone = zone;
2954 INIT_WORK(&drain->work, drain_local_pages_wq);
2955 queue_work_on(cpu, mm_percpu_wq, &drain->work);
2956 }
2957 for_each_cpu(cpu, &cpus_with_pcps)
2958 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
2959
2960 mutex_unlock(&pcpu_drain_mutex);
2961}
2962
2963#ifdef CONFIG_HIBERNATION
2964
2965/*
2966 * Touch the watchdog for every WD_PAGE_COUNT pages.
2967 */
2968#define WD_PAGE_COUNT (128*1024)
2969
2970void mark_free_pages(struct zone *zone)
2971{
2972 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2973 unsigned long flags;
2974 unsigned int order, t;
2975 struct page *page;
2976
2977 if (zone_is_empty(zone))
2978 return;
2979
2980 spin_lock_irqsave(&zone->lock, flags);
2981
2982 max_zone_pfn = zone_end_pfn(zone);
2983 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2984 if (pfn_valid(pfn)) {
2985 page = pfn_to_page(pfn);
2986
2987 if (!--page_count) {
2988 touch_nmi_watchdog();
2989 page_count = WD_PAGE_COUNT;
2990 }
2991
2992 if (page_zone(page) != zone)
2993 continue;
2994
2995 if (!swsusp_page_is_forbidden(page))
2996 swsusp_unset_page_free(page);
2997 }
2998
2999 for_each_migratetype_order(order, t) {
3000 list_for_each_entry(page,
3001 &zone->free_area[order].free_list[t], lru) {
3002 unsigned long i;
3003
3004 pfn = page_to_pfn(page);
3005 for (i = 0; i < (1UL << order); i++) {
3006 if (!--page_count) {
3007 touch_nmi_watchdog();
3008 page_count = WD_PAGE_COUNT;
3009 }
3010 swsusp_set_page_free(pfn_to_page(pfn + i));
3011 }
3012 }
3013 }
3014 spin_unlock_irqrestore(&zone->lock, flags);
3015}
3016#endif /* CONFIG_PM */
3017
3018static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3019{
3020 int migratetype;
3021
3022 if (!free_pcp_prepare(page))
3023 return false;
3024
3025 migratetype = get_pfnblock_migratetype(page, pfn);
3026 set_pcppage_migratetype(page, migratetype);
3027 return true;
3028}
3029
3030static void free_unref_page_commit(struct page *page, unsigned long pfn)
3031{
3032 struct zone *zone = page_zone(page);
3033 struct per_cpu_pages *pcp;
3034 int migratetype;
3035
3036 migratetype = get_pcppage_migratetype(page);
3037 __count_vm_event(PGFREE);
3038
3039 /*
3040 * We only track unmovable, reclaimable and movable on pcp lists.
3041 * Free ISOLATE pages back to the allocator because they are being
3042 * offlined but treat HIGHATOMIC as movable pages so we can get those
3043 * areas back if necessary. Otherwise, we may have to free
3044 * excessively into the page allocator
3045 */
3046 if (migratetype >= MIGRATE_PCPTYPES) {
3047 if (unlikely(is_migrate_isolate(migratetype))) {
3048 free_one_page(zone, page, pfn, 0, migratetype);
3049 return;
3050 }
3051 migratetype = MIGRATE_MOVABLE;
3052 }
3053
3054 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3055 list_add(&page->lru, &pcp->lists[migratetype]);
3056 pcp->count++;
3057 if (pcp->count >= pcp->high) {
3058 unsigned long batch = READ_ONCE(pcp->batch);
3059 free_pcppages_bulk(zone, batch, pcp);
3060 }
3061}
3062
3063/*
3064 * Free a 0-order page
3065 */
3066void free_unref_page(struct page *page)
3067{
3068 unsigned long flags;
3069 unsigned long pfn = page_to_pfn(page);
3070
3071 if (!free_unref_page_prepare(page, pfn))
3072 return;
3073
3074 local_irq_save(flags);
3075 free_unref_page_commit(page, pfn);
3076 local_irq_restore(flags);
3077}
3078
3079/*
3080 * Free a list of 0-order pages
3081 */
3082void free_unref_page_list(struct list_head *list)
3083{
3084 struct page *page, *next;
3085 unsigned long flags, pfn;
3086 int batch_count = 0;
3087
3088 /* Prepare pages for freeing */
3089 list_for_each_entry_safe(page, next, list, lru) {
3090 pfn = page_to_pfn(page);
3091 if (!free_unref_page_prepare(page, pfn))
3092 list_del(&page->lru);
3093 set_page_private(page, pfn);
3094 }
3095
3096 local_irq_save(flags);
3097 list_for_each_entry_safe(page, next, list, lru) {
3098 unsigned long pfn = page_private(page);
3099
3100 set_page_private(page, 0);
3101 trace_mm_page_free_batched(page);
3102 free_unref_page_commit(page, pfn);
3103
3104 /*
3105 * Guard against excessive IRQ disabled times when we get
3106 * a large list of pages to free.
3107 */
3108 if (++batch_count == SWAP_CLUSTER_MAX) {
3109 local_irq_restore(flags);
3110 batch_count = 0;
3111 local_irq_save(flags);
3112 }
3113 }
3114 local_irq_restore(flags);
3115}
3116
3117/*
3118 * split_page takes a non-compound higher-order page, and splits it into
3119 * n (1<<order) sub-pages: page[0..n]
3120 * Each sub-page must be freed individually.
3121 *
3122 * Note: this is probably too low level an operation for use in drivers.
3123 * Please consult with lkml before using this in your driver.
3124 */
3125void split_page(struct page *page, unsigned int order)
3126{
3127 int i;
3128
3129 VM_BUG_ON_PAGE(PageCompound(page), page);
3130 VM_BUG_ON_PAGE(!page_count(page), page);
3131
3132 for (i = 1; i < (1 << order); i++)
3133 set_page_refcounted(page + i);
3134 split_page_owner(page, order);
3135}
3136EXPORT_SYMBOL_GPL(split_page);
3137
3138int __isolate_free_page(struct page *page, unsigned int order)
3139{
3140 struct free_area *area = &page_zone(page)->free_area[order];
3141 unsigned long watermark;
3142 struct zone *zone;
3143 int mt;
3144
3145 BUG_ON(!PageBuddy(page));
3146
3147 zone = page_zone(page);
3148 mt = get_pageblock_migratetype(page);
3149
3150 if (!is_migrate_isolate(mt)) {
3151 /*
3152 * Obey watermarks as if the page was being allocated. We can
3153 * emulate a high-order watermark check with a raised order-0
3154 * watermark, because we already know our high-order page
3155 * exists.
3156 */
3157 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3158 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3159 return 0;
3160
3161 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3162 }
3163
3164 /* Remove page from free list */
3165
3166 del_page_from_free_area(page, area);
3167
3168 /*
3169 * Set the pageblock if the isolated page is at least half of a
3170 * pageblock
3171 */
3172 if (order >= pageblock_order - 1) {
3173 struct page *endpage = page + (1 << order) - 1;
3174 for (; page < endpage; page += pageblock_nr_pages) {
3175 int mt = get_pageblock_migratetype(page);
3176 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3177 && !is_migrate_highatomic(mt))
3178 set_pageblock_migratetype(page,
3179 MIGRATE_MOVABLE);
3180 }
3181 }
3182
3183
3184 return 1UL << order;
3185}
3186
3187/*
3188 * Update NUMA hit/miss statistics
3189 *
3190 * Must be called with interrupts disabled.
3191 */
3192static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3193{
3194#ifdef CONFIG_NUMA
3195 enum numa_stat_item local_stat = NUMA_LOCAL;
3196
3197 /* skip numa counters update if numa stats is disabled */
3198 if (!static_branch_likely(&vm_numa_stat_key))
3199 return;
3200
3201 if (zone_to_nid(z) != numa_node_id())
3202 local_stat = NUMA_OTHER;
3203
3204 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3205 __inc_numa_state(z, NUMA_HIT);
3206 else {
3207 __inc_numa_state(z, NUMA_MISS);
3208 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3209 }
3210 __inc_numa_state(z, local_stat);
3211#endif
3212}
3213
3214/* Remove page from the per-cpu list, caller must protect the list */
3215static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3216 unsigned int alloc_flags,
3217 struct per_cpu_pages *pcp,
3218 struct list_head *list)
3219{
3220 struct page *page;
3221
3222 do {
3223 if (list_empty(list)) {
3224 pcp->count += rmqueue_bulk(zone, 0,
3225 pcp->batch, list,
3226 migratetype, alloc_flags);
3227 if (unlikely(list_empty(list)))
3228 return NULL;
3229 }
3230
3231 page = list_first_entry(list, struct page, lru);
3232 list_del(&page->lru);
3233 pcp->count--;
3234 } while (check_new_pcp(page));
3235
3236 return page;
3237}
3238
3239/* Lock and remove page from the per-cpu list */
3240static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3241 struct zone *zone, gfp_t gfp_flags,
3242 int migratetype, unsigned int alloc_flags)
3243{
3244 struct per_cpu_pages *pcp;
3245 struct list_head *list;
3246 struct page *page;
3247 unsigned long flags;
3248
3249 local_irq_save(flags);
3250 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3251 list = &pcp->lists[migratetype];
3252 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3253 if (page) {
3254 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3255 zone_statistics(preferred_zone, zone);
3256 }
3257 local_irq_restore(flags);
3258 return page;
3259}
3260
3261/*
3262 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3263 */
3264static inline
3265struct page *rmqueue(struct zone *preferred_zone,
3266 struct zone *zone, unsigned int order,
3267 gfp_t gfp_flags, unsigned int alloc_flags,
3268 int migratetype)
3269{
3270 unsigned long flags;
3271 struct page *page;
3272
3273 if (likely(order == 0)) {
3274 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3275 migratetype, alloc_flags);
3276 goto out;
3277 }
3278
3279 /*
3280 * We most definitely don't want callers attempting to
3281 * allocate greater than order-1 page units with __GFP_NOFAIL.
3282 */
3283 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3284 spin_lock_irqsave(&zone->lock, flags);
3285
3286 do {
3287 page = NULL;
3288 if (alloc_flags & ALLOC_HARDER) {
3289 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3290 if (page)
3291 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3292 }
3293 if (!page)
3294 page = __rmqueue(zone, order, migratetype, alloc_flags);
3295 } while (page && check_new_pages(page, order));
3296 spin_unlock(&zone->lock);
3297 if (!page)
3298 goto failed;
3299 __mod_zone_freepage_state(zone, -(1 << order),
3300 get_pcppage_migratetype(page));
3301
3302 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3303 zone_statistics(preferred_zone, zone);
3304 local_irq_restore(flags);
3305
3306out:
3307 /* Separate test+clear to avoid unnecessary atomics */
3308 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3309 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3310 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3311 }
3312
3313 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3314 return page;
3315
3316failed:
3317 local_irq_restore(flags);
3318 return NULL;
3319}
3320
3321#ifdef CONFIG_FAIL_PAGE_ALLOC
3322
3323static struct {
3324 struct fault_attr attr;
3325
3326 bool ignore_gfp_highmem;
3327 bool ignore_gfp_reclaim;
3328 u32 min_order;
3329} fail_page_alloc = {
3330 .attr = FAULT_ATTR_INITIALIZER,
3331 .ignore_gfp_reclaim = true,
3332 .ignore_gfp_highmem = true,
3333 .min_order = 1,
3334};
3335
3336static int __init setup_fail_page_alloc(char *str)
3337{
3338 return setup_fault_attr(&fail_page_alloc.attr, str);
3339}
3340__setup("fail_page_alloc=", setup_fail_page_alloc);
3341
3342static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3343{
3344 if (order < fail_page_alloc.min_order)
3345 return false;
3346 if (gfp_mask & __GFP_NOFAIL)
3347 return false;
3348 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3349 return false;
3350 if (fail_page_alloc.ignore_gfp_reclaim &&
3351 (gfp_mask & __GFP_DIRECT_RECLAIM))
3352 return false;
3353
3354 return should_fail(&fail_page_alloc.attr, 1 << order);
3355}
3356
3357#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3358
3359static int __init fail_page_alloc_debugfs(void)
3360{
3361 umode_t mode = S_IFREG | 0600;
3362 struct dentry *dir;
3363
3364 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3365 &fail_page_alloc.attr);
3366
3367 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3368 &fail_page_alloc.ignore_gfp_reclaim);
3369 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3370 &fail_page_alloc.ignore_gfp_highmem);
3371 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3372
3373 return 0;
3374}
3375
3376late_initcall(fail_page_alloc_debugfs);
3377
3378#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3379
3380#else /* CONFIG_FAIL_PAGE_ALLOC */
3381
3382static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3383{
3384 return false;
3385}
3386
3387#endif /* CONFIG_FAIL_PAGE_ALLOC */
3388
3389static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3390{
3391 return __should_fail_alloc_page(gfp_mask, order);
3392}
3393ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3394
3395/*
3396 * Return true if free base pages are above 'mark'. For high-order checks it
3397 * will return true of the order-0 watermark is reached and there is at least
3398 * one free page of a suitable size. Checking now avoids taking the zone lock
3399 * to check in the allocation paths if no pages are free.
3400 */
3401bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3402 int classzone_idx, unsigned int alloc_flags,
3403 long free_pages)
3404{
3405 long min = mark;
3406 int o;
3407 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3408
3409 /* free_pages may go negative - that's OK */
3410 free_pages -= (1 << order) - 1;
3411
3412 if (alloc_flags & ALLOC_HIGH)
3413 min -= min / 2;
3414
3415 /*
3416 * If the caller does not have rights to ALLOC_HARDER then subtract
3417 * the high-atomic reserves. This will over-estimate the size of the
3418 * atomic reserve but it avoids a search.
3419 */
3420 if (likely(!alloc_harder)) {
3421 free_pages -= z->nr_reserved_highatomic;
3422 } else {
3423 /*
3424 * OOM victims can try even harder than normal ALLOC_HARDER
3425 * users on the grounds that it's definitely going to be in
3426 * the exit path shortly and free memory. Any allocation it
3427 * makes during the free path will be small and short-lived.
3428 */
3429 if (alloc_flags & ALLOC_OOM)
3430 min -= min / 2;
3431 else
3432 min -= min / 4;
3433 }
3434
3435
3436#ifdef CONFIG_CMA
3437 /* If allocation can't use CMA areas don't use free CMA pages */
3438 if (!(alloc_flags & ALLOC_CMA))
3439 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3440#endif
3441
3442 /*
3443 * Check watermarks for an order-0 allocation request. If these
3444 * are not met, then a high-order request also cannot go ahead
3445 * even if a suitable page happened to be free.
3446 */
3447 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3448 return false;
3449
3450 /* If this is an order-0 request then the watermark is fine */
3451 if (!order)
3452 return true;
3453
3454 /* For a high-order request, check at least one suitable page is free */
3455 for (o = order; o < MAX_ORDER; o++) {
3456 struct free_area *area = &z->free_area[o];
3457 int mt;
3458
3459 if (!area->nr_free)
3460 continue;
3461
3462 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3463 if (!free_area_empty(area, mt))
3464 return true;
3465 }
3466
3467#ifdef CONFIG_CMA
3468 if ((alloc_flags & ALLOC_CMA) &&
3469 !free_area_empty(area, MIGRATE_CMA)) {
3470 return true;
3471 }
3472#endif
3473 if (alloc_harder &&
3474 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3475 return true;
3476 }
3477 return false;
3478}
3479
3480bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3481 int classzone_idx, unsigned int alloc_flags)
3482{
3483 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3484 zone_page_state(z, NR_FREE_PAGES));
3485}
3486
3487static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3488 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3489{
3490 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3491 long cma_pages = 0;
3492
3493#ifdef CONFIG_CMA
3494 /* If allocation can't use CMA areas don't use free CMA pages */
3495 if (!(alloc_flags & ALLOC_CMA))
3496 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3497#endif
3498
3499 /*
3500 * Fast check for order-0 only. If this fails then the reserves
3501 * need to be calculated. There is a corner case where the check
3502 * passes but only the high-order atomic reserve are free. If
3503 * the caller is !atomic then it'll uselessly search the free
3504 * list. That corner case is then slower but it is harmless.
3505 */
3506 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3507 return true;
3508
3509 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3510 free_pages);
3511}
3512
3513bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3514 unsigned long mark, int classzone_idx)
3515{
3516 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3517
3518 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3519 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3520
3521 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3522 free_pages);
3523}
3524
3525#ifdef CONFIG_NUMA
3526static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3527{
3528 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3529 node_reclaim_distance;
3530}
3531#else /* CONFIG_NUMA */
3532static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3533{
3534 return true;
3535}
3536#endif /* CONFIG_NUMA */
3537
3538/*
3539 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3540 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3541 * premature use of a lower zone may cause lowmem pressure problems that
3542 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3543 * probably too small. It only makes sense to spread allocations to avoid
3544 * fragmentation between the Normal and DMA32 zones.
3545 */
3546static inline unsigned int
3547alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3548{
3549 unsigned int alloc_flags = 0;
3550
3551 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
3552 alloc_flags |= ALLOC_KSWAPD;
3553
3554#ifdef CONFIG_ZONE_DMA32
3555 if (!zone)
3556 return alloc_flags;
3557
3558 if (zone_idx(zone) != ZONE_NORMAL)
3559 return alloc_flags;
3560
3561 /*
3562 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3563 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3564 * on UMA that if Normal is populated then so is DMA32.
3565 */
3566 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3567 if (nr_online_nodes > 1 && !populated_zone(--zone))
3568 return alloc_flags;
3569
3570 alloc_flags |= ALLOC_NOFRAGMENT;
3571#endif /* CONFIG_ZONE_DMA32 */
3572 return alloc_flags;
3573}
3574
3575/*
3576 * get_page_from_freelist goes through the zonelist trying to allocate
3577 * a page.
3578 */
3579static struct page *
3580get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3581 const struct alloc_context *ac)
3582{
3583 struct zoneref *z;
3584 struct zone *zone;
3585 struct pglist_data *last_pgdat_dirty_limit = NULL;
3586 bool no_fallback;
3587
3588retry:
3589 /*
3590 * Scan zonelist, looking for a zone with enough free.
3591 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3592 */
3593 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3594 z = ac->preferred_zoneref;
3595 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3596 ac->nodemask) {
3597 struct page *page;
3598 unsigned long mark;
3599
3600 if (cpusets_enabled() &&
3601 (alloc_flags & ALLOC_CPUSET) &&
3602 !__cpuset_zone_allowed(zone, gfp_mask))
3603 continue;
3604 /*
3605 * When allocating a page cache page for writing, we
3606 * want to get it from a node that is within its dirty
3607 * limit, such that no single node holds more than its
3608 * proportional share of globally allowed dirty pages.
3609 * The dirty limits take into account the node's
3610 * lowmem reserves and high watermark so that kswapd
3611 * should be able to balance it without having to
3612 * write pages from its LRU list.
3613 *
3614 * XXX: For now, allow allocations to potentially
3615 * exceed the per-node dirty limit in the slowpath
3616 * (spread_dirty_pages unset) before going into reclaim,
3617 * which is important when on a NUMA setup the allowed
3618 * nodes are together not big enough to reach the
3619 * global limit. The proper fix for these situations
3620 * will require awareness of nodes in the
3621 * dirty-throttling and the flusher threads.
3622 */
3623 if (ac->spread_dirty_pages) {
3624 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3625 continue;
3626
3627 if (!node_dirty_ok(zone->zone_pgdat)) {
3628 last_pgdat_dirty_limit = zone->zone_pgdat;
3629 continue;
3630 }
3631 }
3632
3633 if (no_fallback && nr_online_nodes > 1 &&
3634 zone != ac->preferred_zoneref->zone) {
3635 int local_nid;
3636
3637 /*
3638 * If moving to a remote node, retry but allow
3639 * fragmenting fallbacks. Locality is more important
3640 * than fragmentation avoidance.
3641 */
3642 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3643 if (zone_to_nid(zone) != local_nid) {
3644 alloc_flags &= ~ALLOC_NOFRAGMENT;
3645 goto retry;
3646 }
3647 }
3648
3649 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3650 if (!zone_watermark_fast(zone, order, mark,
3651 ac_classzone_idx(ac), alloc_flags)) {
3652 int ret;
3653
3654#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3655 /*
3656 * Watermark failed for this zone, but see if we can
3657 * grow this zone if it contains deferred pages.
3658 */
3659 if (static_branch_unlikely(&deferred_pages)) {
3660 if (_deferred_grow_zone(zone, order))
3661 goto try_this_zone;
3662 }
3663#endif
3664 /* Checked here to keep the fast path fast */
3665 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3666 if (alloc_flags & ALLOC_NO_WATERMARKS)
3667 goto try_this_zone;
3668
3669 if (node_reclaim_mode == 0 ||
3670 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3671 continue;
3672
3673 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3674 switch (ret) {
3675 case NODE_RECLAIM_NOSCAN:
3676 /* did not scan */
3677 continue;
3678 case NODE_RECLAIM_FULL:
3679 /* scanned but unreclaimable */
3680 continue;
3681 default:
3682 /* did we reclaim enough */
3683 if (zone_watermark_ok(zone, order, mark,
3684 ac_classzone_idx(ac), alloc_flags))
3685 goto try_this_zone;
3686
3687 continue;
3688 }
3689 }
3690
3691try_this_zone:
3692 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3693 gfp_mask, alloc_flags, ac->migratetype);
3694 if (page) {
3695 prep_new_page(page, order, gfp_mask, alloc_flags);
3696
3697 /*
3698 * If this is a high-order atomic allocation then check
3699 * if the pageblock should be reserved for the future
3700 */
3701 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3702 reserve_highatomic_pageblock(page, zone, order);
3703
3704 return page;
3705 } else {
3706#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3707 /* Try again if zone has deferred pages */
3708 if (static_branch_unlikely(&deferred_pages)) {
3709 if (_deferred_grow_zone(zone, order))
3710 goto try_this_zone;
3711 }
3712#endif
3713 }
3714 }
3715
3716 /*
3717 * It's possible on a UMA machine to get through all zones that are
3718 * fragmented. If avoiding fragmentation, reset and try again.
