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/memory.h>
55#include <linux/compaction.h>
56#include <trace/events/kmem.h>
57#include <linux/ftrace_event.h>
58#include <linux/memcontrol.h>
59#include <linux/prefetch.h>
60
61#include <asm/tlbflush.h>
62#include <asm/div64.h>
63#include "internal.h"
64
65#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
66DEFINE_PER_CPU(int, numa_node);
67EXPORT_PER_CPU_SYMBOL(numa_node);
68#endif
69
70#ifdef CONFIG_HAVE_MEMORYLESS_NODES
71/*
72 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
73 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
74 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
75 * defined in <linux/topology.h>.
76 */
77DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
78EXPORT_PER_CPU_SYMBOL(_numa_mem_);
79#endif
80
81/*
82 * Array of node states.
83 */
84nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
85 [N_POSSIBLE] = NODE_MASK_ALL,
86 [N_ONLINE] = { { [0] = 1UL } },
87#ifndef CONFIG_NUMA
88 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
89#ifdef CONFIG_HIGHMEM
90 [N_HIGH_MEMORY] = { { [0] = 1UL } },
91#endif
92 [N_CPU] = { { [0] = 1UL } },
93#endif /* NUMA */
94};
95EXPORT_SYMBOL(node_states);
96
97unsigned long totalram_pages __read_mostly;
98unsigned long totalreserve_pages __read_mostly;
99int percpu_pagelist_fraction;
100gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
101
102#ifdef CONFIG_PM_SLEEP
103/*
104 * The following functions are used by the suspend/hibernate code to temporarily
105 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
106 * while devices are suspended. To avoid races with the suspend/hibernate code,
107 * they should always be called with pm_mutex held (gfp_allowed_mask also should
108 * only be modified with pm_mutex held, unless the suspend/hibernate code is
109 * guaranteed not to run in parallel with that modification).
110 */
111
112static gfp_t saved_gfp_mask;
113
114void pm_restore_gfp_mask(void)
115{
116 WARN_ON(!mutex_is_locked(&pm_mutex));
117 if (saved_gfp_mask) {
118 gfp_allowed_mask = saved_gfp_mask;
119 saved_gfp_mask = 0;
120 }
121}
122
123void pm_restrict_gfp_mask(void)
124{
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 WARN_ON(saved_gfp_mask);
127 saved_gfp_mask = gfp_allowed_mask;
128 gfp_allowed_mask &= ~GFP_IOFS;
129}
130#endif /* CONFIG_PM_SLEEP */
131
132#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
133int pageblock_order __read_mostly;
134#endif
135
136static void __free_pages_ok(struct page *page, unsigned int order);
137
138/*
139 * results with 256, 32 in the lowmem_reserve sysctl:
140 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
141 * 1G machine -> (16M dma, 784M normal, 224M high)
142 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
143 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
144 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
145 *
146 * TBD: should special case ZONE_DMA32 machines here - in those we normally
147 * don't need any ZONE_NORMAL reservation
148 */
149int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
150#ifdef CONFIG_ZONE_DMA
151 256,
152#endif
153#ifdef CONFIG_ZONE_DMA32
154 256,
155#endif
156#ifdef CONFIG_HIGHMEM
157 32,
158#endif
159 32,
160};
161
162EXPORT_SYMBOL(totalram_pages);
163
164static char * const zone_names[MAX_NR_ZONES] = {
165#ifdef CONFIG_ZONE_DMA
166 "DMA",
167#endif
168#ifdef CONFIG_ZONE_DMA32
169 "DMA32",
170#endif
171 "Normal",
172#ifdef CONFIG_HIGHMEM
173 "HighMem",
174#endif
175 "Movable",
176};
177
178int min_free_kbytes = 1024;
179
180static unsigned long __meminitdata nr_kernel_pages;
181static unsigned long __meminitdata nr_all_pages;
182static unsigned long __meminitdata dma_reserve;
183
184#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
185 /*
186 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
187 * ranges of memory (RAM) that may be registered with add_active_range().
188 * Ranges passed to add_active_range() will be merged if possible
189 * so the number of times add_active_range() can be called is
190 * related to the number of nodes and the number of holes
191 */
192 #ifdef CONFIG_MAX_ACTIVE_REGIONS
193 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
194 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
195 #else
196 #if MAX_NUMNODES >= 32
197 /* If there can be many nodes, allow up to 50 holes per node */
198 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
199 #else
200 /* By default, allow up to 256 distinct regions */
201 #define MAX_ACTIVE_REGIONS 256
202 #endif
203 #endif
204
205 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
206 static int __meminitdata nr_nodemap_entries;
207 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
208 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
209 static unsigned long __initdata required_kernelcore;
210 static unsigned long __initdata required_movablecore;
211 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
212
213 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
214 int movable_zone;
215 EXPORT_SYMBOL(movable_zone);
216#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
217
218#if MAX_NUMNODES > 1
219int nr_node_ids __read_mostly = MAX_NUMNODES;
220int nr_online_nodes __read_mostly = 1;
221EXPORT_SYMBOL(nr_node_ids);
222EXPORT_SYMBOL(nr_online_nodes);
223#endif
224
225int page_group_by_mobility_disabled __read_mostly;
226
227static void set_pageblock_migratetype(struct page *page, int migratetype)
228{
229
230 if (unlikely(page_group_by_mobility_disabled))
231 migratetype = MIGRATE_UNMOVABLE;
232
233 set_pageblock_flags_group(page, (unsigned long)migratetype,
234 PB_migrate, PB_migrate_end);
235}
236
237bool oom_killer_disabled __read_mostly;
238
239#ifdef CONFIG_DEBUG_VM
240static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
241{
242 int ret = 0;
243 unsigned seq;
244 unsigned long pfn = page_to_pfn(page);
245
246 do {
247 seq = zone_span_seqbegin(zone);
248 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
249 ret = 1;
250 else if (pfn < zone->zone_start_pfn)
251 ret = 1;
252 } while (zone_span_seqretry(zone, seq));
253
254 return ret;
255}
256
257static int page_is_consistent(struct zone *zone, struct page *page)
258{
259 if (!pfn_valid_within(page_to_pfn(page)))
260 return 0;
261 if (zone != page_zone(page))
262 return 0;
263
264 return 1;
265}
266/*
267 * Temporary debugging check for pages not lying within a given zone.
268 */
269static int bad_range(struct zone *zone, struct page *page)
270{
271 if (page_outside_zone_boundaries(zone, page))
272 return 1;
273 if (!page_is_consistent(zone, page))
274 return 1;
275
276 return 0;
277}
278#else
279static inline int bad_range(struct zone *zone, struct page *page)
280{
281 return 0;
282}
283#endif
284
285static void bad_page(struct page *page)
286{
287 static unsigned long resume;
288 static unsigned long nr_shown;
289 static unsigned long nr_unshown;
290
291 /* Don't complain about poisoned pages */
292 if (PageHWPoison(page)) {
293 reset_page_mapcount(page); /* remove PageBuddy */
294 return;
295 }
296
297 /*
298 * Allow a burst of 60 reports, then keep quiet for that minute;
299 * or allow a steady drip of one report per second.
300 */
301 if (nr_shown == 60) {
302 if (time_before(jiffies, resume)) {
303 nr_unshown++;
304 goto out;
305 }
306 if (nr_unshown) {
307 printk(KERN_ALERT
308 "BUG: Bad page state: %lu messages suppressed\n",
309 nr_unshown);
310 nr_unshown = 0;
311 }
312 nr_shown = 0;
313 }
314 if (nr_shown++ == 0)
315 resume = jiffies + 60 * HZ;
316
317 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
318 current->comm, page_to_pfn(page));
319 dump_page(page);
320
321 dump_stack();
322out:
323 /* Leave bad fields for debug, except PageBuddy could make trouble */
324 reset_page_mapcount(page); /* remove PageBuddy */
325 add_taint(TAINT_BAD_PAGE);
326}
327
328/*
329 * Higher-order pages are called "compound pages". They are structured thusly:
330 *
331 * The first PAGE_SIZE page is called the "head page".
332 *
333 * The remaining PAGE_SIZE pages are called "tail pages".
334 *
335 * All pages have PG_compound set. All pages have their ->private pointing at
336 * the head page (even the head page has this).
337 *
338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
341 */
342
343static void free_compound_page(struct page *page)
344{
345 __free_pages_ok(page, compound_order(page));
346}
347
348void prep_compound_page(struct page *page, unsigned long order)
349{
350 int i;
351 int nr_pages = 1 << order;
352
353 set_compound_page_dtor(page, free_compound_page);
354 set_compound_order(page, order);
355 __SetPageHead(page);
356 for (i = 1; i < nr_pages; i++) {
357 struct page *p = page + i;
358
359 __SetPageTail(p);
360 p->first_page = page;
361 }
362}
363
364/* update __split_huge_page_refcount if you change this function */
365static int destroy_compound_page(struct page *page, unsigned long order)
366{
367 int i;
368 int nr_pages = 1 << order;
369 int bad = 0;
370
371 if (unlikely(compound_order(page) != order) ||
372 unlikely(!PageHead(page))) {
373 bad_page(page);
374 bad++;
375 }
376
377 __ClearPageHead(page);
378
379 for (i = 1; i < nr_pages; i++) {
380 struct page *p = page + i;
381
382 if (unlikely(!PageTail(p) || (p->first_page != page))) {
383 bad_page(page);
384 bad++;
385 }
386 __ClearPageTail(p);
387 }
388
389 return bad;
390}
391
392static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
393{
394 int i;
395
396 /*
397 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
398 * and __GFP_HIGHMEM from hard or soft interrupt context.
399 */
400 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
401 for (i = 0; i < (1 << order); i++)
402 clear_highpage(page + i);
403}
404
405static inline void set_page_order(struct page *page, int order)
406{
407 set_page_private(page, order);
408 __SetPageBuddy(page);
409}
410
411static inline void rmv_page_order(struct page *page)
412{
413 __ClearPageBuddy(page);
414 set_page_private(page, 0);
415}
416
417/*
418 * Locate the struct page for both the matching buddy in our
419 * pair (buddy1) and the combined O(n+1) page they form (page).
420 *
421 * 1) Any buddy B1 will have an order O twin B2 which satisfies
422 * the following equation:
423 * B2 = B1 ^ (1 << O)
424 * For example, if the starting buddy (buddy2) is #8 its order
425 * 1 buddy is #10:
426 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
427 *
428 * 2) Any buddy B will have an order O+1 parent P which
429 * satisfies the following equation:
430 * P = B & ~(1 << O)
431 *
432 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
433 */
434static inline unsigned long
435__find_buddy_index(unsigned long page_idx, unsigned int order)
436{
437 return page_idx ^ (1 << order);
438}
439
440/*
441 * This function checks whether a page is free && is the buddy
442 * we can do coalesce a page and its buddy if
443 * (a) the buddy is not in a hole &&
444 * (b) the buddy is in the buddy system &&
445 * (c) a page and its buddy have the same order &&
446 * (d) a page and its buddy are in the same zone.
447 *
448 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
449 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
450 *
451 * For recording page's order, we use page_private(page).
452 */
453static inline int page_is_buddy(struct page *page, struct page *buddy,
454 int order)
455{
456 if (!pfn_valid_within(page_to_pfn(buddy)))
457 return 0;
458
459 if (page_zone_id(page) != page_zone_id(buddy))
460 return 0;
461
462 if (PageBuddy(buddy) && page_order(buddy) == order) {
463 VM_BUG_ON(page_count(buddy) != 0);
464 return 1;
465 }
466 return 0;
467}
468
469/*
470 * Freeing function for a buddy system allocator.
471 *
472 * The concept of a buddy system is to maintain direct-mapped table
473 * (containing bit values) for memory blocks of various "orders".
474 * The bottom level table contains the map for the smallest allocatable
475 * units of memory (here, pages), and each level above it describes
476 * pairs of units from the levels below, hence, "buddies".
477 * At a high level, all that happens here is marking the table entry
478 * at the bottom level available, and propagating the changes upward
479 * as necessary, plus some accounting needed to play nicely with other
480 * parts of the VM system.
481 * At each level, we keep a list of pages, which are heads of continuous
482 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
483 * order is recorded in page_private(page) field.
484 * So when we are allocating or freeing one, we can derive the state of the
485 * other. That is, if we allocate a small block, and both were
486 * free, the remainder of the region must be split into blocks.
487 * If a block is freed, and its buddy is also free, then this
488 * triggers coalescing into a block of larger size.
489 *
490 * -- wli
491 */
492
493static inline void __free_one_page(struct page *page,
494 struct zone *zone, unsigned int order,
495 int migratetype)
496{
497 unsigned long page_idx;
498 unsigned long combined_idx;
499 unsigned long uninitialized_var(buddy_idx);
500 struct page *buddy;
501
502 if (unlikely(PageCompound(page)))
503 if (unlikely(destroy_compound_page(page, order)))
504 return;
505
506 VM_BUG_ON(migratetype == -1);
507
508 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
509
510 VM_BUG_ON(page_idx & ((1 << order) - 1));
511 VM_BUG_ON(bad_range(zone, page));
512
513 while (order < MAX_ORDER-1) {
514 buddy_idx = __find_buddy_index(page_idx, order);
515 buddy = page + (buddy_idx - page_idx);
516 if (!page_is_buddy(page, buddy, order))
517 break;
518
519 /* Our buddy is free, merge with it and move up one order. */
520 list_del(&buddy->lru);
521 zone->free_area[order].nr_free--;
522 rmv_page_order(buddy);
523 combined_idx = buddy_idx & page_idx;
524 page = page + (combined_idx - page_idx);
525 page_idx = combined_idx;
526 order++;
527 }
528 set_page_order(page, order);
529
530 /*
531 * If this is not the largest possible page, check if the buddy
532 * of the next-highest order is free. If it is, it's possible
533 * that pages are being freed that will coalesce soon. In case,
534 * that is happening, add the free page to the tail of the list
535 * so it's less likely to be used soon and more likely to be merged
536 * as a higher order page
537 */
538 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
539 struct page *higher_page, *higher_buddy;
540 combined_idx = buddy_idx & page_idx;
541 higher_page = page + (combined_idx - page_idx);
542 buddy_idx = __find_buddy_index(combined_idx, order + 1);
543 higher_buddy = page + (buddy_idx - combined_idx);
544 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
545 list_add_tail(&page->lru,
546 &zone->free_area[order].free_list[migratetype]);
547 goto out;
548 }
549 }
550
551 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
552out:
553 zone->free_area[order].nr_free++;
554}
555
556/*
557 * free_page_mlock() -- clean up attempts to free and mlocked() page.
558 * Page should not be on lru, so no need to fix that up.
559 * free_pages_check() will verify...
560 */
561static inline void free_page_mlock(struct page *page)
562{
563 __dec_zone_page_state(page, NR_MLOCK);
564 __count_vm_event(UNEVICTABLE_MLOCKFREED);
565}
566
567static inline int free_pages_check(struct page *page)
568{
569 if (unlikely(page_mapcount(page) |
570 (page->mapping != NULL) |
571 (atomic_read(&page->_count) != 0) |
572 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
573 (mem_cgroup_bad_page_check(page)))) {
574 bad_page(page);
575 return 1;
576 }
577 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
578 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
579 return 0;
580}
581
582/*
583 * Frees a number of pages from the PCP lists
584 * Assumes all pages on list are in same zone, and of same order.
585 * count is the number of pages to free.
586 *
587 * If the zone was previously in an "all pages pinned" state then look to
588 * see if this freeing clears that state.
589 *
590 * And clear the zone's pages_scanned counter, to hold off the "all pages are
591 * pinned" detection logic.
592 */
593static void free_pcppages_bulk(struct zone *zone, int count,
594 struct per_cpu_pages *pcp)
595{
596 int migratetype = 0;
597 int batch_free = 0;
598 int to_free = count;
599
600 spin_lock(&zone->lock);
601 zone->all_unreclaimable = 0;
602 zone->pages_scanned = 0;
603
604 while (to_free) {
605 struct page *page;
606 struct list_head *list;
607
608 /*
609 * Remove pages from lists in a round-robin fashion. A
610 * batch_free count is maintained that is incremented when an
611 * empty list is encountered. This is so more pages are freed
612 * off fuller lists instead of spinning excessively around empty
613 * lists
614 */
615 do {
616 batch_free++;
617 if (++migratetype == MIGRATE_PCPTYPES)
618 migratetype = 0;
619 list = &pcp->lists[migratetype];
620 } while (list_empty(list));
621
622 /* This is the only non-empty list. Free them all. */
623 if (batch_free == MIGRATE_PCPTYPES)
624 batch_free = to_free;
625
626 do {
627 page = list_entry(list->prev, struct page, lru);
628 /* must delete as __free_one_page list manipulates */
629 list_del(&page->lru);
630 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
631 __free_one_page(page, zone, 0, page_private(page));
632 trace_mm_page_pcpu_drain(page, 0, page_private(page));
633 } while (--to_free && --batch_free && !list_empty(list));
634 }
635 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
636 spin_unlock(&zone->lock);
637}
638
639static void free_one_page(struct zone *zone, struct page *page, int order,
640 int migratetype)
641{
642 spin_lock(&zone->lock);
643 zone->all_unreclaimable = 0;
644 zone->pages_scanned = 0;
645
646 __free_one_page(page, zone, order, migratetype);
647 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
648 spin_unlock(&zone->lock);
649}
650
651static bool free_pages_prepare(struct page *page, unsigned int order)
652{
653 int i;
654 int bad = 0;
655
656 trace_mm_page_free_direct(page, order);
657 kmemcheck_free_shadow(page, order);
658
659 if (PageAnon(page))
660 page->mapping = NULL;
661 for (i = 0; i < (1 << order); i++)
662 bad += free_pages_check(page + i);
663 if (bad)
664 return false;
665
666 if (!PageHighMem(page)) {
667 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
668 debug_check_no_obj_freed(page_address(page),
669 PAGE_SIZE << order);
670 }
671 arch_free_page(page, order);
672 kernel_map_pages(page, 1 << order, 0);
673
674 return true;
675}
676
677static void __free_pages_ok(struct page *page, unsigned int order)
678{
679 unsigned long flags;
680 int wasMlocked = __TestClearPageMlocked(page);
681
682 if (!free_pages_prepare(page, order))
683 return;
684
685 local_irq_save(flags);
686 if (unlikely(wasMlocked))
687 free_page_mlock(page);
688 __count_vm_events(PGFREE, 1 << order);
689 free_one_page(page_zone(page), page, order,
690 get_pageblock_migratetype(page));
691 local_irq_restore(flags);
692}
693
694/*
695 * permit the bootmem allocator to evade page validation on high-order frees
696 */
697void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
698{
699 if (order == 0) {
700 __ClearPageReserved(page);
701 set_page_count(page, 0);
702 set_page_refcounted(page);
703 __free_page(page);
704 } else {
705 int loop;
706
707 prefetchw(page);
708 for (loop = 0; loop < BITS_PER_LONG; loop++) {
709 struct page *p = &page[loop];
710
711 if (loop + 1 < BITS_PER_LONG)
712 prefetchw(p + 1);
713 __ClearPageReserved(p);
714 set_page_count(p, 0);
715 }
716
717 set_page_refcounted(page);
718 __free_pages(page, order);
719 }
720}
721
722
723/*
724 * The order of subdivision here is critical for the IO subsystem.
725 * Please do not alter this order without good reasons and regression
726 * testing. Specifically, as large blocks of memory are subdivided,
727 * the order in which smaller blocks are delivered depends on the order
728 * they're subdivided in this function. This is the primary factor
729 * influencing the order in which pages are delivered to the IO
730 * subsystem according to empirical testing, and this is also justified
731 * by considering the behavior of a buddy system containing a single
732 * large block of memory acted on by a series of small allocations.
733 * This behavior is a critical factor in sglist merging's success.
734 *
735 * -- wli
736 */
737static inline void expand(struct zone *zone, struct page *page,
738 int low, int high, struct free_area *area,
739 int migratetype)
740{
741 unsigned long size = 1 << high;
742
743 while (high > low) {
744 area--;
745 high--;
746 size >>= 1;
747 VM_BUG_ON(bad_range(zone, &page[size]));
748 list_add(&page[size].lru, &area->free_list[migratetype]);
749 area->nr_free++;
750 set_page_order(&page[size], high);
751 }
752}
753
754/*
755 * This page is about to be returned from the page allocator
756 */
757static inline int check_new_page(struct page *page)
758{
759 if (unlikely(page_mapcount(page) |
760 (page->mapping != NULL) |
761 (atomic_read(&page->_count) != 0) |
762 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
763 (mem_cgroup_bad_page_check(page)))) {
764 bad_page(page);
765 return 1;
766 }
767 return 0;
768}
769
770static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
771{
772 int i;
773
774 for (i = 0; i < (1 << order); i++) {
775 struct page *p = page + i;
776 if (unlikely(check_new_page(p)))
777 return 1;
778 }
779
780 set_page_private(page, 0);
781 set_page_refcounted(page);
782
783 arch_alloc_page(page, order);
784 kernel_map_pages(page, 1 << order, 1);
785
786 if (gfp_flags & __GFP_ZERO)
787 prep_zero_page(page, order, gfp_flags);
788
789 if (order && (gfp_flags & __GFP_COMP))
790 prep_compound_page(page, order);
791
792 return 0;
793}
794
795/*
796 * Go through the free lists for the given migratetype and remove
797 * the smallest available page from the freelists
798 */
799static inline
800struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
801 int migratetype)
802{
803 unsigned int current_order;
804 struct free_area * area;
805 struct page *page;
806
807 /* Find a page of the appropriate size in the preferred list */
808 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
809 area = &(zone->free_area[current_order]);
810 if (list_empty(&area->free_list[migratetype]))
811 continue;
812
813 page = list_entry(area->free_list[migratetype].next,
814 struct page, lru);
815 list_del(&page->lru);
816 rmv_page_order(page);
817 area->nr_free--;
818 expand(zone, page, order, current_order, area, migratetype);
819 return page;
820 }
821
822 return NULL;
823}
824
825
826/*
827 * This array describes the order lists are fallen back to when
828 * the free lists for the desirable migrate type are depleted
829 */
830static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
831 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
832 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
833 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
834 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
835};
836
837/*
838 * Move the free pages in a range to the free lists of the requested type.
839 * Note that start_page and end_pages are not aligned on a pageblock
840 * boundary. If alignment is required, use move_freepages_block()
841 */
842static int move_freepages(struct zone *zone,
843 struct page *start_page, struct page *end_page,
844 int migratetype)
845{
846 struct page *page;
847 unsigned long order;
848 int pages_moved = 0;
849
850#ifndef CONFIG_HOLES_IN_ZONE
851 /*
852 * page_zone is not safe to call in this context when
853 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
854 * anyway as we check zone boundaries in move_freepages_block().
855 * Remove at a later date when no bug reports exist related to
856 * grouping pages by mobility
857 */
858 BUG_ON(page_zone(start_page) != page_zone(end_page));
859#endif
860
861 for (page = start_page; page <= end_page;) {
862 /* Make sure we are not inadvertently changing nodes */
863 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
864
865 if (!pfn_valid_within(page_to_pfn(page))) {
866 page++;
867 continue;
868 }
869
870 if (!PageBuddy(page)) {
871 page++;
872 continue;
873 }
874
875 order = page_order(page);
876 list_move(&page->lru,
877 &zone->free_area[order].free_list[migratetype]);
878 page += 1 << order;
879 pages_moved += 1 << order;
880 }
881
882 return pages_moved;
883}
884
885static int move_freepages_block(struct zone *zone, struct page *page,
886 int migratetype)
887{
888 unsigned long start_pfn, end_pfn;
889 struct page *start_page, *end_page;
890
891 start_pfn = page_to_pfn(page);
892 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
893 start_page = pfn_to_page(start_pfn);
894 end_page = start_page + pageblock_nr_pages - 1;
895 end_pfn = start_pfn + pageblock_nr_pages - 1;
896
897 /* Do not cross zone boundaries */
898 if (start_pfn < zone->zone_start_pfn)
899 start_page = page;
900 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
901 return 0;
902
903 return move_freepages(zone, start_page, end_page, migratetype);
904}
905
906static void change_pageblock_range(struct page *pageblock_page,
907 int start_order, int migratetype)
908{
909 int nr_pageblocks = 1 << (start_order - pageblock_order);
910
911 while (nr_pageblocks--) {
912 set_pageblock_migratetype(pageblock_page, migratetype);
913 pageblock_page += pageblock_nr_pages;
914 }
915}
916
917/* Remove an element from the buddy allocator from the fallback list */
918static inline struct page *
919__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
920{
921 struct free_area * area;
922 int current_order;
923 struct page *page;
924 int migratetype, i;
925
926 /* Find the largest possible block of pages in the other list */
927 for (current_order = MAX_ORDER-1; current_order >= order;
928 --current_order) {
929 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
930 migratetype = fallbacks[start_migratetype][i];
931
932 /* MIGRATE_RESERVE handled later if necessary */
933 if (migratetype == MIGRATE_RESERVE)
934 continue;
935
936 area = &(zone->free_area[current_order]);
937 if (list_empty(&area->free_list[migratetype]))
938 continue;
939
940 page = list_entry(area->free_list[migratetype].next,
941 struct page, lru);
942 area->nr_free--;
943
944 /*
945 * If breaking a large block of pages, move all free
946 * pages to the preferred allocation list. If falling
947 * back for a reclaimable kernel allocation, be more
948 * aggressive about taking ownership of free pages
949 */
950 if (unlikely(current_order >= (pageblock_order >> 1)) ||
951 start_migratetype == MIGRATE_RECLAIMABLE ||
952 page_group_by_mobility_disabled) {
953 unsigned long pages;
954 pages = move_freepages_block(zone, page,
955 start_migratetype);
956
957 /* Claim the whole block if over half of it is free */
958 if (pages >= (1 << (pageblock_order-1)) ||
959 page_group_by_mobility_disabled)
960 set_pageblock_migratetype(page,
961 start_migratetype);
962
963 migratetype = start_migratetype;
964 }
965
966 /* Remove the page from the freelists */
967 list_del(&page->lru);
968 rmv_page_order(page);
969
970 /* Take ownership for orders >= pageblock_order */
971 if (current_order >= pageblock_order)
972 change_pageblock_range(page, current_order,
973 start_migratetype);
974
975 expand(zone, page, order, current_order, area, migratetype);
976
977 trace_mm_page_alloc_extfrag(page, order, current_order,
978 start_migratetype, migratetype);
979
980 return page;
981 }
982 }
983
984 return NULL;
985}
986
987/*
988 * Do the hard work of removing an element from the buddy allocator.
989 * Call me with the zone->lock already held.
990 */
991static struct page *__rmqueue(struct zone *zone, unsigned int order,
992 int migratetype)
993{
994 struct page *page;
995
996retry_reserve:
997 page = __rmqueue_smallest(zone, order, migratetype);
998
999 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1000 page = __rmqueue_fallback(zone, order, migratetype);
1001
1002 /*
1003 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1004 * is used because __rmqueue_smallest is an inline function
1005 * and we want just one call site
1006 */
1007 if (!page) {
1008 migratetype = MIGRATE_RESERVE;
1009 goto retry_reserve;
1010 }
1011 }
1012
1013 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1014 return page;
1015}
1016
1017/*
1018 * Obtain a specified number of elements from the buddy allocator, all under
1019 * a single hold of the lock, for efficiency. Add them to the supplied list.
1020 * Returns the number of new pages which were placed at *list.
1021 */
1022static int rmqueue_bulk(struct zone *zone, unsigned int order,
1023 unsigned long count, struct list_head *list,
1024 int migratetype, int cold)
1025{
1026 int i;
1027
1028 spin_lock(&zone->lock);
1029 for (i = 0; i < count; ++i) {
1030 struct page *page = __rmqueue(zone, order, migratetype);
1031 if (unlikely(page == NULL))
1032 break;
1033
1034 /*
1035 * Split buddy pages returned by expand() are received here
1036 * in physical page order. The page is added to the callers and
1037 * list and the list head then moves forward. From the callers
1038 * perspective, the linked list is ordered by page number in
1039 * some conditions. This is useful for IO devices that can
1040 * merge IO requests if the physical pages are ordered
1041 * properly.
1042 */
1043 if (likely(cold == 0))
1044 list_add(&page->lru, list);
1045 else
1046 list_add_tail(&page->lru, list);
1047 set_page_private(page, migratetype);
1048 list = &page->lru;
1049 }
1050 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1051 spin_unlock(&zone->lock);
1052 return i;
1053}
1054
1055#ifdef CONFIG_NUMA
1056/*
1057 * Called from the vmstat counter updater to drain pagesets of this
1058 * currently executing processor on remote nodes after they have
1059 * expired.
1060 *
1061 * Note that this function must be called with the thread pinned to
1062 * a single processor.
1063 */
1064void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1065{
1066 unsigned long flags;
1067 int to_drain;
1068
1069 local_irq_save(flags);
1070 if (pcp->count >= pcp->batch)
1071 to_drain = pcp->batch;
1072 else
1073 to_drain = pcp->count;
1074 free_pcppages_bulk(zone, to_drain, pcp);
1075 pcp->count -= to_drain;
1076 local_irq_restore(flags);
1077}
1078#endif
1079
1080/*
1081 * Drain pages of the indicated processor.
1082 *
1083 * The processor must either be the current processor and the
1084 * thread pinned to the current processor or a processor that
1085 * is not online.
1086 */
1087static void drain_pages(unsigned int cpu)
1088{
1089 unsigned long flags;
1090 struct zone *zone;
1091
1092 for_each_populated_zone(zone) {
1093 struct per_cpu_pageset *pset;
1094 struct per_cpu_pages *pcp;
1095
1096 local_irq_save(flags);
1097 pset = per_cpu_ptr(zone->pageset, cpu);
1098
1099 pcp = &pset->pcp;
1100 if (pcp->count) {
1101 free_pcppages_bulk(zone, pcp->count, pcp);
1102 pcp->count = 0;
1103 }
1104 local_irq_restore(flags);
1105 }
1106}
1107
1108/*
1109 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1110 */
1111void drain_local_pages(void *arg)
1112{
1113 drain_pages(smp_processor_id());
1114}
1115
1116/*
1117 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1118 */
1119void drain_all_pages(void)
1120{
1121 on_each_cpu(drain_local_pages, NULL, 1);
1122}
1123
1124#ifdef CONFIG_HIBERNATION
1125
1126void mark_free_pages(struct zone *zone)
1127{
1128 unsigned long pfn, max_zone_pfn;
1129 unsigned long flags;
1130 int order, t;
1131 struct list_head *curr;
1132
1133 if (!zone->spanned_pages)
1134 return;
1135
1136 spin_lock_irqsave(&zone->lock, flags);
1137
1138 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1139 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1140 if (pfn_valid(pfn)) {
1141 struct page *page = pfn_to_page(pfn);
1142
1143 if (!swsusp_page_is_forbidden(page))
1144 swsusp_unset_page_free(page);
1145 }
1146
1147 for_each_migratetype_order(order, t) {
1148 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1149 unsigned long i;
1150
1151 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1152 for (i = 0; i < (1UL << order); i++)
1153 swsusp_set_page_free(pfn_to_page(pfn + i));
1154 }
1155 }
1156 spin_unlock_irqrestore(&zone->lock, flags);
1157}
1158#endif /* CONFIG_PM */
1159
1160/*
1161 * Free a 0-order page
1162 * cold == 1 ? free a cold page : free a hot page
1163 */
1164void free_hot_cold_page(struct page *page, int cold)
1165{
1166 struct zone *zone = page_zone(page);
1167 struct per_cpu_pages *pcp;
1168 unsigned long flags;
1169 int migratetype;
1170 int wasMlocked = __TestClearPageMlocked(page);
1171
1172 if (!free_pages_prepare(page, 0))
1173 return;
1174
1175 migratetype = get_pageblock_migratetype(page);
1176 set_page_private(page, migratetype);
1177 local_irq_save(flags);
1178 if (unlikely(wasMlocked))
1179 free_page_mlock(page);
1180 __count_vm_event(PGFREE);
1181
1182 /*
1183 * We only track unmovable, reclaimable and movable on pcp lists.
1184 * Free ISOLATE pages back to the allocator because they are being
1185 * offlined but treat RESERVE as movable pages so we can get those
1186 * areas back if necessary. Otherwise, we may have to free
1187 * excessively into the page allocator
1188 */
1189 if (migratetype >= MIGRATE_PCPTYPES) {
1190 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1191 free_one_page(zone, page, 0, migratetype);
1192 goto out;
1193 }
1194 migratetype = MIGRATE_MOVABLE;
1195 }
1196
1197 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1198 if (cold)
1199 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1200 else
1201 list_add(&page->lru, &pcp->lists[migratetype]);
1202 pcp->count++;
1203 if (pcp->count >= pcp->high) {
1204 free_pcppages_bulk(zone, pcp->batch, pcp);
1205 pcp->count -= pcp->batch;
1206 }
1207
1208out:
1209 local_irq_restore(flags);
1210}
1211
1212/*
1213 * split_page takes a non-compound higher-order page, and splits it into
1214 * n (1<<order) sub-pages: page[0..n]
1215 * Each sub-page must be freed individually.
1216 *
1217 * Note: this is probably too low level an operation for use in drivers.
1218 * Please consult with lkml before using this in your driver.
1219 */
1220void split_page(struct page *page, unsigned int order)
1221{
1222 int i;
1223
1224 VM_BUG_ON(PageCompound(page));
1225 VM_BUG_ON(!page_count(page));
1226
1227#ifdef CONFIG_KMEMCHECK
1228 /*
1229 * Split shadow pages too, because free(page[0]) would
1230 * otherwise free the whole shadow.
1231 */
1232 if (kmemcheck_page_is_tracked(page))
1233 split_page(virt_to_page(page[0].shadow), order);
1234#endif
1235
1236 for (i = 1; i < (1 << order); i++)
1237 set_page_refcounted(page + i);
1238}
1239
1240/*
1241 * Similar to split_page except the page is already free. As this is only
1242 * being used for migration, the migratetype of the block also changes.
1243 * As this is called with interrupts disabled, the caller is responsible
1244 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1245 * are enabled.
1246 *
1247 * Note: this is probably too low level an operation for use in drivers.
1248 * Please consult with lkml before using this in your driver.
1249 */
1250int split_free_page(struct page *page)
1251{
1252 unsigned int order;
1253 unsigned long watermark;
1254 struct zone *zone;
1255
1256 BUG_ON(!PageBuddy(page));
1257
1258 zone = page_zone(page);
1259 order = page_order(page);
1260
1261 /* Obey watermarks as if the page was being allocated */
1262 watermark = low_wmark_pages(zone) + (1 << order);
1263 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1264 return 0;
1265
1266 /* Remove page from free list */
1267 list_del(&page->lru);
1268 zone->free_area[order].nr_free--;
1269 rmv_page_order(page);
1270 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1271
1272 /* Split into individual pages */
1273 set_page_refcounted(page);
1274 split_page(page, order);
1275
1276 if (order >= pageblock_order - 1) {
1277 struct page *endpage = page + (1 << order) - 1;
1278 for (; page < endpage; page += pageblock_nr_pages)
1279 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1280 }
1281
1282 return 1 << order;
1283}
1284
1285/*
1286 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1287 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1288 * or two.
1289 */
1290static inline
1291struct page *buffered_rmqueue(struct zone *preferred_zone,
1292 struct zone *zone, int order, gfp_t gfp_flags,
1293 int migratetype)
1294{
1295 unsigned long flags;
1296 struct page *page;
1297 int cold = !!(gfp_flags & __GFP_COLD);
1298
1299again:
1300 if (likely(order == 0)) {
1301 struct per_cpu_pages *pcp;
1302 struct list_head *list;
1303
1304 local_irq_save(flags);
1305 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1306 list = &pcp->lists[migratetype];
1307 if (list_empty(list)) {
1308 pcp->count += rmqueue_bulk(zone, 0,
1309 pcp->batch, list,
1310 migratetype, cold);
1311 if (unlikely(list_empty(list)))
1312 goto failed;
1313 }
1314
1315 if (cold)
1316 page = list_entry(list->prev, struct page, lru);
1317 else
1318 page = list_entry(list->next, struct page, lru);
1319
1320 list_del(&page->lru);
1321 pcp->count--;
1322 } else {
1323 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1324 /*
1325 * __GFP_NOFAIL is not to be used in new code.
1326 *
1327 * All __GFP_NOFAIL callers should be fixed so that they
1328 * properly detect and handle allocation failures.
1329 *
1330 * We most definitely don't want callers attempting to
1331 * allocate greater than order-1 page units with
1332 * __GFP_NOFAIL.
1333 */
1334 WARN_ON_ONCE(order > 1);
1335 }
1336 spin_lock_irqsave(&zone->lock, flags);
1337 page = __rmqueue(zone, order, migratetype);
1338 spin_unlock(&zone->lock);
1339 if (!page)
1340 goto failed;
1341 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1342 }
1343
1344 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1345 zone_statistics(preferred_zone, zone, gfp_flags);
1346 local_irq_restore(flags);
1347
1348 VM_BUG_ON(bad_range(zone, page));
1349 if (prep_new_page(page, order, gfp_flags))
1350 goto again;
1351 return page;
1352
1353failed:
1354 local_irq_restore(flags);
1355 return NULL;
1356}
1357
1358/* The ALLOC_WMARK bits are used as an index to zone->watermark */
1359#define ALLOC_WMARK_MIN WMARK_MIN
1360#define ALLOC_WMARK_LOW WMARK_LOW
1361#define ALLOC_WMARK_HIGH WMARK_HIGH
1362#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1363
1364/* Mask to get the watermark bits */
1365#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1366
1367#define ALLOC_HARDER 0x10 /* try to alloc harder */
1368#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1369#define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1370
1371#ifdef CONFIG_FAIL_PAGE_ALLOC
1372
1373static struct {
1374 struct fault_attr attr;
1375
1376 u32 ignore_gfp_highmem;
1377 u32 ignore_gfp_wait;
1378 u32 min_order;
1379} fail_page_alloc = {
1380 .attr = FAULT_ATTR_INITIALIZER,
1381 .ignore_gfp_wait = 1,
1382 .ignore_gfp_highmem = 1,
1383 .min_order = 1,
1384};
1385
1386static int __init setup_fail_page_alloc(char *str)
1387{
1388 return setup_fault_attr(&fail_page_alloc.attr, str);
1389}
1390__setup("fail_page_alloc=", setup_fail_page_alloc);
1391
1392static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1393{
1394 if (order < fail_page_alloc.min_order)
1395 return 0;
1396 if (gfp_mask & __GFP_NOFAIL)
1397 return 0;
1398 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1399 return 0;
1400 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1401 return 0;
1402
1403 return should_fail(&fail_page_alloc.attr, 1 << order);
1404}
1405
1406#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1407
1408static int __init fail_page_alloc_debugfs(void)
1409{
1410 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1411 struct dentry *dir;
1412
1413 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1414 &fail_page_alloc.attr);
1415 if (IS_ERR(dir))
1416 return PTR_ERR(dir);
1417
1418 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1419 &fail_page_alloc.ignore_gfp_wait))
1420 goto fail;
1421 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1422 &fail_page_alloc.ignore_gfp_highmem))
1423 goto fail;
1424 if (!debugfs_create_u32("min-order", mode, dir,
1425 &fail_page_alloc.min_order))
1426 goto fail;
1427
1428 return 0;
1429fail:
1430 debugfs_remove_recursive(dir);
1431
1432 return -ENOMEM;
1433}
1434
1435late_initcall(fail_page_alloc_debugfs);
1436
1437#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1438
1439#else /* CONFIG_FAIL_PAGE_ALLOC */
1440
1441static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1442{
1443 return 0;
1444}
1445
1446#endif /* CONFIG_FAIL_PAGE_ALLOC */
1447
1448/*
1449 * Return true if free pages are above 'mark'. This takes into account the order
1450 * of the allocation.
1451 */
1452static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1453 int classzone_idx, int alloc_flags, long free_pages)
1454{
1455 /* free_pages my go negative - that's OK */
1456 long min = mark;
1457 int o;
1458
1459 free_pages -= (1 << order) + 1;
1460 if (alloc_flags & ALLOC_HIGH)
1461 min -= min / 2;
1462 if (alloc_flags & ALLOC_HARDER)
1463 min -= min / 4;
1464
1465 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1466 return false;
1467 for (o = 0; o < order; o++) {
1468 /* At the next order, this order's pages become unavailable */
1469 free_pages -= z->free_area[o].nr_free << o;
1470
1471 /* Require fewer higher order pages to be free */
1472 min >>= 1;
1473
1474 if (free_pages <= min)
1475 return false;
1476 }
1477 return true;
1478}
1479
1480bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1481 int classzone_idx, int alloc_flags)
1482{
1483 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1484 zone_page_state(z, NR_FREE_PAGES));
1485}
1486
1487bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1488 int classzone_idx, int alloc_flags)
1489{
1490 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1491
1492 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1493 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1494
1495 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1496 free_pages);
1497}
1498
1499#ifdef CONFIG_NUMA
1500/*
1501 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1502 * skip over zones that are not allowed by the cpuset, or that have
1503 * been recently (in last second) found to be nearly full. See further
1504 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1505 * that have to skip over a lot of full or unallowed zones.
1506 *
1507 * If the zonelist cache is present in the passed in zonelist, then
1508 * returns a pointer to the allowed node mask (either the current
1509 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1510 *
1511 * If the zonelist cache is not available for this zonelist, does
1512 * nothing and returns NULL.
1513 *
1514 * If the fullzones BITMAP in the zonelist cache is stale (more than
1515 * a second since last zap'd) then we zap it out (clear its bits.)
1516 *
1517 * We hold off even calling zlc_setup, until after we've checked the
1518 * first zone in the zonelist, on the theory that most allocations will
1519 * be satisfied from that first zone, so best to examine that zone as
1520 * quickly as we can.
1521 */
1522static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1523{
1524 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1525 nodemask_t *allowednodes; /* zonelist_cache approximation */
1526
1527 zlc = zonelist->zlcache_ptr;
1528 if (!zlc)
1529 return NULL;
1530
1531 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1532 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1533 zlc->last_full_zap = jiffies;
1534 }
1535
1536 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1537 &cpuset_current_mems_allowed :
1538 &node_states[N_HIGH_MEMORY];
1539 return allowednodes;
1540}
1541
1542/*
1543 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1544 * if it is worth looking at further for free memory:
1545 * 1) Check that the zone isn't thought to be full (doesn't have its
1546 * bit set in the zonelist_cache fullzones BITMAP).
1547 * 2) Check that the zones node (obtained from the zonelist_cache
1548 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1549 * Return true (non-zero) if zone is worth looking at further, or
1550 * else return false (zero) if it is not.
1551 *
1552 * This check -ignores- the distinction between various watermarks,
1553 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1554 * found to be full for any variation of these watermarks, it will
1555 * be considered full for up to one second by all requests, unless
1556 * we are so low on memory on all allowed nodes that we are forced
1557 * into the second scan of the zonelist.
1558 *
1559 * In the second scan we ignore this zonelist cache and exactly
1560 * apply the watermarks to all zones, even it is slower to do so.
1561 * We are low on memory in the second scan, and should leave no stone
1562 * unturned looking for a free page.
1563 */
1564static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1565 nodemask_t *allowednodes)
1566{
1567 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1568 int i; /* index of *z in zonelist zones */
1569 int n; /* node that zone *z is on */
1570
1571 zlc = zonelist->zlcache_ptr;
1572 if (!zlc)
1573 return 1;
1574
1575 i = z - zonelist->_zonerefs;
1576 n = zlc->z_to_n[i];
1577
1578 /* This zone is worth trying if it is allowed but not full */
1579 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1580}
1581
1582/*
1583 * Given 'z' scanning a zonelist, set the corresponding bit in
1584 * zlc->fullzones, so that subsequent attempts to allocate a page
1585 * from that zone don't waste time re-examining it.
1586 */
1587static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1588{
1589 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1590 int i; /* index of *z in zonelist zones */
1591
1592 zlc = zonelist->zlcache_ptr;
1593 if (!zlc)
1594 return;
1595
1596 i = z - zonelist->_zonerefs;
1597
1598 set_bit(i, zlc->fullzones);
1599}
1600
1601/*
1602 * clear all zones full, called after direct reclaim makes progress so that
1603 * a zone that was recently full is not skipped over for up to a second
1604 */
1605static void zlc_clear_zones_full(struct zonelist *zonelist)
1606{
1607 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1608
1609 zlc = zonelist->zlcache_ptr;
1610 if (!zlc)
1611 return;
1612
1613 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1614}
1615
1616#else /* CONFIG_NUMA */
1617
1618static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1619{
1620 return NULL;
1621}
1622
1623static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1624 nodemask_t *allowednodes)
1625{
1626 return 1;
1627}
1628
1629static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1630{
1631}
1632
1633static void zlc_clear_zones_full(struct zonelist *zonelist)
1634{
1635}
1636#endif /* CONFIG_NUMA */
1637
1638/*
1639 * get_page_from_freelist goes through the zonelist trying to allocate
1640 * a page.
1641 */
1642static struct page *
1643get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1644 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1645 struct zone *preferred_zone, int migratetype)
1646{
1647 struct zoneref *z;
1648 struct page *page = NULL;
1649 int classzone_idx;
1650 struct zone *zone;
1651 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1652 int zlc_active = 0; /* set if using zonelist_cache */
1653 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1654
1655 classzone_idx = zone_idx(preferred_zone);
1656zonelist_scan:
1657 /*
1658 * Scan zonelist, looking for a zone with enough free.
1659 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1660 */
1661 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1662 high_zoneidx, nodemask) {
1663 if (NUMA_BUILD && zlc_active &&
1664 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1665 continue;
1666 if ((alloc_flags & ALLOC_CPUSET) &&
1667 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1668 continue;
1669
1670 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1671 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1672 unsigned long mark;
1673 int ret;
1674
1675 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1676 if (zone_watermark_ok(zone, order, mark,
1677 classzone_idx, alloc_flags))
1678 goto try_this_zone;
1679
1680 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1681 /*
1682 * we do zlc_setup if there are multiple nodes
1683 * and before considering the first zone allowed
1684 * by the cpuset.
1685 */
1686 allowednodes = zlc_setup(zonelist, alloc_flags);
1687 zlc_active = 1;
1688 did_zlc_setup = 1;
1689 }
1690
1691 if (zone_reclaim_mode == 0)
1692 goto this_zone_full;
1693
1694 /*
1695 * As we may have just activated ZLC, check if the first
1696 * eligible zone has failed zone_reclaim recently.
1697 */
1698 if (NUMA_BUILD && zlc_active &&
1699 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1700 continue;
1701
1702 ret = zone_reclaim(zone, gfp_mask, order);
1703 switch (ret) {
1704 case ZONE_RECLAIM_NOSCAN:
1705 /* did not scan */
1706 continue;
1707 case ZONE_RECLAIM_FULL:
1708 /* scanned but unreclaimable */
1709 continue;
1710 default:
1711 /* did we reclaim enough */
1712 if (!zone_watermark_ok(zone, order, mark,
1713 classzone_idx, alloc_flags))
1714 goto this_zone_full;
1715 }
1716 }
1717
1718try_this_zone:
1719 page = buffered_rmqueue(preferred_zone, zone, order,
1720 gfp_mask, migratetype);
1721 if (page)
1722 break;
1723this_zone_full:
1724 if (NUMA_BUILD)
1725 zlc_mark_zone_full(zonelist, z);
1726 }
1727
1728 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1729 /* Disable zlc cache for second zonelist scan */
1730 zlc_active = 0;
1731 goto zonelist_scan;
1732 }
1733 return page;
1734}
1735
1736/*
1737 * Large machines with many possible nodes should not always dump per-node
1738 * meminfo in irq context.
1739 */
1740static inline bool should_suppress_show_mem(void)
1741{
1742 bool ret = false;
1743
1744#if NODES_SHIFT > 8
1745 ret = in_interrupt();
1746#endif
1747 return ret;
1748}
1749
1750static DEFINE_RATELIMIT_STATE(nopage_rs,
1751 DEFAULT_RATELIMIT_INTERVAL,
1752 DEFAULT_RATELIMIT_BURST);
1753
1754void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1755{
1756 va_list args;
1757 unsigned int filter = SHOW_MEM_FILTER_NODES;
1758
1759 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1760 return;
1761
1762 /*
1763 * This documents exceptions given to allocations in certain
1764 * contexts that are allowed to allocate outside current's set
1765 * of allowed nodes.
1766 */
1767 if (!(gfp_mask & __GFP_NOMEMALLOC))
1768 if (test_thread_flag(TIF_MEMDIE) ||
1769 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1770 filter &= ~SHOW_MEM_FILTER_NODES;
1771 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1772 filter &= ~SHOW_MEM_FILTER_NODES;
1773
1774 if (fmt) {
1775 printk(KERN_WARNING);
1776 va_start(args, fmt);
1777 vprintk(fmt, args);
1778 va_end(args);
1779 }
1780
1781 pr_warning("%s: page allocation failure: order:%d, mode:0x%x\n",
1782 current->comm, order, gfp_mask);
1783
1784 dump_stack();
1785 if (!should_suppress_show_mem())
1786 show_mem(filter);
1787}
1788
1789static inline int
1790should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1791 unsigned long pages_reclaimed)
1792{
1793 /* Do not loop if specifically requested */
1794 if (gfp_mask & __GFP_NORETRY)
1795 return 0;
1796
1797 /*
1798 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1799 * means __GFP_NOFAIL, but that may not be true in other
1800 * implementations.
1801 */
1802 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1803 return 1;
1804
1805 /*
1806 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1807 * specified, then we retry until we no longer reclaim any pages
1808 * (above), or we've reclaimed an order of pages at least as
1809 * large as the allocation's order. In both cases, if the
1810 * allocation still fails, we stop retrying.
1811 */
1812 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1813 return 1;
1814
1815 /*
1816 * Don't let big-order allocations loop unless the caller
1817 * explicitly requests that.
1818 */
1819 if (gfp_mask & __GFP_NOFAIL)
1820 return 1;
1821
1822 return 0;
1823}
1824
1825static inline struct page *
1826__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1827 struct zonelist *zonelist, enum zone_type high_zoneidx,
1828 nodemask_t *nodemask, struct zone *preferred_zone,
1829 int migratetype)
1830{
1831 struct page *page;
1832
1833 /* Acquire the OOM killer lock for the zones in zonelist */
1834 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1835 schedule_timeout_uninterruptible(1);
1836 return NULL;
1837 }
1838
1839 /*
1840 * Go through the zonelist yet one more time, keep very high watermark
1841 * here, this is only to catch a parallel oom killing, we must fail if
1842 * we're still under heavy pressure.
1843 */
1844 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1845 order, zonelist, high_zoneidx,
1846 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1847 preferred_zone, migratetype);
1848 if (page)
1849 goto out;
1850
1851 if (!(gfp_mask & __GFP_NOFAIL)) {
1852 /* The OOM killer will not help higher order allocs */
1853 if (order > PAGE_ALLOC_COSTLY_ORDER)
1854 goto out;
1855 /* The OOM killer does not needlessly kill tasks for lowmem */
1856 if (high_zoneidx < ZONE_NORMAL)
1857 goto out;
1858 /*
1859 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1860 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1861 * The caller should handle page allocation failure by itself if
1862 * it specifies __GFP_THISNODE.
1863 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1864 */
1865 if (gfp_mask & __GFP_THISNODE)
1866 goto out;
1867 }
1868 /* Exhausted what can be done so it's blamo time */
1869 out_of_memory(zonelist, gfp_mask, order, nodemask);
1870
1871out:
1872 clear_zonelist_oom(zonelist, gfp_mask);
1873 return page;
1874}
1875
1876#ifdef CONFIG_COMPACTION
1877/* Try memory compaction for high-order allocations before reclaim */
1878static struct page *
1879__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1880 struct zonelist *zonelist, enum zone_type high_zoneidx,
1881 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1882 int migratetype, unsigned long *did_some_progress,
1883 bool sync_migration)
1884{
1885 struct page *page;
1886
1887 if (!order || compaction_deferred(preferred_zone))
1888 return NULL;
1889
1890 current->flags |= PF_MEMALLOC;
1891 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1892 nodemask, sync_migration);
1893 current->flags &= ~PF_MEMALLOC;
1894 if (*did_some_progress != COMPACT_SKIPPED) {
1895
1896 /* Page migration frees to the PCP lists but we want merging */
1897 drain_pages(get_cpu());
1898 put_cpu();
1899
1900 page = get_page_from_freelist(gfp_mask, nodemask,
1901 order, zonelist, high_zoneidx,
1902 alloc_flags, preferred_zone,
1903 migratetype);
1904 if (page) {
1905 preferred_zone->compact_considered = 0;
1906 preferred_zone->compact_defer_shift = 0;
1907 count_vm_event(COMPACTSUCCESS);
1908 return page;
1909 }
1910
1911 /*
1912 * It's bad if compaction run occurs and fails.
1913 * The most likely reason is that pages exist,
1914 * but not enough to satisfy watermarks.
1915 */
1916 count_vm_event(COMPACTFAIL);
1917 defer_compaction(preferred_zone);
1918
1919 cond_resched();
1920 }
1921
1922 return NULL;
1923}
1924#else
1925static inline struct page *
1926__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1927 struct zonelist *zonelist, enum zone_type high_zoneidx,
1928 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1929 int migratetype, unsigned long *did_some_progress,
1930 bool sync_migration)
1931{
1932 return NULL;
1933}
1934#endif /* CONFIG_COMPACTION */
1935
1936/* The really slow allocator path where we enter direct reclaim */
1937static inline struct page *
1938__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1939 struct zonelist *zonelist, enum zone_type high_zoneidx,
1940 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1941 int migratetype, unsigned long *did_some_progress)
1942{
1943 struct page *page = NULL;
1944 struct reclaim_state reclaim_state;
1945 bool drained = false;
1946
1947 cond_resched();
1948
1949 /* We now go into synchronous reclaim */
1950 cpuset_memory_pressure_bump();
1951 current->flags |= PF_MEMALLOC;
1952 lockdep_set_current_reclaim_state(gfp_mask);
1953 reclaim_state.reclaimed_slab = 0;
1954 current->reclaim_state = &reclaim_state;
1955
1956 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1957
1958 current->reclaim_state = NULL;
1959 lockdep_clear_current_reclaim_state();
1960 current->flags &= ~PF_MEMALLOC;
1961
1962 cond_resched();
1963
1964 if (unlikely(!(*did_some_progress)))
1965 return NULL;
1966
1967 /* After successful reclaim, reconsider all zones for allocation */
1968 if (NUMA_BUILD)
1969 zlc_clear_zones_full(zonelist);
1970
1971retry:
1972 page = get_page_from_freelist(gfp_mask, nodemask, order,
1973 zonelist, high_zoneidx,
1974 alloc_flags, preferred_zone,
1975 migratetype);
1976
1977 /*
1978 * If an allocation failed after direct reclaim, it could be because
1979 * pages are pinned on the per-cpu lists. Drain them and try again
1980 */
1981 if (!page && !drained) {
1982 drain_all_pages();
1983 drained = true;
1984 goto retry;
1985 }
1986
1987 return page;
1988}
1989
1990/*
1991 * This is called in the allocator slow-path if the allocation request is of
1992 * sufficient urgency to ignore watermarks and take other desperate measures
1993 */
1994static inline struct page *
1995__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1996 struct zonelist *zonelist, enum zone_type high_zoneidx,
1997 nodemask_t *nodemask, struct zone *preferred_zone,
1998 int migratetype)
1999{
2000 struct page *page;
2001
2002 do {
2003 page = get_page_from_freelist(gfp_mask, nodemask, order,
2004 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2005 preferred_zone, migratetype);
2006
2007 if (!page && gfp_mask & __GFP_NOFAIL)
2008 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2009 } while (!page && (gfp_mask & __GFP_NOFAIL));
2010
2011 return page;
2012}
2013
2014static inline
2015void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2016 enum zone_type high_zoneidx,
2017 enum zone_type classzone_idx)
2018{
2019 struct zoneref *z;
2020 struct zone *zone;
2021
2022 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2023 wakeup_kswapd(zone, order, classzone_idx);
2024}
2025
2026static inline int
2027gfp_to_alloc_flags(gfp_t gfp_mask)
2028{
2029 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2030 const gfp_t wait = gfp_mask & __GFP_WAIT;
2031
2032 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2033 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2034
2035 /*
2036 * The caller may dip into page reserves a bit more if the caller
2037 * cannot run direct reclaim, or if the caller has realtime scheduling
2038 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2039 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2040 */
2041 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2042
2043 if (!wait) {
2044 /*
2045 * Not worth trying to allocate harder for
2046 * __GFP_NOMEMALLOC even if it can't schedule.
2047 */
2048 if (!(gfp_mask & __GFP_NOMEMALLOC))
2049 alloc_flags |= ALLOC_HARDER;
2050 /*
2051 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2052 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2053 */
2054 alloc_flags &= ~ALLOC_CPUSET;
2055 } else if (unlikely(rt_task(current)) && !in_interrupt())
2056 alloc_flags |= ALLOC_HARDER;
2057
2058 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2059 if (!in_interrupt() &&
2060 ((current->flags & PF_MEMALLOC) ||
2061 unlikely(test_thread_flag(TIF_MEMDIE))))
2062 alloc_flags |= ALLOC_NO_WATERMARKS;
2063 }
2064
2065 return alloc_flags;
2066}
2067
2068static inline struct page *
2069__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2070 struct zonelist *zonelist, enum zone_type high_zoneidx,
2071 nodemask_t *nodemask, struct zone *preferred_zone,
2072 int migratetype)
2073{
2074 const gfp_t wait = gfp_mask & __GFP_WAIT;
2075 struct page *page = NULL;
2076 int alloc_flags;
2077 unsigned long pages_reclaimed = 0;
2078 unsigned long did_some_progress;
2079 bool sync_migration = false;
2080
2081 /*
2082 * In the slowpath, we sanity check order to avoid ever trying to
2083 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2084 * be using allocators in order of preference for an area that is
2085 * too large.
2086 */
2087 if (order >= MAX_ORDER) {
2088 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2089 return NULL;
2090 }
2091
2092 /*
2093 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2094 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2095 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2096 * using a larger set of nodes after it has established that the
2097 * allowed per node queues are empty and that nodes are
2098 * over allocated.
2099 */
2100 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2101 goto nopage;
2102
2103restart:
2104 if (!(gfp_mask & __GFP_NO_KSWAPD))
2105 wake_all_kswapd(order, zonelist, high_zoneidx,
2106 zone_idx(preferred_zone));
2107
2108 /*
2109 * OK, we're below the kswapd watermark and have kicked background
2110 * reclaim. Now things get more complex, so set up alloc_flags according
2111 * to how we want to proceed.
2112 */
2113 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2114
2115 /*
2116 * Find the true preferred zone if the allocation is unconstrained by
2117 * cpusets.
2118 */
2119 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2120 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2121 &preferred_zone);
2122
2123rebalance:
2124 /* This is the last chance, in general, before the goto nopage. */
2125 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2126 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2127 preferred_zone, migratetype);
2128 if (page)
2129 goto got_pg;
2130
2131 /* Allocate without watermarks if the context allows */
2132 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2133 page = __alloc_pages_high_priority(gfp_mask, order,
2134 zonelist, high_zoneidx, nodemask,
2135 preferred_zone, migratetype);
2136 if (page)
2137 goto got_pg;
2138 }
2139
2140 /* Atomic allocations - we can't balance anything */
2141 if (!wait)
2142 goto nopage;
2143
2144 /* Avoid recursion of direct reclaim */
2145 if (current->flags & PF_MEMALLOC)
2146 goto nopage;
2147
2148 /* Avoid allocations with no watermarks from looping endlessly */
2149 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2150 goto nopage;
2151
2152 /*
2153 * Try direct compaction. The first pass is asynchronous. Subsequent
2154 * attempts after direct reclaim are synchronous
2155 */
2156 page = __alloc_pages_direct_compact(gfp_mask, order,
2157 zonelist, high_zoneidx,
2158 nodemask,
2159 alloc_flags, preferred_zone,
2160 migratetype, &did_some_progress,
2161 sync_migration);
2162 if (page)
2163 goto got_pg;
2164 sync_migration = true;
2165
2166 /* Try direct reclaim and then allocating */
2167 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2168 zonelist, high_zoneidx,
2169 nodemask,
2170 alloc_flags, preferred_zone,
2171 migratetype, &did_some_progress);
2172 if (page)
2173 goto got_pg;
2174
2175 /*
2176 * If we failed to make any progress reclaiming, then we are
2177 * running out of options and have to consider going OOM
2178 */
2179 if (!did_some_progress) {
2180 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2181 if (oom_killer_disabled)
2182 goto nopage;
2183 page = __alloc_pages_may_oom(gfp_mask, order,
2184 zonelist, high_zoneidx,
2185 nodemask, preferred_zone,
2186 migratetype);
2187 if (page)
2188 goto got_pg;
2189
2190 if (!(gfp_mask & __GFP_NOFAIL)) {
2191 /*
2192 * The oom killer is not called for high-order
2193 * allocations that may fail, so if no progress
2194 * is being made, there are no other options and
2195 * retrying is unlikely to help.
2196 */
2197 if (order > PAGE_ALLOC_COSTLY_ORDER)
2198 goto nopage;
2199 /*
2200 * The oom killer is not called for lowmem
2201 * allocations to prevent needlessly killing
2202 * innocent tasks.
2203 */
2204 if (high_zoneidx < ZONE_NORMAL)
2205 goto nopage;
2206 }
2207
2208 goto restart;
2209 }
2210 }
2211
2212 /* Check if we should retry the allocation */
2213 pages_reclaimed += did_some_progress;
2214 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2215 /* Wait for some write requests to complete then retry */
2216 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2217 goto rebalance;
2218 } else {
2219 /*
2220 * High-order allocations do not necessarily loop after
2221 * direct reclaim and reclaim/compaction depends on compaction
2222 * being called after reclaim so call directly if necessary
2223 */
2224 page = __alloc_pages_direct_compact(gfp_mask, order,
2225 zonelist, high_zoneidx,
2226 nodemask,
2227 alloc_flags, preferred_zone,
2228 migratetype, &did_some_progress,
2229 sync_migration);
2230 if (page)
2231 goto got_pg;
2232 }
2233
2234nopage:
2235 warn_alloc_failed(gfp_mask, order, NULL);
2236 return page;
2237got_pg:
2238 if (kmemcheck_enabled)
2239 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2240 return page;
2241
2242}
2243
2244/*
2245 * This is the 'heart' of the zoned buddy allocator.
2246 */
2247struct page *
2248__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2249 struct zonelist *zonelist, nodemask_t *nodemask)
2250{
2251 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2252 struct zone *preferred_zone;
2253 struct page *page;
2254 int migratetype = allocflags_to_migratetype(gfp_mask);
2255
2256 gfp_mask &= gfp_allowed_mask;
2257
2258 lockdep_trace_alloc(gfp_mask);
2259
2260 might_sleep_if(gfp_mask & __GFP_WAIT);
2261
2262 if (should_fail_alloc_page(gfp_mask, order))
2263 return NULL;
2264
2265 /*
2266 * Check the zones suitable for the gfp_mask contain at least one
2267 * valid zone. It's possible to have an empty zonelist as a result
2268 * of GFP_THISNODE and a memoryless node
2269 */
2270 if (unlikely(!zonelist->_zonerefs->zone))
2271 return NULL;
2272
2273 get_mems_allowed();
2274 /* The preferred zone is used for statistics later */
2275 first_zones_zonelist(zonelist, high_zoneidx,
2276 nodemask ? : &cpuset_current_mems_allowed,
2277 &preferred_zone);
2278 if (!preferred_zone) {
2279 put_mems_allowed();
2280 return NULL;
2281 }
2282
2283 /* First allocation attempt */
2284 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2285 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2286 preferred_zone, migratetype);
2287 if (unlikely(!page))
2288 page = __alloc_pages_slowpath(gfp_mask, order,
2289 zonelist, high_zoneidx, nodemask,
2290 preferred_zone, migratetype);
2291 put_mems_allowed();
2292
2293 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2294 return page;
2295}
2296EXPORT_SYMBOL(__alloc_pages_nodemask);
2297
2298/*
2299 * Common helper functions.
2300 */
2301unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2302{
2303 struct page *page;
2304
2305 /*
2306 * __get_free_pages() returns a 32-bit address, which cannot represent
2307 * a highmem page
2308 */
2309 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2310
2311 page = alloc_pages(gfp_mask, order);
2312 if (!page)
2313 return 0;
2314 return (unsigned long) page_address(page);
2315}
2316EXPORT_SYMBOL(__get_free_pages);
2317
2318unsigned long get_zeroed_page(gfp_t gfp_mask)
2319{
2320 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2321}
2322EXPORT_SYMBOL(get_zeroed_page);
2323
2324void __pagevec_free(struct pagevec *pvec)
2325{
2326 int i = pagevec_count(pvec);
2327
2328 while (--i >= 0) {
2329 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2330 free_hot_cold_page(pvec->pages[i], pvec->cold);
2331 }
2332}
2333
2334void __free_pages(struct page *page, unsigned int order)
2335{
2336 if (put_page_testzero(page)) {
2337 if (order == 0)
2338 free_hot_cold_page(page, 0);
2339 else
2340 __free_pages_ok(page, order);
2341 }
2342}
2343
2344EXPORT_SYMBOL(__free_pages);
2345
2346void free_pages(unsigned long addr, unsigned int order)
2347{
2348 if (addr != 0) {
2349 VM_BUG_ON(!virt_addr_valid((void *)addr));
2350 __free_pages(virt_to_page((void *)addr), order);
2351 }
2352}
2353
2354EXPORT_SYMBOL(free_pages);
2355
2356static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2357{
2358 if (addr) {
2359 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2360 unsigned long used = addr + PAGE_ALIGN(size);
2361
2362 split_page(virt_to_page((void *)addr), order);
2363 while (used < alloc_end) {
2364 free_page(used);
2365 used += PAGE_SIZE;
2366 }
2367 }
2368 return (void *)addr;
2369}
2370
2371/**
2372 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2373 * @size: the number of bytes to allocate
2374 * @gfp_mask: GFP flags for the allocation
2375 *
2376 * This function is similar to alloc_pages(), except that it allocates the
2377 * minimum number of pages to satisfy the request. alloc_pages() can only
2378 * allocate memory in power-of-two pages.
2379 *
2380 * This function is also limited by MAX_ORDER.
2381 *
2382 * Memory allocated by this function must be released by free_pages_exact().
2383 */
2384void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2385{
2386 unsigned int order = get_order(size);
2387 unsigned long addr;
2388
2389 addr = __get_free_pages(gfp_mask, order);
2390 return make_alloc_exact(addr, order, size);
2391}
2392EXPORT_SYMBOL(alloc_pages_exact);
2393
2394/**
2395 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2396 * pages on a node.
2397 * @nid: the preferred node ID where memory should be allocated
2398 * @size: the number of bytes to allocate
2399 * @gfp_mask: GFP flags for the allocation
2400 *
2401 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2402 * back.
2403 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2404 * but is not exact.
2405 */
2406void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2407{
2408 unsigned order = get_order(size);
2409 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2410 if (!p)
2411 return NULL;
2412 return make_alloc_exact((unsigned long)page_address(p), order, size);
2413}
2414EXPORT_SYMBOL(alloc_pages_exact_nid);
2415
2416/**
2417 * free_pages_exact - release memory allocated via alloc_pages_exact()
2418 * @virt: the value returned by alloc_pages_exact.
2419 * @size: size of allocation, same value as passed to alloc_pages_exact().
2420 *
2421 * Release the memory allocated by a previous call to alloc_pages_exact.
2422 */
2423void free_pages_exact(void *virt, size_t size)
2424{
2425 unsigned long addr = (unsigned long)virt;
2426 unsigned long end = addr + PAGE_ALIGN(size);
2427
2428 while (addr < end) {
2429 free_page(addr);
2430 addr += PAGE_SIZE;
2431 }
2432}
2433EXPORT_SYMBOL(free_pages_exact);
2434
2435static unsigned int nr_free_zone_pages(int offset)
2436{
2437 struct zoneref *z;
2438 struct zone *zone;
2439
2440 /* Just pick one node, since fallback list is circular */
2441 unsigned int sum = 0;
2442
2443 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2444
2445 for_each_zone_zonelist(zone, z, zonelist, offset) {
2446 unsigned long size = zone->present_pages;
2447 unsigned long high = high_wmark_pages(zone);
2448 if (size > high)
2449 sum += size - high;
2450 }
2451
2452 return sum;
2453}
2454
2455/*
2456 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2457 */
2458unsigned int nr_free_buffer_pages(void)
2459{
2460 return nr_free_zone_pages(gfp_zone(GFP_USER));
2461}
2462EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2463
2464/*
2465 * Amount of free RAM allocatable within all zones
2466 */
2467unsigned int nr_free_pagecache_pages(void)
2468{
2469 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2470}
2471
2472static inline void show_node(struct zone *zone)
2473{
2474 if (NUMA_BUILD)
2475 printk("Node %d ", zone_to_nid(zone));
2476}
2477
2478void si_meminfo(struct sysinfo *val)
2479{
2480 val->totalram = totalram_pages;
2481 val->sharedram = 0;
2482 val->freeram = global_page_state(NR_FREE_PAGES);
2483 val->bufferram = nr_blockdev_pages();
2484 val->totalhigh = totalhigh_pages;
2485 val->freehigh = nr_free_highpages();
2486 val->mem_unit = PAGE_SIZE;
2487}
2488
2489EXPORT_SYMBOL(si_meminfo);
2490
2491#ifdef CONFIG_NUMA
2492void si_meminfo_node(struct sysinfo *val, int nid)
2493{
2494 pg_data_t *pgdat = NODE_DATA(nid);
2495
2496 val->totalram = pgdat->node_present_pages;
2497 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2498#ifdef CONFIG_HIGHMEM
2499 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2500 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2501 NR_FREE_PAGES);
2502#else
2503 val->totalhigh = 0;
2504 val->freehigh = 0;
2505#endif
2506 val->mem_unit = PAGE_SIZE;
2507}
2508#endif
2509
2510/*
2511 * Determine whether the node should be displayed or not, depending on whether
2512 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2513 */
2514bool skip_free_areas_node(unsigned int flags, int nid)
2515{
2516 bool ret = false;
2517
2518 if (!(flags & SHOW_MEM_FILTER_NODES))
2519 goto out;
2520
2521 get_mems_allowed();
2522 ret = !node_isset(nid, cpuset_current_mems_allowed);
2523 put_mems_allowed();
2524out:
2525 return ret;
2526}
2527
2528#define K(x) ((x) << (PAGE_SHIFT-10))
2529
2530/*
2531 * Show free area list (used inside shift_scroll-lock stuff)
2532 * We also calculate the percentage fragmentation. We do this by counting the
2533 * memory on each free list with the exception of the first item on the list.