3719 */
3720 if (no_fallback) {
3721 alloc_flags &= ~ALLOC_NOFRAGMENT;
3722 goto retry;
3723 }
3724
3725 return NULL;
3726}
3727
3728static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3729{
3730 unsigned int filter = SHOW_MEM_FILTER_NODES;
3731
3732 /*
3733 * This documents exceptions given to allocations in certain
3734 * contexts that are allowed to allocate outside current's set
3735 * of allowed nodes.
3736 */
3737 if (!(gfp_mask & __GFP_NOMEMALLOC))
3738 if (tsk_is_oom_victim(current) ||
3739 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3740 filter &= ~SHOW_MEM_FILTER_NODES;
3741 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3742 filter &= ~SHOW_MEM_FILTER_NODES;
3743
3744 show_mem(filter, nodemask);
3745}
3746
3747void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3748{
3749 struct va_format vaf;
3750 va_list args;
3751 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3752
3753 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3754 return;
3755
3756 va_start(args, fmt);
3757 vaf.fmt = fmt;
3758 vaf.va = &args;
3759 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3760 current->comm, &vaf, gfp_mask, &gfp_mask,
3761 nodemask_pr_args(nodemask));
3762 va_end(args);
3763
3764 cpuset_print_current_mems_allowed();
3765 pr_cont("\n");
3766 dump_stack();
3767 warn_alloc_show_mem(gfp_mask, nodemask);
3768}
3769
3770static inline struct page *
3771__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3772 unsigned int alloc_flags,
3773 const struct alloc_context *ac)
3774{
3775 struct page *page;
3776
3777 page = get_page_from_freelist(gfp_mask, order,
3778 alloc_flags|ALLOC_CPUSET, ac);
3779 /*
3780 * fallback to ignore cpuset restriction if our nodes
3781 * are depleted
3782 */
3783 if (!page)
3784 page = get_page_from_freelist(gfp_mask, order,
3785 alloc_flags, ac);
3786
3787 return page;
3788}
3789
3790static inline struct page *
3791__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3792 const struct alloc_context *ac, unsigned long *did_some_progress)
3793{
3794 struct oom_control oc = {
3795 .zonelist = ac->zonelist,
3796 .nodemask = ac->nodemask,
3797 .memcg = NULL,
3798 .gfp_mask = gfp_mask,
3799 .order = order,
3800 };
3801 struct page *page;
3802
3803 *did_some_progress = 0;
3804
3805 /*
3806 * Acquire the oom lock. If that fails, somebody else is
3807 * making progress for us.
3808 */
3809 if (!mutex_trylock(&oom_lock)) {
3810 *did_some_progress = 1;
3811 schedule_timeout_uninterruptible(1);
3812 return NULL;
3813 }
3814
3815 /*
3816 * Go through the zonelist yet one more time, keep very high watermark
3817 * here, this is only to catch a parallel oom killing, we must fail if
3818 * we're still under heavy pressure. But make sure that this reclaim
3819 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3820 * allocation which will never fail due to oom_lock already held.
3821 */
3822 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3823 ~__GFP_DIRECT_RECLAIM, order,
3824 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3825 if (page)
3826 goto out;
3827
3828 /* Coredumps can quickly deplete all memory reserves */
3829 if (current->flags & PF_DUMPCORE)
3830 goto out;
3831 /* The OOM killer will not help higher order allocs */
3832 if (order > PAGE_ALLOC_COSTLY_ORDER)
3833 goto out;
3834 /*
3835 * We have already exhausted all our reclaim opportunities without any
3836 * success so it is time to admit defeat. We will skip the OOM killer
3837 * because it is very likely that the caller has a more reasonable
3838 * fallback than shooting a random task.
3839 */
3840 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3841 goto out;
3842 /* The OOM killer does not needlessly kill tasks for lowmem */
3843 if (ac->high_zoneidx < ZONE_NORMAL)
3844 goto out;
3845 if (pm_suspended_storage())
3846 goto out;
3847 /*
3848 * XXX: GFP_NOFS allocations should rather fail than rely on
3849 * other request to make a forward progress.
3850 * We are in an unfortunate situation where out_of_memory cannot
3851 * do much for this context but let's try it to at least get
3852 * access to memory reserved if the current task is killed (see
3853 * out_of_memory). Once filesystems are ready to handle allocation
3854 * failures more gracefully we should just bail out here.
3855 */
3856
3857 /* The OOM killer may not free memory on a specific node */
3858 if (gfp_mask & __GFP_THISNODE)
3859 goto out;
3860
3861 /* Exhausted what can be done so it's blame time */
3862 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3863 *did_some_progress = 1;
3864
3865 /*
3866 * Help non-failing allocations by giving them access to memory
3867 * reserves
3868 */
3869 if (gfp_mask & __GFP_NOFAIL)
3870 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3871 ALLOC_NO_WATERMARKS, ac);
3872 }
3873out:
3874 mutex_unlock(&oom_lock);
3875 return page;
3876}
3877
3878/*
3879 * Maximum number of compaction retries wit a progress before OOM
3880 * killer is consider as the only way to move forward.
3881 */
3882#define MAX_COMPACT_RETRIES 16
3883
3884#ifdef CONFIG_COMPACTION
3885/* Try memory compaction for high-order allocations before reclaim */
3886static struct page *
3887__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3888 unsigned int alloc_flags, const struct alloc_context *ac,
3889 enum compact_priority prio, enum compact_result *compact_result)
3890{
3891 struct page *page = NULL;
3892 unsigned long pflags;
3893 unsigned int noreclaim_flag;
3894
3895 if (!order)
3896 return NULL;
3897
3898 psi_memstall_enter(&pflags);
3899 noreclaim_flag = memalloc_noreclaim_save();
3900
3901 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3902 prio, &page);
3903
3904 memalloc_noreclaim_restore(noreclaim_flag);
3905 psi_memstall_leave(&pflags);
3906
3907 /*
3908 * At least in one zone compaction wasn't deferred or skipped, so let's
3909 * count a compaction stall
3910 */
3911 count_vm_event(COMPACTSTALL);
3912
3913 /* Prep a captured page if available */
3914 if (page)
3915 prep_new_page(page, order, gfp_mask, alloc_flags);
3916
3917 /* Try get a page from the freelist if available */
3918 if (!page)
3919 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3920
3921 if (page) {
3922 struct zone *zone = page_zone(page);
3923
3924 zone->compact_blockskip_flush = false;
3925 compaction_defer_reset(zone, order, true);
3926 count_vm_event(COMPACTSUCCESS);
3927 return page;
3928 }
3929
3930 /*
3931 * It's bad if compaction run occurs and fails. The most likely reason
3932 * is that pages exist, but not enough to satisfy watermarks.
3933 */
3934 count_vm_event(COMPACTFAIL);
3935
3936 cond_resched();
3937
3938 return NULL;
3939}
3940
3941static inline bool
3942should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3943 enum compact_result compact_result,
3944 enum compact_priority *compact_priority,
3945 int *compaction_retries)
3946{
3947 int max_retries = MAX_COMPACT_RETRIES;
3948 int min_priority;
3949 bool ret = false;
3950 int retries = *compaction_retries;
3951 enum compact_priority priority = *compact_priority;
3952
3953 if (!order)
3954 return false;
3955
3956 if (compaction_made_progress(compact_result))
3957 (*compaction_retries)++;
3958
3959 /*
3960 * compaction considers all the zone as desperately out of memory
3961 * so it doesn't really make much sense to retry except when the
3962 * failure could be caused by insufficient priority
3963 */
3964 if (compaction_failed(compact_result))
3965 goto check_priority;
3966
3967 /*
3968 * compaction was skipped because there are not enough order-0 pages
3969 * to work with, so we retry only if it looks like reclaim can help.
3970 */
3971 if (compaction_needs_reclaim(compact_result)) {
3972 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3973 goto out;
3974 }
3975
3976 /*
3977 * make sure the compaction wasn't deferred or didn't bail out early
3978 * due to locks contention before we declare that we should give up.
3979 * But the next retry should use a higher priority if allowed, so
3980 * we don't just keep bailing out endlessly.
3981 */
3982 if (compaction_withdrawn(compact_result)) {
3983 goto check_priority;
3984 }
3985
3986 /*
3987 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3988 * costly ones because they are de facto nofail and invoke OOM
3989 * killer to move on while costly can fail and users are ready
3990 * to cope with that. 1/4 retries is rather arbitrary but we
3991 * would need much more detailed feedback from compaction to
3992 * make a better decision.
3993 */
3994 if (order > PAGE_ALLOC_COSTLY_ORDER)
3995 max_retries /= 4;
3996 if (*compaction_retries <= max_retries) {
3997 ret = true;
3998 goto out;
3999 }
4000
4001 /*
4002 * Make sure there are attempts at the highest priority if we exhausted
4003 * all retries or failed at the lower priorities.
4004 */
4005check_priority:
4006 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4007 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4008
4009 if (*compact_priority > min_priority) {
4010 (*compact_priority)--;
4011 *compaction_retries = 0;
4012 ret = true;
4013 }
4014out:
4015 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4016 return ret;
4017}
4018#else
4019static inline struct page *
4020__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4021 unsigned int alloc_flags, const struct alloc_context *ac,
4022 enum compact_priority prio, enum compact_result *compact_result)
4023{
4024 *compact_result = COMPACT_SKIPPED;
4025 return NULL;
4026}
4027
4028static inline bool
4029should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4030 enum compact_result compact_result,
4031 enum compact_priority *compact_priority,
4032 int *compaction_retries)
4033{
4034 struct zone *zone;
4035 struct zoneref *z;
4036
4037 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4038 return false;
4039
4040 /*
4041 * There are setups with compaction disabled which would prefer to loop
4042 * inside the allocator rather than hit the oom killer prematurely.
4043 * Let's give them a good hope and keep retrying while the order-0
4044 * watermarks are OK.
4045 */
4046 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4047 ac->nodemask) {
4048 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4049 ac_classzone_idx(ac), alloc_flags))
4050 return true;
4051 }
4052 return false;
4053}
4054#endif /* CONFIG_COMPACTION */
4055
4056#ifdef CONFIG_LOCKDEP
4057static struct lockdep_map __fs_reclaim_map =
4058 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4059
4060static bool __need_fs_reclaim(gfp_t gfp_mask)
4061{
4062 gfp_mask = current_gfp_context(gfp_mask);
4063
4064 /* no reclaim without waiting on it */
4065 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4066 return false;
4067
4068 /* this guy won't enter reclaim */
4069 if (current->flags & PF_MEMALLOC)
4070 return false;
4071
4072 /* We're only interested __GFP_FS allocations for now */
4073 if (!(gfp_mask & __GFP_FS))
4074 return false;
4075
4076 if (gfp_mask & __GFP_NOLOCKDEP)
4077 return false;
4078
4079 return true;
4080}
4081
4082void __fs_reclaim_acquire(void)
4083{
4084 lock_map_acquire(&__fs_reclaim_map);
4085}
4086
4087void __fs_reclaim_release(void)
4088{
4089 lock_map_release(&__fs_reclaim_map);
4090}
4091
4092void fs_reclaim_acquire(gfp_t gfp_mask)
4093{
4094 if (__need_fs_reclaim(gfp_mask))
4095 __fs_reclaim_acquire();
4096}
4097EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4098
4099void fs_reclaim_release(gfp_t gfp_mask)
4100{
4101 if (__need_fs_reclaim(gfp_mask))
4102 __fs_reclaim_release();
4103}
4104EXPORT_SYMBOL_GPL(fs_reclaim_release);
4105#endif
4106
4107/* Perform direct synchronous page reclaim */
4108static int
4109__perform_reclaim(gfp_t gfp_mask, unsigned int order,
4110 const struct alloc_context *ac)
4111{
4112 int progress;
4113 unsigned int noreclaim_flag;
4114 unsigned long pflags;
4115
4116 cond_resched();
4117
4118 /* We now go into synchronous reclaim */
4119 cpuset_memory_pressure_bump();
4120 psi_memstall_enter(&pflags);
4121 fs_reclaim_acquire(gfp_mask);
4122 noreclaim_flag = memalloc_noreclaim_save();
4123
4124 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4125 ac->nodemask);
4126
4127 memalloc_noreclaim_restore(noreclaim_flag);
4128 fs_reclaim_release(gfp_mask);
4129 psi_memstall_leave(&pflags);
4130
4131 cond_resched();
4132
4133 return progress;
4134}
4135
4136/* The really slow allocator path where we enter direct reclaim */
4137static inline struct page *
4138__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4139 unsigned int alloc_flags, const struct alloc_context *ac,
4140 unsigned long *did_some_progress)
4141{
4142 struct page *page = NULL;
4143 bool drained = false;
4144
4145 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4146 if (unlikely(!(*did_some_progress)))
4147 return NULL;
4148
4149retry:
4150 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4151
4152 /*
4153 * If an allocation failed after direct reclaim, it could be because
4154 * pages are pinned on the per-cpu lists or in high alloc reserves.
4155 * Shrink them them and try again
4156 */
4157 if (!page && !drained) {
4158 unreserve_highatomic_pageblock(ac, false);
4159 drain_all_pages(NULL);
4160 drained = true;
4161 goto retry;
4162 }
4163
4164 return page;
4165}
4166
4167static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4168 const struct alloc_context *ac)
4169{
4170 struct zoneref *z;
4171 struct zone *zone;
4172 pg_data_t *last_pgdat = NULL;
4173 enum zone_type high_zoneidx = ac->high_zoneidx;
4174
4175 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
4176 ac->nodemask) {
4177 if (last_pgdat != zone->zone_pgdat)
4178 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
4179 last_pgdat = zone->zone_pgdat;
4180 }
4181}
4182
4183static inline unsigned int
4184gfp_to_alloc_flags(gfp_t gfp_mask)
4185{
4186 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4187
4188 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
4189 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4190
4191 /*
4192 * The caller may dip into page reserves a bit more if the caller
4193 * cannot run direct reclaim, or if the caller has realtime scheduling
4194 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4195 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4196 */
4197 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
4198
4199 if (gfp_mask & __GFP_ATOMIC) {
4200 /*
4201 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4202 * if it can't schedule.
4203 */
4204 if (!(gfp_mask & __GFP_NOMEMALLOC))
4205 alloc_flags |= ALLOC_HARDER;
4206 /*
4207 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4208 * comment for __cpuset_node_allowed().
4209 */
4210 alloc_flags &= ~ALLOC_CPUSET;
4211 } else if (unlikely(rt_task(current)) && !in_interrupt())
4212 alloc_flags |= ALLOC_HARDER;
4213
4214 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4215 alloc_flags |= ALLOC_KSWAPD;
4216
4217#ifdef CONFIG_CMA
4218 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
4219 alloc_flags |= ALLOC_CMA;
4220#endif
4221 return alloc_flags;
4222}
4223
4224static bool oom_reserves_allowed(struct task_struct *tsk)
4225{
4226 if (!tsk_is_oom_victim(tsk))
4227 return false;
4228
4229 /*
4230 * !MMU doesn't have oom reaper so give access to memory reserves
4231 * only to the thread with TIF_MEMDIE set
4232 */
4233 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4234 return false;
4235
4236 return true;
4237}
4238
4239/*
4240 * Distinguish requests which really need access to full memory
4241 * reserves from oom victims which can live with a portion of it
4242 */
4243static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4244{
4245 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4246 return 0;
4247 if (gfp_mask & __GFP_MEMALLOC)
4248 return ALLOC_NO_WATERMARKS;
4249 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4250 return ALLOC_NO_WATERMARKS;
4251 if (!in_interrupt()) {
4252 if (current->flags & PF_MEMALLOC)
4253 return ALLOC_NO_WATERMARKS;
4254 else if (oom_reserves_allowed(current))
4255 return ALLOC_OOM;
4256 }
4257
4258 return 0;
4259}
4260
4261bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4262{
4263 return !!__gfp_pfmemalloc_flags(gfp_mask);
4264}
4265
4266/*
4267 * Checks whether it makes sense to retry the reclaim to make a forward progress
4268 * for the given allocation request.
4269 *
4270 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4271 * without success, or when we couldn't even meet the watermark if we
4272 * reclaimed all remaining pages on the LRU lists.
4273 *
4274 * Returns true if a retry is viable or false to enter the oom path.
4275 */
4276static inline bool
4277should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4278 struct alloc_context *ac, int alloc_flags,
4279 bool did_some_progress, int *no_progress_loops)
4280{
4281 struct zone *zone;
4282 struct zoneref *z;
4283 bool ret = false;
4284
4285 /*
4286 * Costly allocations might have made a progress but this doesn't mean
4287 * their order will become available due to high fragmentation so
4288 * always increment the no progress counter for them
4289 */
4290 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4291 *no_progress_loops = 0;
4292 else
4293 (*no_progress_loops)++;
4294
4295 /*
4296 * Make sure we converge to OOM if we cannot make any progress
4297 * several times in the row.
4298 */
4299 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4300 /* Before OOM, exhaust highatomic_reserve */
4301 return unreserve_highatomic_pageblock(ac, true);
4302 }
4303
4304 /*
4305 * Keep reclaiming pages while there is a chance this will lead
4306 * somewhere. If none of the target zones can satisfy our allocation
4307 * request even if all reclaimable pages are considered then we are
4308 * screwed and have to go OOM.
4309 */
4310 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
4311 ac->nodemask) {
4312 unsigned long available;
4313 unsigned long reclaimable;
4314 unsigned long min_wmark = min_wmark_pages(zone);
4315 bool wmark;
4316
4317 available = reclaimable = zone_reclaimable_pages(zone);
4318 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4319
4320 /*
4321 * Would the allocation succeed if we reclaimed all
4322 * reclaimable pages?
4323 */
4324 wmark = __zone_watermark_ok(zone, order, min_wmark,
4325 ac_classzone_idx(ac), alloc_flags, available);
4326 trace_reclaim_retry_zone(z, order, reclaimable,
4327 available, min_wmark, *no_progress_loops, wmark);
4328 if (wmark) {
4329 /*
4330 * If we didn't make any progress and have a lot of
4331 * dirty + writeback pages then we should wait for
4332 * an IO to complete to slow down the reclaim and
4333 * prevent from pre mature OOM
4334 */
4335 if (!did_some_progress) {
4336 unsigned long write_pending;
4337
4338 write_pending = zone_page_state_snapshot(zone,
4339 NR_ZONE_WRITE_PENDING);
4340
4341 if (2 * write_pending > reclaimable) {
4342 congestion_wait(BLK_RW_ASYNC, HZ/10);
4343 return true;
4344 }
4345 }
4346
4347 ret = true;
4348 goto out;
4349 }
4350 }
4351
4352out:
4353 /*
4354 * Memory allocation/reclaim might be called from a WQ context and the
4355 * current implementation of the WQ concurrency control doesn't
4356 * recognize that a particular WQ is congested if the worker thread is
4357 * looping without ever sleeping. Therefore we have to do a short sleep
4358 * here rather than calling cond_resched().
4359 */
4360 if (current->flags & PF_WQ_WORKER)
4361 schedule_timeout_uninterruptible(1);
4362 else
4363 cond_resched();
4364 return ret;
4365}
4366
4367static inline bool
4368check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4369{
4370 /*
4371 * It's possible that cpuset's mems_allowed and the nodemask from
4372 * mempolicy don't intersect. This should be normally dealt with by
4373 * policy_nodemask(), but it's possible to race with cpuset update in
4374 * such a way the check therein was true, and then it became false
4375 * before we got our cpuset_mems_cookie here.
4376 * This assumes that for all allocations, ac->nodemask can come only
4377 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4378 * when it does not intersect with the cpuset restrictions) or the
4379 * caller can deal with a violated nodemask.
4380 */
4381 if (cpusets_enabled() && ac->nodemask &&
4382 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4383 ac->nodemask = NULL;
4384 return true;
4385 }
4386
4387 /*
4388 * When updating a task's mems_allowed or mempolicy nodemask, it is
4389 * possible to race with parallel threads in such a way that our
4390 * allocation can fail while the mask is being updated. If we are about
4391 * to fail, check if the cpuset changed during allocation and if so,
4392 * retry.
4393 */
4394 if (read_mems_allowed_retry(cpuset_mems_cookie))
4395 return true;
4396
4397 return false;
4398}
4399
4400static inline struct page *
4401__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4402 struct alloc_context *ac)
4403{
4404 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4405 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4406 struct page *page = NULL;
4407 unsigned int alloc_flags;
4408 unsigned long did_some_progress;
4409 enum compact_priority compact_priority;
4410 enum compact_result compact_result;
4411 int compaction_retries;
4412 int no_progress_loops;
4413 unsigned int cpuset_mems_cookie;
4414 int reserve_flags;
4415
4416 /*
4417 * We also sanity check to catch abuse of atomic reserves being used by
4418 * callers that are not in atomic context.
4419 */
4420 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4421 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4422 gfp_mask &= ~__GFP_ATOMIC;
4423
4424retry_cpuset:
4425 compaction_retries = 0;
4426 no_progress_loops = 0;
4427 compact_priority = DEF_COMPACT_PRIORITY;
4428 cpuset_mems_cookie = read_mems_allowed_begin();
4429
4430 /*
4431 * The fast path uses conservative alloc_flags to succeed only until
4432 * kswapd needs to be woken up, and to avoid the cost of setting up
4433 * alloc_flags precisely. So we do that now.
4434 */
4435 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4436
4437 /*
4438 * We need to recalculate the starting point for the zonelist iterator
4439 * because we might have used different nodemask in the fast path, or
4440 * there was a cpuset modification and we are retrying - otherwise we
4441 * could end up iterating over non-eligible zones endlessly.
4442 */
4443 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4444 ac->high_zoneidx, ac->nodemask);
4445 if (!ac->preferred_zoneref->zone)
4446 goto nopage;
4447
4448 if (alloc_flags & ALLOC_KSWAPD)
4449 wake_all_kswapds(order, gfp_mask, ac);
4450
4451 /*
4452 * The adjusted alloc_flags might result in immediate success, so try
4453 * that first
4454 */
4455 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4456 if (page)
4457 goto got_pg;
4458
4459 /*
4460 * For costly allocations, try direct compaction first, as it's likely
4461 * that we have enough base pages and don't need to reclaim. For non-
4462 * movable high-order allocations, do that as well, as compaction will
4463 * try prevent permanent fragmentation by migrating from blocks of the
4464 * same migratetype.
4465 * Don't try this for allocations that are allowed to ignore
4466 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4467 */
4468 if (can_direct_reclaim &&
4469 (costly_order ||
4470 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4471 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4472 page = __alloc_pages_direct_compact(gfp_mask, order,
4473 alloc_flags, ac,
4474 INIT_COMPACT_PRIORITY,
4475 &compact_result);
4476 if (page)
4477 goto got_pg;
4478
4479 if (order >= pageblock_order && (gfp_mask & __GFP_IO) &&
4480 !(gfp_mask & __GFP_RETRY_MAYFAIL)) {
4481 /*
4482 * If allocating entire pageblock(s) and compaction
4483 * failed because all zones are below low watermarks
4484 * or is prohibited because it recently failed at this
4485 * order, fail immediately unless the allocator has
4486 * requested compaction and reclaim retry.
4487 *
4488 * Reclaim is
4489 * - potentially very expensive because zones are far
4490 * below their low watermarks or this is part of very
4491 * bursty high order allocations,
4492 * - not guaranteed to help because isolate_freepages()
4493 * may not iterate over freed pages as part of its
4494 * linear scan, and
4495 * - unlikely to make entire pageblocks free on its
4496 * own.
4497 */
4498 if (compact_result == COMPACT_SKIPPED ||
4499 compact_result == COMPACT_DEFERRED)
4500 goto nopage;
4501 }
4502
4503 /*
4504 * Checks for costly allocations with __GFP_NORETRY, which
4505 * includes THP page fault allocations
4506 */
4507 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4508 /*
4509 * If compaction is deferred for high-order allocations,
4510 * it is because sync compaction recently failed. If
4511 * this is the case and the caller requested a THP
4512 * allocation, we do not want to heavily disrupt the
4513 * system, so we fail the allocation instead of entering
4514 * direct reclaim.
4515 */
4516 if (compact_result == COMPACT_DEFERRED)
4517 goto nopage;
4518
4519 /*
4520 * Looks like reclaim/compaction is worth trying, but
4521 * sync compaction could be very expensive, so keep
4522 * using async compaction.
4523 */
4524 compact_priority = INIT_COMPACT_PRIORITY;
4525 }
4526 }
4527
4528retry:
4529 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4530 if (alloc_flags & ALLOC_KSWAPD)
4531 wake_all_kswapds(order, gfp_mask, ac);
4532
4533 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4534 if (reserve_flags)
4535 alloc_flags = reserve_flags;
4536
4537 /*
4538 * Reset the nodemask and zonelist iterators if memory policies can be
4539 * ignored. These allocations are high priority and system rather than
4540 * user oriented.
4541 */
4542 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4543 ac->nodemask = NULL;
4544 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4545 ac->high_zoneidx, ac->nodemask);
4546 }
4547
4548 /* Attempt with potentially adjusted zonelist and alloc_flags */
4549 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4550 if (page)
4551 goto got_pg;
4552
4553 /* Caller is not willing to reclaim, we can't balance anything */
4554 if (!can_direct_reclaim)
4555 goto nopage;
4556
4557 /* Avoid recursion of direct reclaim */
4558 if (current->flags & PF_MEMALLOC)
4559 goto nopage;
4560
4561 /* Try direct reclaim and then allocating */
4562 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4563 &did_some_progress);
4564 if (page)
4565 goto got_pg;
4566
4567 /* Try direct compaction and then allocating */
4568 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4569 compact_priority, &compact_result);
4570 if (page)
4571 goto got_pg;
4572
4573 /* Do not loop if specifically requested */
4574 if (gfp_mask & __GFP_NORETRY)
4575 goto nopage;
4576
4577 /*
4578 * Do not retry costly high order allocations unless they are
4579 * __GFP_RETRY_MAYFAIL
4580 */
4581 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4582 goto nopage;
4583
4584 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4585 did_some_progress > 0, &no_progress_loops))
4586 goto retry;
4587
4588 /*
4589 * It doesn't make any sense to retry for the compaction if the order-0
4590 * reclaim is not able to make any progress because the current
4591 * implementation of the compaction depends on the sufficient amount
4592 * of free memory (see __compaction_suitable)
4593 */
4594 if (did_some_progress > 0 &&
4595 should_compact_retry(ac, order, alloc_flags,
4596 compact_result, &compact_priority,
4597 &compaction_retries))
4598 goto retry;
4599
4600
4601 /* Deal with possible cpuset update races before we start OOM killing */
4602 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4603 goto retry_cpuset;
4604
4605 /* Reclaim has failed us, start killing things */
4606 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4607 if (page)
4608 goto got_pg;
4609
4610 /* Avoid allocations with no watermarks from looping endlessly */
4611 if (tsk_is_oom_victim(current) &&
4612 (alloc_flags == ALLOC_OOM ||
4613 (gfp_mask & __GFP_NOMEMALLOC)))
4614 goto nopage;
4615
4616 /* Retry as long as the OOM killer is making progress */
4617 if (did_some_progress) {
4618 no_progress_loops = 0;
4619 goto retry;
4620 }
4621
4622nopage:
4623 /* Deal with possible cpuset update races before we fail */
4624 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4625 goto retry_cpuset;
4626
4627 /*
4628 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4629 * we always retry
4630 */
4631 if (gfp_mask & __GFP_NOFAIL) {
4632 /*
4633 * All existing users of the __GFP_NOFAIL are blockable, so warn
4634 * of any new users that actually require GFP_NOWAIT
4635 */
4636 if (WARN_ON_ONCE(!can_direct_reclaim))
4637 goto fail;
4638
4639 /*
4640 * PF_MEMALLOC request from this context is rather bizarre
4641 * because we cannot reclaim anything and only can loop waiting
4642 * for somebody to do a work for us
4643 */
4644 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4645
4646 /*
4647 * non failing costly orders are a hard requirement which we
4648 * are not prepared for much so let's warn about these users
4649 * so that we can identify them and convert them to something
4650 * else.
4651 */
4652 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4653
4654 /*
4655 * Help non-failing allocations by giving them access to memory
4656 * reserves but do not use ALLOC_NO_WATERMARKS because this
4657 * could deplete whole memory reserves which would just make
4658 * the situation worse
4659 */
4660 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4661 if (page)
4662 goto got_pg;
4663
4664 cond_resched();
4665 goto retry;
4666 }
4667fail:
4668 warn_alloc(gfp_mask, ac->nodemask,
4669 "page allocation failure: order:%u", order);
4670got_pg:
4671 return page;
4672}
4673
4674static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4675 int preferred_nid, nodemask_t *nodemask,
4676 struct alloc_context *ac, gfp_t *alloc_mask,
4677 unsigned int *alloc_flags)
4678{
4679 ac->high_zoneidx = gfp_zone(gfp_mask);
4680 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4681 ac->nodemask = nodemask;
4682 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4683
4684 if (cpusets_enabled()) {
4685 *alloc_mask |= __GFP_HARDWALL;
4686 if (!ac->nodemask)
4687 ac->nodemask = &cpuset_current_mems_allowed;
4688 else
4689 *alloc_flags |= ALLOC_CPUSET;
4690 }
4691
4692 fs_reclaim_acquire(gfp_mask);
4693 fs_reclaim_release(gfp_mask);
4694
4695 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4696
4697 if (should_fail_alloc_page(gfp_mask, order))
4698 return false;
4699
4700 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4701 *alloc_flags |= ALLOC_CMA;
4702
4703 return true;
4704}
4705
4706/* Determine whether to spread dirty pages and what the first usable zone */
4707static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4708{
4709 /* Dirty zone balancing only done in the fast path */
4710 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4711
4712 /*
4713 * The preferred zone is used for statistics but crucially it is
4714 * also used as the starting point for the zonelist iterator. It
4715 * may get reset for allocations that ignore memory policies.
4716 */
4717 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4718 ac->high_zoneidx, ac->nodemask);
4719}
4720
4721/*
4722 * This is the 'heart' of the zoned buddy allocator.
4723 */
4724struct page *
4725__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4726 nodemask_t *nodemask)
4727{
4728 struct page *page;
4729 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4730 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4731 struct alloc_context ac = { };
4732
4733 /*
4734 * There are several places where we assume that the order value is sane
4735 * so bail out early if the request is out of bound.
4736 */
4737 if (unlikely(order >= MAX_ORDER)) {
4738 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4739 return NULL;
4740 }
4741
4742 gfp_mask &= gfp_allowed_mask;
4743 alloc_mask = gfp_mask;
4744 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4745 return NULL;
4746
4747 finalise_ac(gfp_mask, &ac);
4748
4749 /*
4750 * Forbid the first pass from falling back to types that fragment
4751 * memory until all local zones are considered.
4752 */
4753 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4754
4755 /* First allocation attempt */
4756 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4757 if (likely(page))
4758 goto out;
4759
4760 /*
4761 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4762 * resp. GFP_NOIO which has to be inherited for all allocation requests
4763 * from a particular context which has been marked by
4764 * memalloc_no{fs,io}_{save,restore}.
4765 */
4766 alloc_mask = current_gfp_context(gfp_mask);
4767 ac.spread_dirty_pages = false;
4768
4769 /*
4770 * Restore the original nodemask if it was potentially replaced with
4771 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4772 */
4773 if (unlikely(ac.nodemask != nodemask))
4774 ac.nodemask = nodemask;
4775
4776 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4777
4778out:
4779 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4780 unlikely(__memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4781 __free_pages(page, order);
4782 page = NULL;
4783 }
4784
4785 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4786
4787 return page;
4788}
4789EXPORT_SYMBOL(__alloc_pages_nodemask);
4790
4791/*
4792 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4793 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4794 * you need to access high mem.
4795 */
4796unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4797{
4798 struct page *page;
4799
4800 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4801 if (!page)
4802 return 0;
4803 return (unsigned long) page_address(page);
4804}
4805EXPORT_SYMBOL(__get_free_pages);
4806
4807unsigned long get_zeroed_page(gfp_t gfp_mask)
4808{
4809 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4810}
4811EXPORT_SYMBOL(get_zeroed_page);
4812
4813static inline void free_the_page(struct page *page, unsigned int order)
4814{
4815 if (order == 0) /* Via pcp? */
4816 free_unref_page(page);
4817 else
4818 __free_pages_ok(page, order);
4819}
4820
4821void __free_pages(struct page *page, unsigned int order)
4822{
4823 if (put_page_testzero(page))
4824 free_the_page(page, order);
4825}
4826EXPORT_SYMBOL(__free_pages);
4827
4828void free_pages(unsigned long addr, unsigned int order)
4829{
4830 if (addr != 0) {
4831 VM_BUG_ON(!virt_addr_valid((void *)addr));
4832 __free_pages(virt_to_page((void *)addr), order);
4833 }
4834}
4835
4836EXPORT_SYMBOL(free_pages);
4837
4838/*
4839 * Page Fragment:
4840 * An arbitrary-length arbitrary-offset area of memory which resides
4841 * within a 0 or higher order page. Multiple fragments within that page
4842 * are individually refcounted, in the page's reference counter.
4843 *
4844 * The page_frag functions below provide a simple allocation framework for
4845 * page fragments. This is used by the network stack and network device
4846 * drivers to provide a backing region of memory for use as either an
4847 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4848 */
4849static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4850 gfp_t gfp_mask)
4851{
4852 struct page *page = NULL;
4853 gfp_t gfp = gfp_mask;
4854
4855#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4856 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4857 __GFP_NOMEMALLOC;
4858 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4859 PAGE_FRAG_CACHE_MAX_ORDER);
4860 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4861#endif
4862 if (unlikely(!page))
4863 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4864
4865 nc->va = page ? page_address(page) : NULL;
4866
4867 return page;
4868}
4869
4870void __page_frag_cache_drain(struct page *page, unsigned int count)
4871{
4872 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4873
4874 if (page_ref_sub_and_test(page, count))
4875 free_the_page(page, compound_order(page));
4876}
4877EXPORT_SYMBOL(__page_frag_cache_drain);
4878
4879void *page_frag_alloc(struct page_frag_cache *nc,
4880 unsigned int fragsz, gfp_t gfp_mask)
4881{
4882 unsigned int size = PAGE_SIZE;
4883 struct page *page;
4884 int offset;
4885
4886 if (unlikely(!nc->va)) {
4887refill:
4888 page = __page_frag_cache_refill(nc, gfp_mask);
4889 if (!page)
4890 return NULL;
4891
4892#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4893 /* if size can vary use size else just use PAGE_SIZE */
4894 size = nc->size;
4895#endif
4896 /* Even if we own the page, we do not use atomic_set().
4897 * This would break get_page_unless_zero() users.
4898 */
4899 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
4900
4901 /* reset page count bias and offset to start of new frag */
4902 nc->pfmemalloc = page_is_pfmemalloc(page);
4903 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4904 nc->offset = size;
4905 }
4906
4907 offset = nc->offset - fragsz;
4908 if (unlikely(offset < 0)) {
4909 page = virt_to_page(nc->va);
4910
4911 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4912 goto refill;
4913
4914#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4915 /* if size can vary use size else just use PAGE_SIZE */
4916 size = nc->size;
4917#endif
4918 /* OK, page count is 0, we can safely set it */
4919 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
4920
4921 /* reset page count bias and offset to start of new frag */
4922 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
4923 offset = size - fragsz;
4924 }
4925
4926 nc->pagecnt_bias--;
4927 nc->offset = offset;
4928
4929 return nc->va + offset;
4930}
4931EXPORT_SYMBOL(page_frag_alloc);
4932
4933/*
4934 * Frees a page fragment allocated out of either a compound or order 0 page.
4935 */
4936void page_frag_free(void *addr)
4937{
4938 struct page *page = virt_to_head_page(addr);
4939
4940 if (unlikely(put_page_testzero(page)))
4941 free_the_page(page, compound_order(page));
4942}
4943EXPORT_SYMBOL(page_frag_free);
4944
4945static void *make_alloc_exact(unsigned long addr, unsigned int order,
4946 size_t size)
4947{
4948 if (addr) {
4949 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4950 unsigned long used = addr + PAGE_ALIGN(size);
4951
4952 split_page(virt_to_page((void *)addr), order);
4953 while (used < alloc_end) {
4954 free_page(used);
4955 used += PAGE_SIZE;
4956 }
4957 }
4958 return (void *)addr;
4959}
4960
4961/**
4962 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4963 * @size: the number of bytes to allocate
4964 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4965 *
4966 * This function is similar to alloc_pages(), except that it allocates the
4967 * minimum number of pages to satisfy the request. alloc_pages() can only
4968 * allocate memory in power-of-two pages.
4969 *
4970 * This function is also limited by MAX_ORDER.
4971 *
4972 * Memory allocated by this function must be released by free_pages_exact().
4973 *
4974 * Return: pointer to the allocated area or %NULL in case of error.
4975 */
4976void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4977{
4978 unsigned int order = get_order(size);
4979 unsigned long addr;
4980
4981 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
4982 gfp_mask &= ~__GFP_COMP;
4983
4984 addr = __get_free_pages(gfp_mask, order);
4985 return make_alloc_exact(addr, order, size);
4986}
4987EXPORT_SYMBOL(alloc_pages_exact);
4988
4989/**
4990 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4991 * pages on a node.
4992 * @nid: the preferred node ID where memory should be allocated
4993 * @size: the number of bytes to allocate
4994 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4995 *
4996 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4997 * back.
4998 *
4999 * Return: pointer to the allocated area or %NULL in case of error.
5000 */
5001void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5002{
5003 unsigned int order = get_order(size);
5004 struct page *p;
5005
5006 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5007 gfp_mask &= ~__GFP_COMP;
5008
5009 p = alloc_pages_node(nid, gfp_mask, order);
5010 if (!p)
5011 return NULL;
5012 return make_alloc_exact((unsigned long)page_address(p), order, size);
5013}
5014
5015/**
5016 * free_pages_exact - release memory allocated via alloc_pages_exact()
5017 * @virt: the value returned by alloc_pages_exact.
5018 * @size: size of allocation, same value as passed to alloc_pages_exact().
5019 *
5020 * Release the memory allocated by a previous call to alloc_pages_exact.
5021 */
5022void free_pages_exact(void *virt, size_t size)
5023{
5024 unsigned long addr = (unsigned long)virt;
5025 unsigned long end = addr + PAGE_ALIGN(size);
5026
5027 while (addr < end) {
5028 free_page(addr);
5029 addr += PAGE_SIZE;
5030 }
5031}
5032EXPORT_SYMBOL(free_pages_exact);
5033
5034/**
5035 * nr_free_zone_pages - count number of pages beyond high watermark
5036 * @offset: The zone index of the highest zone
5037 *
5038 * nr_free_zone_pages() counts the number of pages which are beyond the
5039 * high watermark within all zones at or below a given zone index. For each
5040 * zone, the number of pages is calculated as:
5041 *
5042 * nr_free_zone_pages = managed_pages - high_pages
5043 *
5044 * Return: number of pages beyond high watermark.
5045 */
5046static unsigned long nr_free_zone_pages(int offset)
5047{
5048 struct zoneref *z;
5049 struct zone *zone;
5050
5051 /* Just pick one node, since fallback list is circular */
5052 unsigned long sum = 0;
5053
5054 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5055
5056 for_each_zone_zonelist(zone, z, zonelist, offset) {
5057 unsigned long size = zone_managed_pages(zone);
5058 unsigned long high = high_wmark_pages(zone);
5059 if (size > high)
5060 sum += size - high;
5061 }
5062
5063 return sum;
5064}
5065
5066/**
5067 * nr_free_buffer_pages - count number of pages beyond high watermark
5068 *
5069 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5070 * watermark within ZONE_DMA and ZONE_NORMAL.
5071 *
5072 * Return: number of pages beyond high watermark within ZONE_DMA and
5073 * ZONE_NORMAL.
5074 */
5075unsigned long nr_free_buffer_pages(void)
5076{
5077 return nr_free_zone_pages(gfp_zone(GFP_USER));
5078}
5079EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5080
5081/**
5082 * nr_free_pagecache_pages - count number of pages beyond high watermark
5083 *
5084 * nr_free_pagecache_pages() counts the number of pages which are beyond the
5085 * high watermark within all zones.
5086 *
5087 * Return: number of pages beyond high watermark within all zones.
5088 */
5089unsigned long nr_free_pagecache_pages(void)
5090{
5091 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5092}
5093
5094static inline void show_node(struct zone *zone)
5095{
5096 if (IS_ENABLED(CONFIG_NUMA))
5097 printk("Node %d ", zone_to_nid(zone));
5098}
5099
5100long si_mem_available(void)
5101{
5102 long available;
5103 unsigned long pagecache;
5104 unsigned long wmark_low = 0;
5105 unsigned long pages[NR_LRU_LISTS];
5106 unsigned long reclaimable;
5107 struct zone *zone;
5108 int lru;
5109
5110 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5111 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5112
5113 for_each_zone(zone)
5114 wmark_low += low_wmark_pages(zone);
5115
5116 /*
5117 * Estimate the amount of memory available for userspace allocations,
5118 * without causing swapping.
5119 */
5120 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5121
5122 /*
5123 * Not all the page cache can be freed, otherwise the system will
5124 * start swapping. Assume at least half of the page cache, or the
5125 * low watermark worth of cache, needs to stay.
5126 */
5127 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5128 pagecache -= min(pagecache / 2, wmark_low);
5129 available += pagecache;
5130
5131 /*
5132 * Part of the reclaimable slab and other kernel memory consists of
5133 * items that are in use, and cannot be freed. Cap this estimate at the
5134 * low watermark.