2534 * Suppresses nodes that are not allowed by current's cpuset if
2535 * SHOW_MEM_FILTER_NODES is passed.
2536 */
2537void show_free_areas(unsigned int filter)
2538{
2539 int cpu;
2540 struct zone *zone;
2541
2542 for_each_populated_zone(zone) {
2543 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2544 continue;
2545 show_node(zone);
2546 printk("%s per-cpu:\n", zone->name);
2547
2548 for_each_online_cpu(cpu) {
2549 struct per_cpu_pageset *pageset;
2550
2551 pageset = per_cpu_ptr(zone->pageset, cpu);
2552
2553 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2554 cpu, pageset->pcp.high,
2555 pageset->pcp.batch, pageset->pcp.count);
2556 }
2557 }
2558
2559 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2560 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2561 " unevictable:%lu"
2562 " dirty:%lu writeback:%lu unstable:%lu\n"
2563 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2564 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2565 global_page_state(NR_ACTIVE_ANON),
2566 global_page_state(NR_INACTIVE_ANON),
2567 global_page_state(NR_ISOLATED_ANON),
2568 global_page_state(NR_ACTIVE_FILE),
2569 global_page_state(NR_INACTIVE_FILE),
2570 global_page_state(NR_ISOLATED_FILE),
2571 global_page_state(NR_UNEVICTABLE),
2572 global_page_state(NR_FILE_DIRTY),
2573 global_page_state(NR_WRITEBACK),
2574 global_page_state(NR_UNSTABLE_NFS),
2575 global_page_state(NR_FREE_PAGES),
2576 global_page_state(NR_SLAB_RECLAIMABLE),
2577 global_page_state(NR_SLAB_UNRECLAIMABLE),
2578 global_page_state(NR_FILE_MAPPED),
2579 global_page_state(NR_SHMEM),
2580 global_page_state(NR_PAGETABLE),
2581 global_page_state(NR_BOUNCE));
2582
2583 for_each_populated_zone(zone) {
2584 int i;
2585
2586 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2587 continue;
2588 show_node(zone);
2589 printk("%s"
2590 " free:%lukB"
2591 " min:%lukB"
2592 " low:%lukB"
2593 " high:%lukB"
2594 " active_anon:%lukB"
2595 " inactive_anon:%lukB"
2596 " active_file:%lukB"
2597 " inactive_file:%lukB"
2598 " unevictable:%lukB"
2599 " isolated(anon):%lukB"
2600 " isolated(file):%lukB"
2601 " present:%lukB"
2602 " mlocked:%lukB"
2603 " dirty:%lukB"
2604 " writeback:%lukB"
2605 " mapped:%lukB"
2606 " shmem:%lukB"
2607 " slab_reclaimable:%lukB"
2608 " slab_unreclaimable:%lukB"
2609 " kernel_stack:%lukB"
2610 " pagetables:%lukB"
2611 " unstable:%lukB"
2612 " bounce:%lukB"
2613 " writeback_tmp:%lukB"
2614 " pages_scanned:%lu"
2615 " all_unreclaimable? %s"
2616 "\n",
2617 zone->name,
2618 K(zone_page_state(zone, NR_FREE_PAGES)),
2619 K(min_wmark_pages(zone)),
2620 K(low_wmark_pages(zone)),
2621 K(high_wmark_pages(zone)),
2622 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2623 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2624 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2625 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2626 K(zone_page_state(zone, NR_UNEVICTABLE)),
2627 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2628 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2629 K(zone->present_pages),
2630 K(zone_page_state(zone, NR_MLOCK)),
2631 K(zone_page_state(zone, NR_FILE_DIRTY)),
2632 K(zone_page_state(zone, NR_WRITEBACK)),
2633 K(zone_page_state(zone, NR_FILE_MAPPED)),
2634 K(zone_page_state(zone, NR_SHMEM)),
2635 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2636 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2637 zone_page_state(zone, NR_KERNEL_STACK) *
2638 THREAD_SIZE / 1024,
2639 K(zone_page_state(zone, NR_PAGETABLE)),
2640 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2641 K(zone_page_state(zone, NR_BOUNCE)),
2642 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2643 zone->pages_scanned,
2644 (zone->all_unreclaimable ? "yes" : "no")
2645 );
2646 printk("lowmem_reserve[]:");
2647 for (i = 0; i < MAX_NR_ZONES; i++)
2648 printk(" %lu", zone->lowmem_reserve[i]);
2649 printk("\n");
2650 }
2651
2652 for_each_populated_zone(zone) {
2653 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2654
2655 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2656 continue;
2657 show_node(zone);
2658 printk("%s: ", zone->name);
2659
2660 spin_lock_irqsave(&zone->lock, flags);
2661 for (order = 0; order < MAX_ORDER; order++) {
2662 nr[order] = zone->free_area[order].nr_free;
2663 total += nr[order] << order;
2664 }
2665 spin_unlock_irqrestore(&zone->lock, flags);
2666 for (order = 0; order < MAX_ORDER; order++)
2667 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2668 printk("= %lukB\n", K(total));
2669 }
2670
2671 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2672
2673 show_swap_cache_info();
2674}
2675
2676static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2677{
2678 zoneref->zone = zone;
2679 zoneref->zone_idx = zone_idx(zone);
2680}
2681
2682/*
2683 * Builds allocation fallback zone lists.
2684 *
2685 * Add all populated zones of a node to the zonelist.
2686 */
2687static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2688 int nr_zones, enum zone_type zone_type)
2689{
2690 struct zone *zone;
2691
2692 BUG_ON(zone_type >= MAX_NR_ZONES);
2693 zone_type++;
2694
2695 do {
2696 zone_type--;
2697 zone = pgdat->node_zones + zone_type;
2698 if (populated_zone(zone)) {
2699 zoneref_set_zone(zone,
2700 &zonelist->_zonerefs[nr_zones++]);
2701 check_highest_zone(zone_type);
2702 }
2703
2704 } while (zone_type);
2705 return nr_zones;
2706}
2707
2708
2709/*
2710 * zonelist_order:
2711 * 0 = automatic detection of better ordering.
2712 * 1 = order by ([node] distance, -zonetype)
2713 * 2 = order by (-zonetype, [node] distance)
2714 *
2715 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2716 * the same zonelist. So only NUMA can configure this param.
2717 */
2718#define ZONELIST_ORDER_DEFAULT 0
2719#define ZONELIST_ORDER_NODE 1
2720#define ZONELIST_ORDER_ZONE 2
2721
2722/* zonelist order in the kernel.
2723 * set_zonelist_order() will set this to NODE or ZONE.
2724 */
2725static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2726static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2727
2728
2729#ifdef CONFIG_NUMA
2730/* The value user specified ....changed by config */
2731static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2732/* string for sysctl */
2733#define NUMA_ZONELIST_ORDER_LEN 16
2734char numa_zonelist_order[16] = "default";
2735
2736/*
2737 * interface for configure zonelist ordering.
2738 * command line option "numa_zonelist_order"
2739 * = "[dD]efault - default, automatic configuration.
2740 * = "[nN]ode - order by node locality, then by zone within node
2741 * = "[zZ]one - order by zone, then by locality within zone
2742 */
2743
2744static int __parse_numa_zonelist_order(char *s)
2745{
2746 if (*s == 'd' || *s == 'D') {
2747 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2748 } else if (*s == 'n' || *s == 'N') {
2749 user_zonelist_order = ZONELIST_ORDER_NODE;
2750 } else if (*s == 'z' || *s == 'Z') {
2751 user_zonelist_order = ZONELIST_ORDER_ZONE;
2752 } else {
2753 printk(KERN_WARNING
2754 "Ignoring invalid numa_zonelist_order value: "
2755 "%s\n", s);
2756 return -EINVAL;
2757 }
2758 return 0;
2759}
2760
2761static __init int setup_numa_zonelist_order(char *s)
2762{
2763 int ret;
2764
2765 if (!s)
2766 return 0;
2767
2768 ret = __parse_numa_zonelist_order(s);
2769 if (ret == 0)
2770 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2771
2772 return ret;
2773}
2774early_param("numa_zonelist_order", setup_numa_zonelist_order);
2775
2776/*
2777 * sysctl handler for numa_zonelist_order
2778 */
2779int numa_zonelist_order_handler(ctl_table *table, int write,
2780 void __user *buffer, size_t *length,
2781 loff_t *ppos)
2782{
2783 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2784 int ret;
2785 static DEFINE_MUTEX(zl_order_mutex);
2786
2787 mutex_lock(&zl_order_mutex);
2788 if (write)
2789 strcpy(saved_string, (char*)table->data);
2790 ret = proc_dostring(table, write, buffer, length, ppos);
2791 if (ret)
2792 goto out;
2793 if (write) {
2794 int oldval = user_zonelist_order;
2795 if (__parse_numa_zonelist_order((char*)table->data)) {
2796 /*
2797 * bogus value. restore saved string
2798 */
2799 strncpy((char*)table->data, saved_string,
2800 NUMA_ZONELIST_ORDER_LEN);
2801 user_zonelist_order = oldval;
2802 } else if (oldval != user_zonelist_order) {
2803 mutex_lock(&zonelists_mutex);
2804 build_all_zonelists(NULL);
2805 mutex_unlock(&zonelists_mutex);
2806 }
2807 }
2808out:
2809 mutex_unlock(&zl_order_mutex);
2810 return ret;
2811}
2812
2813
2814#define MAX_NODE_LOAD (nr_online_nodes)
2815static int node_load[MAX_NUMNODES];
2816
2817/**
2818 * find_next_best_node - find the next node that should appear in a given node's fallback list
2819 * @node: node whose fallback list we're appending
2820 * @used_node_mask: nodemask_t of already used nodes
2821 *
2822 * We use a number of factors to determine which is the next node that should
2823 * appear on a given node's fallback list. The node should not have appeared
2824 * already in @node's fallback list, and it should be the next closest node
2825 * according to the distance array (which contains arbitrary distance values
2826 * from each node to each node in the system), and should also prefer nodes
2827 * with no CPUs, since presumably they'll have very little allocation pressure
2828 * on them otherwise.
2829 * It returns -1 if no node is found.
2830 */
2831static int find_next_best_node(int node, nodemask_t *used_node_mask)
2832{
2833 int n, val;
2834 int min_val = INT_MAX;
2835 int best_node = -1;
2836 const struct cpumask *tmp = cpumask_of_node(0);
2837
2838 /* Use the local node if we haven't already */
2839 if (!node_isset(node, *used_node_mask)) {
2840 node_set(node, *used_node_mask);
2841 return node;
2842 }
2843
2844 for_each_node_state(n, N_HIGH_MEMORY) {
2845
2846 /* Don't want a node to appear more than once */
2847 if (node_isset(n, *used_node_mask))
2848 continue;
2849
2850 /* Use the distance array to find the distance */
2851 val = node_distance(node, n);
2852
2853 /* Penalize nodes under us ("prefer the next node") */
2854 val += (n < node);
2855
2856 /* Give preference to headless and unused nodes */
2857 tmp = cpumask_of_node(n);
2858 if (!cpumask_empty(tmp))
2859 val += PENALTY_FOR_NODE_WITH_CPUS;
2860
2861 /* Slight preference for less loaded node */
2862 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2863 val += node_load[n];
2864
2865 if (val < min_val) {
2866 min_val = val;
2867 best_node = n;
2868 }
2869 }
2870
2871 if (best_node >= 0)
2872 node_set(best_node, *used_node_mask);
2873
2874 return best_node;
2875}
2876
2877
2878/*
2879 * Build zonelists ordered by node and zones within node.
2880 * This results in maximum locality--normal zone overflows into local
2881 * DMA zone, if any--but risks exhausting DMA zone.
2882 */
2883static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2884{
2885 int j;
2886 struct zonelist *zonelist;
2887
2888 zonelist = &pgdat->node_zonelists[0];
2889 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2890 ;
2891 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2892 MAX_NR_ZONES - 1);
2893 zonelist->_zonerefs[j].zone = NULL;
2894 zonelist->_zonerefs[j].zone_idx = 0;
2895}
2896
2897/*
2898 * Build gfp_thisnode zonelists
2899 */
2900static void build_thisnode_zonelists(pg_data_t *pgdat)
2901{
2902 int j;
2903 struct zonelist *zonelist;
2904
2905 zonelist = &pgdat->node_zonelists[1];
2906 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2907 zonelist->_zonerefs[j].zone = NULL;
2908 zonelist->_zonerefs[j].zone_idx = 0;
2909}
2910
2911/*
2912 * Build zonelists ordered by zone and nodes within zones.
2913 * This results in conserving DMA zone[s] until all Normal memory is
2914 * exhausted, but results in overflowing to remote node while memory
2915 * may still exist in local DMA zone.
2916 */
2917static int node_order[MAX_NUMNODES];
2918
2919static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2920{
2921 int pos, j, node;
2922 int zone_type; /* needs to be signed */
2923 struct zone *z;
2924 struct zonelist *zonelist;
2925
2926 zonelist = &pgdat->node_zonelists[0];
2927 pos = 0;
2928 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2929 for (j = 0; j < nr_nodes; j++) {
2930 node = node_order[j];
2931 z = &NODE_DATA(node)->node_zones[zone_type];
2932 if (populated_zone(z)) {
2933 zoneref_set_zone(z,
2934 &zonelist->_zonerefs[pos++]);
2935 check_highest_zone(zone_type);
2936 }
2937 }
2938 }
2939 zonelist->_zonerefs[pos].zone = NULL;
2940 zonelist->_zonerefs[pos].zone_idx = 0;
2941}
2942
2943static int default_zonelist_order(void)
2944{
2945 int nid, zone_type;
2946 unsigned long low_kmem_size,total_size;
2947 struct zone *z;
2948 int average_size;
2949 /*
2950 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2951 * If they are really small and used heavily, the system can fall
2952 * into OOM very easily.
2953 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2954 */
2955 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2956 low_kmem_size = 0;
2957 total_size = 0;
2958 for_each_online_node(nid) {
2959 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2960 z = &NODE_DATA(nid)->node_zones[zone_type];
2961 if (populated_zone(z)) {
2962 if (zone_type < ZONE_NORMAL)
2963 low_kmem_size += z->present_pages;
2964 total_size += z->present_pages;
2965 } else if (zone_type == ZONE_NORMAL) {
2966 /*
2967 * If any node has only lowmem, then node order
2968 * is preferred to allow kernel allocations
2969 * locally; otherwise, they can easily infringe
2970 * on other nodes when there is an abundance of
2971 * lowmem available to allocate from.
2972 */
2973 return ZONELIST_ORDER_NODE;
2974 }
2975 }
2976 }
2977 if (!low_kmem_size || /* there are no DMA area. */
2978 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2979 return ZONELIST_ORDER_NODE;
2980 /*
2981 * look into each node's config.
2982 * If there is a node whose DMA/DMA32 memory is very big area on
2983 * local memory, NODE_ORDER may be suitable.
2984 */
2985 average_size = total_size /
2986 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2987 for_each_online_node(nid) {
2988 low_kmem_size = 0;
2989 total_size = 0;
2990 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2991 z = &NODE_DATA(nid)->node_zones[zone_type];
2992 if (populated_zone(z)) {
2993 if (zone_type < ZONE_NORMAL)
2994 low_kmem_size += z->present_pages;
2995 total_size += z->present_pages;
2996 }
2997 }
2998 if (low_kmem_size &&
2999 total_size > average_size && /* ignore small node */
3000 low_kmem_size > total_size * 70/100)
3001 return ZONELIST_ORDER_NODE;
3002 }
3003 return ZONELIST_ORDER_ZONE;
3004}
3005
3006static void set_zonelist_order(void)
3007{
3008 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3009 current_zonelist_order = default_zonelist_order();
3010 else
3011 current_zonelist_order = user_zonelist_order;
3012}
3013
3014static void build_zonelists(pg_data_t *pgdat)
3015{
3016 int j, node, load;
3017 enum zone_type i;
3018 nodemask_t used_mask;
3019 int local_node, prev_node;
3020 struct zonelist *zonelist;
3021 int order = current_zonelist_order;
3022
3023 /* initialize zonelists */
3024 for (i = 0; i < MAX_ZONELISTS; i++) {
3025 zonelist = pgdat->node_zonelists + i;
3026 zonelist->_zonerefs[0].zone = NULL;
3027 zonelist->_zonerefs[0].zone_idx = 0;
3028 }
3029
3030 /* NUMA-aware ordering of nodes */
3031 local_node = pgdat->node_id;
3032 load = nr_online_nodes;
3033 prev_node = local_node;
3034 nodes_clear(used_mask);
3035
3036 memset(node_order, 0, sizeof(node_order));
3037 j = 0;
3038
3039 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3040 int distance = node_distance(local_node, node);
3041
3042 /*
3043 * If another node is sufficiently far away then it is better
3044 * to reclaim pages in a zone before going off node.
3045 */
3046 if (distance > RECLAIM_DISTANCE)
3047 zone_reclaim_mode = 1;
3048
3049 /*
3050 * We don't want to pressure a particular node.
3051 * So adding penalty to the first node in same
3052 * distance group to make it round-robin.
3053 */
3054 if (distance != node_distance(local_node, prev_node))
3055 node_load[node] = load;
3056
3057 prev_node = node;
3058 load--;
3059 if (order == ZONELIST_ORDER_NODE)
3060 build_zonelists_in_node_order(pgdat, node);
3061 else
3062 node_order[j++] = node; /* remember order */
3063 }
3064
3065 if (order == ZONELIST_ORDER_ZONE) {
3066 /* calculate node order -- i.e., DMA last! */
3067 build_zonelists_in_zone_order(pgdat, j);
3068 }
3069
3070 build_thisnode_zonelists(pgdat);
3071}
3072
3073/* Construct the zonelist performance cache - see further mmzone.h */
3074static void build_zonelist_cache(pg_data_t *pgdat)
3075{
3076 struct zonelist *zonelist;
3077 struct zonelist_cache *zlc;
3078 struct zoneref *z;
3079
3080 zonelist = &pgdat->node_zonelists[0];
3081 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3082 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3083 for (z = zonelist->_zonerefs; z->zone; z++)
3084 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3085}
3086
3087#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3088/*
3089 * Return node id of node used for "local" allocations.
3090 * I.e., first node id of first zone in arg node's generic zonelist.
3091 * Used for initializing percpu 'numa_mem', which is used primarily
3092 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3093 */
3094int local_memory_node(int node)
3095{
3096 struct zone *zone;
3097
3098 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3099 gfp_zone(GFP_KERNEL),
3100 NULL,
3101 &zone);
3102 return zone->node;
3103}
3104#endif
3105
3106#else /* CONFIG_NUMA */
3107
3108static void set_zonelist_order(void)
3109{
3110 current_zonelist_order = ZONELIST_ORDER_ZONE;
3111}
3112
3113static void build_zonelists(pg_data_t *pgdat)
3114{
3115 int node, local_node;
3116 enum zone_type j;
3117 struct zonelist *zonelist;
3118
3119 local_node = pgdat->node_id;
3120
3121 zonelist = &pgdat->node_zonelists[0];
3122 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3123
3124 /*
3125 * Now we build the zonelist so that it contains the zones
3126 * of all the other nodes.
3127 * We don't want to pressure a particular node, so when
3128 * building the zones for node N, we make sure that the
3129 * zones coming right after the local ones are those from
3130 * node N+1 (modulo N)
3131 */
3132 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3133 if (!node_online(node))
3134 continue;
3135 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3136 MAX_NR_ZONES - 1);
3137 }
3138 for (node = 0; node < local_node; node++) {
3139 if (!node_online(node))
3140 continue;
3141 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3142 MAX_NR_ZONES - 1);
3143 }
3144
3145 zonelist->_zonerefs[j].zone = NULL;
3146 zonelist->_zonerefs[j].zone_idx = 0;
3147}
3148
3149/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3150static void build_zonelist_cache(pg_data_t *pgdat)
3151{
3152 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3153}
3154
3155#endif /* CONFIG_NUMA */
3156
3157/*
3158 * Boot pageset table. One per cpu which is going to be used for all
3159 * zones and all nodes. The parameters will be set in such a way
3160 * that an item put on a list will immediately be handed over to
3161 * the buddy list. This is safe since pageset manipulation is done
3162 * with interrupts disabled.
3163 *
3164 * The boot_pagesets must be kept even after bootup is complete for
3165 * unused processors and/or zones. They do play a role for bootstrapping
3166 * hotplugged processors.
3167 *
3168 * zoneinfo_show() and maybe other functions do
3169 * not check if the processor is online before following the pageset pointer.
3170 * Other parts of the kernel may not check if the zone is available.
3171 */
3172static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3173static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3174static void setup_zone_pageset(struct zone *zone);
3175
3176/*
3177 * Global mutex to protect against size modification of zonelists
3178 * as well as to serialize pageset setup for the new populated zone.
3179 */
3180DEFINE_MUTEX(zonelists_mutex);
3181
3182/* return values int ....just for stop_machine() */
3183static __init_refok int __build_all_zonelists(void *data)
3184{
3185 int nid;
3186 int cpu;
3187
3188#ifdef CONFIG_NUMA
3189 memset(node_load, 0, sizeof(node_load));
3190#endif
3191 for_each_online_node(nid) {
3192 pg_data_t *pgdat = NODE_DATA(nid);
3193
3194 build_zonelists(pgdat);
3195 build_zonelist_cache(pgdat);
3196 }
3197
3198 /*
3199 * Initialize the boot_pagesets that are going to be used
3200 * for bootstrapping processors. The real pagesets for
3201 * each zone will be allocated later when the per cpu
3202 * allocator is available.
3203 *
3204 * boot_pagesets are used also for bootstrapping offline
3205 * cpus if the system is already booted because the pagesets
3206 * are needed to initialize allocators on a specific cpu too.
3207 * F.e. the percpu allocator needs the page allocator which
3208 * needs the percpu allocator in order to allocate its pagesets
3209 * (a chicken-egg dilemma).
3210 */
3211 for_each_possible_cpu(cpu) {
3212 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3213
3214#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3215 /*
3216 * We now know the "local memory node" for each node--
3217 * i.e., the node of the first zone in the generic zonelist.
3218 * Set up numa_mem percpu variable for on-line cpus. During
3219 * boot, only the boot cpu should be on-line; we'll init the
3220 * secondary cpus' numa_mem as they come on-line. During
3221 * node/memory hotplug, we'll fixup all on-line cpus.
3222 */
3223 if (cpu_online(cpu))
3224 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3225#endif
3226 }
3227
3228 return 0;
3229}
3230
3231/*
3232 * Called with zonelists_mutex held always
3233 * unless system_state == SYSTEM_BOOTING.
3234 */
3235void __ref build_all_zonelists(void *data)
3236{
3237 set_zonelist_order();
3238
3239 if (system_state == SYSTEM_BOOTING) {
3240 __build_all_zonelists(NULL);
3241 mminit_verify_zonelist();
3242 cpuset_init_current_mems_allowed();
3243 } else {
3244 /* we have to stop all cpus to guarantee there is no user
3245 of zonelist */
3246#ifdef CONFIG_MEMORY_HOTPLUG
3247 if (data)
3248 setup_zone_pageset((struct zone *)data);
3249#endif
3250 stop_machine(__build_all_zonelists, NULL, NULL);
3251 /* cpuset refresh routine should be here */
3252 }
3253 vm_total_pages = nr_free_pagecache_pages();
3254 /*
3255 * Disable grouping by mobility if the number of pages in the
3256 * system is too low to allow the mechanism to work. It would be
3257 * more accurate, but expensive to check per-zone. This check is
3258 * made on memory-hotadd so a system can start with mobility
3259 * disabled and enable it later
3260 */
3261 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3262 page_group_by_mobility_disabled = 1;
3263 else
3264 page_group_by_mobility_disabled = 0;
3265
3266 printk("Built %i zonelists in %s order, mobility grouping %s. "
3267 "Total pages: %ld\n",
3268 nr_online_nodes,
3269 zonelist_order_name[current_zonelist_order],
3270 page_group_by_mobility_disabled ? "off" : "on",
3271 vm_total_pages);
3272#ifdef CONFIG_NUMA
3273 printk("Policy zone: %s\n", zone_names[policy_zone]);
3274#endif
3275}
3276
3277/*
3278 * Helper functions to size the waitqueue hash table.
3279 * Essentially these want to choose hash table sizes sufficiently
3280 * large so that collisions trying to wait on pages are rare.
3281 * But in fact, the number of active page waitqueues on typical
3282 * systems is ridiculously low, less than 200. So this is even
3283 * conservative, even though it seems large.
3284 *
3285 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3286 * waitqueues, i.e. the size of the waitq table given the number of pages.
3287 */
3288#define PAGES_PER_WAITQUEUE 256
3289
3290#ifndef CONFIG_MEMORY_HOTPLUG
3291static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3292{
3293 unsigned long size = 1;
3294
3295 pages /= PAGES_PER_WAITQUEUE;
3296
3297 while (size < pages)
3298 size <<= 1;
3299
3300 /*
3301 * Once we have dozens or even hundreds of threads sleeping
3302 * on IO we've got bigger problems than wait queue collision.
3303 * Limit the size of the wait table to a reasonable size.
3304 */
3305 size = min(size, 4096UL);
3306
3307 return max(size, 4UL);
3308}
3309#else
3310/*
3311 * A zone's size might be changed by hot-add, so it is not possible to determine
3312 * a suitable size for its wait_table. So we use the maximum size now.
3313 *
3314 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3315 *
3316 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3317 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3318 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3319 *
3320 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3321 * or more by the traditional way. (See above). It equals:
3322 *
3323 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3324 * ia64(16K page size) : = ( 8G + 4M)byte.
3325 * powerpc (64K page size) : = (32G +16M)byte.
3326 */
3327static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3328{
3329 return 4096UL;
3330}
3331#endif
3332
3333/*
3334 * This is an integer logarithm so that shifts can be used later
3335 * to extract the more random high bits from the multiplicative
3336 * hash function before the remainder is taken.
3337 */
3338static inline unsigned long wait_table_bits(unsigned long size)
3339{
3340 return ffz(~size);
3341}
3342
3343#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3344
3345/*
3346 * Check if a pageblock contains reserved pages
3347 */
3348static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3349{
3350 unsigned long pfn;
3351
3352 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3353 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3354 return 1;
3355 }
3356 return 0;
3357}
3358
3359/*
3360 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3361 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3362 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3363 * higher will lead to a bigger reserve which will get freed as contiguous
3364 * blocks as reclaim kicks in
3365 */
3366static void setup_zone_migrate_reserve(struct zone *zone)
3367{
3368 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3369 struct page *page;
3370 unsigned long block_migratetype;
3371 int reserve;
3372
3373 /* Get the start pfn, end pfn and the number of blocks to reserve */
3374 start_pfn = zone->zone_start_pfn;
3375 end_pfn = start_pfn + zone->spanned_pages;
3376 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3377 pageblock_order;
3378
3379 /*
3380 * Reserve blocks are generally in place to help high-order atomic
3381 * allocations that are short-lived. A min_free_kbytes value that
3382 * would result in more than 2 reserve blocks for atomic allocations
3383 * is assumed to be in place to help anti-fragmentation for the
3384 * future allocation of hugepages at runtime.
3385 */
3386 reserve = min(2, reserve);
3387
3388 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3389 if (!pfn_valid(pfn))
3390 continue;
3391 page = pfn_to_page(pfn);
3392
3393 /* Watch out for overlapping nodes */
3394 if (page_to_nid(page) != zone_to_nid(zone))
3395 continue;
3396
3397 /* Blocks with reserved pages will never free, skip them. */
3398 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3399 if (pageblock_is_reserved(pfn, block_end_pfn))
3400 continue;
3401
3402 block_migratetype = get_pageblock_migratetype(page);
3403
3404 /* If this block is reserved, account for it */
3405 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3406 reserve--;
3407 continue;
3408 }
3409
3410 /* Suitable for reserving if this block is movable */
3411 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3412 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3413 move_freepages_block(zone, page, MIGRATE_RESERVE);
3414 reserve--;
3415 continue;
3416 }
3417
3418 /*
3419 * If the reserve is met and this is a previous reserved block,
3420 * take it back
3421 */
3422 if (block_migratetype == MIGRATE_RESERVE) {
3423 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3424 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3425 }
3426 }
3427}
3428
3429/*
3430 * Initially all pages are reserved - free ones are freed
3431 * up by free_all_bootmem() once the early boot process is
3432 * done. Non-atomic initialization, single-pass.
3433 */
3434void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3435 unsigned long start_pfn, enum memmap_context context)
3436{
3437 struct page *page;
3438 unsigned long end_pfn = start_pfn + size;
3439 unsigned long pfn;
3440 struct zone *z;
3441
3442 if (highest_memmap_pfn < end_pfn - 1)
3443 highest_memmap_pfn = end_pfn - 1;
3444
3445 z = &NODE_DATA(nid)->node_zones[zone];
3446 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3447 /*
3448 * There can be holes in boot-time mem_map[]s
3449 * handed to this function. They do not
3450 * exist on hotplugged memory.
3451 */
3452 if (context == MEMMAP_EARLY) {
3453 if (!early_pfn_valid(pfn))
3454 continue;
3455 if (!early_pfn_in_nid(pfn, nid))
3456 continue;
3457 }
3458 page = pfn_to_page(pfn);
3459 set_page_links(page, zone, nid, pfn);
3460 mminit_verify_page_links(page, zone, nid, pfn);
3461 init_page_count(page);
3462 reset_page_mapcount(page);
3463 SetPageReserved(page);
3464 /*
3465 * Mark the block movable so that blocks are reserved for
3466 * movable at startup. This will force kernel allocations
3467 * to reserve their blocks rather than leaking throughout
3468 * the address space during boot when many long-lived
3469 * kernel allocations are made. Later some blocks near
3470 * the start are marked MIGRATE_RESERVE by
3471 * setup_zone_migrate_reserve()
3472 *
3473 * bitmap is created for zone's valid pfn range. but memmap
3474 * can be created for invalid pages (for alignment)
3475 * check here not to call set_pageblock_migratetype() against
3476 * pfn out of zone.
3477 */
3478 if ((z->zone_start_pfn <= pfn)
3479 && (pfn < z->zone_start_pfn + z->spanned_pages)
3480 && !(pfn & (pageblock_nr_pages - 1)))
3481 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3482
3483 INIT_LIST_HEAD(&page->lru);
3484#ifdef WANT_PAGE_VIRTUAL
3485 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3486 if (!is_highmem_idx(zone))
3487 set_page_address(page, __va(pfn << PAGE_SHIFT));
3488#endif
3489 }
3490}
3491
3492static void __meminit zone_init_free_lists(struct zone *zone)
3493{
3494 int order, t;
3495 for_each_migratetype_order(order, t) {
3496 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3497 zone->free_area[order].nr_free = 0;
3498 }
3499}
3500
3501#ifndef __HAVE_ARCH_MEMMAP_INIT
3502#define memmap_init(size, nid, zone, start_pfn) \
3503 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3504#endif
3505
3506static int zone_batchsize(struct zone *zone)
3507{
3508#ifdef CONFIG_MMU
3509 int batch;
3510
3511 /*
3512 * The per-cpu-pages pools are set to around 1000th of the
3513 * size of the zone. But no more than 1/2 of a meg.
3514 *
3515 * OK, so we don't know how big the cache is. So guess.
3516 */
3517 batch = zone->present_pages / 1024;
3518 if (batch * PAGE_SIZE > 512 * 1024)
3519 batch = (512 * 1024) / PAGE_SIZE;
3520 batch /= 4; /* We effectively *= 4 below */
3521 if (batch < 1)
3522 batch = 1;
3523
3524 /*
3525 * Clamp the batch to a 2^n - 1 value. Having a power
3526 * of 2 value was found to be more likely to have
3527 * suboptimal cache aliasing properties in some cases.
3528 *
3529 * For example if 2 tasks are alternately allocating
3530 * batches of pages, one task can end up with a lot
3531 * of pages of one half of the possible page colors
3532 * and the other with pages of the other colors.
3533 */
3534 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3535
3536 return batch;
3537
3538#else
3539 /* The deferral and batching of frees should be suppressed under NOMMU
3540 * conditions.
3541 *
3542 * The problem is that NOMMU needs to be able to allocate large chunks
3543 * of contiguous memory as there's no hardware page translation to
3544 * assemble apparent contiguous memory from discontiguous pages.
3545 *
3546 * Queueing large contiguous runs of pages for batching, however,
3547 * causes the pages to actually be freed in smaller chunks. As there
3548 * can be a significant delay between the individual batches being
3549 * recycled, this leads to the once large chunks of space being
3550 * fragmented and becoming unavailable for high-order allocations.
3551 */
3552 return 0;
3553#endif
3554}
3555
3556static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3557{
3558 struct per_cpu_pages *pcp;
3559 int migratetype;
3560
3561 memset(p, 0, sizeof(*p));
3562
3563 pcp = &p->pcp;
3564 pcp->count = 0;
3565 pcp->high = 6 * batch;
3566 pcp->batch = max(1UL, 1 * batch);
3567 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3568 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3569}
3570
3571/*
3572 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3573 * to the value high for the pageset p.
3574 */
3575
3576static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3577 unsigned long high)
3578{
3579 struct per_cpu_pages *pcp;
3580
3581 pcp = &p->pcp;
3582 pcp->high = high;
3583 pcp->batch = max(1UL, high/4);
3584 if ((high/4) > (PAGE_SHIFT * 8))
3585 pcp->batch = PAGE_SHIFT * 8;
3586}
3587
3588static void setup_zone_pageset(struct zone *zone)
3589{
3590 int cpu;
3591
3592 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3593
3594 for_each_possible_cpu(cpu) {
3595 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3596
3597 setup_pageset(pcp, zone_batchsize(zone));
3598
3599 if (percpu_pagelist_fraction)
3600 setup_pagelist_highmark(pcp,
3601 (zone->present_pages /
3602 percpu_pagelist_fraction));
3603 }
3604}
3605
3606/*
3607 * Allocate per cpu pagesets and initialize them.
3608 * Before this call only boot pagesets were available.
3609 */
3610void __init setup_per_cpu_pageset(void)
3611{
3612 struct zone *zone;
3613
3614 for_each_populated_zone(zone)
3615 setup_zone_pageset(zone);
3616}
3617
3618static noinline __init_refok
3619int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3620{
3621 int i;
3622 struct pglist_data *pgdat = zone->zone_pgdat;
3623 size_t alloc_size;
3624
3625 /*
3626 * The per-page waitqueue mechanism uses hashed waitqueues
3627 * per zone.
3628 */
3629 zone->wait_table_hash_nr_entries =
3630 wait_table_hash_nr_entries(zone_size_pages);
3631 zone->wait_table_bits =
3632 wait_table_bits(zone->wait_table_hash_nr_entries);
3633 alloc_size = zone->wait_table_hash_nr_entries
3634 * sizeof(wait_queue_head_t);
3635
3636 if (!slab_is_available()) {
3637 zone->wait_table = (wait_queue_head_t *)
3638 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3639 } else {
3640 /*
3641 * This case means that a zone whose size was 0 gets new memory
3642 * via memory hot-add.
3643 * But it may be the case that a new node was hot-added. In
3644 * this case vmalloc() will not be able to use this new node's
3645 * memory - this wait_table must be initialized to use this new
3646 * node itself as well.