5135 */
5136 reclaimable = global_node_page_state(NR_SLAB_RECLAIMABLE) +
5137 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5138 available += reclaimable - min(reclaimable / 2, wmark_low);
5139
5140 if (available < 0)
5141 available = 0;
5142 return available;
5143}
5144EXPORT_SYMBOL_GPL(si_mem_available);
5145
5146void si_meminfo(struct sysinfo *val)
5147{
5148 val->totalram = totalram_pages();
5149 val->sharedram = global_node_page_state(NR_SHMEM);
5150 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5151 val->bufferram = nr_blockdev_pages();
5152 val->totalhigh = totalhigh_pages();
5153 val->freehigh = nr_free_highpages();
5154 val->mem_unit = PAGE_SIZE;
5155}
5156
5157EXPORT_SYMBOL(si_meminfo);
5158
5159#ifdef CONFIG_NUMA
5160void si_meminfo_node(struct sysinfo *val, int nid)
5161{
5162 int zone_type; /* needs to be signed */
5163 unsigned long managed_pages = 0;
5164 unsigned long managed_highpages = 0;
5165 unsigned long free_highpages = 0;
5166 pg_data_t *pgdat = NODE_DATA(nid);
5167
5168 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5169 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5170 val->totalram = managed_pages;
5171 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5172 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5173#ifdef CONFIG_HIGHMEM
5174 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5175 struct zone *zone = &pgdat->node_zones[zone_type];
5176
5177 if (is_highmem(zone)) {
5178 managed_highpages += zone_managed_pages(zone);
5179 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5180 }
5181 }
5182 val->totalhigh = managed_highpages;
5183 val->freehigh = free_highpages;
5184#else
5185 val->totalhigh = managed_highpages;
5186 val->freehigh = free_highpages;
5187#endif
5188 val->mem_unit = PAGE_SIZE;
5189}
5190#endif
5191
5192/*
5193 * Determine whether the node should be displayed or not, depending on whether
5194 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5195 */
5196static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5197{
5198 if (!(flags & SHOW_MEM_FILTER_NODES))
5199 return false;
5200
5201 /*
5202 * no node mask - aka implicit memory numa policy. Do not bother with
5203 * the synchronization - read_mems_allowed_begin - because we do not
5204 * have to be precise here.
5205 */
5206 if (!nodemask)
5207 nodemask = &cpuset_current_mems_allowed;
5208
5209 return !node_isset(nid, *nodemask);
5210}
5211
5212#define K(x) ((x) << (PAGE_SHIFT-10))
5213
5214static void show_migration_types(unsigned char type)
5215{
5216 static const char types[MIGRATE_TYPES] = {
5217 [MIGRATE_UNMOVABLE] = 'U',
5218 [MIGRATE_MOVABLE] = 'M',
5219 [MIGRATE_RECLAIMABLE] = 'E',
5220 [MIGRATE_HIGHATOMIC] = 'H',
5221#ifdef CONFIG_CMA
5222 [MIGRATE_CMA] = 'C',
5223#endif
5224#ifdef CONFIG_MEMORY_ISOLATION
5225 [MIGRATE_ISOLATE] = 'I',
5226#endif
5227 };
5228 char tmp[MIGRATE_TYPES + 1];
5229 char *p = tmp;
5230 int i;
5231
5232 for (i = 0; i < MIGRATE_TYPES; i++) {
5233 if (type & (1 << i))
5234 *p++ = types[i];
5235 }
5236
5237 *p = '\0';
5238 printk(KERN_CONT "(%s) ", tmp);
5239}
5240
5241/*
5242 * Show free area list (used inside shift_scroll-lock stuff)
5243 * We also calculate the percentage fragmentation. We do this by counting the
5244 * memory on each free list with the exception of the first item on the list.
5245 *
5246 * Bits in @filter:
5247 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5248 * cpuset.
5249 */
5250void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5251{
5252 unsigned long free_pcp = 0;
5253 int cpu;
5254 struct zone *zone;
5255 pg_data_t *pgdat;
5256
5257 for_each_populated_zone(zone) {
5258 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5259 continue;
5260
5261 for_each_online_cpu(cpu)
5262 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5263 }
5264
5265 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5266 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5267 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
5268 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5269 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5270 " free:%lu free_pcp:%lu free_cma:%lu\n",
5271 global_node_page_state(NR_ACTIVE_ANON),
5272 global_node_page_state(NR_INACTIVE_ANON),
5273 global_node_page_state(NR_ISOLATED_ANON),
5274 global_node_page_state(NR_ACTIVE_FILE),
5275 global_node_page_state(NR_INACTIVE_FILE),
5276 global_node_page_state(NR_ISOLATED_FILE),
5277 global_node_page_state(NR_UNEVICTABLE),
5278 global_node_page_state(NR_FILE_DIRTY),
5279 global_node_page_state(NR_WRITEBACK),
5280 global_node_page_state(NR_UNSTABLE_NFS),
5281 global_node_page_state(NR_SLAB_RECLAIMABLE),
5282 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
5283 global_node_page_state(NR_FILE_MAPPED),
5284 global_node_page_state(NR_SHMEM),
5285 global_zone_page_state(NR_PAGETABLE),
5286 global_zone_page_state(NR_BOUNCE),
5287 global_zone_page_state(NR_FREE_PAGES),
5288 free_pcp,
5289 global_zone_page_state(NR_FREE_CMA_PAGES));
5290
5291 for_each_online_pgdat(pgdat) {
5292 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5293 continue;
5294
5295 printk("Node %d"
5296 " active_anon:%lukB"
5297 " inactive_anon:%lukB"
5298 " active_file:%lukB"
5299 " inactive_file:%lukB"
5300 " unevictable:%lukB"
5301 " isolated(anon):%lukB"
5302 " isolated(file):%lukB"
5303 " mapped:%lukB"
5304 " dirty:%lukB"
5305 " writeback:%lukB"
5306 " shmem:%lukB"
5307#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5308 " shmem_thp: %lukB"
5309 " shmem_pmdmapped: %lukB"
5310 " anon_thp: %lukB"
5311#endif
5312 " writeback_tmp:%lukB"
5313 " unstable:%lukB"
5314 " all_unreclaimable? %s"
5315 "\n",
5316 pgdat->node_id,
5317 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5318 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5319 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5320 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5321 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5322 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5323 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5324 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5325 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5326 K(node_page_state(pgdat, NR_WRITEBACK)),
5327 K(node_page_state(pgdat, NR_SHMEM)),
5328#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5329 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5330 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5331 * HPAGE_PMD_NR),
5332 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5333#endif
5334 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5335 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
5336 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5337 "yes" : "no");
5338 }
5339
5340 for_each_populated_zone(zone) {
5341 int i;
5342
5343 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5344 continue;
5345
5346 free_pcp = 0;
5347 for_each_online_cpu(cpu)
5348 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5349
5350 show_node(zone);
5351 printk(KERN_CONT
5352 "%s"
5353 " free:%lukB"
5354 " min:%lukB"
5355 " low:%lukB"
5356 " high:%lukB"
5357 " active_anon:%lukB"
5358 " inactive_anon:%lukB"
5359 " active_file:%lukB"
5360 " inactive_file:%lukB"
5361 " unevictable:%lukB"
5362 " writepending:%lukB"
5363 " present:%lukB"
5364 " managed:%lukB"
5365 " mlocked:%lukB"
5366 " kernel_stack:%lukB"
5367 " pagetables:%lukB"
5368 " bounce:%lukB"
5369 " free_pcp:%lukB"
5370 " local_pcp:%ukB"
5371 " free_cma:%lukB"
5372 "\n",
5373 zone->name,
5374 K(zone_page_state(zone, NR_FREE_PAGES)),
5375 K(min_wmark_pages(zone)),
5376 K(low_wmark_pages(zone)),
5377 K(high_wmark_pages(zone)),
5378 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5379 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5380 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5381 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5382 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5383 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5384 K(zone->present_pages),
5385 K(zone_managed_pages(zone)),
5386 K(zone_page_state(zone, NR_MLOCK)),
5387 zone_page_state(zone, NR_KERNEL_STACK_KB),
5388 K(zone_page_state(zone, NR_PAGETABLE)),
5389 K(zone_page_state(zone, NR_BOUNCE)),
5390 K(free_pcp),
5391 K(this_cpu_read(zone->pageset->pcp.count)),
5392 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5393 printk("lowmem_reserve[]:");
5394 for (i = 0; i < MAX_NR_ZONES; i++)
5395 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5396 printk(KERN_CONT "\n");
5397 }
5398
5399 for_each_populated_zone(zone) {
5400 unsigned int order;
5401 unsigned long nr[MAX_ORDER], flags, total = 0;
5402 unsigned char types[MAX_ORDER];
5403
5404 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5405 continue;
5406 show_node(zone);
5407 printk(KERN_CONT "%s: ", zone->name);
5408
5409 spin_lock_irqsave(&zone->lock, flags);
5410 for (order = 0; order < MAX_ORDER; order++) {
5411 struct free_area *area = &zone->free_area[order];
5412 int type;
5413
5414 nr[order] = area->nr_free;
5415 total += nr[order] << order;
5416
5417 types[order] = 0;
5418 for (type = 0; type < MIGRATE_TYPES; type++) {
5419 if (!free_area_empty(area, type))
5420 types[order] |= 1 << type;
5421 }
5422 }
5423 spin_unlock_irqrestore(&zone->lock, flags);
5424 for (order = 0; order < MAX_ORDER; order++) {
5425 printk(KERN_CONT "%lu*%lukB ",
5426 nr[order], K(1UL) << order);
5427 if (nr[order])
5428 show_migration_types(types[order]);
5429 }
5430 printk(KERN_CONT "= %lukB\n", K(total));
5431 }
5432
5433 hugetlb_show_meminfo();
5434
5435 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5436
5437 show_swap_cache_info();
5438}
5439
5440static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5441{
5442 zoneref->zone = zone;
5443 zoneref->zone_idx = zone_idx(zone);
5444}
5445
5446/*
5447 * Builds allocation fallback zone lists.
5448 *
5449 * Add all populated zones of a node to the zonelist.
5450 */
5451static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5452{
5453 struct zone *zone;
5454 enum zone_type zone_type = MAX_NR_ZONES;
5455 int nr_zones = 0;
5456
5457 do {
5458 zone_type--;
5459 zone = pgdat->node_zones + zone_type;
5460 if (managed_zone(zone)) {
5461 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5462 check_highest_zone(zone_type);
5463 }
5464 } while (zone_type);
5465
5466 return nr_zones;
5467}
5468
5469#ifdef CONFIG_NUMA
5470
5471static int __parse_numa_zonelist_order(char *s)
5472{
5473 /*
5474 * We used to support different zonlists modes but they turned
5475 * out to be just not useful. Let's keep the warning in place
5476 * if somebody still use the cmd line parameter so that we do
5477 * not fail it silently
5478 */
5479 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5480 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5481 return -EINVAL;
5482 }
5483 return 0;
5484}
5485
5486static __init int setup_numa_zonelist_order(char *s)
5487{
5488 if (!s)
5489 return 0;
5490
5491 return __parse_numa_zonelist_order(s);
5492}
5493early_param("numa_zonelist_order", setup_numa_zonelist_order);
5494
5495char numa_zonelist_order[] = "Node";
5496
5497/*
5498 * sysctl handler for numa_zonelist_order
5499 */
5500int numa_zonelist_order_handler(struct ctl_table *table, int write,
5501 void __user *buffer, size_t *length,
5502 loff_t *ppos)
5503{
5504 char *str;
5505 int ret;
5506
5507 if (!write)
5508 return proc_dostring(table, write, buffer, length, ppos);
5509 str = memdup_user_nul(buffer, 16);
5510 if (IS_ERR(str))
5511 return PTR_ERR(str);
5512
5513 ret = __parse_numa_zonelist_order(str);
5514 kfree(str);
5515 return ret;
5516}
5517
5518
5519#define MAX_NODE_LOAD (nr_online_nodes)
5520static int node_load[MAX_NUMNODES];
5521
5522/**
5523 * find_next_best_node - find the next node that should appear in a given node's fallback list
5524 * @node: node whose fallback list we're appending
5525 * @used_node_mask: nodemask_t of already used nodes
5526 *
5527 * We use a number of factors to determine which is the next node that should
5528 * appear on a given node's fallback list. The node should not have appeared
5529 * already in @node's fallback list, and it should be the next closest node
5530 * according to the distance array (which contains arbitrary distance values
5531 * from each node to each node in the system), and should also prefer nodes
5532 * with no CPUs, since presumably they'll have very little allocation pressure
5533 * on them otherwise.
5534 *
5535 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5536 */
5537static int find_next_best_node(int node, nodemask_t *used_node_mask)
5538{
5539 int n, val;
5540 int min_val = INT_MAX;
5541 int best_node = NUMA_NO_NODE;
5542 const struct cpumask *tmp = cpumask_of_node(0);
5543
5544 /* Use the local node if we haven't already */
5545 if (!node_isset(node, *used_node_mask)) {
5546 node_set(node, *used_node_mask);
5547 return node;
5548 }
5549
5550 for_each_node_state(n, N_MEMORY) {
5551
5552 /* Don't want a node to appear more than once */
5553 if (node_isset(n, *used_node_mask))
5554 continue;
5555
5556 /* Use the distance array to find the distance */
5557 val = node_distance(node, n);
5558
5559 /* Penalize nodes under us ("prefer the next node") */
5560 val += (n < node);
5561
5562 /* Give preference to headless and unused nodes */
5563 tmp = cpumask_of_node(n);
5564 if (!cpumask_empty(tmp))
5565 val += PENALTY_FOR_NODE_WITH_CPUS;
5566
5567 /* Slight preference for less loaded node */
5568 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5569 val += node_load[n];
5570
5571 if (val < min_val) {
5572 min_val = val;
5573 best_node = n;
5574 }
5575 }
5576
5577 if (best_node >= 0)
5578 node_set(best_node, *used_node_mask);
5579
5580 return best_node;
5581}
5582
5583
5584/*
5585 * Build zonelists ordered by node and zones within node.
5586 * This results in maximum locality--normal zone overflows into local
5587 * DMA zone, if any--but risks exhausting DMA zone.
5588 */
5589static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5590 unsigned nr_nodes)
5591{
5592 struct zoneref *zonerefs;
5593 int i;
5594
5595 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5596
5597 for (i = 0; i < nr_nodes; i++) {
5598 int nr_zones;
5599
5600 pg_data_t *node = NODE_DATA(node_order[i]);
5601
5602 nr_zones = build_zonerefs_node(node, zonerefs);
5603 zonerefs += nr_zones;
5604 }
5605 zonerefs->zone = NULL;
5606 zonerefs->zone_idx = 0;
5607}
5608
5609/*
5610 * Build gfp_thisnode zonelists
5611 */
5612static void build_thisnode_zonelists(pg_data_t *pgdat)
5613{
5614 struct zoneref *zonerefs;
5615 int nr_zones;
5616
5617 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5618 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5619 zonerefs += nr_zones;
5620 zonerefs->zone = NULL;
5621 zonerefs->zone_idx = 0;
5622}
5623
5624/*
5625 * Build zonelists ordered by zone and nodes within zones.
5626 * This results in conserving DMA zone[s] until all Normal memory is
5627 * exhausted, but results in overflowing to remote node while memory
5628 * may still exist in local DMA zone.
5629 */
5630
5631static void build_zonelists(pg_data_t *pgdat)
5632{
5633 static int node_order[MAX_NUMNODES];
5634 int node, load, nr_nodes = 0;
5635 nodemask_t used_mask;
5636 int local_node, prev_node;
5637
5638 /* NUMA-aware ordering of nodes */
5639 local_node = pgdat->node_id;
5640 load = nr_online_nodes;
5641 prev_node = local_node;
5642 nodes_clear(used_mask);
5643
5644 memset(node_order, 0, sizeof(node_order));
5645 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5646 /*
5647 * We don't want to pressure a particular node.
5648 * So adding penalty to the first node in same
5649 * distance group to make it round-robin.
5650 */
5651 if (node_distance(local_node, node) !=
5652 node_distance(local_node, prev_node))
5653 node_load[node] = load;
5654
5655 node_order[nr_nodes++] = node;
5656 prev_node = node;
5657 load--;
5658 }
5659
5660 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5661 build_thisnode_zonelists(pgdat);
5662}
5663
5664#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5665/*
5666 * Return node id of node used for "local" allocations.
5667 * I.e., first node id of first zone in arg node's generic zonelist.
5668 * Used for initializing percpu 'numa_mem', which is used primarily
5669 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5670 */
5671int local_memory_node(int node)
5672{
5673 struct zoneref *z;
5674
5675 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5676 gfp_zone(GFP_KERNEL),
5677 NULL);
5678 return zone_to_nid(z->zone);
5679}
5680#endif
5681
5682static void setup_min_unmapped_ratio(void);
5683static void setup_min_slab_ratio(void);
5684#else /* CONFIG_NUMA */
5685
5686static void build_zonelists(pg_data_t *pgdat)
5687{
5688 int node, local_node;
5689 struct zoneref *zonerefs;
5690 int nr_zones;
5691
5692 local_node = pgdat->node_id;
5693
5694 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5695 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5696 zonerefs += nr_zones;
5697
5698 /*
5699 * Now we build the zonelist so that it contains the zones
5700 * of all the other nodes.
5701 * We don't want to pressure a particular node, so when
5702 * building the zones for node N, we make sure that the
5703 * zones coming right after the local ones are those from
5704 * node N+1 (modulo N)
5705 */
5706 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5707 if (!node_online(node))
5708 continue;
5709 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5710 zonerefs += nr_zones;
5711 }
5712 for (node = 0; node < local_node; node++) {
5713 if (!node_online(node))
5714 continue;
5715 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5716 zonerefs += nr_zones;
5717 }
5718
5719 zonerefs->zone = NULL;
5720 zonerefs->zone_idx = 0;
5721}
5722
5723#endif /* CONFIG_NUMA */
5724
5725/*
5726 * Boot pageset table. One per cpu which is going to be used for all
5727 * zones and all nodes. The parameters will be set in such a way
5728 * that an item put on a list will immediately be handed over to
5729 * the buddy list. This is safe since pageset manipulation is done
5730 * with interrupts disabled.
5731 *
5732 * The boot_pagesets must be kept even after bootup is complete for
5733 * unused processors and/or zones. They do play a role for bootstrapping
5734 * hotplugged processors.
5735 *
5736 * zoneinfo_show() and maybe other functions do
5737 * not check if the processor is online before following the pageset pointer.
5738 * Other parts of the kernel may not check if the zone is available.
5739 */
5740static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5741static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5742static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5743
5744static void __build_all_zonelists(void *data)
5745{
5746 int nid;
5747 int __maybe_unused cpu;
5748 pg_data_t *self = data;
5749 static DEFINE_SPINLOCK(lock);
5750
5751 spin_lock(&lock);
5752
5753#ifdef CONFIG_NUMA
5754 memset(node_load, 0, sizeof(node_load));
5755#endif
5756
5757 /*
5758 * This node is hotadded and no memory is yet present. So just
5759 * building zonelists is fine - no need to touch other nodes.
5760 */
5761 if (self && !node_online(self->node_id)) {
5762 build_zonelists(self);
5763 } else {
5764 for_each_online_node(nid) {
5765 pg_data_t *pgdat = NODE_DATA(nid);
5766
5767 build_zonelists(pgdat);
5768 }
5769
5770#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5771 /*
5772 * We now know the "local memory node" for each node--
5773 * i.e., the node of the first zone in the generic zonelist.
5774 * Set up numa_mem percpu variable for on-line cpus. During
5775 * boot, only the boot cpu should be on-line; we'll init the
5776 * secondary cpus' numa_mem as they come on-line. During
5777 * node/memory hotplug, we'll fixup all on-line cpus.
5778 */
5779 for_each_online_cpu(cpu)
5780 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5781#endif
5782 }
5783
5784 spin_unlock(&lock);
5785}
5786
5787static noinline void __init
5788build_all_zonelists_init(void)
5789{
5790 int cpu;
5791
5792 __build_all_zonelists(NULL);
5793
5794 /*
5795 * Initialize the boot_pagesets that are going to be used
5796 * for bootstrapping processors. The real pagesets for
5797 * each zone will be allocated later when the per cpu
5798 * allocator is available.
5799 *
5800 * boot_pagesets are used also for bootstrapping offline
5801 * cpus if the system is already booted because the pagesets
5802 * are needed to initialize allocators on a specific cpu too.
5803 * F.e. the percpu allocator needs the page allocator which
5804 * needs the percpu allocator in order to allocate its pagesets
5805 * (a chicken-egg dilemma).
5806 */
5807 for_each_possible_cpu(cpu)
5808 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5809
5810 mminit_verify_zonelist();
5811 cpuset_init_current_mems_allowed();
5812}
5813
5814/*
5815 * unless system_state == SYSTEM_BOOTING.
5816 *
5817 * __ref due to call of __init annotated helper build_all_zonelists_init
5818 * [protected by SYSTEM_BOOTING].
5819 */
5820void __ref build_all_zonelists(pg_data_t *pgdat)
5821{
5822 if (system_state == SYSTEM_BOOTING) {
5823 build_all_zonelists_init();
5824 } else {
5825 __build_all_zonelists(pgdat);
5826 /* cpuset refresh routine should be here */
5827 }
5828 vm_total_pages = nr_free_pagecache_pages();
5829 /*
5830 * Disable grouping by mobility if the number of pages in the
5831 * system is too low to allow the mechanism to work. It would be
5832 * more accurate, but expensive to check per-zone. This check is
5833 * made on memory-hotadd so a system can start with mobility
5834 * disabled and enable it later
5835 */
5836 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5837 page_group_by_mobility_disabled = 1;
5838 else
5839 page_group_by_mobility_disabled = 0;
5840
5841 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5842 nr_online_nodes,
5843 page_group_by_mobility_disabled ? "off" : "on",
5844 vm_total_pages);
5845#ifdef CONFIG_NUMA
5846 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5847#endif
5848}
5849
5850/* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5851static bool __meminit
5852overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5853{
5854#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5855 static struct memblock_region *r;
5856
5857 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5858 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5859 for_each_memblock(memory, r) {
5860 if (*pfn < memblock_region_memory_end_pfn(r))
5861 break;
5862 }
5863 }
5864 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5865 memblock_is_mirror(r)) {
5866 *pfn = memblock_region_memory_end_pfn(r);
5867 return true;
5868 }
5869 }
5870#endif
5871 return false;
5872}
5873
5874/*
5875 * Initially all pages are reserved - free ones are freed
5876 * up by memblock_free_all() once the early boot process is
5877 * done. Non-atomic initialization, single-pass.