3647 * To use this new node's memory, further consideration will be
3648 * necessary.
3649 */
3650 zone->wait_table = vmalloc(alloc_size);
3651 }
3652 if (!zone->wait_table)
3653 return -ENOMEM;
3654
3655 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3656 init_waitqueue_head(zone->wait_table + i);
3657
3658 return 0;
3659}
3660
3661static int __zone_pcp_update(void *data)
3662{
3663 struct zone *zone = data;
3664 int cpu;
3665 unsigned long batch = zone_batchsize(zone), flags;
3666
3667 for_each_possible_cpu(cpu) {
3668 struct per_cpu_pageset *pset;
3669 struct per_cpu_pages *pcp;
3670
3671 pset = per_cpu_ptr(zone->pageset, cpu);
3672 pcp = &pset->pcp;
3673
3674 local_irq_save(flags);
3675 free_pcppages_bulk(zone, pcp->count, pcp);
3676 setup_pageset(pset, batch);
3677 local_irq_restore(flags);
3678 }
3679 return 0;
3680}
3681
3682void zone_pcp_update(struct zone *zone)
3683{
3684 stop_machine(__zone_pcp_update, zone, NULL);
3685}
3686
3687static __meminit void zone_pcp_init(struct zone *zone)
3688{
3689 /*
3690 * per cpu subsystem is not up at this point. The following code
3691 * relies on the ability of the linker to provide the
3692 * offset of a (static) per cpu variable into the per cpu area.
3693 */
3694 zone->pageset = &boot_pageset;
3695
3696 if (zone->present_pages)
3697 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3698 zone->name, zone->present_pages,
3699 zone_batchsize(zone));
3700}
3701
3702__meminit int init_currently_empty_zone(struct zone *zone,
3703 unsigned long zone_start_pfn,
3704 unsigned long size,
3705 enum memmap_context context)
3706{
3707 struct pglist_data *pgdat = zone->zone_pgdat;
3708 int ret;
3709 ret = zone_wait_table_init(zone, size);
3710 if (ret)
3711 return ret;
3712 pgdat->nr_zones = zone_idx(zone) + 1;
3713
3714 zone->zone_start_pfn = zone_start_pfn;
3715
3716 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3717 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3718 pgdat->node_id,
3719 (unsigned long)zone_idx(zone),
3720 zone_start_pfn, (zone_start_pfn + size));
3721
3722 zone_init_free_lists(zone);
3723
3724 return 0;
3725}
3726
3727#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3728/*
3729 * Basic iterator support. Return the first range of PFNs for a node
3730 * Note: nid == MAX_NUMNODES returns first region regardless of node
3731 */
3732static int __meminit first_active_region_index_in_nid(int nid)
3733{
3734 int i;
3735
3736 for (i = 0; i < nr_nodemap_entries; i++)
3737 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3738 return i;
3739
3740 return -1;
3741}
3742
3743/*
3744 * Basic iterator support. Return the next active range of PFNs for a node
3745 * Note: nid == MAX_NUMNODES returns next region regardless of node
3746 */
3747static int __meminit next_active_region_index_in_nid(int index, int nid)
3748{
3749 for (index = index + 1; index < nr_nodemap_entries; index++)
3750 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3751 return index;
3752
3753 return -1;
3754}
3755
3756#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3757/*
3758 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3759 * Architectures may implement their own version but if add_active_range()
3760 * was used and there are no special requirements, this is a convenient
3761 * alternative
3762 */
3763int __meminit __early_pfn_to_nid(unsigned long pfn)
3764{
3765 int i;
3766
3767 for (i = 0; i < nr_nodemap_entries; i++) {
3768 unsigned long start_pfn = early_node_map[i].start_pfn;
3769 unsigned long end_pfn = early_node_map[i].end_pfn;
3770
3771 if (start_pfn <= pfn && pfn < end_pfn)
3772 return early_node_map[i].nid;
3773 }
3774 /* This is a memory hole */
3775 return -1;
3776}
3777#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3778
3779int __meminit early_pfn_to_nid(unsigned long pfn)
3780{
3781 int nid;
3782
3783 nid = __early_pfn_to_nid(pfn);
3784 if (nid >= 0)
3785 return nid;
3786 /* just returns 0 */
3787 return 0;
3788}
3789
3790#ifdef CONFIG_NODES_SPAN_OTHER_NODES
3791bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3792{
3793 int nid;
3794
3795 nid = __early_pfn_to_nid(pfn);
3796 if (nid >= 0 && nid != node)
3797 return false;
3798 return true;
3799}
3800#endif
3801
3802/* Basic iterator support to walk early_node_map[] */
3803#define for_each_active_range_index_in_nid(i, nid) \
3804 for (i = first_active_region_index_in_nid(nid); i != -1; \
3805 i = next_active_region_index_in_nid(i, nid))
3806
3807/**
3808 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3809 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3810 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3811 *
3812 * If an architecture guarantees that all ranges registered with
3813 * add_active_ranges() contain no holes and may be freed, this
3814 * this function may be used instead of calling free_bootmem() manually.
3815 */
3816void __init free_bootmem_with_active_regions(int nid,
3817 unsigned long max_low_pfn)
3818{
3819 int i;
3820
3821 for_each_active_range_index_in_nid(i, nid) {
3822 unsigned long size_pages = 0;
3823 unsigned long end_pfn = early_node_map[i].end_pfn;
3824
3825 if (early_node_map[i].start_pfn >= max_low_pfn)
3826 continue;
3827
3828 if (end_pfn > max_low_pfn)
3829 end_pfn = max_low_pfn;
3830
3831 size_pages = end_pfn - early_node_map[i].start_pfn;
3832 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3833 PFN_PHYS(early_node_map[i].start_pfn),
3834 size_pages << PAGE_SHIFT);
3835 }
3836}
3837
3838#ifdef CONFIG_HAVE_MEMBLOCK
3839/*
3840 * Basic iterator support. Return the last range of PFNs for a node
3841 * Note: nid == MAX_NUMNODES returns last region regardless of node
3842 */
3843static int __meminit last_active_region_index_in_nid(int nid)
3844{
3845 int i;
3846
3847 for (i = nr_nodemap_entries - 1; i >= 0; i--)
3848 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3849 return i;
3850
3851 return -1;
3852}
3853
3854/*
3855 * Basic iterator support. Return the previous active range of PFNs for a node
3856 * Note: nid == MAX_NUMNODES returns next region regardless of node
3857 */
3858static int __meminit previous_active_region_index_in_nid(int index, int nid)
3859{
3860 for (index = index - 1; index >= 0; index--)
3861 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3862 return index;
3863
3864 return -1;
3865}
3866
3867#define for_each_active_range_index_in_nid_reverse(i, nid) \
3868 for (i = last_active_region_index_in_nid(nid); i != -1; \
3869 i = previous_active_region_index_in_nid(i, nid))
3870
3871u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3872 u64 goal, u64 limit)
3873{
3874 int i;
3875
3876 /* Need to go over early_node_map to find out good range for node */
3877 for_each_active_range_index_in_nid_reverse(i, nid) {
3878 u64 addr;
3879 u64 ei_start, ei_last;
3880 u64 final_start, final_end;
3881
3882 ei_last = early_node_map[i].end_pfn;
3883 ei_last <<= PAGE_SHIFT;
3884 ei_start = early_node_map[i].start_pfn;
3885 ei_start <<= PAGE_SHIFT;
3886
3887 final_start = max(ei_start, goal);
3888 final_end = min(ei_last, limit);
3889
3890 if (final_start >= final_end)
3891 continue;
3892
3893 addr = memblock_find_in_range(final_start, final_end, size, align);
3894
3895 if (addr == MEMBLOCK_ERROR)
3896 continue;
3897
3898 return addr;
3899 }
3900
3901 return MEMBLOCK_ERROR;
3902}
3903#endif
3904
3905int __init add_from_early_node_map(struct range *range, int az,
3906 int nr_range, int nid)
3907{
3908 int i;
3909 u64 start, end;
3910
3911 /* need to go over early_node_map to find out good range for node */
3912 for_each_active_range_index_in_nid(i, nid) {
3913 start = early_node_map[i].start_pfn;
3914 end = early_node_map[i].end_pfn;
3915 nr_range = add_range(range, az, nr_range, start, end);
3916 }
3917 return nr_range;
3918}
3919
3920void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3921{
3922 int i;
3923 int ret;
3924
3925 for_each_active_range_index_in_nid(i, nid) {
3926 ret = work_fn(early_node_map[i].start_pfn,
3927 early_node_map[i].end_pfn, data);
3928 if (ret)
3929 break;
3930 }
3931}
3932/**
3933 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3934 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3935 *
3936 * If an architecture guarantees that all ranges registered with
3937 * add_active_ranges() contain no holes and may be freed, this
3938 * function may be used instead of calling memory_present() manually.
3939 */
3940void __init sparse_memory_present_with_active_regions(int nid)
3941{
3942 int i;
3943
3944 for_each_active_range_index_in_nid(i, nid)
3945 memory_present(early_node_map[i].nid,
3946 early_node_map[i].start_pfn,
3947 early_node_map[i].end_pfn);
3948}
3949
3950/**
3951 * get_pfn_range_for_nid - Return the start and end page frames for a node
3952 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3953 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3954 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3955 *
3956 * It returns the start and end page frame of a node based on information
3957 * provided by an arch calling add_active_range(). If called for a node
3958 * with no available memory, a warning is printed and the start and end
3959 * PFNs will be 0.
3960 */
3961void __meminit get_pfn_range_for_nid(unsigned int nid,
3962 unsigned long *start_pfn, unsigned long *end_pfn)
3963{
3964 int i;
3965 *start_pfn = -1UL;
3966 *end_pfn = 0;
3967
3968 for_each_active_range_index_in_nid(i, nid) {
3969 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3970 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3971 }
3972
3973 if (*start_pfn == -1UL)
3974 *start_pfn = 0;
3975}
3976
3977/*
3978 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3979 * assumption is made that zones within a node are ordered in monotonic
3980 * increasing memory addresses so that the "highest" populated zone is used
3981 */
3982static void __init find_usable_zone_for_movable(void)
3983{
3984 int zone_index;
3985 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3986 if (zone_index == ZONE_MOVABLE)
3987 continue;
3988
3989 if (arch_zone_highest_possible_pfn[zone_index] >
3990 arch_zone_lowest_possible_pfn[zone_index])
3991 break;
3992 }
3993
3994 VM_BUG_ON(zone_index == -1);
3995 movable_zone = zone_index;
3996}
3997
3998/*
3999 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4000 * because it is sized independent of architecture. Unlike the other zones,
4001 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4002 * in each node depending on the size of each node and how evenly kernelcore
4003 * is distributed. This helper function adjusts the zone ranges
4004 * provided by the architecture for a given node by using the end of the
4005 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4006 * zones within a node are in order of monotonic increases memory addresses
4007 */
4008static void __meminit adjust_zone_range_for_zone_movable(int nid,
4009 unsigned long zone_type,
4010 unsigned long node_start_pfn,
4011 unsigned long node_end_pfn,
4012 unsigned long *zone_start_pfn,
4013 unsigned long *zone_end_pfn)
4014{
4015 /* Only adjust if ZONE_MOVABLE is on this node */
4016 if (zone_movable_pfn[nid]) {
4017 /* Size ZONE_MOVABLE */
4018 if (zone_type == ZONE_MOVABLE) {
4019 *zone_start_pfn = zone_movable_pfn[nid];
4020 *zone_end_pfn = min(node_end_pfn,
4021 arch_zone_highest_possible_pfn[movable_zone]);
4022
4023 /* Adjust for ZONE_MOVABLE starting within this range */
4024 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4025 *zone_end_pfn > zone_movable_pfn[nid]) {
4026 *zone_end_pfn = zone_movable_pfn[nid];
4027
4028 /* Check if this whole range is within ZONE_MOVABLE */
4029 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4030 *zone_start_pfn = *zone_end_pfn;
4031 }
4032}
4033
4034/*
4035 * Return the number of pages a zone spans in a node, including holes
4036 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4037 */
4038static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4039 unsigned long zone_type,
4040 unsigned long *ignored)
4041{
4042 unsigned long node_start_pfn, node_end_pfn;
4043 unsigned long zone_start_pfn, zone_end_pfn;
4044
4045 /* Get the start and end of the node and zone */
4046 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4047 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4048 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4049 adjust_zone_range_for_zone_movable(nid, zone_type,
4050 node_start_pfn, node_end_pfn,
4051 &zone_start_pfn, &zone_end_pfn);
4052
4053 /* Check that this node has pages within the zone's required range */
4054 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4055 return 0;
4056
4057 /* Move the zone boundaries inside the node if necessary */
4058 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4059 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4060
4061 /* Return the spanned pages */
4062 return zone_end_pfn - zone_start_pfn;
4063}
4064
4065/*
4066 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4067 * then all holes in the requested range will be accounted for.
4068 */
4069unsigned long __meminit __absent_pages_in_range(int nid,
4070 unsigned long range_start_pfn,
4071 unsigned long range_end_pfn)
4072{
4073 int i = 0;
4074 unsigned long prev_end_pfn = 0, hole_pages = 0;
4075 unsigned long start_pfn;
4076
4077 /* Find the end_pfn of the first active range of pfns in the node */
4078 i = first_active_region_index_in_nid(nid);
4079 if (i == -1)
4080 return 0;
4081
4082 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4083
4084 /* Account for ranges before physical memory on this node */
4085 if (early_node_map[i].start_pfn > range_start_pfn)
4086 hole_pages = prev_end_pfn - range_start_pfn;
4087
4088 /* Find all holes for the zone within the node */
4089 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4090
4091 /* No need to continue if prev_end_pfn is outside the zone */
4092 if (prev_end_pfn >= range_end_pfn)
4093 break;
4094
4095 /* Make sure the end of the zone is not within the hole */
4096 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4097 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4098
4099 /* Update the hole size cound and move on */
4100 if (start_pfn > range_start_pfn) {
4101 BUG_ON(prev_end_pfn > start_pfn);
4102 hole_pages += start_pfn - prev_end_pfn;
4103 }
4104 prev_end_pfn = early_node_map[i].end_pfn;
4105 }
4106
4107 /* Account for ranges past physical memory on this node */
4108 if (range_end_pfn > prev_end_pfn)
4109 hole_pages += range_end_pfn -
4110 max(range_start_pfn, prev_end_pfn);
4111
4112 return hole_pages;
4113}
4114
4115/**
4116 * absent_pages_in_range - Return number of page frames in holes within a range
4117 * @start_pfn: The start PFN to start searching for holes
4118 * @end_pfn: The end PFN to stop searching for holes
4119 *
4120 * It returns the number of pages frames in memory holes within a range.
4121 */
4122unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4123 unsigned long end_pfn)
4124{
4125 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4126}
4127
4128/* Return the number of page frames in holes in a zone on a node */
4129static unsigned long __meminit zone_absent_pages_in_node(int nid,
4130 unsigned long zone_type,
4131 unsigned long *ignored)
4132{
4133 unsigned long node_start_pfn, node_end_pfn;
4134 unsigned long zone_start_pfn, zone_end_pfn;
4135
4136 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4137 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4138 node_start_pfn);
4139 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4140 node_end_pfn);
4141
4142 adjust_zone_range_for_zone_movable(nid, zone_type,
4143 node_start_pfn, node_end_pfn,
4144 &zone_start_pfn, &zone_end_pfn);
4145 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4146}
4147
4148#else
4149static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4150 unsigned long zone_type,
4151 unsigned long *zones_size)
4152{
4153 return zones_size[zone_type];
4154}
4155
4156static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4157 unsigned long zone_type,
4158 unsigned long *zholes_size)
4159{
4160 if (!zholes_size)
4161 return 0;
4162
4163 return zholes_size[zone_type];
4164}
4165
4166#endif
4167
4168static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4169 unsigned long *zones_size, unsigned long *zholes_size)
4170{
4171 unsigned long realtotalpages, totalpages = 0;
4172 enum zone_type i;
4173
4174 for (i = 0; i < MAX_NR_ZONES; i++)
4175 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4176 zones_size);
4177 pgdat->node_spanned_pages = totalpages;
4178
4179 realtotalpages = totalpages;
4180 for (i = 0; i < MAX_NR_ZONES; i++)
4181 realtotalpages -=
4182 zone_absent_pages_in_node(pgdat->node_id, i,
4183 zholes_size);
4184 pgdat->node_present_pages = realtotalpages;
4185 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4186 realtotalpages);
4187}
4188
4189#ifndef CONFIG_SPARSEMEM
4190/*
4191 * Calculate the size of the zone->blockflags rounded to an unsigned long
4192 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4193 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4194 * round what is now in bits to nearest long in bits, then return it in
4195 * bytes.
4196 */
4197static unsigned long __init usemap_size(unsigned long zonesize)
4198{
4199 unsigned long usemapsize;
4200
4201 usemapsize = roundup(zonesize, pageblock_nr_pages);
4202 usemapsize = usemapsize >> pageblock_order;
4203 usemapsize *= NR_PAGEBLOCK_BITS;
4204 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4205
4206 return usemapsize / 8;
4207}
4208
4209static void __init setup_usemap(struct pglist_data *pgdat,
4210 struct zone *zone, unsigned long zonesize)
4211{
4212 unsigned long usemapsize = usemap_size(zonesize);
4213 zone->pageblock_flags = NULL;
4214 if (usemapsize)
4215 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4216 usemapsize);
4217}
4218#else
4219static inline void setup_usemap(struct pglist_data *pgdat,
4220 struct zone *zone, unsigned long zonesize) {}
4221#endif /* CONFIG_SPARSEMEM */
4222
4223#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4224
4225/* Return a sensible default order for the pageblock size. */
4226static inline int pageblock_default_order(void)
4227{
4228 if (HPAGE_SHIFT > PAGE_SHIFT)
4229 return HUGETLB_PAGE_ORDER;
4230
4231 return MAX_ORDER-1;
4232}
4233
4234/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4235static inline void __init set_pageblock_order(unsigned int order)
4236{
4237 /* Check that pageblock_nr_pages has not already been setup */
4238 if (pageblock_order)
4239 return;
4240
4241 /*
4242 * Assume the largest contiguous order of interest is a huge page.
4243 * This value may be variable depending on boot parameters on IA64
4244 */
4245 pageblock_order = order;
4246}
4247#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4248
4249/*
4250 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4251 * and pageblock_default_order() are unused as pageblock_order is set
4252 * at compile-time. See include/linux/pageblock-flags.h for the values of
4253 * pageblock_order based on the kernel config
4254 */
4255static inline int pageblock_default_order(unsigned int order)
4256{
4257 return MAX_ORDER-1;
4258}
4259#define set_pageblock_order(x) do {} while (0)
4260
4261#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4262
4263/*
4264 * Set up the zone data structures:
4265 * - mark all pages reserved
4266 * - mark all memory queues empty
4267 * - clear the memory bitmaps
4268 */
4269static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4270 unsigned long *zones_size, unsigned long *zholes_size)
4271{
4272 enum zone_type j;
4273 int nid = pgdat->node_id;
4274 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4275 int ret;
4276
4277 pgdat_resize_init(pgdat);
4278 pgdat->nr_zones = 0;
4279 init_waitqueue_head(&pgdat->kswapd_wait);
4280 pgdat->kswapd_max_order = 0;
4281 pgdat_page_cgroup_init(pgdat);
4282
4283 for (j = 0; j < MAX_NR_ZONES; j++) {
4284 struct zone *zone = pgdat->node_zones + j;
4285 unsigned long size, realsize, memmap_pages;
4286 enum lru_list l;
4287
4288 size = zone_spanned_pages_in_node(nid, j, zones_size);
4289 realsize = size - zone_absent_pages_in_node(nid, j,
4290 zholes_size);
4291
4292 /*
4293 * Adjust realsize so that it accounts for how much memory
4294 * is used by this zone for memmap. This affects the watermark
4295 * and per-cpu initialisations
4296 */
4297 memmap_pages =
4298 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4299 if (realsize >= memmap_pages) {
4300 realsize -= memmap_pages;
4301 if (memmap_pages)
4302 printk(KERN_DEBUG
4303 " %s zone: %lu pages used for memmap\n",
4304 zone_names[j], memmap_pages);
4305 } else
4306 printk(KERN_WARNING
4307 " %s zone: %lu pages exceeds realsize %lu\n",
4308 zone_names[j], memmap_pages, realsize);
4309
4310 /* Account for reserved pages */
4311 if (j == 0 && realsize > dma_reserve) {
4312 realsize -= dma_reserve;
4313 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4314 zone_names[0], dma_reserve);
4315 }
4316
4317 if (!is_highmem_idx(j))
4318 nr_kernel_pages += realsize;
4319 nr_all_pages += realsize;
4320
4321 zone->spanned_pages = size;
4322 zone->present_pages = realsize;
4323#ifdef CONFIG_NUMA
4324 zone->node = nid;
4325 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4326 / 100;
4327 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4328#endif
4329 zone->name = zone_names[j];
4330 spin_lock_init(&zone->lock);
4331 spin_lock_init(&zone->lru_lock);
4332 zone_seqlock_init(zone);
4333 zone->zone_pgdat = pgdat;
4334
4335 zone_pcp_init(zone);
4336 for_each_lru(l)
4337 INIT_LIST_HEAD(&zone->lru[l].list);
4338 zone->reclaim_stat.recent_rotated[0] = 0;
4339 zone->reclaim_stat.recent_rotated[1] = 0;
4340 zone->reclaim_stat.recent_scanned[0] = 0;
4341 zone->reclaim_stat.recent_scanned[1] = 0;
4342 zap_zone_vm_stats(zone);
4343 zone->flags = 0;
4344 if (!size)
4345 continue;
4346
4347 set_pageblock_order(pageblock_default_order());
4348 setup_usemap(pgdat, zone, size);
4349 ret = init_currently_empty_zone(zone, zone_start_pfn,
4350 size, MEMMAP_EARLY);
4351 BUG_ON(ret);
4352 memmap_init(size, nid, j, zone_start_pfn);
4353 zone_start_pfn += size;
4354 }
4355}
4356
4357static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4358{
4359 /* Skip empty nodes */
4360 if (!pgdat->node_spanned_pages)
4361 return;
4362
4363#ifdef CONFIG_FLAT_NODE_MEM_MAP
4364 /* ia64 gets its own node_mem_map, before this, without bootmem */
4365 if (!pgdat->node_mem_map) {
4366 unsigned long size, start, end;
4367 struct page *map;
4368
4369 /*
4370 * The zone's endpoints aren't required to be MAX_ORDER
4371 * aligned but the node_mem_map endpoints must be in order
4372 * for the buddy allocator to function correctly.
4373 */
4374 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4375 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4376 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4377 size = (end - start) * sizeof(struct page);
4378 map = alloc_remap(pgdat->node_id, size);
4379 if (!map)
4380 map = alloc_bootmem_node_nopanic(pgdat, size);
4381 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4382 }
4383#ifndef CONFIG_NEED_MULTIPLE_NODES
4384 /*
4385 * With no DISCONTIG, the global mem_map is just set as node 0's
4386 */
4387 if (pgdat == NODE_DATA(0)) {
4388 mem_map = NODE_DATA(0)->node_mem_map;
4389#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4390 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4391 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4392#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4393 }
4394#endif
4395#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4396}
4397
4398void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4399 unsigned long node_start_pfn, unsigned long *zholes_size)
4400{
4401 pg_data_t *pgdat = NODE_DATA(nid);
4402
4403 pgdat->node_id = nid;
4404 pgdat->node_start_pfn = node_start_pfn;
4405 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4406
4407 alloc_node_mem_map(pgdat);
4408#ifdef CONFIG_FLAT_NODE_MEM_MAP
4409 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4410 nid, (unsigned long)pgdat,
4411 (unsigned long)pgdat->node_mem_map);
4412#endif
4413
4414 free_area_init_core(pgdat, zones_size, zholes_size);
4415}
4416
4417#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4418
4419#if MAX_NUMNODES > 1
4420/*
4421 * Figure out the number of possible node ids.
4422 */
4423static void __init setup_nr_node_ids(void)
4424{
4425 unsigned int node;
4426 unsigned int highest = 0;
4427
4428 for_each_node_mask(node, node_possible_map)
4429 highest = node;
4430 nr_node_ids = highest + 1;
4431}
4432#else
4433static inline void setup_nr_node_ids(void)
4434{
4435}
4436#endif
4437
4438/**
4439 * add_active_range - Register a range of PFNs backed by physical memory
4440 * @nid: The node ID the range resides on
4441 * @start_pfn: The start PFN of the available physical memory
4442 * @end_pfn: The end PFN of the available physical memory
4443 *
4444 * These ranges are stored in an early_node_map[] and later used by
4445 * free_area_init_nodes() to calculate zone sizes and holes. If the
4446 * range spans a memory hole, it is up to the architecture to ensure
4447 * the memory is not freed by the bootmem allocator. If possible
4448 * the range being registered will be merged with existing ranges.
4449 */
4450void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4451 unsigned long end_pfn)
4452{
4453 int i;
4454
4455 mminit_dprintk(MMINIT_TRACE, "memory_register",
4456 "Entering add_active_range(%d, %#lx, %#lx) "
4457 "%d entries of %d used\n",
4458 nid, start_pfn, end_pfn,
4459 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4460
4461 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4462
4463 /* Merge with existing active regions if possible */
4464 for (i = 0; i < nr_nodemap_entries; i++) {
4465 if (early_node_map[i].nid != nid)
4466 continue;
4467
4468 /* Skip if an existing region covers this new one */
4469 if (start_pfn >= early_node_map[i].start_pfn &&
4470 end_pfn <= early_node_map[i].end_pfn)
4471 return;
4472
4473 /* Merge forward if suitable */
4474 if (start_pfn <= early_node_map[i].end_pfn &&
4475 end_pfn > early_node_map[i].end_pfn) {
4476 early_node_map[i].end_pfn = end_pfn;
4477 return;
4478 }
4479
4480 /* Merge backward if suitable */
4481 if (start_pfn < early_node_map[i].start_pfn &&
4482 end_pfn >= early_node_map[i].start_pfn) {
4483 early_node_map[i].start_pfn = start_pfn;
4484 return;
4485 }
4486 }
4487
4488 /* Check that early_node_map is large enough */
4489 if (i >= MAX_ACTIVE_REGIONS) {
4490 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4491 MAX_ACTIVE_REGIONS);
4492 return;
4493 }
4494
4495 early_node_map[i].nid = nid;
4496 early_node_map[i].start_pfn = start_pfn;
4497 early_node_map[i].end_pfn = end_pfn;
4498 nr_nodemap_entries = i + 1;
4499}
4500
4501/**
4502 * remove_active_range - Shrink an existing registered range of PFNs
4503 * @nid: The node id the range is on that should be shrunk
4504 * @start_pfn: The new PFN of the range
4505 * @end_pfn: The new PFN of the range
4506 *
4507 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4508 * The map is kept near the end physical page range that has already been
4509 * registered. This function allows an arch to shrink an existing registered
4510 * range.
4511 */
4512void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4513 unsigned long end_pfn)
4514{
4515 int i, j;
4516 int removed = 0;
4517
4518 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4519 nid, start_pfn, end_pfn);
4520
4521 /* Find the old active region end and shrink */
4522 for_each_active_range_index_in_nid(i, nid) {
4523 if (early_node_map[i].start_pfn >= start_pfn &&
4524 early_node_map[i].end_pfn <= end_pfn) {
4525 /* clear it */
4526 early_node_map[i].start_pfn = 0;
4527 early_node_map[i].end_pfn = 0;
4528 removed = 1;
4529 continue;
4530 }
4531 if (early_node_map[i].start_pfn < start_pfn &&
4532 early_node_map[i].end_pfn > start_pfn) {
4533 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4534 early_node_map[i].end_pfn = start_pfn;
4535 if (temp_end_pfn > end_pfn)
4536 add_active_range(nid, end_pfn, temp_end_pfn);
4537 continue;
4538 }
4539 if (early_node_map[i].start_pfn >= start_pfn &&
4540 early_node_map[i].end_pfn > end_pfn &&
4541 early_node_map[i].start_pfn < end_pfn) {
4542 early_node_map[i].start_pfn = end_pfn;
4543 continue;
4544 }
4545 }
4546
4547 if (!removed)
4548 return;
4549
4550 /* remove the blank ones */
4551 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4552 if (early_node_map[i].nid != nid)
4553 continue;
4554 if (early_node_map[i].end_pfn)
4555 continue;
4556 /* we found it, get rid of it */
4557 for (j = i; j < nr_nodemap_entries - 1; j++)
4558 memcpy(&early_node_map[j], &early_node_map[j+1],
4559 sizeof(early_node_map[j]));
4560 j = nr_nodemap_entries - 1;
4561 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4562 nr_nodemap_entries--;
4563 }
4564}
4565
4566/**
4567 * remove_all_active_ranges - Remove all currently registered regions
4568 *
4569 * During discovery, it may be found that a table like SRAT is invalid
4570 * and an alternative discovery method must be used. This function removes
4571 * all currently registered regions.
4572 */
4573void __init remove_all_active_ranges(void)
4574{
4575 memset(early_node_map, 0, sizeof(early_node_map));
4576 nr_nodemap_entries = 0;
4577}
4578
4579/* Compare two active node_active_regions */
4580static int __init cmp_node_active_region(const void *a, const void *b)
4581{
4582 struct node_active_region *arange = (struct node_active_region *)a;
4583 struct node_active_region *brange = (struct node_active_region *)b;
4584
4585 /* Done this way to avoid overflows */
4586 if (arange->start_pfn > brange->start_pfn)
4587 return 1;
4588 if (arange->start_pfn < brange->start_pfn)
4589 return -1;
4590
4591 return 0;
4592}
4593
4594/* sort the node_map by start_pfn */
4595void __init sort_node_map(void)
4596{
4597 sort(early_node_map, (size_t)nr_nodemap_entries,
4598 sizeof(struct node_active_region),
4599 cmp_node_active_region, NULL);
4600}
4601
4602/**
4603 * node_map_pfn_alignment - determine the maximum internode alignment
4604 *
4605 * This function should be called after node map is populated and sorted.
4606 * It calculates the maximum power of two alignment which can distinguish
4607 * all the nodes.
4608 *
4609 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4610 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4611 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4612 * shifted, 1GiB is enough and this function will indicate so.
4613 *
4614 * This is used to test whether pfn -> nid mapping of the chosen memory
4615 * model has fine enough granularity to avoid incorrect mapping for the
4616 * populated node map.
4617 *
4618 * Returns the determined alignment in pfn's. 0 if there is no alignment
4619 * requirement (single node).
4620 */
4621unsigned long __init node_map_pfn_alignment(void)
4622{
4623 unsigned long accl_mask = 0, last_end = 0;
4624 int last_nid = -1;
4625 int i;
4626
4627 for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4628 int nid = early_node_map[i].nid;
4629 unsigned long start = early_node_map[i].start_pfn;
4630 unsigned long end = early_node_map[i].end_pfn;
4631 unsigned long mask;
4632
4633 if (!start || last_nid < 0 || last_nid == nid) {
4634 last_nid = nid;
4635 last_end = end;
4636 continue;
4637 }
4638
4639 /*
4640 * Start with a mask granular enough to pin-point to the
4641 * start pfn and tick off bits one-by-one until it becomes
4642 * too coarse to separate the current node from the last.
4643 */
4644 mask = ~((1 << __ffs(start)) - 1);
4645 while (mask && last_end <= (start & (mask << 1)))
4646 mask <<= 1;
4647
4648 /* accumulate all internode masks */
4649 accl_mask |= mask;
4650 }
4651
4652 /* convert mask to number of pages */
4653 return ~accl_mask + 1;
4654}
4655
4656/* Find the lowest pfn for a node */
4657static unsigned long __init find_min_pfn_for_node(int nid)
4658{
4659 int i;
4660 unsigned long min_pfn = ULONG_MAX;
4661
4662 /* Assuming a sorted map, the first range found has the starting pfn */
4663 for_each_active_range_index_in_nid(i, nid)
4664 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4665
4666 if (min_pfn == ULONG_MAX) {
4667 printk(KERN_WARNING
4668 "Could not find start_pfn for node %d\n", nid);
4669 return 0;
4670 }
4671
4672 return min_pfn;
4673}
4674
4675/**
4676 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4677 *
4678 * It returns the minimum PFN based on information provided via
4679 * add_active_range().
4680 */
4681unsigned long __init find_min_pfn_with_active_regions(void)
4682{
4683 return find_min_pfn_for_node(MAX_NUMNODES);
4684}
4685
4686/*
4687 * early_calculate_totalpages()
4688 * Sum pages in active regions for movable zone.
4689 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4690 */
4691static unsigned long __init early_calculate_totalpages(void)
4692{
4693 int i;
4694 unsigned long totalpages = 0;
4695
4696 for (i = 0; i < nr_nodemap_entries; i++) {
4697 unsigned long pages = early_node_map[i].end_pfn -
4698 early_node_map[i].start_pfn;
4699 totalpages += pages;
4700 if (pages)
4701 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4702 }
4703 return totalpages;
4704}
4705
4706/*
4707 * Find the PFN the Movable zone begins in each node. Kernel memory
4708 * is spread evenly between nodes as long as the nodes have enough
4709 * memory. When they don't, some nodes will have more kernelcore than
4710 * others
4711 */
4712static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4713{
4714 int i, nid;
4715 unsigned long usable_startpfn;
4716 unsigned long kernelcore_node, kernelcore_remaining;
4717 /* save the state before borrow the nodemask */
4718 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4719 unsigned long totalpages = early_calculate_totalpages();
4720 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4721
4722 /*
4723 * If movablecore was specified, calculate what size of
4724 * kernelcore that corresponds so that memory usable for
4725 * any allocation type is evenly spread. If both kernelcore
4726 * and movablecore are specified, then the value of kernelcore
4727 * will be used for required_kernelcore if it's greater than
4728 * what movablecore would have allowed.