5878 */
5879void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5880 unsigned long start_pfn, enum memmap_context context,
5881 struct vmem_altmap *altmap)
5882{
5883 unsigned long pfn, end_pfn = start_pfn + size;
5884 struct page *page;
5885
5886 if (highest_memmap_pfn < end_pfn - 1)
5887 highest_memmap_pfn = end_pfn - 1;
5888
5889#ifdef CONFIG_ZONE_DEVICE
5890 /*
5891 * Honor reservation requested by the driver for this ZONE_DEVICE
5892 * memory. We limit the total number of pages to initialize to just
5893 * those that might contain the memory mapping. We will defer the
5894 * ZONE_DEVICE page initialization until after we have released
5895 * the hotplug lock.
5896 */
5897 if (zone == ZONE_DEVICE) {
5898 if (!altmap)
5899 return;
5900
5901 if (start_pfn == altmap->base_pfn)
5902 start_pfn += altmap->reserve;
5903 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5904 }
5905#endif
5906
5907 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5908 /*
5909 * There can be holes in boot-time mem_map[]s handed to this
5910 * function. They do not exist on hotplugged memory.
5911 */
5912 if (context == MEMMAP_EARLY) {
5913 if (!early_pfn_valid(pfn))
5914 continue;
5915 if (!early_pfn_in_nid(pfn, nid))
5916 continue;
5917 if (overlap_memmap_init(zone, &pfn))
5918 continue;
5919 if (defer_init(nid, pfn, end_pfn))
5920 break;
5921 }
5922
5923 page = pfn_to_page(pfn);
5924 __init_single_page(page, pfn, zone, nid);
5925 if (context == MEMMAP_HOTPLUG)
5926 __SetPageReserved(page);
5927
5928 /*
5929 * Mark the block movable so that blocks are reserved for
5930 * movable at startup. This will force kernel allocations
5931 * to reserve their blocks rather than leaking throughout
5932 * the address space during boot when many long-lived
5933 * kernel allocations are made.
5934 *
5935 * bitmap is created for zone's valid pfn range. but memmap
5936 * can be created for invalid pages (for alignment)
5937 * check here not to call set_pageblock_migratetype() against
5938 * pfn out of zone.
5939 */
5940 if (!(pfn & (pageblock_nr_pages - 1))) {
5941 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5942 cond_resched();
5943 }
5944 }
5945}
5946
5947#ifdef CONFIG_ZONE_DEVICE
5948void __ref memmap_init_zone_device(struct zone *zone,
5949 unsigned long start_pfn,
5950 unsigned long size,
5951 struct dev_pagemap *pgmap)
5952{
5953 unsigned long pfn, end_pfn = start_pfn + size;
5954 struct pglist_data *pgdat = zone->zone_pgdat;
5955 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
5956 unsigned long zone_idx = zone_idx(zone);
5957 unsigned long start = jiffies;
5958 int nid = pgdat->node_id;
5959
5960 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
5961 return;
5962
5963 /*
5964 * The call to memmap_init_zone should have already taken care
5965 * of the pages reserved for the memmap, so we can just jump to
5966 * the end of that region and start processing the device pages.
5967 */
5968 if (altmap) {
5969 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
5970 size = end_pfn - start_pfn;
5971 }
5972
5973 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5974 struct page *page = pfn_to_page(pfn);
5975
5976 __init_single_page(page, pfn, zone_idx, nid);
5977
5978 /*
5979 * Mark page reserved as it will need to wait for onlining
5980 * phase for it to be fully associated with a zone.
5981 *
5982 * We can use the non-atomic __set_bit operation for setting
5983 * the flag as we are still initializing the pages.
5984 */
5985 __SetPageReserved(page);
5986
5987 /*
5988 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
5989 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
5990 * ever freed or placed on a driver-private list.
5991 */
5992 page->pgmap = pgmap;
5993 page->zone_device_data = NULL;
5994
5995 /*
5996 * Mark the block movable so that blocks are reserved for
5997 * movable at startup. This will force kernel allocations
5998 * to reserve their blocks rather than leaking throughout
5999 * the address space during boot when many long-lived
6000 * kernel allocations are made.
6001 *
6002 * bitmap is created for zone's valid pfn range. but memmap
6003 * can be created for invalid pages (for alignment)
6004 * check here not to call set_pageblock_migratetype() against
6005 * pfn out of zone.
6006 *
6007 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
6008 * because this is done early in section_activate()
6009 */
6010 if (!(pfn & (pageblock_nr_pages - 1))) {
6011 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6012 cond_resched();
6013 }
6014 }
6015
6016 pr_info("%s initialised %lu pages in %ums\n", __func__,
6017 size, jiffies_to_msecs(jiffies - start));
6018}
6019
6020#endif
6021static void __meminit zone_init_free_lists(struct zone *zone)
6022{
6023 unsigned int order, t;
6024 for_each_migratetype_order(order, t) {
6025 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6026 zone->free_area[order].nr_free = 0;
6027 }
6028}
6029
6030void __meminit __weak memmap_init(unsigned long size, int nid,
6031 unsigned long zone, unsigned long start_pfn)
6032{
6033 memmap_init_zone(size, nid, zone, start_pfn, MEMMAP_EARLY, NULL);
6034}
6035
6036static int zone_batchsize(struct zone *zone)
6037{
6038#ifdef CONFIG_MMU
6039 int batch;
6040
6041 /*
6042 * The per-cpu-pages pools are set to around 1000th of the
6043 * size of the zone.
6044 */
6045 batch = zone_managed_pages(zone) / 1024;
6046 /* But no more than a meg. */
6047 if (batch * PAGE_SIZE > 1024 * 1024)
6048 batch = (1024 * 1024) / PAGE_SIZE;
6049 batch /= 4; /* We effectively *= 4 below */
6050 if (batch < 1)
6051 batch = 1;
6052
6053 /*
6054 * Clamp the batch to a 2^n - 1 value. Having a power
6055 * of 2 value was found to be more likely to have
6056 * suboptimal cache aliasing properties in some cases.
6057 *
6058 * For example if 2 tasks are alternately allocating
6059 * batches of pages, one task can end up with a lot
6060 * of pages of one half of the possible page colors
6061 * and the other with pages of the other colors.
6062 */
6063 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6064
6065 return batch;
6066
6067#else
6068 /* The deferral and batching of frees should be suppressed under NOMMU
6069 * conditions.
6070 *
6071 * The problem is that NOMMU needs to be able to allocate large chunks
6072 * of contiguous memory as there's no hardware page translation to
6073 * assemble apparent contiguous memory from discontiguous pages.
6074 *
6075 * Queueing large contiguous runs of pages for batching, however,
6076 * causes the pages to actually be freed in smaller chunks. As there
6077 * can be a significant delay between the individual batches being
6078 * recycled, this leads to the once large chunks of space being
6079 * fragmented and becoming unavailable for high-order allocations.
6080 */
6081 return 0;
6082#endif
6083}
6084
6085/*
6086 * pcp->high and pcp->batch values are related and dependent on one another:
6087 * ->batch must never be higher then ->high.
6088 * The following function updates them in a safe manner without read side
6089 * locking.
6090 *
6091 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6092 * those fields changing asynchronously (acording the the above rule).
6093 *
6094 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6095 * outside of boot time (or some other assurance that no concurrent updaters
6096 * exist).
6097 */
6098static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6099 unsigned long batch)
6100{
6101 /* start with a fail safe value for batch */
6102 pcp->batch = 1;
6103 smp_wmb();
6104
6105 /* Update high, then batch, in order */
6106 pcp->high = high;
6107 smp_wmb();
6108
6109 pcp->batch = batch;
6110}
6111
6112/* a companion to pageset_set_high() */
6113static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6114{
6115 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6116}
6117
6118static void pageset_init(struct per_cpu_pageset *p)
6119{
6120 struct per_cpu_pages *pcp;
6121 int migratetype;
6122
6123 memset(p, 0, sizeof(*p));
6124
6125 pcp = &p->pcp;
6126 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6127 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6128}
6129
6130static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6131{
6132 pageset_init(p);
6133 pageset_set_batch(p, batch);
6134}
6135
6136/*
6137 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6138 * to the value high for the pageset p.
6139 */
6140static void pageset_set_high(struct per_cpu_pageset *p,
6141 unsigned long high)
6142{
6143 unsigned long batch = max(1UL, high / 4);
6144 if ((high / 4) > (PAGE_SHIFT * 8))
6145 batch = PAGE_SHIFT * 8;
6146
6147 pageset_update(&p->pcp, high, batch);
6148}
6149
6150static void pageset_set_high_and_batch(struct zone *zone,
6151 struct per_cpu_pageset *pcp)
6152{
6153 if (percpu_pagelist_fraction)
6154 pageset_set_high(pcp,
6155 (zone_managed_pages(zone) /
6156 percpu_pagelist_fraction));
6157 else
6158 pageset_set_batch(pcp, zone_batchsize(zone));
6159}
6160
6161static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6162{
6163 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6164
6165 pageset_init(pcp);
6166 pageset_set_high_and_batch(zone, pcp);
6167}
6168
6169void __meminit setup_zone_pageset(struct zone *zone)
6170{
6171 int cpu;
6172 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6173 for_each_possible_cpu(cpu)
6174 zone_pageset_init(zone, cpu);
6175}
6176
6177/*
6178 * Allocate per cpu pagesets and initialize them.
6179 * Before this call only boot pagesets were available.
6180 */
6181void __init setup_per_cpu_pageset(void)
6182{
6183 struct pglist_data *pgdat;
6184 struct zone *zone;
6185
6186 for_each_populated_zone(zone)
6187 setup_zone_pageset(zone);
6188
6189 for_each_online_pgdat(pgdat)
6190 pgdat->per_cpu_nodestats =
6191 alloc_percpu(struct per_cpu_nodestat);
6192}
6193
6194static __meminit void zone_pcp_init(struct zone *zone)
6195{
6196 /*
6197 * per cpu subsystem is not up at this point. The following code
6198 * relies on the ability of the linker to provide the
6199 * offset of a (static) per cpu variable into the per cpu area.
6200 */
6201 zone->pageset = &boot_pageset;
6202
6203 if (populated_zone(zone))
6204 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6205 zone->name, zone->present_pages,
6206 zone_batchsize(zone));
6207}
6208
6209void __meminit init_currently_empty_zone(struct zone *zone,
6210 unsigned long zone_start_pfn,
6211 unsigned long size)
6212{
6213 struct pglist_data *pgdat = zone->zone_pgdat;
6214 int zone_idx = zone_idx(zone) + 1;
6215
6216 if (zone_idx > pgdat->nr_zones)
6217 pgdat->nr_zones = zone_idx;
6218
6219 zone->zone_start_pfn = zone_start_pfn;
6220
6221 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6222 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6223 pgdat->node_id,
6224 (unsigned long)zone_idx(zone),
6225 zone_start_pfn, (zone_start_pfn + size));
6226
6227 zone_init_free_lists(zone);
6228 zone->initialized = 1;
6229}
6230
6231#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6232#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
6233
6234/*
6235 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
6236 */
6237int __meminit __early_pfn_to_nid(unsigned long pfn,
6238 struct mminit_pfnnid_cache *state)
6239{
6240 unsigned long start_pfn, end_pfn;
6241 int nid;
6242
6243 if (state->last_start <= pfn && pfn < state->last_end)
6244 return state->last_nid;
6245
6246 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
6247 if (nid != NUMA_NO_NODE) {
6248 state->last_start = start_pfn;
6249 state->last_end = end_pfn;
6250 state->last_nid = nid;
6251 }
6252
6253 return nid;
6254}
6255#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
6256
6257/**
6258 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
6259 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
6260 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
6261 *
6262 * If an architecture guarantees that all ranges registered contain no holes
6263 * and may be freed, this this function may be used instead of calling
6264 * memblock_free_early_nid() manually.
6265 */
6266void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
6267{
6268 unsigned long start_pfn, end_pfn;
6269 int i, this_nid;
6270
6271 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
6272 start_pfn = min(start_pfn, max_low_pfn);
6273 end_pfn = min(end_pfn, max_low_pfn);
6274
6275 if (start_pfn < end_pfn)
6276 memblock_free_early_nid(PFN_PHYS(start_pfn),
6277 (end_pfn - start_pfn) << PAGE_SHIFT,
6278 this_nid);
6279 }
6280}
6281
6282/**
6283 * sparse_memory_present_with_active_regions - Call memory_present for each active range
6284 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
6285 *
6286 * If an architecture guarantees that all ranges registered contain no holes and may
6287 * be freed, this function may be used instead of calling memory_present() manually.
6288 */
6289void __init sparse_memory_present_with_active_regions(int nid)
6290{
6291 unsigned long start_pfn, end_pfn;
6292 int i, this_nid;
6293
6294 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
6295 memory_present(this_nid, start_pfn, end_pfn);
6296}
6297
6298/**
6299 * get_pfn_range_for_nid - Return the start and end page frames for a node
6300 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6301 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6302 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6303 *
6304 * It returns the start and end page frame of a node based on information
6305 * provided by memblock_set_node(). If called for a node
6306 * with no available memory, a warning is printed and the start and end
6307 * PFNs will be 0.
6308 */
6309void __init get_pfn_range_for_nid(unsigned int nid,
6310 unsigned long *start_pfn, unsigned long *end_pfn)
6311{
6312 unsigned long this_start_pfn, this_end_pfn;
6313 int i;
6314
6315 *start_pfn = -1UL;
6316 *end_pfn = 0;
6317
6318 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6319 *start_pfn = min(*start_pfn, this_start_pfn);
6320 *end_pfn = max(*end_pfn, this_end_pfn);
6321 }
6322
6323 if (*start_pfn == -1UL)
6324 *start_pfn = 0;
6325}
6326
6327/*
6328 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6329 * assumption is made that zones within a node are ordered in monotonic
6330 * increasing memory addresses so that the "highest" populated zone is used
6331 */
6332static void __init find_usable_zone_for_movable(void)
6333{
6334 int zone_index;
6335 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6336 if (zone_index == ZONE_MOVABLE)
6337 continue;
6338
6339 if (arch_zone_highest_possible_pfn[zone_index] >
6340 arch_zone_lowest_possible_pfn[zone_index])
6341 break;
6342 }
6343
6344 VM_BUG_ON(zone_index == -1);
6345 movable_zone = zone_index;
6346}
6347
6348/*
6349 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6350 * because it is sized independent of architecture. Unlike the other zones,
6351 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6352 * in each node depending on the size of each node and how evenly kernelcore
6353 * is distributed. This helper function adjusts the zone ranges
6354 * provided by the architecture for a given node by using the end of the
6355 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6356 * zones within a node are in order of monotonic increases memory addresses
6357 */
6358static void __init adjust_zone_range_for_zone_movable(int nid,
6359 unsigned long zone_type,
6360 unsigned long node_start_pfn,
6361 unsigned long node_end_pfn,
6362 unsigned long *zone_start_pfn,
6363 unsigned long *zone_end_pfn)
6364{
6365 /* Only adjust if ZONE_MOVABLE is on this node */
6366 if (zone_movable_pfn[nid]) {
6367 /* Size ZONE_MOVABLE */
6368 if (zone_type == ZONE_MOVABLE) {
6369 *zone_start_pfn = zone_movable_pfn[nid];
6370 *zone_end_pfn = min(node_end_pfn,
6371 arch_zone_highest_possible_pfn[movable_zone]);
6372
6373 /* Adjust for ZONE_MOVABLE starting within this range */
6374 } else if (!mirrored_kernelcore &&
6375 *zone_start_pfn < zone_movable_pfn[nid] &&
6376 *zone_end_pfn > zone_movable_pfn[nid]) {
6377 *zone_end_pfn = zone_movable_pfn[nid];
6378
6379 /* Check if this whole range is within ZONE_MOVABLE */
6380 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6381 *zone_start_pfn = *zone_end_pfn;
6382 }
6383}
6384
6385/*
6386 * Return the number of pages a zone spans in a node, including holes
6387 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6388 */
6389static unsigned long __init zone_spanned_pages_in_node(int nid,
6390 unsigned long zone_type,
6391 unsigned long node_start_pfn,
6392 unsigned long node_end_pfn,
6393 unsigned long *zone_start_pfn,
6394 unsigned long *zone_end_pfn,
6395 unsigned long *ignored)
6396{
6397 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6398 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6399 /* When hotadd a new node from cpu_up(), the node should be empty */
6400 if (!node_start_pfn && !node_end_pfn)
6401 return 0;
6402
6403 /* Get the start and end of the zone */
6404 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6405 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6406 adjust_zone_range_for_zone_movable(nid, zone_type,
6407 node_start_pfn, node_end_pfn,
6408 zone_start_pfn, zone_end_pfn);
6409
6410 /* Check that this node has pages within the zone's required range */
6411 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6412 return 0;
6413
6414 /* Move the zone boundaries inside the node if necessary */
6415 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6416 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6417
6418 /* Return the spanned pages */
6419 return *zone_end_pfn - *zone_start_pfn;
6420}
6421
6422/*
6423 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6424 * then all holes in the requested range will be accounted for.
6425 */
6426unsigned long __init __absent_pages_in_range(int nid,
6427 unsigned long range_start_pfn,
6428 unsigned long range_end_pfn)
6429{
6430 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6431 unsigned long start_pfn, end_pfn;
6432 int i;
6433
6434 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6435 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6436 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6437 nr_absent -= end_pfn - start_pfn;
6438 }
6439 return nr_absent;
6440}
6441
6442/**
6443 * absent_pages_in_range - Return number of page frames in holes within a range
6444 * @start_pfn: The start PFN to start searching for holes
6445 * @end_pfn: The end PFN to stop searching for holes
6446 *
6447 * Return: the number of pages frames in memory holes within a range.
6448 */
6449unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6450 unsigned long end_pfn)
6451{
6452 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6453}
6454
6455/* Return the number of page frames in holes in a zone on a node */
6456static unsigned long __init zone_absent_pages_in_node(int nid,
6457 unsigned long zone_type,
6458 unsigned long node_start_pfn,
6459 unsigned long node_end_pfn,
6460 unsigned long *ignored)
6461{
6462 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6463 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6464 unsigned long zone_start_pfn, zone_end_pfn;
6465 unsigned long nr_absent;
6466
6467 /* When hotadd a new node from cpu_up(), the node should be empty */
6468 if (!node_start_pfn && !node_end_pfn)
6469 return 0;
6470
6471 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6472 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6473
6474 adjust_zone_range_for_zone_movable(nid, zone_type,
6475 node_start_pfn, node_end_pfn,
6476 &zone_start_pfn, &zone_end_pfn);
6477 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6478
6479 /*
6480 * ZONE_MOVABLE handling.
6481 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6482 * and vice versa.
6483 */
6484 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6485 unsigned long start_pfn, end_pfn;
6486 struct memblock_region *r;
6487
6488 for_each_memblock(memory, r) {
6489 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6490 zone_start_pfn, zone_end_pfn);
6491 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6492 zone_start_pfn, zone_end_pfn);
6493
6494 if (zone_type == ZONE_MOVABLE &&
6495 memblock_is_mirror(r))
6496 nr_absent += end_pfn - start_pfn;
6497
6498 if (zone_type == ZONE_NORMAL &&
6499 !memblock_is_mirror(r))
6500 nr_absent += end_pfn - start_pfn;
6501 }
6502 }
6503
6504 return nr_absent;
6505}
6506
6507#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6508static inline unsigned long __init zone_spanned_pages_in_node(int nid,
6509 unsigned long zone_type,
6510 unsigned long node_start_pfn,
6511 unsigned long node_end_pfn,
6512 unsigned long *zone_start_pfn,
6513 unsigned long *zone_end_pfn,
6514 unsigned long *zones_size)
6515{
6516 unsigned int zone;
6517
6518 *zone_start_pfn = node_start_pfn;
6519 for (zone = 0; zone < zone_type; zone++)
6520 *zone_start_pfn += zones_size[zone];
6521
6522 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6523
6524 return zones_size[zone_type];
6525}
6526
6527static inline unsigned long __init zone_absent_pages_in_node(int nid,
6528 unsigned long zone_type,
6529 unsigned long node_start_pfn,
6530 unsigned long node_end_pfn,
6531 unsigned long *zholes_size)
6532{
6533 if (!zholes_size)
6534 return 0;
6535
6536 return zholes_size[zone_type];
6537}
6538
6539#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6540
6541static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6542 unsigned long node_start_pfn,
6543 unsigned long node_end_pfn,
6544 unsigned long *zones_size,
6545 unsigned long *zholes_size)
6546{
6547 unsigned long realtotalpages = 0, totalpages = 0;
6548 enum zone_type i;
6549
6550 for (i = 0; i < MAX_NR_ZONES; i++) {
6551 struct zone *zone = pgdat->node_zones + i;
6552 unsigned long zone_start_pfn, zone_end_pfn;
6553 unsigned long size, real_size;
6554
6555 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6556 node_start_pfn,
6557 node_end_pfn,
6558 &zone_start_pfn,
6559 &zone_end_pfn,
6560 zones_size);
6561 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6562 node_start_pfn, node_end_pfn,
6563 zholes_size);
6564 if (size)
6565 zone->zone_start_pfn = zone_start_pfn;
6566 else
6567 zone->zone_start_pfn = 0;
6568 zone->spanned_pages = size;
6569 zone->present_pages = real_size;
6570
6571 totalpages += size;
6572 realtotalpages += real_size;
6573 }
6574
6575 pgdat->node_spanned_pages = totalpages;
6576 pgdat->node_present_pages = realtotalpages;
6577 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6578 realtotalpages);
6579}
6580
6581#ifndef CONFIG_SPARSEMEM
6582/*
6583 * Calculate the size of the zone->blockflags rounded to an unsigned long
6584 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6585 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6586 * round what is now in bits to nearest long in bits, then return it in
6587 * bytes.