4729 */
4730 if (required_movablecore) {
4731 unsigned long corepages;
4732
4733 /*
4734 * Round-up so that ZONE_MOVABLE is at least as large as what
4735 * was requested by the user
4736 */
4737 required_movablecore =
4738 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4739 corepages = totalpages - required_movablecore;
4740
4741 required_kernelcore = max(required_kernelcore, corepages);
4742 }
4743
4744 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4745 if (!required_kernelcore)
4746 goto out;
4747
4748 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4749 find_usable_zone_for_movable();
4750 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4751
4752restart:
4753 /* Spread kernelcore memory as evenly as possible throughout nodes */
4754 kernelcore_node = required_kernelcore / usable_nodes;
4755 for_each_node_state(nid, N_HIGH_MEMORY) {
4756 /*
4757 * Recalculate kernelcore_node if the division per node
4758 * now exceeds what is necessary to satisfy the requested
4759 * amount of memory for the kernel
4760 */
4761 if (required_kernelcore < kernelcore_node)
4762 kernelcore_node = required_kernelcore / usable_nodes;
4763
4764 /*
4765 * As the map is walked, we track how much memory is usable
4766 * by the kernel using kernelcore_remaining. When it is
4767 * 0, the rest of the node is usable by ZONE_MOVABLE
4768 */
4769 kernelcore_remaining = kernelcore_node;
4770
4771 /* Go through each range of PFNs within this node */
4772 for_each_active_range_index_in_nid(i, nid) {
4773 unsigned long start_pfn, end_pfn;
4774 unsigned long size_pages;
4775
4776 start_pfn = max(early_node_map[i].start_pfn,
4777 zone_movable_pfn[nid]);
4778 end_pfn = early_node_map[i].end_pfn;
4779 if (start_pfn >= end_pfn)
4780 continue;
4781
4782 /* Account for what is only usable for kernelcore */
4783 if (start_pfn < usable_startpfn) {
4784 unsigned long kernel_pages;
4785 kernel_pages = min(end_pfn, usable_startpfn)
4786 - start_pfn;
4787
4788 kernelcore_remaining -= min(kernel_pages,
4789 kernelcore_remaining);
4790 required_kernelcore -= min(kernel_pages,
4791 required_kernelcore);
4792
4793 /* Continue if range is now fully accounted */
4794 if (end_pfn <= usable_startpfn) {
4795
4796 /*
4797 * Push zone_movable_pfn to the end so
4798 * that if we have to rebalance
4799 * kernelcore across nodes, we will
4800 * not double account here
4801 */
4802 zone_movable_pfn[nid] = end_pfn;
4803 continue;
4804 }
4805 start_pfn = usable_startpfn;
4806 }
4807
4808 /*
4809 * The usable PFN range for ZONE_MOVABLE is from
4810 * start_pfn->end_pfn. Calculate size_pages as the
4811 * number of pages used as kernelcore
4812 */
4813 size_pages = end_pfn - start_pfn;
4814 if (size_pages > kernelcore_remaining)
4815 size_pages = kernelcore_remaining;
4816 zone_movable_pfn[nid] = start_pfn + size_pages;
4817
4818 /*
4819 * Some kernelcore has been met, update counts and
4820 * break if the kernelcore for this node has been
4821 * satisified
4822 */
4823 required_kernelcore -= min(required_kernelcore,
4824 size_pages);
4825 kernelcore_remaining -= size_pages;
4826 if (!kernelcore_remaining)
4827 break;
4828 }
4829 }
4830
4831 /*
4832 * If there is still required_kernelcore, we do another pass with one
4833 * less node in the count. This will push zone_movable_pfn[nid] further
4834 * along on the nodes that still have memory until kernelcore is
4835 * satisified
4836 */
4837 usable_nodes--;
4838 if (usable_nodes && required_kernelcore > usable_nodes)
4839 goto restart;
4840
4841 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4842 for (nid = 0; nid < MAX_NUMNODES; nid++)
4843 zone_movable_pfn[nid] =
4844 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4845
4846out:
4847 /* restore the node_state */
4848 node_states[N_HIGH_MEMORY] = saved_node_state;
4849}
4850
4851/* Any regular memory on that node ? */
4852static void check_for_regular_memory(pg_data_t *pgdat)
4853{
4854#ifdef CONFIG_HIGHMEM
4855 enum zone_type zone_type;
4856
4857 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4858 struct zone *zone = &pgdat->node_zones[zone_type];
4859 if (zone->present_pages)
4860 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4861 }
4862#endif
4863}
4864
4865/**
4866 * free_area_init_nodes - Initialise all pg_data_t and zone data
4867 * @max_zone_pfn: an array of max PFNs for each zone
4868 *
4869 * This will call free_area_init_node() for each active node in the system.
4870 * Using the page ranges provided by add_active_range(), the size of each
4871 * zone in each node and their holes is calculated. If the maximum PFN
4872 * between two adjacent zones match, it is assumed that the zone is empty.
4873 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4874 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4875 * starts where the previous one ended. For example, ZONE_DMA32 starts
4876 * at arch_max_dma_pfn.
4877 */
4878void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4879{
4880 unsigned long nid;
4881 int i;
4882
4883 /* Sort early_node_map as initialisation assumes it is sorted */
4884 sort_node_map();
4885
4886 /* Record where the zone boundaries are */
4887 memset(arch_zone_lowest_possible_pfn, 0,
4888 sizeof(arch_zone_lowest_possible_pfn));
4889 memset(arch_zone_highest_possible_pfn, 0,
4890 sizeof(arch_zone_highest_possible_pfn));
4891 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4892 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4893 for (i = 1; i < MAX_NR_ZONES; i++) {
4894 if (i == ZONE_MOVABLE)
4895 continue;
4896 arch_zone_lowest_possible_pfn[i] =
4897 arch_zone_highest_possible_pfn[i-1];
4898 arch_zone_highest_possible_pfn[i] =
4899 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4900 }
4901 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4902 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4903
4904 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4905 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4906 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4907
4908 /* Print out the zone ranges */
4909 printk("Zone PFN ranges:\n");
4910 for (i = 0; i < MAX_NR_ZONES; i++) {
4911 if (i == ZONE_MOVABLE)
4912 continue;
4913 printk(" %-8s ", zone_names[i]);
4914 if (arch_zone_lowest_possible_pfn[i] ==
4915 arch_zone_highest_possible_pfn[i])
4916 printk("empty\n");
4917 else
4918 printk("%0#10lx -> %0#10lx\n",
4919 arch_zone_lowest_possible_pfn[i],
4920 arch_zone_highest_possible_pfn[i]);
4921 }
4922
4923 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4924 printk("Movable zone start PFN for each node\n");
4925 for (i = 0; i < MAX_NUMNODES; i++) {
4926 if (zone_movable_pfn[i])
4927 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4928 }
4929
4930 /* Print out the early_node_map[] */
4931 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4932 for (i = 0; i < nr_nodemap_entries; i++)
4933 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4934 early_node_map[i].start_pfn,
4935 early_node_map[i].end_pfn);
4936
4937 /* Initialise every node */
4938 mminit_verify_pageflags_layout();
4939 setup_nr_node_ids();
4940 for_each_online_node(nid) {
4941 pg_data_t *pgdat = NODE_DATA(nid);
4942 free_area_init_node(nid, NULL,
4943 find_min_pfn_for_node(nid), NULL);
4944
4945 /* Any memory on that node */
4946 if (pgdat->node_present_pages)
4947 node_set_state(nid, N_HIGH_MEMORY);
4948 check_for_regular_memory(pgdat);
4949 }
4950}
4951
4952static int __init cmdline_parse_core(char *p, unsigned long *core)
4953{
4954 unsigned long long coremem;
4955 if (!p)
4956 return -EINVAL;
4957
4958 coremem = memparse(p, &p);
4959 *core = coremem >> PAGE_SHIFT;
4960
4961 /* Paranoid check that UL is enough for the coremem value */
4962 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4963
4964 return 0;
4965}
4966
4967/*
4968 * kernelcore=size sets the amount of memory for use for allocations that
4969 * cannot be reclaimed or migrated.
4970 */
4971static int __init cmdline_parse_kernelcore(char *p)
4972{
4973 return cmdline_parse_core(p, &required_kernelcore);
4974}
4975
4976/*
4977 * movablecore=size sets the amount of memory for use for allocations that
4978 * can be reclaimed or migrated.
4979 */
4980static int __init cmdline_parse_movablecore(char *p)
4981{
4982 return cmdline_parse_core(p, &required_movablecore);
4983}
4984
4985early_param("kernelcore", cmdline_parse_kernelcore);
4986early_param("movablecore", cmdline_parse_movablecore);
4987
4988#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4989
4990/**
4991 * set_dma_reserve - set the specified number of pages reserved in the first zone
4992 * @new_dma_reserve: The number of pages to mark reserved
4993 *
4994 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4995 * In the DMA zone, a significant percentage may be consumed by kernel image
4996 * and other unfreeable allocations which can skew the watermarks badly. This
4997 * function may optionally be used to account for unfreeable pages in the
4998 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4999 * smaller per-cpu batchsize.
5000 */
5001void __init set_dma_reserve(unsigned long new_dma_reserve)
5002{
5003 dma_reserve = new_dma_reserve;
5004}
5005
5006void __init free_area_init(unsigned long *zones_size)
5007{
5008 free_area_init_node(0, zones_size,
5009 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5010}
5011
5012static int page_alloc_cpu_notify(struct notifier_block *self,
5013 unsigned long action, void *hcpu)
5014{
5015 int cpu = (unsigned long)hcpu;
5016
5017 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5018 drain_pages(cpu);
5019
5020 /*
5021 * Spill the event counters of the dead processor
5022 * into the current processors event counters.
5023 * This artificially elevates the count of the current
5024 * processor.
5025 */
5026 vm_events_fold_cpu(cpu);
5027
5028 /*
5029 * Zero the differential counters of the dead processor
5030 * so that the vm statistics are consistent.
5031 *
5032 * This is only okay since the processor is dead and cannot
5033 * race with what we are doing.
5034 */
5035 refresh_cpu_vm_stats(cpu);
5036 }
5037 return NOTIFY_OK;
5038}
5039
5040void __init page_alloc_init(void)
5041{
5042 hotcpu_notifier(page_alloc_cpu_notify, 0);
5043}
5044
5045/*
5046 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5047 * or min_free_kbytes changes.
5048 */
5049static void calculate_totalreserve_pages(void)
5050{
5051 struct pglist_data *pgdat;
5052 unsigned long reserve_pages = 0;
5053 enum zone_type i, j;
5054
5055 for_each_online_pgdat(pgdat) {
5056 for (i = 0; i < MAX_NR_ZONES; i++) {
5057 struct zone *zone = pgdat->node_zones + i;
5058 unsigned long max = 0;
5059
5060 /* Find valid and maximum lowmem_reserve in the zone */
5061 for (j = i; j < MAX_NR_ZONES; j++) {
5062 if (zone->lowmem_reserve[j] > max)
5063 max = zone->lowmem_reserve[j];
5064 }
5065
5066 /* we treat the high watermark as reserved pages. */
5067 max += high_wmark_pages(zone);
5068
5069 if (max > zone->present_pages)
5070 max = zone->present_pages;
5071 reserve_pages += max;
5072 }
5073 }
5074 totalreserve_pages = reserve_pages;
5075}
5076
5077/*
5078 * setup_per_zone_lowmem_reserve - called whenever
5079 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5080 * has a correct pages reserved value, so an adequate number of
5081 * pages are left in the zone after a successful __alloc_pages().
5082 */
5083static void setup_per_zone_lowmem_reserve(void)
5084{
5085 struct pglist_data *pgdat;
5086 enum zone_type j, idx;
5087
5088 for_each_online_pgdat(pgdat) {
5089 for (j = 0; j < MAX_NR_ZONES; j++) {
5090 struct zone *zone = pgdat->node_zones + j;
5091 unsigned long present_pages = zone->present_pages;
5092
5093 zone->lowmem_reserve[j] = 0;
5094
5095 idx = j;
5096 while (idx) {
5097 struct zone *lower_zone;
5098
5099 idx--;
5100
5101 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5102 sysctl_lowmem_reserve_ratio[idx] = 1;
5103
5104 lower_zone = pgdat->node_zones + idx;
5105 lower_zone->lowmem_reserve[j] = present_pages /
5106 sysctl_lowmem_reserve_ratio[idx];
5107 present_pages += lower_zone->present_pages;
5108 }
5109 }
5110 }
5111
5112 /* update totalreserve_pages */
5113 calculate_totalreserve_pages();
5114}
5115
5116/**
5117 * setup_per_zone_wmarks - called when min_free_kbytes changes
5118 * or when memory is hot-{added|removed}
5119 *
5120 * Ensures that the watermark[min,low,high] values for each zone are set
5121 * correctly with respect to min_free_kbytes.
5122 */
5123void setup_per_zone_wmarks(void)
5124{
5125 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5126 unsigned long lowmem_pages = 0;
5127 struct zone *zone;
5128 unsigned long flags;
5129
5130 /* Calculate total number of !ZONE_HIGHMEM pages */
5131 for_each_zone(zone) {
5132 if (!is_highmem(zone))
5133 lowmem_pages += zone->present_pages;
5134 }
5135
5136 for_each_zone(zone) {
5137 u64 tmp;
5138
5139 spin_lock_irqsave(&zone->lock, flags);
5140 tmp = (u64)pages_min * zone->present_pages;
5141 do_div(tmp, lowmem_pages);
5142 if (is_highmem(zone)) {
5143 /*
5144 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5145 * need highmem pages, so cap pages_min to a small
5146 * value here.
5147 *
5148 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5149 * deltas controls asynch page reclaim, and so should
5150 * not be capped for highmem.
5151 */
5152 int min_pages;
5153
5154 min_pages = zone->present_pages / 1024;
5155 if (min_pages < SWAP_CLUSTER_MAX)
5156 min_pages = SWAP_CLUSTER_MAX;
5157 if (min_pages > 128)
5158 min_pages = 128;
5159 zone->watermark[WMARK_MIN] = min_pages;
5160 } else {
5161 /*
5162 * If it's a lowmem zone, reserve a number of pages
5163 * proportionate to the zone's size.
5164 */
5165 zone->watermark[WMARK_MIN] = tmp;
5166 }
5167
5168 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5169 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5170 setup_zone_migrate_reserve(zone);
5171 spin_unlock_irqrestore(&zone->lock, flags);
5172 }
5173
5174 /* update totalreserve_pages */
5175 calculate_totalreserve_pages();
5176}
5177
5178/*
5179 * The inactive anon list should be small enough that the VM never has to
5180 * do too much work, but large enough that each inactive page has a chance
5181 * to be referenced again before it is swapped out.
5182 *
5183 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5184 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5185 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5186 * the anonymous pages are kept on the inactive list.
5187 *
5188 * total target max
5189 * memory ratio inactive anon
5190 * -------------------------------------
5191 * 10MB 1 5MB
5192 * 100MB 1 50MB
5193 * 1GB 3 250MB
5194 * 10GB 10 0.9GB
5195 * 100GB 31 3GB
5196 * 1TB 101 10GB
5197 * 10TB 320 32GB
5198 */
5199static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5200{
5201 unsigned int gb, ratio;
5202
5203 /* Zone size in gigabytes */
5204 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5205 if (gb)
5206 ratio = int_sqrt(10 * gb);
5207 else
5208 ratio = 1;
5209
5210 zone->inactive_ratio = ratio;
5211}
5212
5213static void __meminit setup_per_zone_inactive_ratio(void)
5214{
5215 struct zone *zone;
5216
5217 for_each_zone(zone)
5218 calculate_zone_inactive_ratio(zone);
5219}
5220
5221/*
5222 * Initialise min_free_kbytes.
5223 *
5224 * For small machines we want it small (128k min). For large machines
5225 * we want it large (64MB max). But it is not linear, because network
5226 * bandwidth does not increase linearly with machine size. We use
5227 *
5228 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5229 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5230 *
5231 * which yields
5232 *
5233 * 16MB: 512k
5234 * 32MB: 724k
5235 * 64MB: 1024k
5236 * 128MB: 1448k
5237 * 256MB: 2048k
5238 * 512MB: 2896k
5239 * 1024MB: 4096k
5240 * 2048MB: 5792k
5241 * 4096MB: 8192k
5242 * 8192MB: 11584k
5243 * 16384MB: 16384k
5244 */
5245int __meminit init_per_zone_wmark_min(void)
5246{
5247 unsigned long lowmem_kbytes;
5248
5249 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5250
5251 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5252 if (min_free_kbytes < 128)
5253 min_free_kbytes = 128;
5254 if (min_free_kbytes > 65536)
5255 min_free_kbytes = 65536;
5256 setup_per_zone_wmarks();
5257 refresh_zone_stat_thresholds();
5258 setup_per_zone_lowmem_reserve();
5259 setup_per_zone_inactive_ratio();
5260 return 0;
5261}
5262module_init(init_per_zone_wmark_min)
5263
5264/*
5265 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5266 * that we can call two helper functions whenever min_free_kbytes
5267 * changes.
5268 */
5269int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5270 void __user *buffer, size_t *length, loff_t *ppos)
5271{
5272 proc_dointvec(table, write, buffer, length, ppos);
5273 if (write)
5274 setup_per_zone_wmarks();
5275 return 0;
5276}
5277
5278#ifdef CONFIG_NUMA
5279int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5280 void __user *buffer, size_t *length, loff_t *ppos)
5281{
5282 struct zone *zone;
5283 int rc;
5284
5285 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5286 if (rc)
5287 return rc;
5288
5289 for_each_zone(zone)
5290 zone->min_unmapped_pages = (zone->present_pages *
5291 sysctl_min_unmapped_ratio) / 100;
5292 return 0;
5293}
5294
5295int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5296 void __user *buffer, size_t *length, loff_t *ppos)
5297{
5298 struct zone *zone;
5299 int rc;
5300
5301 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5302 if (rc)
5303 return rc;
5304
5305 for_each_zone(zone)
5306 zone->min_slab_pages = (zone->present_pages *
5307 sysctl_min_slab_ratio) / 100;
5308 return 0;
5309}
5310#endif
5311
5312/*
5313 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5314 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5315 * whenever sysctl_lowmem_reserve_ratio changes.
5316 *
5317 * The reserve ratio obviously has absolutely no relation with the
5318 * minimum watermarks. The lowmem reserve ratio can only make sense
5319 * if in function of the boot time zone sizes.
5320 */
5321int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5322 void __user *buffer, size_t *length, loff_t *ppos)
5323{
5324 proc_dointvec_minmax(table, write, buffer, length, ppos);
5325 setup_per_zone_lowmem_reserve();
5326 return 0;
5327}
5328
5329/*
5330 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5331 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5332 * can have before it gets flushed back to buddy allocator.
5333 */
5334
5335int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5336 void __user *buffer, size_t *length, loff_t *ppos)
5337{
5338 struct zone *zone;
5339 unsigned int cpu;
5340 int ret;
5341
5342 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5343 if (!write || (ret == -EINVAL))
5344 return ret;
5345 for_each_populated_zone(zone) {
5346 for_each_possible_cpu(cpu) {
5347 unsigned long high;
5348 high = zone->present_pages / percpu_pagelist_fraction;
5349 setup_pagelist_highmark(
5350 per_cpu_ptr(zone->pageset, cpu), high);
5351 }
5352 }
5353 return 0;
5354}
5355
5356int hashdist = HASHDIST_DEFAULT;
5357
5358#ifdef CONFIG_NUMA
5359static int __init set_hashdist(char *str)
5360{
5361 if (!str)
5362 return 0;
5363 hashdist = simple_strtoul(str, &str, 0);
5364 return 1;
5365}
5366__setup("hashdist=", set_hashdist);
5367#endif
5368
5369/*
5370 * allocate a large system hash table from bootmem
5371 * - it is assumed that the hash table must contain an exact power-of-2
5372 * quantity of entries
5373 * - limit is the number of hash buckets, not the total allocation size
5374 */
5375void *__init alloc_large_system_hash(const char *tablename,
5376 unsigned long bucketsize,
5377 unsigned long numentries,
5378 int scale,
5379 int flags,
5380 unsigned int *_hash_shift,
5381 unsigned int *_hash_mask,
5382 unsigned long limit)
5383{
5384 unsigned long long max = limit;
5385 unsigned long log2qty, size;
5386 void *table = NULL;
5387
5388 /* allow the kernel cmdline to have a say */
5389 if (!numentries) {
5390 /* round applicable memory size up to nearest megabyte */
5391 numentries = nr_kernel_pages;
5392 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5393 numentries >>= 20 - PAGE_SHIFT;
5394 numentries <<= 20 - PAGE_SHIFT;
5395
5396 /* limit to 1 bucket per 2^scale bytes of low memory */
5397 if (scale > PAGE_SHIFT)
5398 numentries >>= (scale - PAGE_SHIFT);
5399 else
5400 numentries <<= (PAGE_SHIFT - scale);
5401
5402 /* Make sure we've got at least a 0-order allocation.. */
5403 if (unlikely(flags & HASH_SMALL)) {
5404 /* Makes no sense without HASH_EARLY */
5405 WARN_ON(!(flags & HASH_EARLY));
5406 if (!(numentries >> *_hash_shift)) {
5407 numentries = 1UL << *_hash_shift;
5408 BUG_ON(!numentries);
5409 }
5410 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5411 numentries = PAGE_SIZE / bucketsize;
5412 }
5413 numentries = roundup_pow_of_two(numentries);
5414
5415 /* limit allocation size to 1/16 total memory by default */
5416 if (max == 0) {
5417 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5418 do_div(max, bucketsize);
5419 }
5420
5421 if (numentries > max)
5422 numentries = max;
5423
5424 log2qty = ilog2(numentries);
5425
5426 do {
5427 size = bucketsize << log2qty;
5428 if (flags & HASH_EARLY)
5429 table = alloc_bootmem_nopanic(size);
5430 else if (hashdist)
5431 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5432 else {
5433 /*
5434 * If bucketsize is not a power-of-two, we may free
5435 * some pages at the end of hash table which
5436 * alloc_pages_exact() automatically does
5437 */
5438 if (get_order(size) < MAX_ORDER) {
5439 table = alloc_pages_exact(size, GFP_ATOMIC);
5440 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5441 }
5442 }
5443 } while (!table && size > PAGE_SIZE && --log2qty);
5444
5445 if (!table)
5446 panic("Failed to allocate %s hash table\n", tablename);
5447
5448 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5449 tablename,
5450 (1UL << log2qty),
5451 ilog2(size) - PAGE_SHIFT,
5452 size);
5453
5454 if (_hash_shift)
5455 *_hash_shift = log2qty;
5456 if (_hash_mask)
5457 *_hash_mask = (1 << log2qty) - 1;
5458
5459 return table;
5460}
5461
5462/* Return a pointer to the bitmap storing bits affecting a block of pages */
5463static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5464 unsigned long pfn)
5465{
5466#ifdef CONFIG_SPARSEMEM
5467 return __pfn_to_section(pfn)->pageblock_flags;
5468#else
5469 return zone->pageblock_flags;
5470#endif /* CONFIG_SPARSEMEM */
5471}
5472
5473static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5474{
5475#ifdef CONFIG_SPARSEMEM
5476 pfn &= (PAGES_PER_SECTION-1);
5477 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5478#else
5479 pfn = pfn - zone->zone_start_pfn;
5480 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5481#endif /* CONFIG_SPARSEMEM */
5482}
5483
5484/**
5485 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5486 * @page: The page within the block of interest
5487 * @start_bitidx: The first bit of interest to retrieve
5488 * @end_bitidx: The last bit of interest
5489 * returns pageblock_bits flags
5490 */
5491unsigned long get_pageblock_flags_group(struct page *page,
5492 int start_bitidx, int end_bitidx)
5493{
5494 struct zone *zone;
5495 unsigned long *bitmap;
5496 unsigned long pfn, bitidx;
5497 unsigned long flags = 0;
5498 unsigned long value = 1;
5499
5500 zone = page_zone(page);
5501 pfn = page_to_pfn(page);
5502 bitmap = get_pageblock_bitmap(zone, pfn);
5503 bitidx = pfn_to_bitidx(zone, pfn);
5504
5505 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5506 if (test_bit(bitidx + start_bitidx, bitmap))
5507 flags |= value;
5508
5509 return flags;
5510}
5511
5512/**
5513 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5514 * @page: The page within the block of interest
5515 * @start_bitidx: The first bit of interest
5516 * @end_bitidx: The last bit of interest
5517 * @flags: The flags to set
5518 */
5519void set_pageblock_flags_group(struct page *page, unsigned long flags,
5520 int start_bitidx, int end_bitidx)
5521{
5522 struct zone *zone;
5523 unsigned long *bitmap;
5524 unsigned long pfn, bitidx;
5525 unsigned long value = 1;
5526
5527 zone = page_zone(page);
5528 pfn = page_to_pfn(page);
5529 bitmap = get_pageblock_bitmap(zone, pfn);
5530 bitidx = pfn_to_bitidx(zone, pfn);
5531 VM_BUG_ON(pfn < zone->zone_start_pfn);
5532 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5533
5534 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5535 if (flags & value)
5536 __set_bit(bitidx + start_bitidx, bitmap);
5537 else
5538 __clear_bit(bitidx + start_bitidx, bitmap);
5539}
5540
5541/*
5542 * This is designed as sub function...plz see page_isolation.c also.
5543 * set/clear page block's type to be ISOLATE.
5544 * page allocater never alloc memory from ISOLATE block.
5545 */
5546
5547static int
5548__count_immobile_pages(struct zone *zone, struct page *page, int count)
5549{
5550 unsigned long pfn, iter, found;
5551 /*
5552 * For avoiding noise data, lru_add_drain_all() should be called
5553 * If ZONE_MOVABLE, the zone never contains immobile pages
5554 */
5555 if (zone_idx(zone) == ZONE_MOVABLE)
5556 return true;
5557
5558 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5559 return true;
5560
5561 pfn = page_to_pfn(page);
5562 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5563 unsigned long check = pfn + iter;
5564
5565 if (!pfn_valid_within(check))
5566 continue;
5567
5568 page = pfn_to_page(check);
5569 if (!page_count(page)) {
5570 if (PageBuddy(page))
5571 iter += (1 << page_order(page)) - 1;
5572 continue;
5573 }
5574 if (!PageLRU(page))
5575 found++;
5576 /*
5577 * If there are RECLAIMABLE pages, we need to check it.
5578 * But now, memory offline itself doesn't call shrink_slab()
5579 * and it still to be fixed.
5580 */
5581 /*
5582 * If the page is not RAM, page_count()should be 0.
5583 * we don't need more check. This is an _used_ not-movable page.
5584 *
5585 * The problematic thing here is PG_reserved pages. PG_reserved
5586 * is set to both of a memory hole page and a _used_ kernel
5587 * page at boot.
5588 */
5589 if (found > count)
5590 return false;
5591 }
5592 return true;
5593}
5594
5595bool is_pageblock_removable_nolock(struct page *page)
5596{
5597 struct zone *zone = page_zone(page);
5598 return __count_immobile_pages(zone, page, 0);
5599}
5600
5601int set_migratetype_isolate(struct page *page)
5602{
5603 struct zone *zone;
5604 unsigned long flags, pfn;
5605 struct memory_isolate_notify arg;
5606 int notifier_ret;
5607 int ret = -EBUSY;
5608
5609 zone = page_zone(page);
5610
5611 spin_lock_irqsave(&zone->lock, flags);
5612
5613 pfn = page_to_pfn(page);
5614 arg.start_pfn = pfn;
5615 arg.nr_pages = pageblock_nr_pages;
5616 arg.pages_found = 0;
5617
5618 /*
5619 * It may be possible to isolate a pageblock even if the
5620 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5621 * notifier chain is used by balloon drivers to return the
5622 * number of pages in a range that are held by the balloon
5623 * driver to shrink memory. If all the pages are accounted for
5624 * by balloons, are free, or on the LRU, isolation can continue.
5625 * Later, for example, when memory hotplug notifier runs, these
5626 * pages reported as "can be isolated" should be isolated(freed)
5627 * by the balloon driver through the memory notifier chain.
5628 */
5629 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5630 notifier_ret = notifier_to_errno(notifier_ret);
5631 if (notifier_ret)
5632 goto out;
5633 /*
5634 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5635 * We just check MOVABLE pages.
5636 */
5637 if (__count_immobile_pages(zone, page, arg.pages_found))
5638 ret = 0;
5639
5640 /*
5641 * immobile means "not-on-lru" paes. If immobile is larger than
5642 * removable-by-driver pages reported by notifier, we'll fail.
5643 */
5644
5645out:
5646 if (!ret) {
5647 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5648 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5649 }
5650
5651 spin_unlock_irqrestore(&zone->lock, flags);
5652 if (!ret)
5653 drain_all_pages();
5654 return ret;
5655}
5656
5657void unset_migratetype_isolate(struct page *page)
5658{
5659 struct zone *zone;
5660 unsigned long flags;
5661 zone = page_zone(page);
5662 spin_lock_irqsave(&zone->lock, flags);
5663 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5664 goto out;
5665 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5666 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5667out:
5668 spin_unlock_irqrestore(&zone->lock, flags);
5669}
5670
5671#ifdef CONFIG_MEMORY_HOTREMOVE
5672/*
5673 * All pages in the range must be isolated before calling this.
5674 */
5675void
5676__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5677{
5678 struct page *page;
5679 struct zone *zone;
5680 int order, i;
5681 unsigned long pfn;
5682 unsigned long flags;
5683 /* find the first valid pfn */
5684 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5685 if (pfn_valid(pfn))
5686 break;
5687 if (pfn == end_pfn)
5688 return;
5689 zone = page_zone(pfn_to_page(pfn));
5690 spin_lock_irqsave(&zone->lock, flags);
5691 pfn = start_pfn;
5692 while (pfn < end_pfn) {
5693 if (!pfn_valid(pfn)) {
5694 pfn++;
5695 continue;
5696 }
5697 page = pfn_to_page(pfn);
5698 BUG_ON(page_count(page));
5699 BUG_ON(!PageBuddy(page));
5700 order = page_order(page);
5701#ifdef CONFIG_DEBUG_VM
5702 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5703 pfn, 1 << order, end_pfn);
5704#endif
5705 list_del(&page->lru);
5706 rmv_page_order(page);
5707 zone->free_area[order].nr_free--;
5708 __mod_zone_page_state(zone, NR_FREE_PAGES,
5709 - (1UL << order));
5710 for (i = 0; i < (1 << order); i++)
5711 SetPageReserved((page+i));
5712 pfn += (1 << order);
5713 }
5714 spin_unlock_irqrestore(&zone->lock, flags);
5715}
5716#endif
5717
5718#ifdef CONFIG_MEMORY_FAILURE
5719bool is_free_buddy_page(struct page *page)
5720{
5721 struct zone *zone = page_zone(page);
5722 unsigned long pfn = page_to_pfn(page);
5723 unsigned long flags;
5724 int order;
5725
5726 spin_lock_irqsave(&zone->lock, flags);
5727 for (order = 0; order < MAX_ORDER; order++) {
5728 struct page *page_head = page - (pfn & ((1 << order) - 1));
5729
5730 if (PageBuddy(page_head) && page_order(page_head) >= order)
5731 break;
5732 }
5733 spin_unlock_irqrestore(&zone->lock, flags);
5734
5735 return order < MAX_ORDER;
5736}
5737#endif
5738
5739static struct trace_print_flags pageflag_names[] = {
5740 {1UL << PG_locked, "locked" },
5741 {1UL << PG_error, "error" },
5742 {1UL << PG_referenced, "referenced" },
5743 {1UL << PG_uptodate, "uptodate" },
5744 {1UL << PG_dirty, "dirty" },
5745 {1UL << PG_lru, "lru" },
5746 {1UL << PG_active, "active" },
5747 {1UL << PG_slab, "slab" },
5748 {1UL << PG_owner_priv_1, "owner_priv_1" },
5749 {1UL << PG_arch_1, "arch_1" },
5750 {1UL << PG_reserved, "reserved" },
5751 {1UL << PG_private, "private" },
5752 {1UL << PG_private_2, "private_2" },
5753 {1UL << PG_writeback, "writeback" },
5754#ifdef CONFIG_PAGEFLAGS_EXTENDED
5755 {1UL << PG_head, "head" },
5756 {1UL << PG_tail, "tail" },
5757#else
5758 {1UL << PG_compound, "compound" },
5759#endif
5760 {1UL << PG_swapcache, "swapcache" },
5761 {1UL << PG_mappedtodisk, "mappedtodisk" },
5762 {1UL << PG_reclaim, "reclaim" },
5763 {1UL << PG_swapbacked, "swapbacked" },
5764 {1UL << PG_unevictable, "unevictable" },
5765#ifdef CONFIG_MMU
5766 {1UL << PG_mlocked, "mlocked" },
5767#endif
5768#ifdef CONFIG_ARCH_USES_PG_UNCACHED
5769 {1UL << PG_uncached, "uncached" },
5770#endif
5771#ifdef CONFIG_MEMORY_FAILURE
5772 {1UL << PG_hwpoison, "hwpoison" },
5773#endif
5774 {-1UL, NULL },
5775};
5776
5777static void dump_page_flags(unsigned long flags)
5778{
5779 const char *delim = "";
5780 unsigned long mask;
5781 int i;
5782
5783 printk(KERN_ALERT "page flags: %#lx(", flags);
5784
5785 /* remove zone id */
5786 flags &= (1UL << NR_PAGEFLAGS) - 1;
5787
5788 for (i = 0; pageflag_names[i].name && flags; i++) {
5789
5790 mask = pageflag_names[i].mask;
5791 if ((flags & mask) != mask)
5792 continue;
5793
5794 flags &= ~mask;
5795 printk("%s%s", delim, pageflag_names[i].name);
5796 delim = "|";
5797 }
5798
5799 /* check for left over flags */
5800 if (flags)
5801 printk("%s%#lx", delim, flags);
5802
5803 printk(")\n");
5804}
5805
5806void dump_page(struct page *page)
5807{
5808 printk(KERN_ALERT
5809 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5810 page, atomic_read(&page->_count), page_mapcount(page),
5811 page->mapping, page->index);
5812 dump_page_flags(page->flags);
5813 mem_cgroup_print_bad_page(page);
5814}
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#include <linux/padata.h>
72#include <linux/khugepaged.h>
73
74#include <asm/sections.h>
75#include <asm/tlbflush.h>
76#include <asm/div64.h>
77#include "internal.h"
78#include "shuffle.h"
79#include "page_reporting.h"
80
81/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
82static DEFINE_MUTEX(pcp_batch_high_lock);
83#define MIN_PERCPU_PAGELIST_FRACTION (8)
84
85#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
86DEFINE_PER_CPU(int, numa_node);
87EXPORT_PER_CPU_SYMBOL(numa_node);
88#endif
89
90DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
91
92#ifdef CONFIG_HAVE_MEMORYLESS_NODES
93/*
94 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
95 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
96 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
97 * defined in <linux/topology.h>.
98 */
99DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
100EXPORT_PER_CPU_SYMBOL(_numa_mem_);
101#endif
102
103/* work_structs for global per-cpu drains */
104struct pcpu_drain {
105 struct zone *zone;
106 struct work_struct work;
107};
108static DEFINE_MUTEX(pcpu_drain_mutex);
109static DEFINE_PER_CPU(struct pcpu_drain, pcpu_drain);
110
111#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
112volatile unsigned long latent_entropy __latent_entropy;
113EXPORT_SYMBOL(latent_entropy);
114#endif
115
116/*
117 * Array of node states.
118 */
119nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
120 [N_POSSIBLE] = NODE_MASK_ALL,
121 [N_ONLINE] = { { [0] = 1UL } },
122#ifndef CONFIG_NUMA
123 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
124#ifdef CONFIG_HIGHMEM
125 [N_HIGH_MEMORY] = { { [0] = 1UL } },
126#endif
127 [N_MEMORY] = { { [0] = 1UL } },
128 [N_CPU] = { { [0] = 1UL } },
129#endif /* NUMA */
130};
131EXPORT_SYMBOL(node_states);
132
133atomic_long_t _totalram_pages __read_mostly;
134EXPORT_SYMBOL(_totalram_pages);
135unsigned long totalreserve_pages __read_mostly;
136unsigned long totalcma_pages __read_mostly;
137
138int percpu_pagelist_fraction;
139gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
140#ifdef CONFIG_INIT_ON_ALLOC_DEFAULT_ON
141DEFINE_STATIC_KEY_TRUE(init_on_alloc);
142#else
143DEFINE_STATIC_KEY_FALSE(init_on_alloc);
144#endif
145EXPORT_SYMBOL(init_on_alloc);
146
147#ifdef CONFIG_INIT_ON_FREE_DEFAULT_ON
148DEFINE_STATIC_KEY_TRUE(init_on_free);
149#else
150DEFINE_STATIC_KEY_FALSE(init_on_free);
151#endif
152EXPORT_SYMBOL(init_on_free);
153
154static int __init early_init_on_alloc(char *buf)
155{
156 int ret;
157 bool bool_result;
158
159 if (!buf)
160 return -EINVAL;
161 ret = kstrtobool(buf, &bool_result);
162 if (bool_result && page_poisoning_enabled())
163 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_alloc\n");
164 if (bool_result)
165 static_branch_enable(&init_on_alloc);
166 else
167 static_branch_disable(&init_on_alloc);
168 return ret;
169}
170early_param("init_on_alloc", early_init_on_alloc);
171
172static int __init early_init_on_free(char *buf)
173{
174 int ret;
175 bool bool_result;
176
177 if (!buf)
178 return -EINVAL;
179 ret = kstrtobool(buf, &bool_result);
180 if (bool_result && page_poisoning_enabled())
181 pr_info("mem auto-init: CONFIG_PAGE_POISONING is on, will take precedence over init_on_free\n");
182 if (bool_result)
183 static_branch_enable(&init_on_free);
184 else
185 static_branch_disable(&init_on_free);
186 return ret;
187}
188early_param("init_on_free", early_init_on_free);
189
190/*
191 * A cached value of the page's pageblock's migratetype, used when the page is
192 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
193 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
194 * Also the migratetype set in the page does not necessarily match the pcplist
195 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
196 * other index - this ensures that it will be put on the correct CMA freelist.