6588 */
6589static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6590{
6591 unsigned long usemapsize;
6592
6593 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6594 usemapsize = roundup(zonesize, pageblock_nr_pages);
6595 usemapsize = usemapsize >> pageblock_order;
6596 usemapsize *= NR_PAGEBLOCK_BITS;
6597 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6598
6599 return usemapsize / 8;
6600}
6601
6602static void __ref setup_usemap(struct pglist_data *pgdat,
6603 struct zone *zone,
6604 unsigned long zone_start_pfn,
6605 unsigned long zonesize)
6606{
6607 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6608 zone->pageblock_flags = NULL;
6609 if (usemapsize) {
6610 zone->pageblock_flags =
6611 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6612 pgdat->node_id);
6613 if (!zone->pageblock_flags)
6614 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6615 usemapsize, zone->name, pgdat->node_id);
6616 }
6617}
6618#else
6619static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6620 unsigned long zone_start_pfn, unsigned long zonesize) {}
6621#endif /* CONFIG_SPARSEMEM */
6622
6623#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6624
6625/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6626void __init set_pageblock_order(void)
6627{
6628 unsigned int order;
6629
6630 /* Check that pageblock_nr_pages has not already been setup */
6631 if (pageblock_order)
6632 return;
6633
6634 if (HPAGE_SHIFT > PAGE_SHIFT)
6635 order = HUGETLB_PAGE_ORDER;
6636 else
6637 order = MAX_ORDER - 1;
6638
6639 /*
6640 * Assume the largest contiguous order of interest is a huge page.
6641 * This value may be variable depending on boot parameters on IA64 and
6642 * powerpc.
6643 */
6644 pageblock_order = order;
6645}
6646#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6647
6648/*
6649 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6650 * is unused as pageblock_order is set at compile-time. See
6651 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6652 * the kernel config
6653 */
6654void __init set_pageblock_order(void)
6655{
6656}
6657
6658#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6659
6660static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6661 unsigned long present_pages)
6662{
6663 unsigned long pages = spanned_pages;
6664
6665 /*
6666 * Provide a more accurate estimation if there are holes within
6667 * the zone and SPARSEMEM is in use. If there are holes within the
6668 * zone, each populated memory region may cost us one or two extra
6669 * memmap pages due to alignment because memmap pages for each
6670 * populated regions may not be naturally aligned on page boundary.
6671 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6672 */
6673 if (spanned_pages > present_pages + (present_pages >> 4) &&
6674 IS_ENABLED(CONFIG_SPARSEMEM))
6675 pages = present_pages;
6676
6677 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6678}
6679
6680#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6681static void pgdat_init_split_queue(struct pglist_data *pgdat)
6682{
6683 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6684
6685 spin_lock_init(&ds_queue->split_queue_lock);
6686 INIT_LIST_HEAD(&ds_queue->split_queue);
6687 ds_queue->split_queue_len = 0;
6688}
6689#else
6690static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6691#endif
6692
6693#ifdef CONFIG_COMPACTION
6694static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6695{
6696 init_waitqueue_head(&pgdat->kcompactd_wait);
6697}
6698#else
6699static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6700#endif
6701
6702static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6703{
6704 pgdat_resize_init(pgdat);
6705
6706 pgdat_init_split_queue(pgdat);
6707 pgdat_init_kcompactd(pgdat);
6708
6709 init_waitqueue_head(&pgdat->kswapd_wait);
6710 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6711
6712 pgdat_page_ext_init(pgdat);
6713 spin_lock_init(&pgdat->lru_lock);
6714 lruvec_init(node_lruvec(pgdat));
6715}
6716
6717static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6718 unsigned long remaining_pages)
6719{
6720 atomic_long_set(&zone->managed_pages, remaining_pages);
6721 zone_set_nid(zone, nid);
6722 zone->name = zone_names[idx];
6723 zone->zone_pgdat = NODE_DATA(nid);
6724 spin_lock_init(&zone->lock);
6725 zone_seqlock_init(zone);
6726 zone_pcp_init(zone);
6727}
6728
6729/*
6730 * Set up the zone data structures
6731 * - init pgdat internals
6732 * - init all zones belonging to this node
6733 *
6734 * NOTE: this function is only called during memory hotplug
6735 */
6736#ifdef CONFIG_MEMORY_HOTPLUG
6737void __ref free_area_init_core_hotplug(int nid)
6738{
6739 enum zone_type z;
6740 pg_data_t *pgdat = NODE_DATA(nid);
6741
6742 pgdat_init_internals(pgdat);
6743 for (z = 0; z < MAX_NR_ZONES; z++)
6744 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6745}
6746#endif
6747
6748/*
6749 * Set up the zone data structures:
6750 * - mark all pages reserved
6751 * - mark all memory queues empty
6752 * - clear the memory bitmaps
6753 *
6754 * NOTE: pgdat should get zeroed by caller.
6755 * NOTE: this function is only called during early init.
6756 */
6757static void __init free_area_init_core(struct pglist_data *pgdat)
6758{
6759 enum zone_type j;
6760 int nid = pgdat->node_id;
6761
6762 pgdat_init_internals(pgdat);
6763 pgdat->per_cpu_nodestats = &boot_nodestats;
6764
6765 for (j = 0; j < MAX_NR_ZONES; j++) {
6766 struct zone *zone = pgdat->node_zones + j;
6767 unsigned long size, freesize, memmap_pages;
6768 unsigned long zone_start_pfn = zone->zone_start_pfn;
6769
6770 size = zone->spanned_pages;
6771 freesize = zone->present_pages;
6772
6773 /*
6774 * Adjust freesize so that it accounts for how much memory
6775 * is used by this zone for memmap. This affects the watermark
6776 * and per-cpu initialisations
6777 */
6778 memmap_pages = calc_memmap_size(size, freesize);
6779 if (!is_highmem_idx(j)) {
6780 if (freesize >= memmap_pages) {
6781 freesize -= memmap_pages;
6782 if (memmap_pages)
6783 printk(KERN_DEBUG
6784 " %s zone: %lu pages used for memmap\n",
6785 zone_names[j], memmap_pages);
6786 } else
6787 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6788 zone_names[j], memmap_pages, freesize);
6789 }
6790
6791 /* Account for reserved pages */
6792 if (j == 0 && freesize > dma_reserve) {
6793 freesize -= dma_reserve;
6794 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6795 zone_names[0], dma_reserve);
6796 }
6797
6798 if (!is_highmem_idx(j))
6799 nr_kernel_pages += freesize;
6800 /* Charge for highmem memmap if there are enough kernel pages */
6801 else if (nr_kernel_pages > memmap_pages * 2)
6802 nr_kernel_pages -= memmap_pages;
6803 nr_all_pages += freesize;
6804
6805 /*
6806 * Set an approximate value for lowmem here, it will be adjusted
6807 * when the bootmem allocator frees pages into the buddy system.
6808 * And all highmem pages will be managed by the buddy system.
6809 */
6810 zone_init_internals(zone, j, nid, freesize);
6811
6812 if (!size)
6813 continue;
6814
6815 set_pageblock_order();
6816 setup_usemap(pgdat, zone, zone_start_pfn, size);
6817 init_currently_empty_zone(zone, zone_start_pfn, size);
6818 memmap_init(size, nid, j, zone_start_pfn);
6819 }
6820}
6821
6822#ifdef CONFIG_FLAT_NODE_MEM_MAP
6823static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6824{
6825 unsigned long __maybe_unused start = 0;
6826 unsigned long __maybe_unused offset = 0;
6827
6828 /* Skip empty nodes */
6829 if (!pgdat->node_spanned_pages)
6830 return;
6831
6832 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6833 offset = pgdat->node_start_pfn - start;
6834 /* ia64 gets its own node_mem_map, before this, without bootmem */
6835 if (!pgdat->node_mem_map) {
6836 unsigned long size, end;
6837 struct page *map;
6838
6839 /*
6840 * The zone's endpoints aren't required to be MAX_ORDER
6841 * aligned but the node_mem_map endpoints must be in order
6842 * for the buddy allocator to function correctly.
6843 */
6844 end = pgdat_end_pfn(pgdat);
6845 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6846 size = (end - start) * sizeof(struct page);
6847 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6848 pgdat->node_id);
6849 if (!map)
6850 panic("Failed to allocate %ld bytes for node %d memory map\n",
6851 size, pgdat->node_id);
6852 pgdat->node_mem_map = map + offset;
6853 }
6854 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6855 __func__, pgdat->node_id, (unsigned long)pgdat,
6856 (unsigned long)pgdat->node_mem_map);
6857#ifndef CONFIG_NEED_MULTIPLE_NODES
6858 /*
6859 * With no DISCONTIG, the global mem_map is just set as node 0's
6860 */
6861 if (pgdat == NODE_DATA(0)) {
6862 mem_map = NODE_DATA(0)->node_mem_map;
6863#if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6864 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6865 mem_map -= offset;
6866#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6867 }
6868#endif
6869}
6870#else
6871static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6872#endif /* CONFIG_FLAT_NODE_MEM_MAP */
6873
6874#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6875static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6876{
6877 pgdat->first_deferred_pfn = ULONG_MAX;
6878}
6879#else
6880static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6881#endif
6882
6883void __init free_area_init_node(int nid, unsigned long *zones_size,
6884 unsigned long node_start_pfn,
6885 unsigned long *zholes_size)
6886{
6887 pg_data_t *pgdat = NODE_DATA(nid);
6888 unsigned long start_pfn = 0;
6889 unsigned long end_pfn = 0;
6890
6891 /* pg_data_t should be reset to zero when it's allocated */
6892 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6893
6894 pgdat->node_id = nid;
6895 pgdat->node_start_pfn = node_start_pfn;
6896 pgdat->per_cpu_nodestats = NULL;
6897#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6898 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6899 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6900 (u64)start_pfn << PAGE_SHIFT,
6901 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6902#else
6903 start_pfn = node_start_pfn;
6904#endif
6905 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6906 zones_size, zholes_size);
6907
6908 alloc_node_mem_map(pgdat);
6909 pgdat_set_deferred_range(pgdat);
6910
6911 free_area_init_core(pgdat);
6912}
6913
6914#if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6915/*
6916 * Zero all valid struct pages in range [spfn, epfn), return number of struct
6917 * pages zeroed
6918 */
6919static u64 zero_pfn_range(unsigned long spfn, unsigned long epfn)
6920{
6921 unsigned long pfn;
6922 u64 pgcnt = 0;
6923
6924 for (pfn = spfn; pfn < epfn; pfn++) {
6925 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6926 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6927 + pageblock_nr_pages - 1;
6928 continue;
6929 }
6930 mm_zero_struct_page(pfn_to_page(pfn));
6931 pgcnt++;
6932 }
6933
6934 return pgcnt;
6935}
6936
6937/*
6938 * Only struct pages that are backed by physical memory are zeroed and
6939 * initialized by going through __init_single_page(). But, there are some
6940 * struct pages which are reserved in memblock allocator and their fields
6941 * may be accessed (for example page_to_pfn() on some configuration accesses
6942 * flags). We must explicitly zero those struct pages.
6943 *
6944 * This function also addresses a similar issue where struct pages are left
6945 * uninitialized because the physical address range is not covered by
6946 * memblock.memory or memblock.reserved. That could happen when memblock
6947 * layout is manually configured via memmap=.
6948 */
6949void __init zero_resv_unavail(void)
6950{
6951 phys_addr_t start, end;
6952 u64 i, pgcnt;
6953 phys_addr_t next = 0;
6954
6955 /*
6956 * Loop through unavailable ranges not covered by memblock.memory.
6957 */
6958 pgcnt = 0;
6959 for_each_mem_range(i, &memblock.memory, NULL,
6960 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
6961 if (next < start)
6962 pgcnt += zero_pfn_range(PFN_DOWN(next), PFN_UP(start));
6963 next = end;
6964 }
6965 pgcnt += zero_pfn_range(PFN_DOWN(next), max_pfn);
6966
6967 /*
6968 * Struct pages that do not have backing memory. This could be because
6969 * firmware is using some of this memory, or for some other reasons.
6970 */
6971 if (pgcnt)
6972 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
6973}
6974#endif /* !CONFIG_FLAT_NODE_MEM_MAP */
6975
6976#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6977
6978#if MAX_NUMNODES > 1
6979/*
6980 * Figure out the number of possible node ids.
6981 */
6982void __init setup_nr_node_ids(void)
6983{
6984 unsigned int highest;
6985
6986 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6987 nr_node_ids = highest + 1;
6988}
6989#endif
6990
6991/**
6992 * node_map_pfn_alignment - determine the maximum internode alignment
6993 *
6994 * This function should be called after node map is populated and sorted.
6995 * It calculates the maximum power of two alignment which can distinguish
6996 * all the nodes.
6997 *
6998 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6999 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7000 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7001 * shifted, 1GiB is enough and this function will indicate so.
7002 *
7003 * This is used to test whether pfn -> nid mapping of the chosen memory
7004 * model has fine enough granularity to avoid incorrect mapping for the
7005 * populated node map.
7006 *
7007 * Return: the determined alignment in pfn's. 0 if there is no alignment
7008 * requirement (single node).
7009 */
7010unsigned long __init node_map_pfn_alignment(void)
7011{
7012 unsigned long accl_mask = 0, last_end = 0;
7013 unsigned long start, end, mask;
7014 int last_nid = NUMA_NO_NODE;
7015 int i, nid;
7016
7017 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7018 if (!start || last_nid < 0 || last_nid == nid) {
7019 last_nid = nid;
7020 last_end = end;
7021 continue;
7022 }
7023
7024 /*
7025 * Start with a mask granular enough to pin-point to the
7026 * start pfn and tick off bits one-by-one until it becomes
7027 * too coarse to separate the current node from the last.
7028 */
7029 mask = ~((1 << __ffs(start)) - 1);
7030 while (mask && last_end <= (start & (mask << 1)))
7031 mask <<= 1;
7032
7033 /* accumulate all internode masks */
7034 accl_mask |= mask;
7035 }
7036
7037 /* convert mask to number of pages */
7038 return ~accl_mask + 1;
7039}
7040
7041/* Find the lowest pfn for a node */
7042static unsigned long __init find_min_pfn_for_node(int nid)
7043{
7044 unsigned long min_pfn = ULONG_MAX;
7045 unsigned long start_pfn;
7046 int i;
7047
7048 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
7049 min_pfn = min(min_pfn, start_pfn);
7050
7051 if (min_pfn == ULONG_MAX) {
7052 pr_warn("Could not find start_pfn for node %d\n", nid);
7053 return 0;
7054 }
7055
7056 return min_pfn;
7057}
7058
7059/**
7060 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7061 *
7062 * Return: the minimum PFN based on information provided via
7063 * memblock_set_node().
7064 */
7065unsigned long __init find_min_pfn_with_active_regions(void)
7066{
7067 return find_min_pfn_for_node(MAX_NUMNODES);
7068}
7069
7070/*
7071 * early_calculate_totalpages()
7072 * Sum pages in active regions for movable zone.
7073 * Populate N_MEMORY for calculating usable_nodes.
7074 */
7075static unsigned long __init early_calculate_totalpages(void)
7076{
7077 unsigned long totalpages = 0;
7078 unsigned long start_pfn, end_pfn;
7079 int i, nid;
7080
7081 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7082 unsigned long pages = end_pfn - start_pfn;
7083
7084 totalpages += pages;
7085 if (pages)
7086 node_set_state(nid, N_MEMORY);
7087 }
7088 return totalpages;
7089}
7090
7091/*
7092 * Find the PFN the Movable zone begins in each node. Kernel memory
7093 * is spread evenly between nodes as long as the nodes have enough
7094 * memory. When they don't, some nodes will have more kernelcore than
7095 * others
7096 */
7097static void __init find_zone_movable_pfns_for_nodes(void)
7098{
7099 int i, nid;
7100 unsigned long usable_startpfn;
7101 unsigned long kernelcore_node, kernelcore_remaining;
7102 /* save the state before borrow the nodemask */
7103 nodemask_t saved_node_state = node_states[N_MEMORY];
7104 unsigned long totalpages = early_calculate_totalpages();
7105 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7106 struct memblock_region *r;
7107
7108 /* Need to find movable_zone earlier when movable_node is specified. */
7109 find_usable_zone_for_movable();
7110
7111 /*
7112 * If movable_node is specified, ignore kernelcore and movablecore
7113 * options.
7114 */
7115 if (movable_node_is_enabled()) {
7116 for_each_memblock(memory, r) {
7117 if (!memblock_is_hotpluggable(r))
7118 continue;
7119
7120 nid = r->nid;
7121
7122 usable_startpfn = PFN_DOWN(r->base);
7123 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7124 min(usable_startpfn, zone_movable_pfn[nid]) :
7125 usable_startpfn;
7126 }
7127
7128 goto out2;
7129 }
7130
7131 /*
7132 * If kernelcore=mirror is specified, ignore movablecore option
7133 */
7134 if (mirrored_kernelcore) {
7135 bool mem_below_4gb_not_mirrored = false;
7136
7137 for_each_memblock(memory, r) {
7138 if (memblock_is_mirror(r))
7139 continue;
7140
7141 nid = r->nid;
7142
7143 usable_startpfn = memblock_region_memory_base_pfn(r);
7144
7145 if (usable_startpfn < 0x100000) {
7146 mem_below_4gb_not_mirrored = true;
7147 continue;
7148 }
7149
7150 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7151 min(usable_startpfn, zone_movable_pfn[nid]) :
7152 usable_startpfn;
7153 }
7154
7155 if (mem_below_4gb_not_mirrored)
7156 pr_warn("This configuration results in unmirrored kernel memory.");
7157
7158 goto out2;
7159 }
7160
7161 /*
7162 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7163 * amount of necessary memory.
7164 */
7165 if (required_kernelcore_percent)
7166 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7167 10000UL;
7168 if (required_movablecore_percent)
7169 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7170 10000UL;
7171
7172 /*
7173 * If movablecore= was specified, calculate what size of
7174 * kernelcore that corresponds so that memory usable for
7175 * any allocation type is evenly spread. If both kernelcore
7176 * and movablecore are specified, then the value of kernelcore
7177 * will be used for required_kernelcore if it's greater than
7178 * what movablecore would have allowed.
7179 */
7180 if (required_movablecore) {
7181 unsigned long corepages;
7182
7183 /*
7184 * Round-up so that ZONE_MOVABLE is at least as large as what
7185 * was requested by the user
7186 */
7187 required_movablecore =
7188 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7189 required_movablecore = min(totalpages, required_movablecore);
7190 corepages = totalpages - required_movablecore;
7191
7192 required_kernelcore = max(required_kernelcore, corepages);
7193 }
7194
7195 /*
7196 * If kernelcore was not specified or kernelcore size is larger
7197 * than totalpages, there is no ZONE_MOVABLE.
7198 */
7199 if (!required_kernelcore || required_kernelcore >= totalpages)
7200 goto out;
7201
7202 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7203 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7204
7205restart:
7206 /* Spread kernelcore memory as evenly as possible throughout nodes */
7207 kernelcore_node = required_kernelcore / usable_nodes;
7208 for_each_node_state(nid, N_MEMORY) {
7209 unsigned long start_pfn, end_pfn;
7210
7211 /*
7212 * Recalculate kernelcore_node if the division per node
7213 * now exceeds what is necessary to satisfy the requested
7214 * amount of memory for the kernel
7215 */
7216 if (required_kernelcore < kernelcore_node)
7217 kernelcore_node = required_kernelcore / usable_nodes;
7218
7219 /*
7220 * As the map is walked, we track how much memory is usable
7221 * by the kernel using kernelcore_remaining. When it is
7222 * 0, the rest of the node is usable by ZONE_MOVABLE
7223 */
7224 kernelcore_remaining = kernelcore_node;
7225
7226 /* Go through each range of PFNs within this node */
7227 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7228 unsigned long size_pages;
7229
7230 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7231 if (start_pfn >= end_pfn)
7232 continue;
7233
7234 /* Account for what is only usable for kernelcore */
7235 if (start_pfn < usable_startpfn) {
7236 unsigned long kernel_pages;
7237 kernel_pages = min(end_pfn, usable_startpfn)
7238 - start_pfn;
7239
7240 kernelcore_remaining -= min(kernel_pages,
7241 kernelcore_remaining);
7242 required_kernelcore -= min(kernel_pages,
7243 required_kernelcore);
7244
7245 /* Continue if range is now fully accounted */
7246 if (end_pfn <= usable_startpfn) {
7247
7248 /*
7249 * Push zone_movable_pfn to the end so
7250 * that if we have to rebalance
7251 * kernelcore across nodes, we will
7252 * not double account here
7253 */
7254 zone_movable_pfn[nid] = end_pfn;
7255 continue;
7256 }
7257 start_pfn = usable_startpfn;
7258 }
7259
7260 /*
7261 * The usable PFN range for ZONE_MOVABLE is from
7262 * start_pfn->end_pfn. Calculate size_pages as the
7263 * number of pages used as kernelcore
7264 */
7265 size_pages = end_pfn - start_pfn;
7266 if (size_pages > kernelcore_remaining)
7267 size_pages = kernelcore_remaining;
7268 zone_movable_pfn[nid] = start_pfn + size_pages;
7269
7270 /*
7271 * Some kernelcore has been met, update counts and
7272 * break if the kernelcore for this node has been
7273 * satisfied
7274 */
7275 required_kernelcore -= min(required_kernelcore,
7276 size_pages);
7277 kernelcore_remaining -= size_pages;
7278 if (!kernelcore_remaining)
7279 break;
7280 }
7281 }
7282
7283 /*
7284 * If there is still required_kernelcore, we do another pass with one
7285 * less node in the count. This will push zone_movable_pfn[nid] further
7286 * along on the nodes that still have memory until kernelcore is
7287 * satisfied
7288 */
7289 usable_nodes--;
7290 if (usable_nodes && required_kernelcore > usable_nodes)
7291 goto restart;
7292
7293out2:
7294 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7295 for (nid = 0; nid < MAX_NUMNODES; nid++)
7296 zone_movable_pfn[nid] =
7297 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7298
7299out:
7300 /* restore the node_state */
7301 node_states[N_MEMORY] = saved_node_state;
7302}
7303
7304/* Any regular or high memory on that node ? */
7305static void check_for_memory(pg_data_t *pgdat, int nid)
7306{
7307 enum zone_type zone_type;
7308
7309 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7310 struct zone *zone = &pgdat->node_zones[zone_type];
7311 if (populated_zone(zone)) {
7312 if (IS_ENABLED(CONFIG_HIGHMEM))
7313 node_set_state(nid, N_HIGH_MEMORY);
7314 if (zone_type <= ZONE_NORMAL)
7315 node_set_state(nid, N_NORMAL_MEMORY);
7316 break;
7317 }
7318 }
7319}
7320
7321/**
7322 * free_area_init_nodes - Initialise all pg_data_t and zone data
7323 * @max_zone_pfn: an array of max PFNs for each zone
7324 *
7325 * This will call free_area_init_node() for each active node in the system.