197 */
198static inline int get_pcppage_migratetype(struct page *page)
199{
200 return page->index;
201}
202
203static inline void set_pcppage_migratetype(struct page *page, int migratetype)
204{
205 page->index = migratetype;
206}
207
208#ifdef CONFIG_PM_SLEEP
209/*
210 * The following functions are used by the suspend/hibernate code to temporarily
211 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
212 * while devices are suspended. To avoid races with the suspend/hibernate code,
213 * they should always be called with system_transition_mutex held
214 * (gfp_allowed_mask also should only be modified with system_transition_mutex
215 * held, unless the suspend/hibernate code is guaranteed not to run in parallel
216 * with that modification).
217 */
218
219static gfp_t saved_gfp_mask;
220
221void pm_restore_gfp_mask(void)
222{
223 WARN_ON(!mutex_is_locked(&system_transition_mutex));
224 if (saved_gfp_mask) {
225 gfp_allowed_mask = saved_gfp_mask;
226 saved_gfp_mask = 0;
227 }
228}
229
230void pm_restrict_gfp_mask(void)
231{
232 WARN_ON(!mutex_is_locked(&system_transition_mutex));
233 WARN_ON(saved_gfp_mask);
234 saved_gfp_mask = gfp_allowed_mask;
235 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
236}
237
238bool pm_suspended_storage(void)
239{
240 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
241 return false;
242 return true;
243}
244#endif /* CONFIG_PM_SLEEP */
245
246#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
247unsigned int pageblock_order __read_mostly;
248#endif
249
250static void __free_pages_ok(struct page *page, unsigned int order);
251
252/*
253 * results with 256, 32 in the lowmem_reserve sysctl:
254 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
255 * 1G machine -> (16M dma, 784M normal, 224M high)
256 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
257 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
258 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
259 *
260 * TBD: should special case ZONE_DMA32 machines here - in those we normally
261 * don't need any ZONE_NORMAL reservation
262 */
263int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
264#ifdef CONFIG_ZONE_DMA
265 [ZONE_DMA] = 256,
266#endif
267#ifdef CONFIG_ZONE_DMA32
268 [ZONE_DMA32] = 256,
269#endif
270 [ZONE_NORMAL] = 32,
271#ifdef CONFIG_HIGHMEM
272 [ZONE_HIGHMEM] = 0,
273#endif
274 [ZONE_MOVABLE] = 0,
275};
276
277static char * const zone_names[MAX_NR_ZONES] = {
278#ifdef CONFIG_ZONE_DMA
279 "DMA",
280#endif
281#ifdef CONFIG_ZONE_DMA32
282 "DMA32",
283#endif
284 "Normal",
285#ifdef CONFIG_HIGHMEM
286 "HighMem",
287#endif
288 "Movable",
289#ifdef CONFIG_ZONE_DEVICE
290 "Device",
291#endif
292};
293
294const char * const migratetype_names[MIGRATE_TYPES] = {
295 "Unmovable",
296 "Movable",
297 "Reclaimable",
298 "HighAtomic",
299#ifdef CONFIG_CMA
300 "CMA",
301#endif
302#ifdef CONFIG_MEMORY_ISOLATION
303 "Isolate",
304#endif
305};
306
307compound_page_dtor * const compound_page_dtors[NR_COMPOUND_DTORS] = {
308 [NULL_COMPOUND_DTOR] = NULL,
309 [COMPOUND_PAGE_DTOR] = free_compound_page,
310#ifdef CONFIG_HUGETLB_PAGE
311 [HUGETLB_PAGE_DTOR] = free_huge_page,
312#endif
313#ifdef CONFIG_TRANSPARENT_HUGEPAGE
314 [TRANSHUGE_PAGE_DTOR] = free_transhuge_page,
315#endif
316};
317
318int min_free_kbytes = 1024;
319int user_min_free_kbytes = -1;
320#ifdef CONFIG_DISCONTIGMEM
321/*
322 * DiscontigMem defines memory ranges as separate pg_data_t even if the ranges
323 * are not on separate NUMA nodes. Functionally this works but with
324 * watermark_boost_factor, it can reclaim prematurely as the ranges can be
325 * quite small. By default, do not boost watermarks on discontigmem as in
326 * many cases very high-order allocations like THP are likely to be
327 * unsupported and the premature reclaim offsets the advantage of long-term
328 * fragmentation avoidance.
329 */
330int watermark_boost_factor __read_mostly;
331#else
332int watermark_boost_factor __read_mostly = 15000;
333#endif
334int watermark_scale_factor = 10;
335
336static unsigned long nr_kernel_pages __initdata;
337static unsigned long nr_all_pages __initdata;
338static unsigned long dma_reserve __initdata;
339
340static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __initdata;
341static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __initdata;
342static unsigned long required_kernelcore __initdata;
343static unsigned long required_kernelcore_percent __initdata;
344static unsigned long required_movablecore __initdata;
345static unsigned long required_movablecore_percent __initdata;
346static unsigned long zone_movable_pfn[MAX_NUMNODES] __initdata;
347static bool mirrored_kernelcore __meminitdata;
348
349/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
350int movable_zone;
351EXPORT_SYMBOL(movable_zone);
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#else
464 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
465#endif /* CONFIG_SPARSEMEM */
466 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
467}
468
469/**
470 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
471 * @page: The page within the block of interest
472 * @pfn: The target page frame number
473 * @mask: mask of bits that the caller is interested in
474 *
475 * Return: pageblock_bits flags
476 */
477static __always_inline
478unsigned long __get_pfnblock_flags_mask(struct page *page,
479 unsigned long pfn,
480 unsigned long mask)
481{
482 unsigned long *bitmap;
483 unsigned long bitidx, word_bitidx;
484 unsigned long word;
485
486 bitmap = get_pageblock_bitmap(page, pfn);
487 bitidx = pfn_to_bitidx(page, pfn);
488 word_bitidx = bitidx / BITS_PER_LONG;
489 bitidx &= (BITS_PER_LONG-1);
490
491 word = bitmap[word_bitidx];
492 return (word >> bitidx) & mask;
493}
494
495unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
496 unsigned long mask)
497{
498 return __get_pfnblock_flags_mask(page, pfn, mask);
499}
500
501static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
502{
503 return __get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
504}
505
506/**
507 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
508 * @page: The page within the block of interest
509 * @flags: The flags to set
510 * @pfn: The target page frame number
511 * @mask: mask of bits that the caller is interested in
512 */
513void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
514 unsigned long pfn,
515 unsigned long mask)
516{
517 unsigned long *bitmap;
518 unsigned long bitidx, word_bitidx;
519 unsigned long old_word, word;
520
521 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
522 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
523
524 bitmap = get_pageblock_bitmap(page, pfn);
525 bitidx = pfn_to_bitidx(page, pfn);
526 word_bitidx = bitidx / BITS_PER_LONG;
527 bitidx &= (BITS_PER_LONG-1);
528
529 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
530
531 mask <<= bitidx;
532 flags <<= bitidx;
533
534 word = READ_ONCE(bitmap[word_bitidx]);
535 for (;;) {
536 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
537 if (word == old_word)
538 break;
539 word = old_word;
540 }
541}
542
543void set_pageblock_migratetype(struct page *page, int migratetype)
544{
545 if (unlikely(page_group_by_mobility_disabled &&
546 migratetype < MIGRATE_PCPTYPES))
547 migratetype = MIGRATE_UNMOVABLE;
548
549 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
550 page_to_pfn(page), MIGRATETYPE_MASK);
551}
552
553#ifdef CONFIG_DEBUG_VM
554static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
555{
556 int ret = 0;
557 unsigned seq;
558 unsigned long pfn = page_to_pfn(page);
559 unsigned long sp, start_pfn;
560
561 do {
562 seq = zone_span_seqbegin(zone);
563 start_pfn = zone->zone_start_pfn;
564 sp = zone->spanned_pages;
565 if (!zone_spans_pfn(zone, pfn))
566 ret = 1;
567 } while (zone_span_seqretry(zone, seq));
568
569 if (ret)
570 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
571 pfn, zone_to_nid(zone), zone->name,
572 start_pfn, start_pfn + sp);
573
574 return ret;
575}
576
577static int page_is_consistent(struct zone *zone, struct page *page)
578{
579 if (!pfn_valid_within(page_to_pfn(page)))
580 return 0;
581 if (zone != page_zone(page))
582 return 0;
583
584 return 1;
585}
586/*
587 * Temporary debugging check for pages not lying within a given zone.
588 */
589static int __maybe_unused bad_range(struct zone *zone, struct page *page)
590{
591 if (page_outside_zone_boundaries(zone, page))
592 return 1;
593 if (!page_is_consistent(zone, page))
594 return 1;
595
596 return 0;
597}
598#else
599static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
600{
601 return 0;
602}
603#endif
604
605static void bad_page(struct page *page, const char *reason)
606{
607 static unsigned long resume;
608 static unsigned long nr_shown;
609 static unsigned long nr_unshown;
610
611 /*
612 * Allow a burst of 60 reports, then keep quiet for that minute;
613 * or allow a steady drip of one report per second.
614 */
615 if (nr_shown == 60) {
616 if (time_before(jiffies, resume)) {
617 nr_unshown++;
618 goto out;
619 }
620 if (nr_unshown) {
621 pr_alert(
622 "BUG: Bad page state: %lu messages suppressed\n",
623 nr_unshown);
624 nr_unshown = 0;
625 }
626 nr_shown = 0;
627 }
628 if (nr_shown++ == 0)
629 resume = jiffies + 60 * HZ;
630
631 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
632 current->comm, page_to_pfn(page));
633 __dump_page(page, reason);
634 dump_page_owner(page);
635
636 print_modules();
637 dump_stack();
638out:
639 /* Leave bad fields for debug, except PageBuddy could make trouble */
640 page_mapcount_reset(page); /* remove PageBuddy */
641 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
642}
643
644/*
645 * Higher-order pages are called "compound pages". They are structured thusly:
646 *
647 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
648 *
649 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
650 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
651 *
652 * The first tail page's ->compound_dtor holds the offset in array of compound
653 * page destructors. See compound_page_dtors.
654 *
655 * The first tail page's ->compound_order holds the order of allocation.
656 * This usage means that zero-order pages may not be compound.
657 */
658
659void free_compound_page(struct page *page)
660{
661 mem_cgroup_uncharge(page);
662 __free_pages_ok(page, compound_order(page));
663}
664
665void prep_compound_page(struct page *page, unsigned int order)
666{
667 int i;
668 int nr_pages = 1 << order;
669
670 __SetPageHead(page);
671 for (i = 1; i < nr_pages; i++) {
672 struct page *p = page + i;
673 set_page_count(p, 0);
674 p->mapping = TAIL_MAPPING;
675 set_compound_head(p, page);
676 }
677
678 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
679 set_compound_order(page, order);
680 atomic_set(compound_mapcount_ptr(page), -1);
681 if (hpage_pincount_available(page))
682 atomic_set(compound_pincount_ptr(page), 0);
683}
684
685#ifdef CONFIG_DEBUG_PAGEALLOC
686unsigned int _debug_guardpage_minorder;
687
688bool _debug_pagealloc_enabled_early __read_mostly
689 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
690EXPORT_SYMBOL(_debug_pagealloc_enabled_early);
691DEFINE_STATIC_KEY_FALSE(_debug_pagealloc_enabled);
692EXPORT_SYMBOL(_debug_pagealloc_enabled);
693
694DEFINE_STATIC_KEY_FALSE(_debug_guardpage_enabled);
695
696static int __init early_debug_pagealloc(char *buf)
697{
698 return kstrtobool(buf, &_debug_pagealloc_enabled_early);
699}
700early_param("debug_pagealloc", early_debug_pagealloc);
701
702void init_debug_pagealloc(void)
703{
704 if (!debug_pagealloc_enabled())
705 return;
706
707 static_branch_enable(&_debug_pagealloc_enabled);
708
709 if (!debug_guardpage_minorder())
710 return;
711
712 static_branch_enable(&_debug_guardpage_enabled);
713}
714
715static int __init debug_guardpage_minorder_setup(char *buf)
716{
717 unsigned long res;
718
719 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
720 pr_err("Bad debug_guardpage_minorder value\n");
721 return 0;
722 }
723 _debug_guardpage_minorder = res;
724 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
725 return 0;
726}
727early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
728
729static inline bool set_page_guard(struct zone *zone, struct page *page,
730 unsigned int order, int migratetype)
731{
732 if (!debug_guardpage_enabled())
733 return false;
734
735 if (order >= debug_guardpage_minorder())
736 return false;
737
738 __SetPageGuard(page);
739 INIT_LIST_HEAD(&page->lru);
740 set_page_private(page, order);
741 /* Guard pages are not available for any usage */
742 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
743
744 return true;
745}
746
747static inline void clear_page_guard(struct zone *zone, struct page *page,
748 unsigned int order, int migratetype)
749{
750 if (!debug_guardpage_enabled())
751 return;
752
753 __ClearPageGuard(page);
754
755 set_page_private(page, 0);
756 if (!is_migrate_isolate(migratetype))
757 __mod_zone_freepage_state(zone, (1 << order), migratetype);
758}
759#else
760static inline bool set_page_guard(struct zone *zone, struct page *page,
761 unsigned int order, int migratetype) { return false; }
762static inline void clear_page_guard(struct zone *zone, struct page *page,
763 unsigned int order, int migratetype) {}
764#endif
765
766static inline void set_page_order(struct page *page, unsigned int order)
767{
768 set_page_private(page, order);
769 __SetPageBuddy(page);
770}
771
772/*
773 * This function checks whether a page is free && is the buddy
774 * we can coalesce a page and its buddy if
775 * (a) the buddy is not in a hole (check before calling!) &&
776 * (b) the buddy is in the buddy system &&
777 * (c) a page and its buddy have the same order &&
778 * (d) a page and its buddy are in the same zone.
779 *
780 * For recording whether a page is in the buddy system, we set PageBuddy.
781 * Setting, clearing, and testing PageBuddy is serialized by zone->lock.
782 *
783 * For recording page's order, we use page_private(page).
784 */
785static inline bool page_is_buddy(struct page *page, struct page *buddy,
786 unsigned int order)
787{
788 if (!page_is_guard(buddy) && !PageBuddy(buddy))
789 return false;
790
791 if (page_order(buddy) != order)
792 return false;
793
794 /*
795 * zone check is done late to avoid uselessly calculating
796 * zone/node ids for pages that could never merge.
797 */
798 if (page_zone_id(page) != page_zone_id(buddy))
799 return false;
800
801 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
802
803 return true;
804}
805
806#ifdef CONFIG_COMPACTION
807static inline struct capture_control *task_capc(struct zone *zone)
808{
809 struct capture_control *capc = current->capture_control;
810
811 return unlikely(capc) &&
812 !(current->flags & PF_KTHREAD) &&
813 !capc->page &&
814 capc->cc->zone == zone ? capc : NULL;
815}
816
817static inline bool
818compaction_capture(struct capture_control *capc, struct page *page,
819 int order, int migratetype)
820{
821 if (!capc || order != capc->cc->order)
822 return false;
823
824 /* Do not accidentally pollute CMA or isolated regions*/
825 if (is_migrate_cma(migratetype) ||
826 is_migrate_isolate(migratetype))
827 return false;
828
829 /*
830 * Do not let lower order allocations polluate a movable pageblock.
831 * This might let an unmovable request use a reclaimable pageblock
832 * and vice-versa but no more than normal fallback logic which can
833 * have trouble finding a high-order free page.
834 */
835 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
836 return false;
837
838 capc->page = page;
839 return true;
840}
841
842#else
843static inline struct capture_control *task_capc(struct zone *zone)
844{
845 return NULL;
846}
847
848static inline bool
849compaction_capture(struct capture_control *capc, struct page *page,
850 int order, int migratetype)
851{
852 return false;
853}
854#endif /* CONFIG_COMPACTION */
855
856/* Used for pages not on another list */
857static inline void add_to_free_list(struct page *page, struct zone *zone,
858 unsigned int order, int migratetype)
859{
860 struct free_area *area = &zone->free_area[order];
861
862 list_add(&page->lru, &area->free_list[migratetype]);
863 area->nr_free++;
864}
865
866/* Used for pages not on another list */
867static inline void add_to_free_list_tail(struct page *page, struct zone *zone,
868 unsigned int order, int migratetype)
869{
870 struct free_area *area = &zone->free_area[order];
871
872 list_add_tail(&page->lru, &area->free_list[migratetype]);
873 area->nr_free++;
874}
875
876/* Used for pages which are on another list */
877static inline void move_to_free_list(struct page *page, struct zone *zone,
878 unsigned int order, int migratetype)
879{
880 struct free_area *area = &zone->free_area[order];
881
882 list_move(&page->lru, &area->free_list[migratetype]);
883}
884
885static inline void del_page_from_free_list(struct page *page, struct zone *zone,
886 unsigned int order)
887{
888 /* clear reported state and update reported page count */
889 if (page_reported(page))
890 __ClearPageReported(page);
891
892 list_del(&page->lru);
893 __ClearPageBuddy(page);
894 set_page_private(page, 0);
895 zone->free_area[order].nr_free--;
896}
897
898/*
899 * If this is not the largest possible page, check if the buddy
900 * of the next-highest order is free. If it is, it's possible
901 * that pages are being freed that will coalesce soon. In case,
902 * that is happening, add the free page to the tail of the list
903 * so it's less likely to be used soon and more likely to be merged
904 * as a higher order page
905 */
906static inline bool
907buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
908 struct page *page, unsigned int order)
909{
910 struct page *higher_page, *higher_buddy;
911 unsigned long combined_pfn;
912
913 if (order >= MAX_ORDER - 2)
914 return false;
915
916 if (!pfn_valid_within(buddy_pfn))
917 return false;
918
919 combined_pfn = buddy_pfn & pfn;
920 higher_page = page + (combined_pfn - pfn);
921 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
922 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
923
924 return pfn_valid_within(buddy_pfn) &&
925 page_is_buddy(higher_page, higher_buddy, order + 1);
926}
927
928/*
929 * Freeing function for a buddy system allocator.
930 *
931 * The concept of a buddy system is to maintain direct-mapped table
932 * (containing bit values) for memory blocks of various "orders".
933 * The bottom level table contains the map for the smallest allocatable
934 * units of memory (here, pages), and each level above it describes
935 * pairs of units from the levels below, hence, "buddies".
936 * At a high level, all that happens here is marking the table entry
937 * at the bottom level available, and propagating the changes upward
938 * as necessary, plus some accounting needed to play nicely with other
939 * parts of the VM system.
940 * At each level, we keep a list of pages, which are heads of continuous
941 * free pages of length of (1 << order) and marked with PageBuddy.
942 * Page's order is recorded in page_private(page) field.
943 * So when we are allocating or freeing one, we can derive the state of the
944 * other. That is, if we allocate a small block, and both were
945 * free, the remainder of the region must be split into blocks.
946 * If a block is freed, and its buddy is also free, then this
947 * triggers coalescing into a block of larger size.
948 *
949 * -- nyc
950 */
951
952static inline void __free_one_page(struct page *page,
953 unsigned long pfn,
954 struct zone *zone, unsigned int order,
955 int migratetype, bool report)
956{
957 struct capture_control *capc = task_capc(zone);
958 unsigned long buddy_pfn;
959 unsigned long combined_pfn;
960 unsigned int max_order;
961 struct page *buddy;
962 bool to_tail;
963
964 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
965
966 VM_BUG_ON(!zone_is_initialized(zone));
967 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
968
969 VM_BUG_ON(migratetype == -1);
970 if (likely(!is_migrate_isolate(migratetype)))
971 __mod_zone_freepage_state(zone, 1 << order, migratetype);
972
973 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
974 VM_BUG_ON_PAGE(bad_range(zone, page), page);
975
976continue_merging:
977 while (order < max_order - 1) {
978 if (compaction_capture(capc, page, order, migratetype)) {
979 __mod_zone_freepage_state(zone, -(1 << order),
980 migratetype);
981 return;
982 }
983 buddy_pfn = __find_buddy_pfn(pfn, order);
984 buddy = page + (buddy_pfn - pfn);
985
986 if (!pfn_valid_within(buddy_pfn))
987 goto done_merging;
988 if (!page_is_buddy(page, buddy, order))
989 goto done_merging;
990 /*
991 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
992 * merge with it and move up one order.
993 */
994 if (page_is_guard(buddy))
995 clear_page_guard(zone, buddy, order, migratetype);
996 else
997 del_page_from_free_list(buddy, zone, order);
998 combined_pfn = buddy_pfn & pfn;
999 page = page + (combined_pfn - pfn);
1000 pfn = combined_pfn;
1001 order++;
1002 }
1003 if (max_order < MAX_ORDER) {
1004 /* If we are here, it means order is >= pageblock_order.
1005 * We want to prevent merge between freepages on isolate
1006 * pageblock and normal pageblock. Without this, pageblock
1007 * isolation could cause incorrect freepage or CMA accounting.
1008 *
1009 * We don't want to hit this code for the more frequent
1010 * low-order merging.
1011 */
1012 if (unlikely(has_isolate_pageblock(zone))) {
1013 int buddy_mt;
1014
1015 buddy_pfn = __find_buddy_pfn(pfn, order);
1016 buddy = page + (buddy_pfn - pfn);
1017 buddy_mt = get_pageblock_migratetype(buddy);
1018
1019 if (migratetype != buddy_mt
1020 && (is_migrate_isolate(migratetype) ||
1021 is_migrate_isolate(buddy_mt)))
1022 goto done_merging;
1023 }
1024 max_order++;
1025 goto continue_merging;
1026 }
1027
1028done_merging:
1029 set_page_order(page, order);
1030
1031 if (is_shuffle_order(order))
1032 to_tail = shuffle_pick_tail();
1033 else
1034 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
1035
1036 if (to_tail)
1037 add_to_free_list_tail(page, zone, order, migratetype);
1038 else
1039 add_to_free_list(page, zone, order, migratetype);
1040
1041 /* Notify page reporting subsystem of freed page */
1042 if (report)
1043 page_reporting_notify_free(order);
1044}
1045
1046/*
1047 * A bad page could be due to a number of fields. Instead of multiple branches,
1048 * try and check multiple fields with one check. The caller must do a detailed
1049 * check if necessary.
1050 */
1051static inline bool page_expected_state(struct page *page,
1052 unsigned long check_flags)
1053{
1054 if (unlikely(atomic_read(&page->_mapcount) != -1))
1055 return false;
1056
1057 if (unlikely((unsigned long)page->mapping |
1058 page_ref_count(page) |
1059#ifdef CONFIG_MEMCG
1060 (unsigned long)page->mem_cgroup |
1061#endif
1062 (page->flags & check_flags)))
1063 return false;
1064
1065 return true;
1066}
1067
1068static const char *page_bad_reason(struct page *page, unsigned long flags)
1069{
1070 const char *bad_reason = NULL;
1071
1072 if (unlikely(atomic_read(&page->_mapcount) != -1))
1073 bad_reason = "nonzero mapcount";
1074 if (unlikely(page->mapping != NULL))
1075 bad_reason = "non-NULL mapping";
1076 if (unlikely(page_ref_count(page) != 0))
1077 bad_reason = "nonzero _refcount";
1078 if (unlikely(page->flags & flags)) {
1079 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
1080 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
1081 else
1082 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
1083 }
1084#ifdef CONFIG_MEMCG
1085 if (unlikely(page->mem_cgroup))
1086 bad_reason = "page still charged to cgroup";
1087#endif
1088 return bad_reason;
1089}
1090
1091static void check_free_page_bad(struct page *page)
1092{
1093 bad_page(page,
1094 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
1095}
1096
1097static inline int check_free_page(struct page *page)
1098{
1099 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
1100 return 0;
1101
1102 /* Something has gone sideways, find it */
1103 check_free_page_bad(page);
1104 return 1;
1105}
1106
1107static int free_tail_pages_check(struct page *head_page, struct page *page)
1108{
1109 int ret = 1;
1110
1111 /*
1112 * We rely page->lru.next never has bit 0 set, unless the page
1113 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
1114 */
1115 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
1116
1117 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
1118 ret = 0;
1119 goto out;
1120 }
1121 switch (page - head_page) {
1122 case 1:
1123 /* the first tail page: ->mapping may be compound_mapcount() */
1124 if (unlikely(compound_mapcount(page))) {
1125 bad_page(page, "nonzero compound_mapcount");
1126 goto out;
1127 }
1128 break;
1129 case 2:
1130 /*
1131 * the second tail page: ->mapping is
1132 * deferred_list.next -- ignore value.
1133 */
1134 break;
1135 default:
1136 if (page->mapping != TAIL_MAPPING) {
1137 bad_page(page, "corrupted mapping in tail page");
1138 goto out;
1139 }
1140 break;
1141 }
1142 if (unlikely(!PageTail(page))) {
1143 bad_page(page, "PageTail not set");
1144 goto out;
1145 }
1146 if (unlikely(compound_head(page) != head_page)) {
1147 bad_page(page, "compound_head not consistent");
1148 goto out;
1149 }
1150 ret = 0;
1151out:
1152 page->mapping = NULL;
1153 clear_compound_head(page);
1154 return ret;
1155}
1156
1157static void kernel_init_free_pages(struct page *page, int numpages)
1158{
1159 int i;
1160
1161 /* s390's use of memset() could override KASAN redzones. */
1162 kasan_disable_current();
1163 for (i = 0; i < numpages; i++)
1164 clear_highpage(page + i);
1165 kasan_enable_current();
1166}
1167
1168static __always_inline bool free_pages_prepare(struct page *page,
1169 unsigned int order, bool check_free)
1170{
1171 int bad = 0;
1172
1173 VM_BUG_ON_PAGE(PageTail(page), page);
1174
1175 trace_mm_page_free(page, order);
1176
1177 /*
1178 * Check tail pages before head page information is cleared to
1179 * avoid checking PageCompound for order-0 pages.
1180 */
1181 if (unlikely(order)) {
1182 bool compound = PageCompound(page);
1183 int i;
1184
1185 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1186
1187 if (compound)
1188 ClearPageDoubleMap(page);
1189 for (i = 1; i < (1 << order); i++) {
1190 if (compound)
1191 bad += free_tail_pages_check(page, page + i);
1192 if (unlikely(check_free_page(page + i))) {
1193 bad++;
1194 continue;
1195 }
1196 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1197 }
1198 }
1199 if (PageMappingFlags(page))
1200 page->mapping = NULL;
1201 if (memcg_kmem_enabled() && PageKmemcg(page))
1202 __memcg_kmem_uncharge_page(page, order);
1203 if (check_free)
1204 bad += check_free_page(page);
1205 if (bad)
1206 return false;
1207
1208 page_cpupid_reset_last(page);
1209 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1210 reset_page_owner(page, order);
1211
1212 if (!PageHighMem(page)) {
1213 debug_check_no_locks_freed(page_address(page),
1214 PAGE_SIZE << order);
1215 debug_check_no_obj_freed(page_address(page),
1216 PAGE_SIZE << order);
1217 }
1218 if (want_init_on_free())
1219 kernel_init_free_pages(page, 1 << order);
1220
1221 kernel_poison_pages(page, 1 << order, 0);
1222 /*
1223 * arch_free_page() can make the page's contents inaccessible. s390
1224 * does this. So nothing which can access the page's contents should
1225 * happen after this.
1226 */
1227 arch_free_page(page, order);
1228
1229 if (debug_pagealloc_enabled_static())
1230 kernel_map_pages(page, 1 << order, 0);
1231
1232 kasan_free_nondeferred_pages(page, order);
1233
1234 return true;
1235}
1236
1237#ifdef CONFIG_DEBUG_VM
1238/*
1239 * With DEBUG_VM enabled, order-0 pages are checked immediately when being freed
1240 * to pcp lists. With debug_pagealloc also enabled, they are also rechecked when
1241 * moved from pcp lists to free lists.
1242 */
1243static bool free_pcp_prepare(struct page *page)
1244{
1245 return free_pages_prepare(page, 0, true);
1246}
1247
1248static bool bulkfree_pcp_prepare(struct page *page)
1249{
1250 if (debug_pagealloc_enabled_static())
1251 return check_free_page(page);
1252 else
1253 return false;
1254}
1255#else
1256/*
1257 * With DEBUG_VM disabled, order-0 pages being freed are checked only when
1258 * moving from pcp lists to free list in order to reduce overhead. With
1259 * debug_pagealloc enabled, they are checked also immediately when being freed
1260 * to the pcp lists.
1261 */
1262static bool free_pcp_prepare(struct page *page)
1263{
1264 if (debug_pagealloc_enabled_static())
1265 return free_pages_prepare(page, 0, true);
1266 else
1267 return free_pages_prepare(page, 0, false);
1268}
1269
1270static bool bulkfree_pcp_prepare(struct page *page)
1271{
1272 return check_free_page(page);
1273}
1274#endif /* CONFIG_DEBUG_VM */
1275
1276static inline void prefetch_buddy(struct page *page)
1277{
1278 unsigned long pfn = page_to_pfn(page);
1279 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1280 struct page *buddy = page + (buddy_pfn - pfn);
1281
1282 prefetch(buddy);
1283}
1284
1285/*
1286 * Frees a number of pages from the PCP lists
1287 * Assumes all pages on list are in same zone, and of same order.
1288 * count is the number of pages to free.
1289 *
1290 * If the zone was previously in an "all pages pinned" state then look to
1291 * see if this freeing clears that state.
1292 *
1293 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1294 * pinned" detection logic.
1295 */
1296static void free_pcppages_bulk(struct zone *zone, int count,
1297 struct per_cpu_pages *pcp)
1298{
1299 int migratetype = 0;
1300 int batch_free = 0;
1301 int prefetch_nr = 0;
1302 bool isolated_pageblocks;
1303 struct page *page, *tmp;
1304 LIST_HEAD(head);
1305
1306 /*
1307 * Ensure proper count is passed which otherwise would stuck in the
1308 * below while (list_empty(list)) loop.
1309 */
1310 count = min(pcp->count, count);
1311 while (count) {
1312 struct list_head *list;
1313
1314 /*
1315 * Remove pages from lists in a round-robin fashion. A
1316 * batch_free count is maintained that is incremented when an
1317 * empty list is encountered. This is so more pages are freed
1318 * off fuller lists instead of spinning excessively around empty
1319 * lists
1320 */
1321 do {
1322 batch_free++;
1323 if (++migratetype == MIGRATE_PCPTYPES)
1324 migratetype = 0;
1325 list = &pcp->lists[migratetype];
1326 } while (list_empty(list));
1327
1328 /* This is the only non-empty list. Free them all. */
1329 if (batch_free == MIGRATE_PCPTYPES)
1330 batch_free = count;
1331
1332 do {
1333 page = list_last_entry(list, struct page, lru);
1334 /* must delete to avoid corrupting pcp list */
1335 list_del(&page->lru);
1336 pcp->count--;
1337
1338 if (bulkfree_pcp_prepare(page))
1339 continue;
1340
1341 list_add_tail(&page->lru, &head);
1342
1343 /*
1344 * We are going to put the page back to the global
1345 * pool, prefetch its buddy to speed up later access
1346 * under zone->lock. It is believed the overhead of
1347 * an additional test and calculating buddy_pfn here
1348 * can be offset by reduced memory latency later. To
1349 * avoid excessive prefetching due to large count, only
1350 * prefetch buddy for the first pcp->batch nr of pages.
1351 */
1352 if (prefetch_nr++ < pcp->batch)
1353 prefetch_buddy(page);
1354 } while (--count && --batch_free && !list_empty(list));
1355 }
1356
1357 spin_lock(&zone->lock);
1358 isolated_pageblocks = has_isolate_pageblock(zone);
1359
1360 /*
1361 * Use safe version since after __free_one_page(),
1362 * page->lru.next will not point to original list.
1363 */
1364 list_for_each_entry_safe(page, tmp, &head, lru) {
1365 int mt = get_pcppage_migratetype(page);
1366 /* MIGRATE_ISOLATE page should not go to pcplists */
1367 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1368 /* Pageblock could have been isolated meanwhile */
1369 if (unlikely(isolated_pageblocks))
1370 mt = get_pageblock_migratetype(page);
1371
1372 __free_one_page(page, page_to_pfn(page), zone, 0, mt, true);
1373 trace_mm_page_pcpu_drain(page, 0, mt);
1374 }
1375 spin_unlock(&zone->lock);
1376}
1377
1378static void free_one_page(struct zone *zone,
1379 struct page *page, unsigned long pfn,
1380 unsigned int order,
1381 int migratetype)
1382{
1383 spin_lock(&zone->lock);
1384 if (unlikely(has_isolate_pageblock(zone) ||
1385 is_migrate_isolate(migratetype))) {
1386 migratetype = get_pfnblock_migratetype(page, pfn);
1387 }
1388 __free_one_page(page, pfn, zone, order, migratetype, true);
1389 spin_unlock(&zone->lock);
1390}
1391
1392static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1393 unsigned long zone, int nid)
1394{
1395 mm_zero_struct_page(page);
1396 set_page_links(page, zone, nid, pfn);
1397 init_page_count(page);
1398 page_mapcount_reset(page);
1399 page_cpupid_reset_last(page);
1400 page_kasan_tag_reset(page);
1401
1402 INIT_LIST_HEAD(&page->lru);
1403#ifdef WANT_PAGE_VIRTUAL
1404 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1405 if (!is_highmem_idx(zone))
1406 set_page_address(page, __va(pfn << PAGE_SHIFT));
1407#endif
1408}
1409
1410#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1411static void __meminit init_reserved_page(unsigned long pfn)
1412{
1413 pg_data_t *pgdat;
1414 int nid, zid;
1415
1416 if (!early_page_uninitialised(pfn))
1417 return;
1418
1419 nid = early_pfn_to_nid(pfn);
1420 pgdat = NODE_DATA(nid);
1421
1422 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1423 struct zone *zone = &pgdat->node_zones[zid];
1424
1425 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1426 break;
1427 }
1428 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1429}
1430#else
1431static inline void init_reserved_page(unsigned long pfn)
1432{
1433}
1434#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1435
1436/*
1437 * Initialised pages do not have PageReserved set. This function is
1438 * called for each range allocated by the bootmem allocator and
1439 * marks the pages PageReserved. The remaining valid pages are later
1440 * sent to the buddy page allocator.