7326 * Using the page ranges provided by memblock_set_node(), the size of each
7327 * zone in each node and their holes is calculated. If the maximum PFN
7328 * between two adjacent zones match, it is assumed that the zone is empty.
7329 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7330 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7331 * starts where the previous one ended. For example, ZONE_DMA32 starts
7332 * at arch_max_dma_pfn.
7333 */
7334void __init free_area_init_nodes(unsigned long *max_zone_pfn)
7335{
7336 unsigned long start_pfn, end_pfn;
7337 int i, nid;
7338
7339 /* Record where the zone boundaries are */
7340 memset(arch_zone_lowest_possible_pfn, 0,
7341 sizeof(arch_zone_lowest_possible_pfn));
7342 memset(arch_zone_highest_possible_pfn, 0,
7343 sizeof(arch_zone_highest_possible_pfn));
7344
7345 start_pfn = find_min_pfn_with_active_regions();
7346
7347 for (i = 0; i < MAX_NR_ZONES; i++) {
7348 if (i == ZONE_MOVABLE)
7349 continue;
7350
7351 end_pfn = max(max_zone_pfn[i], start_pfn);
7352 arch_zone_lowest_possible_pfn[i] = start_pfn;
7353 arch_zone_highest_possible_pfn[i] = end_pfn;
7354
7355 start_pfn = end_pfn;
7356 }
7357
7358 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7359 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7360 find_zone_movable_pfns_for_nodes();
7361
7362 /* Print out the zone ranges */
7363 pr_info("Zone ranges:\n");
7364 for (i = 0; i < MAX_NR_ZONES; i++) {
7365 if (i == ZONE_MOVABLE)
7366 continue;
7367 pr_info(" %-8s ", zone_names[i]);
7368 if (arch_zone_lowest_possible_pfn[i] ==
7369 arch_zone_highest_possible_pfn[i])
7370 pr_cont("empty\n");
7371 else
7372 pr_cont("[mem %#018Lx-%#018Lx]\n",
7373 (u64)arch_zone_lowest_possible_pfn[i]
7374 << PAGE_SHIFT,
7375 ((u64)arch_zone_highest_possible_pfn[i]
7376 << PAGE_SHIFT) - 1);
7377 }
7378
7379 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7380 pr_info("Movable zone start for each node\n");
7381 for (i = 0; i < MAX_NUMNODES; i++) {
7382 if (zone_movable_pfn[i])
7383 pr_info(" Node %d: %#018Lx\n", i,
7384 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7385 }
7386
7387 /*
7388 * Print out the early node map, and initialize the
7389 * subsection-map relative to active online memory ranges to
7390 * enable future "sub-section" extensions of the memory map.
7391 */
7392 pr_info("Early memory node ranges\n");
7393 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7394 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7395 (u64)start_pfn << PAGE_SHIFT,
7396 ((u64)end_pfn << PAGE_SHIFT) - 1);
7397 subsection_map_init(start_pfn, end_pfn - start_pfn);
7398 }
7399
7400 /* Initialise every node */
7401 mminit_verify_pageflags_layout();
7402 setup_nr_node_ids();
7403 zero_resv_unavail();
7404 for_each_online_node(nid) {
7405 pg_data_t *pgdat = NODE_DATA(nid);
7406 free_area_init_node(nid, NULL,
7407 find_min_pfn_for_node(nid), NULL);
7408
7409 /* Any memory on that node */
7410 if (pgdat->node_present_pages)
7411 node_set_state(nid, N_MEMORY);
7412 check_for_memory(pgdat, nid);
7413 }
7414}
7415
7416static int __init cmdline_parse_core(char *p, unsigned long *core,
7417 unsigned long *percent)
7418{
7419 unsigned long long coremem;
7420 char *endptr;
7421
7422 if (!p)
7423 return -EINVAL;
7424
7425 /* Value may be a percentage of total memory, otherwise bytes */
7426 coremem = simple_strtoull(p, &endptr, 0);
7427 if (*endptr == '%') {
7428 /* Paranoid check for percent values greater than 100 */
7429 WARN_ON(coremem > 100);
7430
7431 *percent = coremem;
7432 } else {
7433 coremem = memparse(p, &p);
7434 /* Paranoid check that UL is enough for the coremem value */
7435 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7436
7437 *core = coremem >> PAGE_SHIFT;
7438 *percent = 0UL;
7439 }
7440 return 0;
7441}
7442
7443/*
7444 * kernelcore=size sets the amount of memory for use for allocations that
7445 * cannot be reclaimed or migrated.
7446 */
7447static int __init cmdline_parse_kernelcore(char *p)
7448{
7449 /* parse kernelcore=mirror */
7450 if (parse_option_str(p, "mirror")) {
7451 mirrored_kernelcore = true;
7452 return 0;
7453 }
7454
7455 return cmdline_parse_core(p, &required_kernelcore,
7456 &required_kernelcore_percent);
7457}
7458
7459/*
7460 * movablecore=size sets the amount of memory for use for allocations that
7461 * can be reclaimed or migrated.
7462 */
7463static int __init cmdline_parse_movablecore(char *p)
7464{
7465 return cmdline_parse_core(p, &required_movablecore,
7466 &required_movablecore_percent);
7467}
7468
7469early_param("kernelcore", cmdline_parse_kernelcore);
7470early_param("movablecore", cmdline_parse_movablecore);
7471
7472#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
7473
7474void adjust_managed_page_count(struct page *page, long count)
7475{
7476 atomic_long_add(count, &page_zone(page)->managed_pages);
7477 totalram_pages_add(count);
7478#ifdef CONFIG_HIGHMEM
7479 if (PageHighMem(page))
7480 totalhigh_pages_add(count);
7481#endif
7482}
7483EXPORT_SYMBOL(adjust_managed_page_count);
7484
7485unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7486{
7487 void *pos;
7488 unsigned long pages = 0;
7489
7490 start = (void *)PAGE_ALIGN((unsigned long)start);
7491 end = (void *)((unsigned long)end & PAGE_MASK);
7492 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7493 struct page *page = virt_to_page(pos);
7494 void *direct_map_addr;
7495
7496 /*
7497 * 'direct_map_addr' might be different from 'pos'
7498 * because some architectures' virt_to_page()
7499 * work with aliases. Getting the direct map
7500 * address ensures that we get a _writeable_
7501 * alias for the memset().
7502 */
7503 direct_map_addr = page_address(page);
7504 if ((unsigned int)poison <= 0xFF)
7505 memset(direct_map_addr, poison, PAGE_SIZE);
7506
7507 free_reserved_page(page);
7508 }
7509
7510 if (pages && s)
7511 pr_info("Freeing %s memory: %ldK\n",
7512 s, pages << (PAGE_SHIFT - 10));
7513
7514 return pages;
7515}
7516
7517#ifdef CONFIG_HIGHMEM
7518void free_highmem_page(struct page *page)
7519{
7520 __free_reserved_page(page);
7521 totalram_pages_inc();
7522 atomic_long_inc(&page_zone(page)->managed_pages);
7523 totalhigh_pages_inc();
7524}
7525#endif
7526
7527
7528void __init mem_init_print_info(const char *str)
7529{
7530 unsigned long physpages, codesize, datasize, rosize, bss_size;
7531 unsigned long init_code_size, init_data_size;
7532
7533 physpages = get_num_physpages();
7534 codesize = _etext - _stext;
7535 datasize = _edata - _sdata;
7536 rosize = __end_rodata - __start_rodata;
7537 bss_size = __bss_stop - __bss_start;
7538 init_data_size = __init_end - __init_begin;
7539 init_code_size = _einittext - _sinittext;
7540
7541 /*
7542 * Detect special cases and adjust section sizes accordingly:
7543 * 1) .init.* may be embedded into .data sections
7544 * 2) .init.text.* may be out of [__init_begin, __init_end],
7545 * please refer to arch/tile/kernel/vmlinux.lds.S.
7546 * 3) .rodata.* may be embedded into .text or .data sections.
7547 */
7548#define adj_init_size(start, end, size, pos, adj) \
7549 do { \
7550 if (start <= pos && pos < end && size > adj) \
7551 size -= adj; \
7552 } while (0)
7553
7554 adj_init_size(__init_begin, __init_end, init_data_size,
7555 _sinittext, init_code_size);
7556 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7557 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7558 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7559 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7560
7561#undef adj_init_size
7562
7563 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7564#ifdef CONFIG_HIGHMEM
7565 ", %luK highmem"
7566#endif
7567 "%s%s)\n",
7568 nr_free_pages() << (PAGE_SHIFT - 10),
7569 physpages << (PAGE_SHIFT - 10),
7570 codesize >> 10, datasize >> 10, rosize >> 10,
7571 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7572 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7573 totalcma_pages << (PAGE_SHIFT - 10),
7574#ifdef CONFIG_HIGHMEM
7575 totalhigh_pages() << (PAGE_SHIFT - 10),
7576#endif
7577 str ? ", " : "", str ? str : "");
7578}
7579
7580/**
7581 * set_dma_reserve - set the specified number of pages reserved in the first zone
7582 * @new_dma_reserve: The number of pages to mark reserved
7583 *
7584 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7585 * In the DMA zone, a significant percentage may be consumed by kernel image
7586 * and other unfreeable allocations which can skew the watermarks badly. This
7587 * function may optionally be used to account for unfreeable pages in the
7588 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7589 * smaller per-cpu batchsize.
7590 */
7591void __init set_dma_reserve(unsigned long new_dma_reserve)
7592{
7593 dma_reserve = new_dma_reserve;
7594}
7595
7596void __init free_area_init(unsigned long *zones_size)
7597{
7598 zero_resv_unavail();
7599 free_area_init_node(0, zones_size,
7600 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7601}
7602
7603static int page_alloc_cpu_dead(unsigned int cpu)
7604{
7605
7606 lru_add_drain_cpu(cpu);
7607 drain_pages(cpu);
7608
7609 /*
7610 * Spill the event counters of the dead processor
7611 * into the current processors event counters.
7612 * This artificially elevates the count of the current
7613 * processor.
7614 */
7615 vm_events_fold_cpu(cpu);
7616
7617 /*
7618 * Zero the differential counters of the dead processor
7619 * so that the vm statistics are consistent.
7620 *
7621 * This is only okay since the processor is dead and cannot
7622 * race with what we are doing.
7623 */
7624 cpu_vm_stats_fold(cpu);
7625 return 0;
7626}
7627
7628#ifdef CONFIG_NUMA
7629int hashdist = HASHDIST_DEFAULT;
7630
7631static int __init set_hashdist(char *str)
7632{
7633 if (!str)
7634 return 0;
7635 hashdist = simple_strtoul(str, &str, 0);
7636 return 1;
7637}
7638__setup("hashdist=", set_hashdist);
7639#endif
7640
7641void __init page_alloc_init(void)
7642{
7643 int ret;
7644
7645#ifdef CONFIG_NUMA
7646 if (num_node_state(N_MEMORY) == 1)
7647 hashdist = 0;
7648#endif
7649
7650 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7651 "mm/page_alloc:dead", NULL,
7652 page_alloc_cpu_dead);
7653 WARN_ON(ret < 0);
7654}
7655
7656/*
7657 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7658 * or min_free_kbytes changes.
7659 */
7660static void calculate_totalreserve_pages(void)
7661{
7662 struct pglist_data *pgdat;
7663 unsigned long reserve_pages = 0;
7664 enum zone_type i, j;
7665
7666 for_each_online_pgdat(pgdat) {
7667
7668 pgdat->totalreserve_pages = 0;
7669
7670 for (i = 0; i < MAX_NR_ZONES; i++) {
7671 struct zone *zone = pgdat->node_zones + i;
7672 long max = 0;
7673 unsigned long managed_pages = zone_managed_pages(zone);
7674
7675 /* Find valid and maximum lowmem_reserve in the zone */
7676 for (j = i; j < MAX_NR_ZONES; j++) {
7677 if (zone->lowmem_reserve[j] > max)
7678 max = zone->lowmem_reserve[j];
7679 }
7680
7681 /* we treat the high watermark as reserved pages. */
7682 max += high_wmark_pages(zone);
7683
7684 if (max > managed_pages)
7685 max = managed_pages;
7686
7687 pgdat->totalreserve_pages += max;
7688
7689 reserve_pages += max;
7690 }
7691 }
7692 totalreserve_pages = reserve_pages;
7693}
7694
7695/*
7696 * setup_per_zone_lowmem_reserve - called whenever
7697 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7698 * has a correct pages reserved value, so an adequate number of
7699 * pages are left in the zone after a successful __alloc_pages().
7700 */
7701static void setup_per_zone_lowmem_reserve(void)
7702{
7703 struct pglist_data *pgdat;
7704 enum zone_type j, idx;
7705
7706 for_each_online_pgdat(pgdat) {
7707 for (j = 0; j < MAX_NR_ZONES; j++) {
7708 struct zone *zone = pgdat->node_zones + j;
7709 unsigned long managed_pages = zone_managed_pages(zone);
7710
7711 zone->lowmem_reserve[j] = 0;
7712
7713 idx = j;
7714 while (idx) {
7715 struct zone *lower_zone;
7716
7717 idx--;
7718 lower_zone = pgdat->node_zones + idx;
7719
7720 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7721 sysctl_lowmem_reserve_ratio[idx] = 0;
7722 lower_zone->lowmem_reserve[j] = 0;
7723 } else {
7724 lower_zone->lowmem_reserve[j] =
7725 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7726 }
7727 managed_pages += zone_managed_pages(lower_zone);
7728 }
7729 }
7730 }
7731
7732 /* update totalreserve_pages */
7733 calculate_totalreserve_pages();
7734}
7735
7736static void __setup_per_zone_wmarks(void)
7737{
7738 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7739 unsigned long lowmem_pages = 0;
7740 struct zone *zone;
7741 unsigned long flags;
7742
7743 /* Calculate total number of !ZONE_HIGHMEM pages */
7744 for_each_zone(zone) {
7745 if (!is_highmem(zone))
7746 lowmem_pages += zone_managed_pages(zone);
7747 }
7748
7749 for_each_zone(zone) {
7750 u64 tmp;
7751
7752 spin_lock_irqsave(&zone->lock, flags);
7753 tmp = (u64)pages_min * zone_managed_pages(zone);
7754 do_div(tmp, lowmem_pages);
7755 if (is_highmem(zone)) {
7756 /*
7757 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7758 * need highmem pages, so cap pages_min to a small
7759 * value here.
7760 *
7761 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7762 * deltas control async page reclaim, and so should
7763 * not be capped for highmem.
7764 */
7765 unsigned long min_pages;
7766
7767 min_pages = zone_managed_pages(zone) / 1024;
7768 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7769 zone->_watermark[WMARK_MIN] = min_pages;
7770 } else {
7771 /*
7772 * If it's a lowmem zone, reserve a number of pages
7773 * proportionate to the zone's size.
7774 */
7775 zone->_watermark[WMARK_MIN] = tmp;
7776 }
7777
7778 /*
7779 * Set the kswapd watermarks distance according to the
7780 * scale factor in proportion to available memory, but
7781 * ensure a minimum size on small systems.
7782 */
7783 tmp = max_t(u64, tmp >> 2,
7784 mult_frac(zone_managed_pages(zone),
7785 watermark_scale_factor, 10000));
7786
7787 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7788 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7789 zone->watermark_boost = 0;
7790
7791 spin_unlock_irqrestore(&zone->lock, flags);
7792 }
7793
7794 /* update totalreserve_pages */
7795 calculate_totalreserve_pages();
7796}
7797
7798/**
7799 * setup_per_zone_wmarks - called when min_free_kbytes changes
7800 * or when memory is hot-{added|removed}
7801 *
7802 * Ensures that the watermark[min,low,high] values for each zone are set
7803 * correctly with respect to min_free_kbytes.
7804 */
7805void setup_per_zone_wmarks(void)
7806{
7807 static DEFINE_SPINLOCK(lock);
7808
7809 spin_lock(&lock);
7810 __setup_per_zone_wmarks();
7811 spin_unlock(&lock);
7812}
7813
7814/*
7815 * Initialise min_free_kbytes.
7816 *
7817 * For small machines we want it small (128k min). For large machines
7818 * we want it large (64MB max). But it is not linear, because network
7819 * bandwidth does not increase linearly with machine size. We use
7820 *
7821 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7822 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7823 *
7824 * which yields
7825 *
7826 * 16MB: 512k
7827 * 32MB: 724k
7828 * 64MB: 1024k
7829 * 128MB: 1448k
7830 * 256MB: 2048k
7831 * 512MB: 2896k
7832 * 1024MB: 4096k
7833 * 2048MB: 5792k
7834 * 4096MB: 8192k
7835 * 8192MB: 11584k
7836 * 16384MB: 16384k
7837 */
7838int __meminit init_per_zone_wmark_min(void)
7839{
7840 unsigned long lowmem_kbytes;
7841 int new_min_free_kbytes;
7842
7843 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7844 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7845
7846 if (new_min_free_kbytes > user_min_free_kbytes) {
7847 min_free_kbytes = new_min_free_kbytes;
7848 if (min_free_kbytes < 128)
7849 min_free_kbytes = 128;
7850 if (min_free_kbytes > 65536)
7851 min_free_kbytes = 65536;
7852 } else {
7853 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7854 new_min_free_kbytes, user_min_free_kbytes);
7855 }
7856 setup_per_zone_wmarks();
7857 refresh_zone_stat_thresholds();
7858 setup_per_zone_lowmem_reserve();
7859
7860#ifdef CONFIG_NUMA
7861 setup_min_unmapped_ratio();
7862 setup_min_slab_ratio();
7863#endif
7864
7865 return 0;
7866}
7867core_initcall(init_per_zone_wmark_min)
7868
7869/*
7870 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7871 * that we can call two helper functions whenever min_free_kbytes
7872 * changes.
7873 */
7874int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7875 void __user *buffer, size_t *length, loff_t *ppos)
7876{
7877 int rc;
7878
7879 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7880 if (rc)
7881 return rc;
7882
7883 if (write) {
7884 user_min_free_kbytes = min_free_kbytes;
7885 setup_per_zone_wmarks();
7886 }
7887 return 0;
7888}
7889
7890int watermark_boost_factor_sysctl_handler(struct ctl_table *table, int write,
7891 void __user *buffer, size_t *length, loff_t *ppos)
7892{
7893 int rc;
7894
7895 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7896 if (rc)
7897 return rc;
7898
7899 return 0;
7900}
7901
7902int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7903 void __user *buffer, size_t *length, loff_t *ppos)
7904{
7905 int rc;
7906
7907 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7908 if (rc)
7909 return rc;
7910
7911 if (write)
7912 setup_per_zone_wmarks();
7913
7914 return 0;
7915}
7916
7917#ifdef CONFIG_NUMA
7918static void setup_min_unmapped_ratio(void)
7919{
7920 pg_data_t *pgdat;
7921 struct zone *zone;
7922
7923 for_each_online_pgdat(pgdat)
7924 pgdat->min_unmapped_pages = 0;
7925
7926 for_each_zone(zone)
7927 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7928 sysctl_min_unmapped_ratio) / 100;
7929}
7930
7931
7932int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7933 void __user *buffer, size_t *length, loff_t *ppos)
7934{
7935 int rc;
7936
7937 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7938 if (rc)
7939 return rc;
7940
7941 setup_min_unmapped_ratio();
7942
7943 return 0;
7944}
7945
7946static void setup_min_slab_ratio(void)
7947{
7948 pg_data_t *pgdat;
7949 struct zone *zone;
7950
7951 for_each_online_pgdat(pgdat)
7952 pgdat->min_slab_pages = 0;
7953
7954 for_each_zone(zone)
7955 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7956 sysctl_min_slab_ratio) / 100;
7957}
7958
7959int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7960 void __user *buffer, size_t *length, loff_t *ppos)
7961{
7962 int rc;
7963
7964 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7965 if (rc)
7966 return rc;
7967
7968 setup_min_slab_ratio();
7969
7970 return 0;
7971}
7972#endif
7973
7974/*
7975 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7976 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7977 * whenever sysctl_lowmem_reserve_ratio changes.
7978 *
7979 * The reserve ratio obviously has absolutely no relation with the
7980 * minimum watermarks. The lowmem reserve ratio can only make sense
7981 * if in function of the boot time zone sizes.
7982 */
7983int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7984 void __user *buffer, size_t *length, loff_t *ppos)
7985{
7986 proc_dointvec_minmax(table, write, buffer, length, ppos);
7987 setup_per_zone_lowmem_reserve();
7988 return 0;
7989}
7990
7991/*
7992 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7993 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7994 * pagelist can have before it gets flushed back to buddy allocator.
7995 */
7996int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7997 void __user *buffer, size_t *length, loff_t *ppos)
7998{
7999 struct zone *zone;
8000 int old_percpu_pagelist_fraction;
8001 int ret;
8002
8003 mutex_lock(&pcp_batch_high_lock);
8004 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8005
8006 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8007 if (!write || ret < 0)
8008 goto out;
8009
8010 /* Sanity checking to avoid pcp imbalance */
8011 if (percpu_pagelist_fraction &&
8012 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8013 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8014 ret = -EINVAL;
8015 goto out;
8016 }
8017
8018 /* No change? */
8019 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8020 goto out;
8021
8022 for_each_populated_zone(zone) {
8023 unsigned int cpu;
8024
8025 for_each_possible_cpu(cpu)
8026 pageset_set_high_and_batch(zone,
8027 per_cpu_ptr(zone->pageset, cpu));
8028 }
8029out:
8030 mutex_unlock(&pcp_batch_high_lock);
8031 return ret;
8032}
8033
8034#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8035/*
8036 * Returns the number of pages that arch has reserved but
8037 * is not known to alloc_large_system_hash().
8038 */
8039static unsigned long __init arch_reserved_kernel_pages(void)
8040{
8041 return 0;
8042}
8043#endif
8044
8045/*
8046 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8047 * machines. As memory size is increased the scale is also increased but at
8048 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8049 * quadruples the scale is increased by one, which means the size of hash table
8050 * only doubles, instead of quadrupling as well.
8051 * Because 32-bit systems cannot have large physical memory, where this scaling
8052 * makes sense, it is disabled on such platforms.
8053 */
8054#if __BITS_PER_LONG > 32
8055#define ADAPT_SCALE_BASE (64ul << 30)
8056#define ADAPT_SCALE_SHIFT 2
8057#define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8058#endif
8059
8060/*
8061 * allocate a large system hash table from bootmem
8062 * - it is assumed that the hash table must contain an exact power-of-2
8063 * quantity of entries
8064 * - limit is the number of hash buckets, not the total allocation size
8065 */
8066void *__init alloc_large_system_hash(const char *tablename,
8067 unsigned long bucketsize,
8068 unsigned long numentries,
8069 int scale,
8070 int flags,
8071 unsigned int *_hash_shift,
8072 unsigned int *_hash_mask,
8073 unsigned long low_limit,
8074 unsigned long high_limit)
8075{
8076 unsigned long long max = high_limit;
8077 unsigned long log2qty, size;
8078 void *table = NULL;
8079 gfp_t gfp_flags;
8080 bool virt;
8081
8082 /* allow the kernel cmdline to have a say */
8083 if (!numentries) {
8084 /* round applicable memory size up to nearest megabyte */
8085 numentries = nr_kernel_pages;
8086 numentries -= arch_reserved_kernel_pages();
8087
8088 /* It isn't necessary when PAGE_SIZE >= 1MB */
8089 if (PAGE_SHIFT < 20)
8090 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8091
8092#if __BITS_PER_LONG > 32
8093 if (!high_limit) {
8094 unsigned long adapt;
8095
8096 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8097 adapt <<= ADAPT_SCALE_SHIFT)
8098 scale++;
8099 }
8100#endif
8101
8102 /* limit to 1 bucket per 2^scale bytes of low memory */
8103 if (scale > PAGE_SHIFT)
8104 numentries >>= (scale - PAGE_SHIFT);
8105 else
8106 numentries <<= (PAGE_SHIFT - scale);
8107
8108 /* Make sure we've got at least a 0-order allocation.. */
8109 if (unlikely(flags & HASH_SMALL)) {
8110 /* Makes no sense without HASH_EARLY */
8111 WARN_ON(!(flags & HASH_EARLY));
8112 if (!(numentries >> *_hash_shift)) {
8113 numentries = 1UL << *_hash_shift;
8114 BUG_ON(!numentries);
8115 }
8116 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8117 numentries = PAGE_SIZE / bucketsize;
8118 }
8119 numentries = roundup_pow_of_two(numentries);
8120
8121 /* limit allocation size to 1/16 total memory by default */
8122 if (max == 0) {
8123 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8124 do_div(max, bucketsize);
8125 }
8126 max = min(max, 0x80000000ULL);
8127
8128 if (numentries < low_limit)
8129 numentries = low_limit;
8130 if (numentries > max)
8131 numentries = max;
8132
8133 log2qty = ilog2(numentries);
8134
8135 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8136 do {
8137 virt = false;
8138 size = bucketsize << log2qty;
8139 if (flags & HASH_EARLY) {
8140 if (flags & HASH_ZERO)
8141 table = memblock_alloc(size, SMP_CACHE_BYTES);
8142 else
8143 table = memblock_alloc_raw(size,
8144 SMP_CACHE_BYTES);
8145 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8146 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
8147 virt = true;
8148 } else {
8149 /*
8150 * If bucketsize is not a power-of-two, we may free
8151 * some pages at the end of hash table which
8152 * alloc_pages_exact() automatically does
8153 */
8154 table = alloc_pages_exact(size, gfp_flags);
8155 kmemleak_alloc(table, size, 1, gfp_flags);
8156 }
8157 } while (!table && size > PAGE_SIZE && --log2qty);
8158
8159 if (!table)
8160 panic("Failed to allocate %s hash table\n", tablename);
8161
8162 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8163 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8164 virt ? "vmalloc" : "linear");
8165
8166 if (_hash_shift)
8167 *_hash_shift = log2qty;
8168 if (_hash_mask)
8169 *_hash_mask = (1 << log2qty) - 1;
8170
8171 return table;
8172}
8173
8174/*
8175 * This function checks whether pageblock includes unmovable pages or not.
8176 * If @count is not zero, it is okay to include less @count unmovable pages
8177 *
8178 * PageLRU check without isolation or lru_lock could race so that
8179 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8180 * check without lock_page also may miss some movable non-lru pages at
8181 * race condition. So you can't expect this function should be exact.
8182 */
8183bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
8184 int migratetype, int flags)
8185{
8186 unsigned long found;
8187 unsigned long iter = 0;
8188 unsigned long pfn = page_to_pfn(page);
8189 const char *reason = "unmovable page";
8190
8191 /*
8192 * TODO we could make this much more efficient by not checking every
8193 * page in the range if we know all of them are in MOVABLE_ZONE and
8194 * that the movable zone guarantees that pages are migratable but
8195 * the later is not the case right now unfortunatelly. E.g. movablecore
8196 * can still lead to having bootmem allocations in zone_movable.
8197 */
8198
8199 if (is_migrate_cma_page(page)) {
8200 /*
8201 * CMA allocations (alloc_contig_range) really need to mark
8202 * isolate CMA pageblocks even when they are not movable in fact
8203 * so consider them movable here.
8204 */
8205 if (is_migrate_cma(migratetype))
8206 return false;
8207
8208 reason = "CMA page";
8209 goto unmovable;
8210 }
8211
8212 for (found = 0; iter < pageblock_nr_pages; iter++) {
8213 unsigned long check = pfn + iter;
8214
8215 if (!pfn_valid_within(check))
8216 continue;
8217
8218 page = pfn_to_page(check);
8219
8220 if (PageReserved(page))
8221 goto unmovable;
8222
8223 /*
8224 * If the zone is movable and we have ruled out all reserved
8225 * pages then it should be reasonably safe to assume the rest
8226 * is movable.
8227 */
8228 if (zone_idx(zone) == ZONE_MOVABLE)
8229 continue;
8230
8231 /*
8232 * Hugepages are not in LRU lists, but they're movable.
8233 * We need not scan over tail pages because we don't
8234 * handle each tail page individually in migration.
8235 */
8236 if (PageHuge(page)) {
8237 struct page *head = compound_head(page);
8238 unsigned int skip_pages;
8239
8240 if (!hugepage_migration_supported(page_hstate(head)))
8241 goto unmovable;
8242
8243 skip_pages = compound_nr(head) - (page - head);
8244 iter += skip_pages - 1;
8245 continue;
8246 }
8247
8248 /*
8249 * We can't use page_count without pin a page
8250 * because another CPU can free compound page.
8251 * This check already skips compound tails of THP
8252 * because their page->_refcount is zero at all time.
8253 */
8254 if (!page_ref_count(page)) {
8255 if (PageBuddy(page))
8256 iter += (1 << page_order(page)) - 1;
8257 continue;
8258 }
8259
8260 /*
8261 * The HWPoisoned page may be not in buddy system, and
8262 * page_count() is not 0.
8263 */
8264 if ((flags & SKIP_HWPOISON) && PageHWPoison(page))
8265 continue;
8266
8267 if (__PageMovable(page))
8268 continue;
8269
8270 if (!PageLRU(page))
8271 found++;
8272 /*
8273 * If there are RECLAIMABLE pages, we need to check
8274 * it. But now, memory offline itself doesn't call
8275 * shrink_node_slabs() and it still to be fixed.
8276 */
8277 /*
8278 * If the page is not RAM, page_count()should be 0.
8279 * we don't need more check. This is an _used_ not-movable page.
8280 *
8281 * The problematic thing here is PG_reserved pages. PG_reserved
8282 * is set to both of a memory hole page and a _used_ kernel
8283 * page at boot.
8284 */
8285 if (found > count)
8286 goto unmovable;
8287 }
8288 return false;
8289unmovable:
8290 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
8291 if (flags & REPORT_FAILURE)
8292 dump_page(pfn_to_page(pfn + iter), reason);
8293 return true;
8294}
8295
8296#ifdef CONFIG_CONTIG_ALLOC
8297static unsigned long pfn_max_align_down(unsigned long pfn)
8298{
8299 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8300 pageblock_nr_pages) - 1);
8301}
8302
8303static unsigned long pfn_max_align_up(unsigned long pfn)
8304{
8305 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8306 pageblock_nr_pages));
8307}
8308
8309/* [start, end) must belong to a single zone. */
8310static int __alloc_contig_migrate_range(struct compact_control *cc,
8311 unsigned long start, unsigned long end)
8312{
8313 /* This function is based on compact_zone() from compaction.c. */
8314 unsigned long nr_reclaimed;
8315 unsigned long pfn = start;
8316 unsigned int tries = 0;
8317 int ret = 0;
8318
8319 migrate_prep();
8320
8321 while (pfn < end || !list_empty(&cc->migratepages)) {
8322 if (fatal_signal_pending(current)) {
8323 ret = -EINTR;
8324 break;
8325 }
8326
8327 if (list_empty(&cc->migratepages)) {
8328 cc->nr_migratepages = 0;
8329 pfn = isolate_migratepages_range(cc, pfn, end);
8330 if (!pfn) {
8331 ret = -EINTR;
8332 break;
8333 }
8334 tries = 0;
8335 } else if (++tries == 5) {
8336 ret = ret < 0 ? ret : -EBUSY;
8337 break;
8338 }
8339
8340 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8341 &cc->migratepages);
8342 cc->nr_migratepages -= nr_reclaimed;
8343
8344 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
8345 NULL, 0, cc->mode, MR_CONTIG_RANGE);
8346 }
8347 if (ret < 0) {
8348 putback_movable_pages(&cc->migratepages);
8349 return ret;
8350 }
8351 return 0;
8352}
8353
8354/**
8355 * alloc_contig_range() -- tries to allocate given range of pages
8356 * @start: start PFN to allocate
8357 * @end: one-past-the-last PFN to allocate
8358 * @migratetype: migratetype of the underlaying pageblocks (either
8359 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8360 * in range must have the same migratetype and it must
8361 * be either of the two.
8362 * @gfp_mask: GFP mask to use during compaction
8363 *
8364 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8365 * aligned. The PFN range must belong to a single zone.
8366 *
8367 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8368 * pageblocks in the range. Once isolated, the pageblocks should not
8369 * be modified by others.
8370 *
8371 * Return: zero on success or negative error code. On success all
8372 * pages which PFN is in [start, end) are allocated for the caller and
8373 * need to be freed with free_contig_range().
8374 */
8375int alloc_contig_range(unsigned long start, unsigned long end,
8376 unsigned migratetype, gfp_t gfp_mask)
8377{
8378 unsigned long outer_start, outer_end;
8379 unsigned int order;
8380 int ret = 0;
8381
8382 struct compact_control cc = {
8383 .nr_migratepages = 0,
8384 .order = -1,
8385 .zone = page_zone(pfn_to_page(start)),
8386 .mode = MIGRATE_SYNC,
8387 .ignore_skip_hint = true,
8388 .no_set_skip_hint = true,
8389 .gfp_mask = current_gfp_context(gfp_mask),
8390 };
8391 INIT_LIST_HEAD(&cc.migratepages);
8392
8393 /*
8394 * What we do here is we mark all pageblocks in range as
8395 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8396 * have different sizes, and due to the way page allocator
8397 * work, we align the range to biggest of the two pages so
8398 * that page allocator won't try to merge buddies from
8399 * different pageblocks and change MIGRATE_ISOLATE to some
8400 * other migration type.
8401 *
8402 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8403 * migrate the pages from an unaligned range (ie. pages that
8404 * we are interested in). This will put all the pages in
8405 * range back to page allocator as MIGRATE_ISOLATE.
8406 *
8407 * When this is done, we take the pages in range from page
8408 * allocator removing them from the buddy system. This way
8409 * page allocator will never consider using them.
8410 *
8411 * This lets us mark the pageblocks back as
8412 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8413 * aligned range but not in the unaligned, original range are
8414 * put back to page allocator so that buddy can use them.
8415 */
8416
8417 ret = start_isolate_page_range(pfn_max_align_down(start),
8418 pfn_max_align_up(end), migratetype, 0);
8419 if (ret < 0)
8420 return ret;
8421
8422 /*
8423 * In case of -EBUSY, we'd like to know which page causes problem.
8424 * So, just fall through. test_pages_isolated() has a tracepoint
8425 * which will report the busy page.
8426 *
8427 * It is possible that busy pages could become available before
8428 * the call to test_pages_isolated, and the range will actually be
8429 * allocated. So, if we fall through be sure to clear ret so that
8430 * -EBUSY is not accidentally used or returned to caller.
8431 */
8432 ret = __alloc_contig_migrate_range(&cc, start, end);
8433 if (ret && ret != -EBUSY)
8434 goto done;
8435 ret =0;
8436
8437 /*
8438 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8439 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8440 * more, all pages in [start, end) are free in page allocator.
8441 * What we are going to do is to allocate all pages from
8442 * [start, end) (that is remove them from page allocator).
8443 *
8444 * The only problem is that pages at the beginning and at the
8445 * end of interesting range may be not aligned with pages that
8446 * page allocator holds, ie. they can be part of higher order
8447 * pages. Because of this, we reserve the bigger range and
8448 * once this is done free the pages we are not interested in.
8449 *
8450 * We don't have to hold zone->lock here because the pages are
8451 * isolated thus they won't get removed from buddy.
8452 */
8453
8454 lru_add_drain_all();
8455
8456 order = 0;
8457 outer_start = start;
8458 while (!PageBuddy(pfn_to_page(outer_start))) {
8459 if (++order >= MAX_ORDER) {
8460 outer_start = start;
8461 break;
8462 }
8463 outer_start &= ~0UL << order;
8464 }
8465
8466 if (outer_start != start) {
8467 order = page_order(pfn_to_page(outer_start));
8468
8469 /*
8470 * outer_start page could be small order buddy page and
8471 * it doesn't include start page. Adjust outer_start
8472 * in this case to report failed page properly
8473 * on tracepoint in test_pages_isolated()
8474 */
8475 if (outer_start + (1UL << order) <= start)
8476 outer_start = start;
8477 }
8478
8479 /* Make sure the range is really isolated. */
8480 if (test_pages_isolated(outer_start, end, false)) {
8481 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8482 __func__, outer_start, end);
8483 ret = -EBUSY;
8484 goto done;
8485 }
8486
8487 /* Grab isolated pages from freelists. */
8488 outer_end = isolate_freepages_range(&cc, outer_start, end);
8489 if (!outer_end) {
8490 ret = -EBUSY;
8491 goto done;
8492 }
8493
8494 /* Free head and tail (if any) */
8495 if (start != outer_start)
8496 free_contig_range(outer_start, start - outer_start);
8497 if (end != outer_end)
8498 free_contig_range(end, outer_end - end);
8499
8500done:
8501 undo_isolate_page_range(pfn_max_align_down(start),
8502 pfn_max_align_up(end), migratetype);
8503 return ret;
8504}
8505#endif /* CONFIG_CONTIG_ALLOC */
8506
8507void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8508{
8509 unsigned int count = 0;
8510
8511 for (; nr_pages--; pfn++) {
8512 struct page *page = pfn_to_page(pfn);
8513
8514 count += page_count(page) != 1;
8515 __free_page(page);
8516 }
8517 WARN(count != 0, "%d pages are still in use!\n", count);
8518}
8519
8520/*
8521 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8522 * page high values need to be recalulated.
8523 */
8524void __meminit zone_pcp_update(struct zone *zone)
8525{
8526 unsigned cpu;
8527 mutex_lock(&pcp_batch_high_lock);
8528 for_each_possible_cpu(cpu)
8529 pageset_set_high_and_batch(zone,
8530 per_cpu_ptr(zone->pageset, cpu));
8531 mutex_unlock(&pcp_batch_high_lock);
8532}
8533
8534void zone_pcp_reset(struct zone *zone)
8535{
8536 unsigned long flags;
8537 int cpu;
8538 struct per_cpu_pageset *pset;
8539
8540 /* avoid races with drain_pages() */
8541 local_irq_save(flags);
8542 if (zone->pageset != &boot_pageset) {
8543 for_each_online_cpu(cpu) {
8544 pset = per_cpu_ptr(zone->pageset, cpu);
8545 drain_zonestat(zone, pset);
8546 }
8547 free_percpu(zone->pageset);
8548 zone->pageset = &boot_pageset;
8549 }
8550 local_irq_restore(flags);
8551}
8552
8553#ifdef CONFIG_MEMORY_HOTREMOVE
8554/*
8555 * All pages in the range must be in a single zone and isolated
8556 * before calling this.
8557 */
8558unsigned long
8559__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8560{
8561 struct page *page;
8562 struct zone *zone;
8563 unsigned int order, i;
8564 unsigned long pfn;
8565 unsigned long flags;
8566 unsigned long offlined_pages = 0;
8567
8568 /* find the first valid pfn */
8569 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8570 if (pfn_valid(pfn))
8571 break;
8572 if (pfn == end_pfn)
8573 return offlined_pages;
8574
8575 offline_mem_sections(pfn, end_pfn);
8576 zone = page_zone(pfn_to_page(pfn));
8577 spin_lock_irqsave(&zone->lock, flags);
8578 pfn = start_pfn;
8579 while (pfn < end_pfn) {
8580 if (!pfn_valid(pfn)) {
8581 pfn++;
8582 continue;
8583 }
8584 page = pfn_to_page(pfn);
8585 /*
8586 * The HWPoisoned page may be not in buddy system, and
8587 * page_count() is not 0.
8588 */
8589 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8590 pfn++;
8591 SetPageReserved(page);
8592 offlined_pages++;
8593 continue;
8594 }
8595
8596 BUG_ON(page_count(page));
8597 BUG_ON(!PageBuddy(page));
8598 order = page_order(page);
8599 offlined_pages += 1 << order;
8600#ifdef CONFIG_DEBUG_VM
8601 pr_info("remove from free list %lx %d %lx\n",
8602 pfn, 1 << order, end_pfn);
8603#endif
8604 del_page_from_free_area(page, &zone->free_area[order]);
8605 for (i = 0; i < (1 << order); i++)
8606 SetPageReserved((page+i));
8607 pfn += (1 << order);
8608 }
8609 spin_unlock_irqrestore(&zone->lock, flags);
8610
8611 return offlined_pages;
8612}
8613#endif
8614
8615bool is_free_buddy_page(struct page *page)
8616{
8617 struct zone *zone = page_zone(page);
8618 unsigned long pfn = page_to_pfn(page);
8619 unsigned long flags;
8620 unsigned int order;
8621
8622 spin_lock_irqsave(&zone->lock, flags);
8623 for (order = 0; order < MAX_ORDER; order++) {
8624 struct page *page_head = page - (pfn & ((1 << order) - 1));
8625
8626 if (PageBuddy(page_head) && page_order(page_head) >= order)
8627 break;
8628 }
8629 spin_unlock_irqrestore(&zone->lock, flags);
8630
8631 return order < MAX_ORDER;
8632}
8633
8634#ifdef CONFIG_MEMORY_FAILURE
8635/*
8636 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8637 * test is performed under the zone lock to prevent a race against page
8638 * allocation.
8639 */
8640bool set_hwpoison_free_buddy_page(struct page *page)
8641{
8642 struct zone *zone = page_zone(page);
8643 unsigned long pfn = page_to_pfn(page);
8644 unsigned long flags;
8645 unsigned int order;
8646 bool hwpoisoned = false;
8647
8648 spin_lock_irqsave(&zone->lock, flags);
8649 for (order = 0; order < MAX_ORDER; order++) {
8650 struct page *page_head = page - (pfn & ((1 << order) - 1));
8651
8652 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8653 if (!TestSetPageHWPoison(page))
8654 hwpoisoned = true;
8655 break;
8656 }
8657 }
8658 spin_unlock_irqrestore(&zone->lock, flags);
8659
8660 return hwpoisoned;
8661}
8662#endif