1441 */
1442void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1443{
1444 unsigned long start_pfn = PFN_DOWN(start);
1445 unsigned long end_pfn = PFN_UP(end);
1446
1447 for (; start_pfn < end_pfn; start_pfn++) {
1448 if (pfn_valid(start_pfn)) {
1449 struct page *page = pfn_to_page(start_pfn);
1450
1451 init_reserved_page(start_pfn);
1452
1453 /* Avoid false-positive PageTail() */
1454 INIT_LIST_HEAD(&page->lru);
1455
1456 /*
1457 * no need for atomic set_bit because the struct
1458 * page is not visible yet so nobody should
1459 * access it yet.
1460 */
1461 __SetPageReserved(page);
1462 }
1463 }
1464}
1465
1466static void __free_pages_ok(struct page *page, unsigned int order)
1467{
1468 unsigned long flags;
1469 int migratetype;
1470 unsigned long pfn = page_to_pfn(page);
1471
1472 if (!free_pages_prepare(page, order, true))
1473 return;
1474
1475 migratetype = get_pfnblock_migratetype(page, pfn);
1476 local_irq_save(flags);
1477 __count_vm_events(PGFREE, 1 << order);
1478 free_one_page(page_zone(page), page, pfn, order, migratetype);
1479 local_irq_restore(flags);
1480}
1481
1482void __free_pages_core(struct page *page, unsigned int order)
1483{
1484 unsigned int nr_pages = 1 << order;
1485 struct page *p = page;
1486 unsigned int loop;
1487
1488 prefetchw(p);
1489 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1490 prefetchw(p + 1);
1491 __ClearPageReserved(p);
1492 set_page_count(p, 0);
1493 }
1494 __ClearPageReserved(p);
1495 set_page_count(p, 0);
1496
1497 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1498 set_page_refcounted(page);
1499 __free_pages(page, order);
1500}
1501
1502#ifdef CONFIG_NEED_MULTIPLE_NODES
1503
1504static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1505
1506#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
1507
1508/*
1509 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
1510 */
1511int __meminit __early_pfn_to_nid(unsigned long pfn,
1512 struct mminit_pfnnid_cache *state)
1513{
1514 unsigned long start_pfn, end_pfn;
1515 int nid;
1516
1517 if (state->last_start <= pfn && pfn < state->last_end)
1518 return state->last_nid;
1519
1520 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
1521 if (nid != NUMA_NO_NODE) {
1522 state->last_start = start_pfn;
1523 state->last_end = end_pfn;
1524 state->last_nid = nid;
1525 }
1526
1527 return nid;
1528}
1529#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
1530
1531int __meminit early_pfn_to_nid(unsigned long pfn)
1532{
1533 static DEFINE_SPINLOCK(early_pfn_lock);
1534 int nid;
1535
1536 spin_lock(&early_pfn_lock);
1537 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1538 if (nid < 0)
1539 nid = first_online_node;
1540 spin_unlock(&early_pfn_lock);
1541
1542 return nid;
1543}
1544#endif /* CONFIG_NEED_MULTIPLE_NODES */
1545
1546void __init memblock_free_pages(struct page *page, unsigned long pfn,
1547 unsigned int order)
1548{
1549 if (early_page_uninitialised(pfn))
1550 return;
1551 __free_pages_core(page, order);
1552}
1553
1554/*
1555 * Check that the whole (or subset of) a pageblock given by the interval of
1556 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1557 * with the migration of free compaction scanner. The scanners then need to
1558 * use only pfn_valid_within() check for arches that allow holes within
1559 * pageblocks.
1560 *
1561 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1562 *
1563 * It's possible on some configurations to have a setup like node0 node1 node0
1564 * i.e. it's possible that all pages within a zones range of pages do not
1565 * belong to a single zone. We assume that a border between node0 and node1
1566 * can occur within a single pageblock, but not a node0 node1 node0
1567 * interleaving within a single pageblock. It is therefore sufficient to check
1568 * the first and last page of a pageblock and avoid checking each individual
1569 * page in a pageblock.
1570 */
1571struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1572 unsigned long end_pfn, struct zone *zone)
1573{
1574 struct page *start_page;
1575 struct page *end_page;
1576
1577 /* end_pfn is one past the range we are checking */
1578 end_pfn--;
1579
1580 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1581 return NULL;
1582
1583 start_page = pfn_to_online_page(start_pfn);
1584 if (!start_page)
1585 return NULL;
1586
1587 if (page_zone(start_page) != zone)
1588 return NULL;
1589
1590 end_page = pfn_to_page(end_pfn);
1591
1592 /* This gives a shorter code than deriving page_zone(end_page) */
1593 if (page_zone_id(start_page) != page_zone_id(end_page))
1594 return NULL;
1595
1596 return start_page;
1597}
1598
1599void set_zone_contiguous(struct zone *zone)
1600{
1601 unsigned long block_start_pfn = zone->zone_start_pfn;
1602 unsigned long block_end_pfn;
1603
1604 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1605 for (; block_start_pfn < zone_end_pfn(zone);
1606 block_start_pfn = block_end_pfn,
1607 block_end_pfn += pageblock_nr_pages) {
1608
1609 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1610
1611 if (!__pageblock_pfn_to_page(block_start_pfn,
1612 block_end_pfn, zone))
1613 return;
1614 cond_resched();
1615 }
1616
1617 /* We confirm that there is no hole */
1618 zone->contiguous = true;
1619}
1620
1621void clear_zone_contiguous(struct zone *zone)
1622{
1623 zone->contiguous = false;
1624}
1625
1626#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1627static void __init deferred_free_range(unsigned long pfn,
1628 unsigned long nr_pages)
1629{
1630 struct page *page;
1631 unsigned long i;
1632
1633 if (!nr_pages)
1634 return;
1635
1636 page = pfn_to_page(pfn);
1637
1638 /* Free a large naturally-aligned chunk if possible */
1639 if (nr_pages == pageblock_nr_pages &&
1640 (pfn & (pageblock_nr_pages - 1)) == 0) {
1641 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1642 __free_pages_core(page, pageblock_order);
1643 return;
1644 }
1645
1646 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1647 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1648 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1649 __free_pages_core(page, 0);
1650 }
1651}
1652
1653/* Completion tracking for deferred_init_memmap() threads */
1654static atomic_t pgdat_init_n_undone __initdata;
1655static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1656
1657static inline void __init pgdat_init_report_one_done(void)
1658{
1659 if (atomic_dec_and_test(&pgdat_init_n_undone))
1660 complete(&pgdat_init_all_done_comp);
1661}
1662
1663/*
1664 * Returns true if page needs to be initialized or freed to buddy allocator.
1665 *
1666 * First we check if pfn is valid on architectures where it is possible to have
1667 * holes within pageblock_nr_pages. On systems where it is not possible, this
1668 * function is optimized out.
1669 *
1670 * Then, we check if a current large page is valid by only checking the validity
1671 * of the head pfn.
1672 */
1673static inline bool __init deferred_pfn_valid(unsigned long pfn)
1674{
1675 if (!pfn_valid_within(pfn))
1676 return false;
1677 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1678 return false;
1679 return true;
1680}
1681
1682/*
1683 * Free pages to buddy allocator. Try to free aligned pages in
1684 * pageblock_nr_pages sizes.
1685 */
1686static void __init deferred_free_pages(unsigned long pfn,
1687 unsigned long end_pfn)
1688{
1689 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1690 unsigned long nr_free = 0;
1691
1692 for (; pfn < end_pfn; pfn++) {
1693 if (!deferred_pfn_valid(pfn)) {
1694 deferred_free_range(pfn - nr_free, nr_free);
1695 nr_free = 0;
1696 } else if (!(pfn & nr_pgmask)) {
1697 deferred_free_range(pfn - nr_free, nr_free);
1698 nr_free = 1;
1699 } else {
1700 nr_free++;
1701 }
1702 }
1703 /* Free the last block of pages to allocator */
1704 deferred_free_range(pfn - nr_free, nr_free);
1705}
1706
1707/*
1708 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1709 * by performing it only once every pageblock_nr_pages.
1710 * Return number of pages initialized.
1711 */
1712static unsigned long __init deferred_init_pages(struct zone *zone,
1713 unsigned long pfn,
1714 unsigned long end_pfn)
1715{
1716 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1717 int nid = zone_to_nid(zone);
1718 unsigned long nr_pages = 0;
1719 int zid = zone_idx(zone);
1720 struct page *page = NULL;
1721
1722 for (; pfn < end_pfn; pfn++) {
1723 if (!deferred_pfn_valid(pfn)) {
1724 page = NULL;
1725 continue;
1726 } else if (!page || !(pfn & nr_pgmask)) {
1727 page = pfn_to_page(pfn);
1728 } else {
1729 page++;
1730 }
1731 __init_single_page(page, pfn, zid, nid);
1732 nr_pages++;
1733 }
1734 return (nr_pages);
1735}
1736
1737/*
1738 * This function is meant to pre-load the iterator for the zone init.
1739 * Specifically it walks through the ranges until we are caught up to the
1740 * first_init_pfn value and exits there. If we never encounter the value we
1741 * return false indicating there are no valid ranges left.
1742 */
1743static bool __init
1744deferred_init_mem_pfn_range_in_zone(u64 *i, struct zone *zone,
1745 unsigned long *spfn, unsigned long *epfn,
1746 unsigned long first_init_pfn)
1747{
1748 u64 j;
1749
1750 /*
1751 * Start out by walking through the ranges in this zone that have
1752 * already been initialized. We don't need to do anything with them
1753 * so we just need to flush them out of the system.
1754 */
1755 for_each_free_mem_pfn_range_in_zone(j, zone, spfn, epfn) {
1756 if (*epfn <= first_init_pfn)
1757 continue;
1758 if (*spfn < first_init_pfn)
1759 *spfn = first_init_pfn;
1760 *i = j;
1761 return true;
1762 }
1763
1764 return false;
1765}
1766
1767/*
1768 * Initialize and free pages. We do it in two loops: first we initialize
1769 * struct page, then free to buddy allocator, because while we are
1770 * freeing pages we can access pages that are ahead (computing buddy
1771 * page in __free_one_page()).
1772 *
1773 * In order to try and keep some memory in the cache we have the loop
1774 * broken along max page order boundaries. This way we will not cause
1775 * any issues with the buddy page computation.
1776 */
1777static unsigned long __init
1778deferred_init_maxorder(u64 *i, struct zone *zone, unsigned long *start_pfn,
1779 unsigned long *end_pfn)
1780{
1781 unsigned long mo_pfn = ALIGN(*start_pfn + 1, MAX_ORDER_NR_PAGES);
1782 unsigned long spfn = *start_pfn, epfn = *end_pfn;
1783 unsigned long nr_pages = 0;
1784 u64 j = *i;
1785
1786 /* First we loop through and initialize the page values */
1787 for_each_free_mem_pfn_range_in_zone_from(j, zone, start_pfn, end_pfn) {
1788 unsigned long t;
1789
1790 if (mo_pfn <= *start_pfn)
1791 break;
1792
1793 t = min(mo_pfn, *end_pfn);
1794 nr_pages += deferred_init_pages(zone, *start_pfn, t);
1795
1796 if (mo_pfn < *end_pfn) {
1797 *start_pfn = mo_pfn;
1798 break;
1799 }
1800 }
1801
1802 /* Reset values and now loop through freeing pages as needed */
1803 swap(j, *i);
1804
1805 for_each_free_mem_pfn_range_in_zone_from(j, zone, &spfn, &epfn) {
1806 unsigned long t;
1807
1808 if (mo_pfn <= spfn)
1809 break;
1810
1811 t = min(mo_pfn, epfn);
1812 deferred_free_pages(spfn, t);
1813
1814 if (mo_pfn <= epfn)
1815 break;
1816 }
1817
1818 return nr_pages;
1819}
1820
1821static void __init
1822deferred_init_memmap_chunk(unsigned long start_pfn, unsigned long end_pfn,
1823 void *arg)
1824{
1825 unsigned long spfn, epfn;
1826 struct zone *zone = arg;
1827 u64 i;
1828
1829 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn, start_pfn);
1830
1831 /*
1832 * Initialize and free pages in MAX_ORDER sized increments so that we
1833 * can avoid introducing any issues with the buddy allocator.
1834 */
1835 while (spfn < end_pfn) {
1836 deferred_init_maxorder(&i, zone, &spfn, &epfn);
1837 cond_resched();
1838 }
1839}
1840
1841/* An arch may override for more concurrency. */
1842__weak int __init
1843deferred_page_init_max_threads(const struct cpumask *node_cpumask)
1844{
1845 return 1;
1846}
1847
1848/* Initialise remaining memory on a node */
1849static int __init deferred_init_memmap(void *data)
1850{
1851 pg_data_t *pgdat = data;
1852 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1853 unsigned long spfn = 0, epfn = 0;
1854 unsigned long first_init_pfn, flags;
1855 unsigned long start = jiffies;
1856 struct zone *zone;
1857 int zid, max_threads;
1858 u64 i;
1859
1860 /* Bind memory initialisation thread to a local node if possible */
1861 if (!cpumask_empty(cpumask))
1862 set_cpus_allowed_ptr(current, cpumask);
1863
1864 pgdat_resize_lock(pgdat, &flags);
1865 first_init_pfn = pgdat->first_deferred_pfn;
1866 if (first_init_pfn == ULONG_MAX) {
1867 pgdat_resize_unlock(pgdat, &flags);
1868 pgdat_init_report_one_done();
1869 return 0;
1870 }
1871
1872 /* Sanity check boundaries */
1873 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1874 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1875 pgdat->first_deferred_pfn = ULONG_MAX;
1876
1877 /*
1878 * Once we unlock here, the zone cannot be grown anymore, thus if an
1879 * interrupt thread must allocate this early in boot, zone must be
1880 * pre-grown prior to start of deferred page initialization.
1881 */
1882 pgdat_resize_unlock(pgdat, &flags);
1883
1884 /* Only the highest zone is deferred so find it */
1885 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1886 zone = pgdat->node_zones + zid;
1887 if (first_init_pfn < zone_end_pfn(zone))
1888 break;
1889 }
1890
1891 /* If the zone is empty somebody else may have cleared out the zone */
1892 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1893 first_init_pfn))
1894 goto zone_empty;
1895
1896 max_threads = deferred_page_init_max_threads(cpumask);
1897
1898 while (spfn < epfn) {
1899 unsigned long epfn_align = ALIGN(epfn, PAGES_PER_SECTION);
1900 struct padata_mt_job job = {
1901 .thread_fn = deferred_init_memmap_chunk,
1902 .fn_arg = zone,
1903 .start = spfn,
1904 .size = epfn_align - spfn,
1905 .align = PAGES_PER_SECTION,
1906 .min_chunk = PAGES_PER_SECTION,
1907 .max_threads = max_threads,
1908 };
1909
1910 padata_do_multithreaded(&job);
1911 deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1912 epfn_align);
1913 }
1914zone_empty:
1915 /* Sanity check that the next zone really is unpopulated */
1916 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1917
1918 pr_info("node %d deferred pages initialised in %ums\n",
1919 pgdat->node_id, jiffies_to_msecs(jiffies - start));
1920
1921 pgdat_init_report_one_done();
1922 return 0;
1923}
1924
1925/*
1926 * If this zone has deferred pages, try to grow it by initializing enough
1927 * deferred pages to satisfy the allocation specified by order, rounded up to
1928 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1929 * of SECTION_SIZE bytes by initializing struct pages in increments of
1930 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1931 *
1932 * Return true when zone was grown, otherwise return false. We return true even
1933 * when we grow less than requested, to let the caller decide if there are
1934 * enough pages to satisfy the allocation.
1935 *
1936 * Note: We use noinline because this function is needed only during boot, and
1937 * it is called from a __ref function _deferred_grow_zone. This way we are
1938 * making sure that it is not inlined into permanent text section.
1939 */
1940static noinline bool __init
1941deferred_grow_zone(struct zone *zone, unsigned int order)
1942{
1943 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1944 pg_data_t *pgdat = zone->zone_pgdat;
1945 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1946 unsigned long spfn, epfn, flags;
1947 unsigned long nr_pages = 0;
1948 u64 i;
1949
1950 /* Only the last zone may have deferred pages */
1951 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1952 return false;
1953
1954 pgdat_resize_lock(pgdat, &flags);
1955
1956 /*
1957 * If someone grew this zone while we were waiting for spinlock, return
1958 * true, as there might be enough pages already.
1959 */
1960 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1961 pgdat_resize_unlock(pgdat, &flags);
1962 return true;
1963 }
1964
1965 /* If the zone is empty somebody else may have cleared out the zone */
1966 if (!deferred_init_mem_pfn_range_in_zone(&i, zone, &spfn, &epfn,
1967 first_deferred_pfn)) {
1968 pgdat->first_deferred_pfn = ULONG_MAX;
1969 pgdat_resize_unlock(pgdat, &flags);
1970 /* Retry only once. */
1971 return first_deferred_pfn != ULONG_MAX;
1972 }
1973
1974 /*
1975 * Initialize and free pages in MAX_ORDER sized increments so
1976 * that we can avoid introducing any issues with the buddy
1977 * allocator.
1978 */
1979 while (spfn < epfn) {
1980 /* update our first deferred PFN for this section */
1981 first_deferred_pfn = spfn;
1982
1983 nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
1984 touch_nmi_watchdog();
1985
1986 /* We should only stop along section boundaries */
1987 if ((first_deferred_pfn ^ spfn) < PAGES_PER_SECTION)
1988 continue;
1989
1990 /* If our quota has been met we can stop here */
1991 if (nr_pages >= nr_pages_needed)
1992 break;
1993 }
1994
1995 pgdat->first_deferred_pfn = spfn;
1996 pgdat_resize_unlock(pgdat, &flags);
1997
1998 return nr_pages > 0;
1999}
2000
2001/*
2002 * deferred_grow_zone() is __init, but it is called from
2003 * get_page_from_freelist() during early boot until deferred_pages permanently
2004 * disables this call. This is why we have refdata wrapper to avoid warning,
2005 * and to ensure that the function body gets unloaded.
2006 */
2007static bool __ref
2008_deferred_grow_zone(struct zone *zone, unsigned int order)
2009{
2010 return deferred_grow_zone(zone, order);
2011}
2012
2013#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
2014
2015void __init page_alloc_init_late(void)
2016{
2017 struct zone *zone;
2018 int nid;
2019
2020#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
2021
2022 /* There will be num_node_state(N_MEMORY) threads */
2023 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
2024 for_each_node_state(nid, N_MEMORY) {
2025 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
2026 }
2027
2028 /* Block until all are initialised */
2029 wait_for_completion(&pgdat_init_all_done_comp);
2030
2031 /*
2032 * The number of managed pages has changed due to the initialisation
2033 * so the pcpu batch and high limits needs to be updated or the limits
2034 * will be artificially small.
2035 */
2036 for_each_populated_zone(zone)
2037 zone_pcp_update(zone);
2038
2039 /*
2040 * We initialized the rest of the deferred pages. Permanently disable
2041 * on-demand struct page initialization.
2042 */
2043 static_branch_disable(&deferred_pages);
2044
2045 /* Reinit limits that are based on free pages after the kernel is up */
2046 files_maxfiles_init();
2047#endif
2048
2049 /* Discard memblock private memory */
2050 memblock_discard();
2051
2052 for_each_node_state(nid, N_MEMORY)
2053 shuffle_free_memory(NODE_DATA(nid));
2054
2055 for_each_populated_zone(zone)
2056 set_zone_contiguous(zone);
2057}
2058
2059#ifdef CONFIG_CMA
2060/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
2061void __init init_cma_reserved_pageblock(struct page *page)
2062{
2063 unsigned i = pageblock_nr_pages;
2064 struct page *p = page;
2065
2066 do {
2067 __ClearPageReserved(p);
2068 set_page_count(p, 0);
2069 } while (++p, --i);
2070
2071 set_pageblock_migratetype(page, MIGRATE_CMA);
2072
2073 if (pageblock_order >= MAX_ORDER) {
2074 i = pageblock_nr_pages;
2075 p = page;
2076 do {
2077 set_page_refcounted(p);
2078 __free_pages(p, MAX_ORDER - 1);
2079 p += MAX_ORDER_NR_PAGES;
2080 } while (i -= MAX_ORDER_NR_PAGES);
2081 } else {
2082 set_page_refcounted(page);
2083 __free_pages(page, pageblock_order);
2084 }
2085
2086 adjust_managed_page_count(page, pageblock_nr_pages);
2087}
2088#endif
2089
2090/*
2091 * The order of subdivision here is critical for the IO subsystem.
2092 * Please do not alter this order without good reasons and regression
2093 * testing. Specifically, as large blocks of memory are subdivided,
2094 * the order in which smaller blocks are delivered depends on the order
2095 * they're subdivided in this function. This is the primary factor
2096 * influencing the order in which pages are delivered to the IO
2097 * subsystem according to empirical testing, and this is also justified
2098 * by considering the behavior of a buddy system containing a single
2099 * large block of memory acted on by a series of small allocations.
2100 * This behavior is a critical factor in sglist merging's success.
2101 *
2102 * -- nyc
2103 */
2104static inline void expand(struct zone *zone, struct page *page,
2105 int low, int high, int migratetype)
2106{
2107 unsigned long size = 1 << high;
2108
2109 while (high > low) {
2110 high--;
2111 size >>= 1;
2112 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
2113
2114 /*
2115 * Mark as guard pages (or page), that will allow to
2116 * merge back to allocator when buddy will be freed.
2117 * Corresponding page table entries will not be touched,
2118 * pages will stay not present in virtual address space
2119 */
2120 if (set_page_guard(zone, &page[size], high, migratetype))
2121 continue;
2122
2123 add_to_free_list(&page[size], zone, high, migratetype);
2124 set_page_order(&page[size], high);
2125 }
2126}
2127
2128static void check_new_page_bad(struct page *page)
2129{
2130 if (unlikely(page->flags & __PG_HWPOISON)) {
2131 /* Don't complain about hwpoisoned pages */
2132 page_mapcount_reset(page); /* remove PageBuddy */
2133 return;
2134 }
2135
2136 bad_page(page,
2137 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
2138}
2139
2140/*
2141 * This page is about to be returned from the page allocator
2142 */
2143static inline int check_new_page(struct page *page)
2144{
2145 if (likely(page_expected_state(page,
2146 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
2147 return 0;
2148
2149 check_new_page_bad(page);
2150 return 1;
2151}
2152
2153static inline bool free_pages_prezeroed(void)
2154{
2155 return (IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
2156 page_poisoning_enabled()) || want_init_on_free();
2157}
2158
2159#ifdef CONFIG_DEBUG_VM
2160/*
2161 * With DEBUG_VM enabled, order-0 pages are checked for expected state when
2162 * being allocated from pcp lists. With debug_pagealloc also enabled, they are
2163 * also checked when pcp lists are refilled from the free lists.
2164 */
2165static inline bool check_pcp_refill(struct page *page)
2166{
2167 if (debug_pagealloc_enabled_static())
2168 return check_new_page(page);
2169 else
2170 return false;
2171}
2172
2173static inline bool check_new_pcp(struct page *page)
2174{
2175 return check_new_page(page);
2176}
2177#else
2178/*
2179 * With DEBUG_VM disabled, free order-0 pages are checked for expected state
2180 * when pcp lists are being refilled from the free lists. With debug_pagealloc
2181 * enabled, they are also checked when being allocated from the pcp lists.
2182 */
2183static inline bool check_pcp_refill(struct page *page)
2184{
2185 return check_new_page(page);
2186}
2187static inline bool check_new_pcp(struct page *page)
2188{
2189 if (debug_pagealloc_enabled_static())
2190 return check_new_page(page);
2191 else
2192 return false;
2193}
2194#endif /* CONFIG_DEBUG_VM */
2195
2196static bool check_new_pages(struct page *page, unsigned int order)
2197{
2198 int i;
2199 for (i = 0; i < (1 << order); i++) {
2200 struct page *p = page + i;
2201
2202 if (unlikely(check_new_page(p)))
2203 return true;
2204 }
2205
2206 return false;
2207}
2208
2209inline void post_alloc_hook(struct page *page, unsigned int order,
2210 gfp_t gfp_flags)
2211{
2212 set_page_private(page, 0);
2213 set_page_refcounted(page);
2214
2215 arch_alloc_page(page, order);
2216 if (debug_pagealloc_enabled_static())
2217 kernel_map_pages(page, 1 << order, 1);
2218 kasan_alloc_pages(page, order);
2219 kernel_poison_pages(page, 1 << order, 1);
2220 set_page_owner(page, order, gfp_flags);
2221}
2222
2223static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
2224 unsigned int alloc_flags)
2225{
2226 post_alloc_hook(page, order, gfp_flags);
2227
2228 if (!free_pages_prezeroed() && want_init_on_alloc(gfp_flags))
2229 kernel_init_free_pages(page, 1 << order);
2230
2231 if (order && (gfp_flags & __GFP_COMP))
2232 prep_compound_page(page, order);
2233
2234 /*
2235 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
2236 * allocate the page. The expectation is that the caller is taking
2237 * steps that will free more memory. The caller should avoid the page
2238 * being used for !PFMEMALLOC purposes.
2239 */
2240 if (alloc_flags & ALLOC_NO_WATERMARKS)
2241 set_page_pfmemalloc(page);
2242 else
2243 clear_page_pfmemalloc(page);
2244}
2245
2246/*
2247 * Go through the free lists for the given migratetype and remove
2248 * the smallest available page from the freelists
2249 */
2250static __always_inline
2251struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
2252 int migratetype)
2253{
2254 unsigned int current_order;
2255 struct free_area *area;
2256 struct page *page;
2257
2258 /* Find a page of the appropriate size in the preferred list */
2259 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
2260 area = &(zone->free_area[current_order]);
2261 page = get_page_from_free_area(area, migratetype);
2262 if (!page)
2263 continue;
2264 del_page_from_free_list(page, zone, current_order);
2265 expand(zone, page, order, current_order, migratetype);
2266 set_pcppage_migratetype(page, migratetype);
2267 return page;
2268 }
2269
2270 return NULL;
2271}
2272
2273
2274/*
2275 * This array describes the order lists are fallen back to when
2276 * the free lists for the desirable migrate type are depleted
2277 */
2278static int fallbacks[MIGRATE_TYPES][3] = {
2279 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2280 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
2281 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
2282#ifdef CONFIG_CMA
2283 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
2284#endif
2285#ifdef CONFIG_MEMORY_ISOLATION
2286 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
2287#endif
2288};
2289
2290#ifdef CONFIG_CMA
2291static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2292 unsigned int order)
2293{
2294 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
2295}
2296#else
2297static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
2298 unsigned int order) { return NULL; }
2299#endif
2300
2301/*
2302 * Move the free pages in a range to the free lists of the requested type.
2303 * Note that start_page and end_pages are not aligned on a pageblock
2304 * boundary. If alignment is required, use move_freepages_block()
2305 */
2306static int move_freepages(struct zone *zone,
2307 struct page *start_page, struct page *end_page,
2308 int migratetype, int *num_movable)
2309{
2310 struct page *page;
2311 unsigned int order;
2312 int pages_moved = 0;
2313
2314 for (page = start_page; page <= end_page;) {
2315 if (!pfn_valid_within(page_to_pfn(page))) {
2316 page++;
2317 continue;
2318 }
2319
2320 if (!PageBuddy(page)) {
2321 /*
2322 * We assume that pages that could be isolated for
2323 * migration are movable. But we don't actually try
2324 * isolating, as that would be expensive.
2325 */
2326 if (num_movable &&
2327 (PageLRU(page) || __PageMovable(page)))
2328 (*num_movable)++;
2329
2330 page++;
2331 continue;
2332 }
2333
2334 /* Make sure we are not inadvertently changing nodes */
2335 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2336 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
2337
2338 order = page_order(page);
2339 move_to_free_list(page, zone, order, migratetype);
2340 page += 1 << order;
2341 pages_moved += 1 << order;
2342 }
2343
2344 return pages_moved;
2345}
2346
2347int move_freepages_block(struct zone *zone, struct page *page,
2348 int migratetype, int *num_movable)
2349{
2350 unsigned long start_pfn, end_pfn;
2351 struct page *start_page, *end_page;
2352
2353 if (num_movable)
2354 *num_movable = 0;
2355
2356 start_pfn = page_to_pfn(page);
2357 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2358 start_page = pfn_to_page(start_pfn);
2359 end_page = start_page + pageblock_nr_pages - 1;
2360 end_pfn = start_pfn + pageblock_nr_pages - 1;
2361
2362 /* Do not cross zone boundaries */
2363 if (!zone_spans_pfn(zone, start_pfn))
2364 start_page = page;
2365 if (!zone_spans_pfn(zone, end_pfn))
2366 return 0;
2367
2368 return move_freepages(zone, start_page, end_page, migratetype,
2369 num_movable);
2370}
2371
2372static void change_pageblock_range(struct page *pageblock_page,
2373 int start_order, int migratetype)
2374{
2375 int nr_pageblocks = 1 << (start_order - pageblock_order);
2376
2377 while (nr_pageblocks--) {
2378 set_pageblock_migratetype(pageblock_page, migratetype);
2379 pageblock_page += pageblock_nr_pages;
2380 }
2381}
2382
2383/*
2384 * When we are falling back to another migratetype during allocation, try to
2385 * steal extra free pages from the same pageblocks to satisfy further
2386 * allocations, instead of polluting multiple pageblocks.
2387 *
2388 * If we are stealing a relatively large buddy page, it is likely there will
2389 * be more free pages in the pageblock, so try to steal them all. For
2390 * reclaimable and unmovable allocations, we steal regardless of page size,
2391 * as fragmentation caused by those allocations polluting movable pageblocks
2392 * is worse than movable allocations stealing from unmovable and reclaimable
2393 * pageblocks.
2394 */
2395static bool can_steal_fallback(unsigned int order, int start_mt)
2396{
2397 /*
2398 * Leaving this order check is intended, although there is
2399 * relaxed order check in next check. The reason is that
2400 * we can actually steal whole pageblock if this condition met,
2401 * but, below check doesn't guarantee it and that is just heuristic
2402 * so could be changed anytime.
2403 */
2404 if (order >= pageblock_order)
2405 return true;
2406
2407 if (order >= pageblock_order / 2 ||
2408 start_mt == MIGRATE_RECLAIMABLE ||
2409 start_mt == MIGRATE_UNMOVABLE ||
2410 page_group_by_mobility_disabled)
2411 return true;
2412
2413 return false;
2414}
2415
2416static inline void boost_watermark(struct zone *zone)
2417{
2418 unsigned long max_boost;
2419
2420 if (!watermark_boost_factor)
2421 return;
2422 /*
2423 * Don't bother in zones that are unlikely to produce results.
2424 * On small machines, including kdump capture kernels running
2425 * in a small area, boosting the watermark can cause an out of
2426 * memory situation immediately.
2427 */
2428 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
2429 return;
2430
2431 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
2432 watermark_boost_factor, 10000);
2433
2434 /*
2435 * high watermark may be uninitialised if fragmentation occurs
2436 * very early in boot so do not boost. We do not fall
2437 * through and boost by pageblock_nr_pages as failing
2438 * allocations that early means that reclaim is not going
2439 * to help and it may even be impossible to reclaim the
2440 * boosted watermark resulting in a hang.
2441 */
2442 if (!max_boost)
2443 return;
2444
2445 max_boost = max(pageblock_nr_pages, max_boost);
2446
2447 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
2448 max_boost);
2449}
2450
2451/*
2452 * This function implements actual steal behaviour. If order is large enough,
2453 * we can steal whole pageblock. If not, we first move freepages in this
2454 * pageblock to our migratetype and determine how many already-allocated pages
2455 * are there in the pageblock with a compatible migratetype. If at least half
2456 * of pages are free or compatible, we can change migratetype of the pageblock
2457 * itself, so pages freed in the future will be put on the correct free list.
2458 */
2459static void steal_suitable_fallback(struct zone *zone, struct page *page,
2460 unsigned int alloc_flags, int start_type, bool whole_block)
2461{
2462 unsigned int current_order = page_order(page);
2463 int free_pages, movable_pages, alike_pages;
2464 int old_block_type;
2465
2466 old_block_type = get_pageblock_migratetype(page);
2467
2468 /*
2469 * This can happen due to races and we want to prevent broken
2470 * highatomic accounting.
2471 */
2472 if (is_migrate_highatomic(old_block_type))
2473 goto single_page;
2474
2475 /* Take ownership for orders >= pageblock_order */
2476 if (current_order >= pageblock_order) {
2477 change_pageblock_range(page, current_order, start_type);
2478 goto single_page;
2479 }
2480
2481 /*
2482 * Boost watermarks to increase reclaim pressure to reduce the
2483 * likelihood of future fallbacks. Wake kswapd now as the node
2484 * may be balanced overall and kswapd will not wake naturally.
2485 */
2486 boost_watermark(zone);
2487 if (alloc_flags & ALLOC_KSWAPD)
2488 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
2489
2490 /* We are not allowed to try stealing from the whole block */
2491 if (!whole_block)
2492 goto single_page;
2493
2494 free_pages = move_freepages_block(zone, page, start_type,
2495 &movable_pages);
2496 /*
2497 * Determine how many pages are compatible with our allocation.
2498 * For movable allocation, it's the number of movable pages which
2499 * we just obtained. For other types it's a bit more tricky.
2500 */
2501 if (start_type == MIGRATE_MOVABLE) {
2502 alike_pages = movable_pages;
2503 } else {
2504 /*
2505 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2506 * to MOVABLE pageblock, consider all non-movable pages as
2507 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2508 * vice versa, be conservative since we can't distinguish the
2509 * exact migratetype of non-movable pages.
2510 */
2511 if (old_block_type == MIGRATE_MOVABLE)
2512 alike_pages = pageblock_nr_pages
2513 - (free_pages + movable_pages);
2514 else
2515 alike_pages = 0;
2516 }
2517
2518 /* moving whole block can fail due to zone boundary conditions */
2519 if (!free_pages)
2520 goto single_page;
2521
2522 /*
2523 * If a sufficient number of pages in the block are either free or of
2524 * comparable migratability as our allocation, claim the whole block.
2525 */
2526 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2527 page_group_by_mobility_disabled)
2528 set_pageblock_migratetype(page, start_type);
2529
2530 return;
2531
2532single_page:
2533 move_to_free_list(page, zone, current_order, start_type);
2534}
2535
2536/*
2537 * Check whether there is a suitable fallback freepage with requested order.
2538 * If only_stealable is true, this function returns fallback_mt only if
2539 * we can steal other freepages all together. This would help to reduce
2540 * fragmentation due to mixed migratetype pages in one pageblock.
2541 */
2542int find_suitable_fallback(struct free_area *area, unsigned int order,
2543 int migratetype, bool only_stealable, bool *can_steal)
2544{
2545 int i;
2546 int fallback_mt;
2547
2548 if (area->nr_free == 0)
2549 return -1;
2550
2551 *can_steal = false;
2552 for (i = 0;; i++) {
2553 fallback_mt = fallbacks[migratetype][i];
2554 if (fallback_mt == MIGRATE_TYPES)
2555 break;
2556
2557 if (free_area_empty(area, fallback_mt))
2558 continue;
2559
2560 if (can_steal_fallback(order, migratetype))
2561 *can_steal = true;
2562
2563 if (!only_stealable)
2564 return fallback_mt;
2565
2566 if (*can_steal)
2567 return fallback_mt;
2568 }
2569
2570 return -1;
2571}
2572
2573/*
2574 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2575 * there are no empty page blocks that contain a page with a suitable order
2576 */
2577static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2578 unsigned int alloc_order)
2579{
2580 int mt;
2581 unsigned long max_managed, flags;
2582
2583 /*
2584 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2585 * Check is race-prone but harmless.
2586 */
2587 max_managed = (zone_managed_pages(zone) / 100) + pageblock_nr_pages;
2588 if (zone->nr_reserved_highatomic >= max_managed)
2589 return;
2590
2591 spin_lock_irqsave(&zone->lock, flags);
2592
2593 /* Recheck the nr_reserved_highatomic limit under the lock */
2594 if (zone->nr_reserved_highatomic >= max_managed)
2595 goto out_unlock;
2596
2597 /* Yoink! */
2598 mt = get_pageblock_migratetype(page);
2599 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2600 && !is_migrate_cma(mt)) {
2601 zone->nr_reserved_highatomic += pageblock_nr_pages;
2602 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2603 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2604 }
2605
2606out_unlock:
2607 spin_unlock_irqrestore(&zone->lock, flags);
2608}
2609
2610/*
2611 * Used when an allocation is about to fail under memory pressure. This
2612 * potentially hurts the reliability of high-order allocations when under
2613 * intense memory pressure but failed atomic allocations should be easier
2614 * to recover from than an OOM.
2615 *
2616 * If @force is true, try to unreserve a pageblock even though highatomic
2617 * pageblock is exhausted.
2618 */
2619static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2620 bool force)
2621{
2622 struct zonelist *zonelist = ac->zonelist;
2623 unsigned long flags;
2624 struct zoneref *z;
2625 struct zone *zone;
2626 struct page *page;
2627 int order;
2628 bool ret;
2629
2630 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2631 ac->nodemask) {
2632 /*
2633 * Preserve at least one pageblock unless memory pressure
2634 * is really high.
2635 */
2636 if (!force && zone->nr_reserved_highatomic <=
2637 pageblock_nr_pages)
2638 continue;
2639
2640 spin_lock_irqsave(&zone->lock, flags);
2641 for (order = 0; order < MAX_ORDER; order++) {
2642 struct free_area *area = &(zone->free_area[order]);
2643
2644 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2645 if (!page)
2646 continue;
2647
2648 /*
2649 * In page freeing path, migratetype change is racy so
2650 * we can counter several free pages in a pageblock
2651 * in this loop althoug we changed the pageblock type
2652 * from highatomic to ac->migratetype. So we should
2653 * adjust the count once.
2654 */
2655 if (is_migrate_highatomic_page(page)) {
2656 /*
2657 * It should never happen but changes to
2658 * locking could inadvertently allow a per-cpu
2659 * drain to add pages to MIGRATE_HIGHATOMIC
2660 * while unreserving so be safe and watch for
2661 * underflows.
2662 */
2663 zone->nr_reserved_highatomic -= min(
2664 pageblock_nr_pages,
2665 zone->nr_reserved_highatomic);
2666 }
2667
2668 /*
2669 * Convert to ac->migratetype and avoid the normal
2670 * pageblock stealing heuristics. Minimally, the caller
2671 * is doing the work and needs the pages. More
2672 * importantly, if the block was always converted to
2673 * MIGRATE_UNMOVABLE or another type then the number
2674 * of pageblocks that cannot be completely freed
2675 * may increase.
2676 */
2677 set_pageblock_migratetype(page, ac->migratetype);
2678 ret = move_freepages_block(zone, page, ac->migratetype,
2679 NULL);
2680 if (ret) {
2681 spin_unlock_irqrestore(&zone->lock, flags);
2682 return ret;
2683 }
2684 }
2685 spin_unlock_irqrestore(&zone->lock, flags);
2686 }
2687
2688 return false;
2689}
2690
2691/*
2692 * Try finding a free buddy page on the fallback list and put it on the free
2693 * list of requested migratetype, possibly along with other pages from the same
2694 * block, depending on fragmentation avoidance heuristics. Returns true if
2695 * fallback was found so that __rmqueue_smallest() can grab it.
2696 *
2697 * The use of signed ints for order and current_order is a deliberate
2698 * deviation from the rest of this file, to make the for loop
2699 * condition simpler.
2700 */
2701static __always_inline bool
2702__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2703 unsigned int alloc_flags)
2704{
2705 struct free_area *area;
2706 int current_order;
2707 int min_order = order;
2708 struct page *page;
2709 int fallback_mt;
2710 bool can_steal;
2711
2712 /*
2713 * Do not steal pages from freelists belonging to other pageblocks
2714 * i.e. orders < pageblock_order. If there are no local zones free,
2715 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2716 */
2717 if (alloc_flags & ALLOC_NOFRAGMENT)
2718 min_order = pageblock_order;
2719
2720 /*
2721 * Find the largest available free page in the other list. This roughly
2722 * approximates finding the pageblock with the most free pages, which
2723 * would be too costly to do exactly.
2724 */
2725 for (current_order = MAX_ORDER - 1; current_order >= min_order;
2726 --current_order) {
2727 area = &(zone->free_area[current_order]);
2728 fallback_mt = find_suitable_fallback(area, current_order,
2729 start_migratetype, false, &can_steal);
2730 if (fallback_mt == -1)
2731 continue;
2732
2733 /*
2734 * We cannot steal all free pages from the pageblock and the
2735 * requested migratetype is movable. In that case it's better to
2736 * steal and split the smallest available page instead of the
2737 * largest available page, because even if the next movable
2738 * allocation falls back into a different pageblock than this
2739 * one, it won't cause permanent fragmentation.
2740 */
2741 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2742 && current_order > order)
2743 goto find_smallest;
2744
2745 goto do_steal;
2746 }
2747
2748 return false;
2749
2750find_smallest:
2751 for (current_order = order; current_order < MAX_ORDER;
2752 current_order++) {
2753 area = &(zone->free_area[current_order]);
2754 fallback_mt = find_suitable_fallback(area, current_order,
2755 start_migratetype, false, &can_steal);
2756 if (fallback_mt != -1)
2757 break;
2758 }
2759
2760 /*
2761 * This should not happen - we already found a suitable fallback
2762 * when looking for the largest page.
2763 */
2764 VM_BUG_ON(current_order == MAX_ORDER);
2765
2766do_steal:
2767 page = get_page_from_free_area(area, fallback_mt);
2768
2769 steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
2770 can_steal);
2771
2772 trace_mm_page_alloc_extfrag(page, order, current_order,
2773 start_migratetype, fallback_mt);
2774
2775 return true;
2776
2777}
2778
2779/*
2780 * Do the hard work of removing an element from the buddy allocator.
2781 * Call me with the zone->lock already held.
2782 */
2783static __always_inline struct page *
2784__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2785 unsigned int alloc_flags)
2786{
2787 struct page *page;
2788
2789#ifdef CONFIG_CMA
2790 /*
2791 * Balance movable allocations between regular and CMA areas by
2792 * allocating from CMA when over half of the zone's free memory
2793 * is in the CMA area.
2794 */
2795 if (alloc_flags & ALLOC_CMA &&
2796 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2797 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2798 page = __rmqueue_cma_fallback(zone, order);
2799 if (page)
2800 return page;
2801 }
2802#endif
2803retry:
2804 page = __rmqueue_smallest(zone, order, migratetype);
2805 if (unlikely(!page)) {
2806 if (alloc_flags & ALLOC_CMA)
2807 page = __rmqueue_cma_fallback(zone, order);
2808
2809 if (!page && __rmqueue_fallback(zone, order, migratetype,
2810 alloc_flags))
2811 goto retry;
2812 }
2813
2814 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2815 return page;
2816}
2817
2818/*
2819 * Obtain a specified number of elements from the buddy allocator, all under
2820 * a single hold of the lock, for efficiency. Add them to the supplied list.
2821 * Returns the number of new pages which were placed at *list.
2822 */
2823static int rmqueue_bulk(struct zone *zone, unsigned int order,
2824 unsigned long count, struct list_head *list,
2825 int migratetype, unsigned int alloc_flags)
2826{
2827 int i, alloced = 0;
2828
2829 spin_lock(&zone->lock);
2830 for (i = 0; i < count; ++i) {
2831 struct page *page = __rmqueue(zone, order, migratetype,
2832 alloc_flags);
2833 if (unlikely(page == NULL))
2834 break;
2835
2836 if (unlikely(check_pcp_refill(page)))
2837 continue;
2838
2839 /*
2840 * Split buddy pages returned by expand() are received here in
2841 * physical page order. The page is added to the tail of
2842 * caller's list. From the callers perspective, the linked list
2843 * is ordered by page number under some conditions. This is
2844 * useful for IO devices that can forward direction from the
2845 * head, thus also in the physical page order. This is useful
2846 * for IO devices that can merge IO requests if the physical
2847 * pages are ordered properly.
2848 */
2849 list_add_tail(&page->lru, list);
2850 alloced++;
2851 if (is_migrate_cma(get_pcppage_migratetype(page)))
2852 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2853 -(1 << order));
2854 }
2855
2856 /*
2857 * i pages were removed from the buddy list even if some leak due
2858 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2859 * on i. Do not confuse with 'alloced' which is the number of
2860 * pages added to the pcp list.
2861 */
2862 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2863 spin_unlock(&zone->lock);
2864 return alloced;
2865}
2866
2867#ifdef CONFIG_NUMA
2868/*
2869 * Called from the vmstat counter updater to drain pagesets of this
2870 * currently executing processor on remote nodes after they have
2871 * expired.
2872 *
2873 * Note that this function must be called with the thread pinned to
2874 * a single processor.
2875 */
2876void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2877{
2878 unsigned long flags;
2879 int to_drain, batch;
2880
2881 local_irq_save(flags);
2882 batch = READ_ONCE(pcp->batch);
2883 to_drain = min(pcp->count, batch);
2884 if (to_drain > 0)
2885 free_pcppages_bulk(zone, to_drain, pcp);
2886 local_irq_restore(flags);
2887}
2888#endif
2889
2890/*
2891 * Drain pcplists of the indicated processor and zone.
2892 *
2893 * The processor must either be the current processor and the
2894 * thread pinned to the current processor or a processor that
2895 * is not online.
2896 */
2897static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2898{
2899 unsigned long flags;
2900 struct per_cpu_pageset *pset;
2901 struct per_cpu_pages *pcp;
2902
2903 local_irq_save(flags);
2904 pset = per_cpu_ptr(zone->pageset, cpu);
2905
2906 pcp = &pset->pcp;
2907 if (pcp->count)
2908 free_pcppages_bulk(zone, pcp->count, pcp);
2909 local_irq_restore(flags);
2910}
2911
2912/*
2913 * Drain pcplists of all zones on the indicated processor.
2914 *
2915 * The processor must either be the current processor and the
2916 * thread pinned to the current processor or a processor that
2917 * is not online.
2918 */
2919static void drain_pages(unsigned int cpu)
2920{
2921 struct zone *zone;
2922
2923 for_each_populated_zone(zone) {
2924 drain_pages_zone(cpu, zone);
2925 }
2926}
2927
2928/*
2929 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2930 *
2931 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2932 * the single zone's pages.
2933 */
2934void drain_local_pages(struct zone *zone)
2935{
2936 int cpu = smp_processor_id();
2937
2938 if (zone)
2939 drain_pages_zone(cpu, zone);
2940 else
2941 drain_pages(cpu);
2942}
2943
2944static void drain_local_pages_wq(struct work_struct *work)
2945{
2946 struct pcpu_drain *drain;
2947
2948 drain = container_of(work, struct pcpu_drain, work);
2949
2950 /*
2951 * drain_all_pages doesn't use proper cpu hotplug protection so
2952 * we can race with cpu offline when the WQ can move this from
2953 * a cpu pinned worker to an unbound one. We can operate on a different
2954 * cpu which is allright but we also have to make sure to not move to
2955 * a different one.
2956 */
2957 preempt_disable();
2958 drain_local_pages(drain->zone);
2959 preempt_enable();
2960}
2961
2962/*
2963 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2964 *
2965 * When zone parameter is non-NULL, spill just the single zone's pages.
2966 *
2967 * Note that this can be extremely slow as the draining happens in a workqueue.
2968 */
2969void drain_all_pages(struct zone *zone)
2970{
2971 int cpu;
2972
2973 /*
2974 * Allocate in the BSS so we wont require allocation in
2975 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2976 */
2977 static cpumask_t cpus_with_pcps;
2978
2979 /*
2980 * Make sure nobody triggers this path before mm_percpu_wq is fully
2981 * initialized.
2982 */
2983 if (WARN_ON_ONCE(!mm_percpu_wq))
2984 return;
2985
2986 /*
2987 * Do not drain if one is already in progress unless it's specific to
2988 * a zone. Such callers are primarily CMA and memory hotplug and need
2989 * the drain to be complete when the call returns.
2990 */
2991 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2992 if (!zone)
2993 return;
2994 mutex_lock(&pcpu_drain_mutex);
2995 }
2996
2997 /*
2998 * We don't care about racing with CPU hotplug event
2999 * as offline notification will cause the notified
3000 * cpu to drain that CPU pcps and on_each_cpu_mask
3001 * disables preemption as part of its processing
3002 */
3003 for_each_online_cpu(cpu) {
3004 struct per_cpu_pageset *pcp;
3005 struct zone *z;
3006 bool has_pcps = false;
3007
3008 if (zone) {
3009 pcp = per_cpu_ptr(zone->pageset, cpu);
3010 if (pcp->pcp.count)
3011 has_pcps = true;
3012 } else {
3013 for_each_populated_zone(z) {
3014 pcp = per_cpu_ptr(z->pageset, cpu);
3015 if (pcp->pcp.count) {
3016 has_pcps = true;
3017 break;
3018 }
3019 }
3020 }
3021
3022 if (has_pcps)
3023 cpumask_set_cpu(cpu, &cpus_with_pcps);
3024 else
3025 cpumask_clear_cpu(cpu, &cpus_with_pcps);
3026 }
3027
3028 for_each_cpu(cpu, &cpus_with_pcps) {
3029 struct pcpu_drain *drain = per_cpu_ptr(&pcpu_drain, cpu);
3030
3031 drain->zone = zone;
3032 INIT_WORK(&drain->work, drain_local_pages_wq);
3033 queue_work_on(cpu, mm_percpu_wq, &drain->work);
3034 }
3035 for_each_cpu(cpu, &cpus_with_pcps)
3036 flush_work(&per_cpu_ptr(&pcpu_drain, cpu)->work);
3037
3038 mutex_unlock(&pcpu_drain_mutex);
3039}
3040
3041#ifdef CONFIG_HIBERNATION
3042
3043/*
3044 * Touch the watchdog for every WD_PAGE_COUNT pages.
3045 */
3046#define WD_PAGE_COUNT (128*1024)
3047
3048void mark_free_pages(struct zone *zone)
3049{
3050 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
3051 unsigned long flags;
3052 unsigned int order, t;
3053 struct page *page;
3054
3055 if (zone_is_empty(zone))
3056 return;
3057
3058 spin_lock_irqsave(&zone->lock, flags);
3059
3060 max_zone_pfn = zone_end_pfn(zone);
3061 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
3062 if (pfn_valid(pfn)) {
3063 page = pfn_to_page(pfn);
3064
3065 if (!--page_count) {
3066 touch_nmi_watchdog();
3067 page_count = WD_PAGE_COUNT;
3068 }
3069
3070 if (page_zone(page) != zone)
3071 continue;
3072
3073 if (!swsusp_page_is_forbidden(page))
3074 swsusp_unset_page_free(page);
3075 }
3076
3077 for_each_migratetype_order(order, t) {
3078 list_for_each_entry(page,
3079 &zone->free_area[order].free_list[t], lru) {
3080 unsigned long i;
3081
3082 pfn = page_to_pfn(page);
3083 for (i = 0; i < (1UL << order); i++) {
3084 if (!--page_count) {
3085 touch_nmi_watchdog();
3086 page_count = WD_PAGE_COUNT;
3087 }
3088 swsusp_set_page_free(pfn_to_page(pfn + i));
3089 }
3090 }
3091 }
3092 spin_unlock_irqrestore(&zone->lock, flags);
3093}
3094#endif /* CONFIG_PM */
3095
3096static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
3097{
3098 int migratetype;
3099
3100 if (!free_pcp_prepare(page))
3101 return false;
3102
3103 migratetype = get_pfnblock_migratetype(page, pfn);
3104 set_pcppage_migratetype(page, migratetype);
3105 return true;
3106}
3107
3108static void free_unref_page_commit(struct page *page, unsigned long pfn)
3109{
3110 struct zone *zone = page_zone(page);
3111 struct per_cpu_pages *pcp;
3112 int migratetype;
3113
3114 migratetype = get_pcppage_migratetype(page);
3115 __count_vm_event(PGFREE);
3116
3117 /*
3118 * We only track unmovable, reclaimable and movable on pcp lists.
3119 * Free ISOLATE pages back to the allocator because they are being
3120 * offlined but treat HIGHATOMIC as movable pages so we can get those
3121 * areas back if necessary. Otherwise, we may have to free
3122 * excessively into the page allocator
3123 */
3124 if (migratetype >= MIGRATE_PCPTYPES) {
3125 if (unlikely(is_migrate_isolate(migratetype))) {
3126 free_one_page(zone, page, pfn, 0, migratetype);
3127 return;
3128 }
3129 migratetype = MIGRATE_MOVABLE;
3130 }
3131
3132 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3133 list_add(&page->lru, &pcp->lists[migratetype]);
3134 pcp->count++;
3135 if (pcp->count >= pcp->high) {
3136 unsigned long batch = READ_ONCE(pcp->batch);
3137 free_pcppages_bulk(zone, batch, pcp);
3138 }
3139}
3140
3141/*
3142 * Free a 0-order page
3143 */
3144void free_unref_page(struct page *page)
3145{
3146 unsigned long flags;
3147 unsigned long pfn = page_to_pfn(page);
3148
3149 if (!free_unref_page_prepare(page, pfn))
3150 return;
3151
3152 local_irq_save(flags);
3153 free_unref_page_commit(page, pfn);
3154 local_irq_restore(flags);
3155}
3156
3157/*
3158 * Free a list of 0-order pages
3159 */
3160void free_unref_page_list(struct list_head *list)
3161{
3162 struct page *page, *next;
3163 unsigned long flags, pfn;
3164 int batch_count = 0;
3165
3166 /* Prepare pages for freeing */
3167 list_for_each_entry_safe(page, next, list, lru) {
3168 pfn = page_to_pfn(page);
3169 if (!free_unref_page_prepare(page, pfn))
3170 list_del(&page->lru);
3171 set_page_private(page, pfn);
3172 }
3173
3174 local_irq_save(flags);
3175 list_for_each_entry_safe(page, next, list, lru) {
3176 unsigned long pfn = page_private(page);
3177
3178 set_page_private(page, 0);
3179 trace_mm_page_free_batched(page);
3180 free_unref_page_commit(page, pfn);
3181
3182 /*
3183 * Guard against excessive IRQ disabled times when we get
3184 * a large list of pages to free.
3185 */
3186 if (++batch_count == SWAP_CLUSTER_MAX) {
3187 local_irq_restore(flags);
3188 batch_count = 0;
3189 local_irq_save(flags);
3190 }
3191 }
3192 local_irq_restore(flags);
3193}
3194
3195/*
3196 * split_page takes a non-compound higher-order page, and splits it into
3197 * n (1<<order) sub-pages: page[0..n]
3198 * Each sub-page must be freed individually.
3199 *
3200 * Note: this is probably too low level an operation for use in drivers.
3201 * Please consult with lkml before using this in your driver.
3202 */
3203void split_page(struct page *page, unsigned int order)
3204{
3205 int i;
3206
3207 VM_BUG_ON_PAGE(PageCompound(page), page);
3208 VM_BUG_ON_PAGE(!page_count(page), page);
3209
3210 for (i = 1; i < (1 << order); i++)
3211 set_page_refcounted(page + i);
3212 split_page_owner(page, order);
3213}
3214EXPORT_SYMBOL_GPL(split_page);
3215
3216int __isolate_free_page(struct page *page, unsigned int order)
3217{
3218 unsigned long watermark;
3219 struct zone *zone;
3220 int mt;
3221
3222 BUG_ON(!PageBuddy(page));
3223
3224 zone = page_zone(page);
3225 mt = get_pageblock_migratetype(page);
3226
3227 if (!is_migrate_isolate(mt)) {
3228 /*
3229 * Obey watermarks as if the page was being allocated. We can
3230 * emulate a high-order watermark check with a raised order-0
3231 * watermark, because we already know our high-order page
3232 * exists.
3233 */
3234 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
3235 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
3236 return 0;
3237
3238 __mod_zone_freepage_state(zone, -(1UL << order), mt);
3239 }
3240
3241 /* Remove page from free list */
3242
3243 del_page_from_free_list(page, zone, order);
3244
3245 /*
3246 * Set the pageblock if the isolated page is at least half of a
3247 * pageblock
3248 */
3249 if (order >= pageblock_order - 1) {
3250 struct page *endpage = page + (1 << order) - 1;
3251 for (; page < endpage; page += pageblock_nr_pages) {
3252 int mt = get_pageblock_migratetype(page);
3253 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
3254 && !is_migrate_highatomic(mt))
3255 set_pageblock_migratetype(page,
3256 MIGRATE_MOVABLE);
3257 }
3258 }
3259
3260
3261 return 1UL << order;
3262}
3263
3264/**
3265 * __putback_isolated_page - Return a now-isolated page back where we got it
3266 * @page: Page that was isolated
3267 * @order: Order of the isolated page
3268 * @mt: The page's pageblock's migratetype
3269 *
3270 * This function is meant to return a page pulled from the free lists via
3271 * __isolate_free_page back to the free lists they were pulled from.
3272 */
3273void __putback_isolated_page(struct page *page, unsigned int order, int mt)
3274{
3275 struct zone *zone = page_zone(page);
3276
3277 /* zone lock should be held when this function is called */
3278 lockdep_assert_held(&zone->lock);
3279
3280 /* Return isolated page to tail of freelist. */
3281 __free_one_page(page, page_to_pfn(page), zone, order, mt, false);
3282}
3283
3284/*
3285 * Update NUMA hit/miss statistics
3286 *
3287 * Must be called with interrupts disabled.
3288 */
3289static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
3290{
3291#ifdef CONFIG_NUMA
3292 enum numa_stat_item local_stat = NUMA_LOCAL;
3293
3294 /* skip numa counters update if numa stats is disabled */
3295 if (!static_branch_likely(&vm_numa_stat_key))
3296 return;
3297
3298 if (zone_to_nid(z) != numa_node_id())
3299 local_stat = NUMA_OTHER;
3300
3301 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
3302 __inc_numa_state(z, NUMA_HIT);
3303 else {
3304 __inc_numa_state(z, NUMA_MISS);
3305 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
3306 }
3307 __inc_numa_state(z, local_stat);
3308#endif
3309}
3310
3311/* Remove page from the per-cpu list, caller must protect the list */
3312static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
3313 unsigned int alloc_flags,
3314 struct per_cpu_pages *pcp,
3315 struct list_head *list)
3316{
3317 struct page *page;
3318
3319 do {
3320 if (list_empty(list)) {
3321 pcp->count += rmqueue_bulk(zone, 0,
3322 pcp->batch, list,
3323 migratetype, alloc_flags);
3324 if (unlikely(list_empty(list)))
3325 return NULL;
3326 }
3327
3328 page = list_first_entry(list, struct page, lru);
3329 list_del(&page->lru);
3330 pcp->count--;
3331 } while (check_new_pcp(page));
3332
3333 return page;
3334}
3335
3336/* Lock and remove page from the per-cpu list */
3337static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3338 struct zone *zone, gfp_t gfp_flags,
3339 int migratetype, unsigned int alloc_flags)
3340{
3341 struct per_cpu_pages *pcp;
3342 struct list_head *list;
3343 struct page *page;
3344 unsigned long flags;
3345
3346 local_irq_save(flags);
3347 pcp = &this_cpu_ptr(zone->pageset)->pcp;
3348 list = &pcp->lists[migratetype];
3349 page = __rmqueue_pcplist(zone, migratetype, alloc_flags, pcp, list);
3350 if (page) {
3351 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1);
3352 zone_statistics(preferred_zone, zone);
3353 }
3354 local_irq_restore(flags);
3355 return page;
3356}
3357
3358/*
3359 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
3360 */
3361static inline
3362struct page *rmqueue(struct zone *preferred_zone,
3363 struct zone *zone, unsigned int order,
3364 gfp_t gfp_flags, unsigned int alloc_flags,
3365 int migratetype)
3366{
3367 unsigned long flags;
3368 struct page *page;
3369
3370 if (likely(order == 0)) {
3371 /*
3372 * MIGRATE_MOVABLE pcplist could have the pages on CMA area and
3373 * we need to skip it when CMA area isn't allowed.
3374 */
3375 if (!IS_ENABLED(CONFIG_CMA) || alloc_flags & ALLOC_CMA ||
3376 migratetype != MIGRATE_MOVABLE) {
3377 page = rmqueue_pcplist(preferred_zone, zone, gfp_flags,
3378 migratetype, alloc_flags);
3379 goto out;
3380 }
3381 }
3382
3383 /*
3384 * We most definitely don't want callers attempting to
3385 * allocate greater than order-1 page units with __GFP_NOFAIL.
3386 */
3387 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
3388 spin_lock_irqsave(&zone->lock, flags);
3389
3390 do {
3391 page = NULL;
3392 /*
3393 * order-0 request can reach here when the pcplist is skipped
3394 * due to non-CMA allocation context. HIGHATOMIC area is
3395 * reserved for high-order atomic allocation, so order-0
3396 * request should skip it.
3397 */
3398 if (order > 0 && alloc_flags & ALLOC_HARDER) {
3399 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3400 if (page)
3401 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3402 }
3403 if (!page)
3404 page = __rmqueue(zone, order, migratetype, alloc_flags);
3405 } while (page && check_new_pages(page, order));
3406 spin_unlock(&zone->lock);
3407 if (!page)
3408 goto failed;
3409 __mod_zone_freepage_state(zone, -(1 << order),
3410 get_pcppage_migratetype(page));
3411
3412 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3413 zone_statistics(preferred_zone, zone);
3414 local_irq_restore(flags);
3415
3416out:
3417 /* Separate test+clear to avoid unnecessary atomics */
3418 if (test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags)) {
3419 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3420 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3421 }
3422
3423 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3424 return page;
3425
3426failed:
3427 local_irq_restore(flags);
3428 return NULL;
3429}
3430
3431#ifdef CONFIG_FAIL_PAGE_ALLOC
3432
3433static struct {
3434 struct fault_attr attr;
3435
3436 bool ignore_gfp_highmem;
3437 bool ignore_gfp_reclaim;
3438 u32 min_order;
3439} fail_page_alloc = {
3440 .attr = FAULT_ATTR_INITIALIZER,
3441 .ignore_gfp_reclaim = true,
3442 .ignore_gfp_highmem = true,
3443 .min_order = 1,
3444};
3445
3446static int __init setup_fail_page_alloc(char *str)
3447{
3448 return setup_fault_attr(&fail_page_alloc.attr, str);
3449}
3450__setup("fail_page_alloc=", setup_fail_page_alloc);
3451
3452static bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3453{
3454 if (order < fail_page_alloc.min_order)
3455 return false;
3456 if (gfp_mask & __GFP_NOFAIL)
3457 return false;
3458 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3459 return false;
3460 if (fail_page_alloc.ignore_gfp_reclaim &&
3461 (gfp_mask & __GFP_DIRECT_RECLAIM))
3462 return false;
3463
3464 return should_fail(&fail_page_alloc.attr, 1 << order);
3465}
3466
3467#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3468
3469static int __init fail_page_alloc_debugfs(void)
3470{
3471 umode_t mode = S_IFREG | 0600;
3472 struct dentry *dir;
3473
3474 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3475 &fail_page_alloc.attr);
3476
3477 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3478 &fail_page_alloc.ignore_gfp_reclaim);
3479 debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3480 &fail_page_alloc.ignore_gfp_highmem);
3481 debugfs_create_u32("min-order", mode, dir, &fail_page_alloc.min_order);
3482
3483 return 0;
3484}
3485
3486late_initcall(fail_page_alloc_debugfs);
3487
3488#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3489
3490#else /* CONFIG_FAIL_PAGE_ALLOC */
3491
3492static inline bool __should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3493{
3494 return false;
3495}
3496
3497#endif /* CONFIG_FAIL_PAGE_ALLOC */
3498
3499static noinline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3500{
3501 return __should_fail_alloc_page(gfp_mask, order);
3502}
3503ALLOW_ERROR_INJECTION(should_fail_alloc_page, TRUE);
3504
3505static inline long __zone_watermark_unusable_free(struct zone *z,
3506 unsigned int order, unsigned int alloc_flags)
3507{
3508 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3509 long unusable_free = (1 << order) - 1;
3510
3511 /*
3512 * If the caller does not have rights to ALLOC_HARDER then subtract
3513 * the high-atomic reserves. This will over-estimate the size of the
3514 * atomic reserve but it avoids a search.
3515 */
3516 if (likely(!alloc_harder))
3517 unusable_free += z->nr_reserved_highatomic;
3518
3519#ifdef CONFIG_CMA
3520 /* If allocation can't use CMA areas don't use free CMA pages */
3521 if (!(alloc_flags & ALLOC_CMA))
3522 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3523#endif
3524
3525 return unusable_free;
3526}
3527
3528/*
3529 * Return true if free base pages are above 'mark'. For high-order checks it
3530 * will return true of the order-0 watermark is reached and there is at least
3531 * one free page of a suitable size. Checking now avoids taking the zone lock
3532 * to check in the allocation paths if no pages are free.
3533 */
3534bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3535 int highest_zoneidx, unsigned int alloc_flags,
3536 long free_pages)
3537{
3538 long min = mark;
3539 int o;
3540 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3541
3542 /* free_pages may go negative - that's OK */
3543 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3544
3545 if (alloc_flags & ALLOC_HIGH)
3546 min -= min / 2;
3547
3548 if (unlikely(alloc_harder)) {
3549 /*
3550 * OOM victims can try even harder than normal ALLOC_HARDER
3551 * users on the grounds that it's definitely going to be in
3552 * the exit path shortly and free memory. Any allocation it
3553 * makes during the free path will be small and short-lived.
3554 */
3555 if (alloc_flags & ALLOC_OOM)
3556 min -= min / 2;
3557 else
3558 min -= min / 4;
3559 }
3560
3561 /*
3562 * Check watermarks for an order-0 allocation request. If these
3563 * are not met, then a high-order request also cannot go ahead
3564 * even if a suitable page happened to be free.
3565 */
3566 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3567 return false;
3568
3569 /* If this is an order-0 request then the watermark is fine */
3570 if (!order)
3571 return true;
3572
3573 /* For a high-order request, check at least one suitable page is free */
3574 for (o = order; o < MAX_ORDER; o++) {
3575 struct free_area *area = &z->free_area[o];
3576 int mt;
3577
3578 if (!area->nr_free)
3579 continue;
3580
3581 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3582 if (!free_area_empty(area, mt))
3583 return true;
3584 }
3585
3586#ifdef CONFIG_CMA
3587 if ((alloc_flags & ALLOC_CMA) &&
3588 !free_area_empty(area, MIGRATE_CMA)) {
3589 return true;
3590 }
3591#endif
3592 if (alloc_harder && !free_area_empty(area, MIGRATE_HIGHATOMIC))
3593 return true;
3594 }
3595 return false;
3596}
3597
3598bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3599 int highest_zoneidx, unsigned int alloc_flags)
3600{
3601 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3602 zone_page_state(z, NR_FREE_PAGES));
3603}
3604
3605static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3606 unsigned long mark, int highest_zoneidx,
3607 unsigned int alloc_flags, gfp_t gfp_mask)
3608{
3609 long free_pages;
3610
3611 free_pages = zone_page_state(z, NR_FREE_PAGES);
3612
3613 /*
3614 * Fast check for order-0 only. If this fails then the reserves
3615 * need to be calculated.
3616 */
3617 if (!order) {
3618 long fast_free;
3619
3620 fast_free = free_pages;
3621 fast_free -= __zone_watermark_unusable_free(z, 0, alloc_flags);
3622 if (fast_free > mark + z->lowmem_reserve[highest_zoneidx])
3623 return true;
3624 }
3625
3626 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3627 free_pages))
3628 return true;
3629 /*
3630 * Ignore watermark boosting for GFP_ATOMIC order-0 allocations
3631 * when checking the min watermark. The min watermark is the
3632 * point where boosting is ignored so that kswapd is woken up
3633 * when below the low watermark.
3634 */
3635 if (unlikely(!order && (gfp_mask & __GFP_ATOMIC) && z->watermark_boost
3636 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3637 mark = z->_watermark[WMARK_MIN];
3638 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3639 alloc_flags, free_pages);
3640 }
3641
3642 return false;
3643}
3644
3645bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3646 unsigned long mark, int highest_zoneidx)
3647{
3648 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3649
3650 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3651 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3652
3653 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3654 free_pages);
3655}
3656
3657#ifdef CONFIG_NUMA
3658static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3659{
3660 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3661 node_reclaim_distance;
3662}
3663#else /* CONFIG_NUMA */
3664static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3665{
3666 return true;
3667}
3668#endif /* CONFIG_NUMA */
3669
3670/*
3671 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3672 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3673 * premature use of a lower zone may cause lowmem pressure problems that
3674 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3675 * probably too small. It only makes sense to spread allocations to avoid
3676 * fragmentation between the Normal and DMA32 zones.
3677 */
3678static inline unsigned int
3679alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3680{
3681 unsigned int alloc_flags;
3682
3683 /*
3684 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3685 * to save a branch.
3686 */
3687 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3688
3689#ifdef CONFIG_ZONE_DMA32
3690 if (!zone)
3691 return alloc_flags;
3692
3693 if (zone_idx(zone) != ZONE_NORMAL)
3694 return alloc_flags;
3695
3696 /*
3697 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3698 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3699 * on UMA that if Normal is populated then so is DMA32.
3700 */
3701 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3702 if (nr_online_nodes > 1 && !populated_zone(--zone))
3703 return alloc_flags;
3704
3705 alloc_flags |= ALLOC_NOFRAGMENT;
3706#endif /* CONFIG_ZONE_DMA32 */
3707 return alloc_flags;
3708}
3709
3710static inline unsigned int current_alloc_flags(gfp_t gfp_mask,
3711 unsigned int alloc_flags)
3712{
3713#ifdef CONFIG_CMA
3714 unsigned int pflags = current->flags;
3715
3716 if (!(pflags & PF_MEMALLOC_NOCMA) &&
3717 gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3718 alloc_flags |= ALLOC_CMA;
3719
3720#endif
3721 return alloc_flags;
3722}
3723
3724/*
3725 * get_page_from_freelist goes through the zonelist trying to allocate
3726 * a page.
3727 */
3728static struct page *
3729get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3730 const struct alloc_context *ac)
3731{
3732 struct zoneref *z;
3733 struct zone *zone;
3734 struct pglist_data *last_pgdat_dirty_limit = NULL;
3735 bool no_fallback;
3736
3737retry:
3738 /*
3739 * Scan zonelist, looking for a zone with enough free.
3740 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3741 */
3742 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3743 z = ac->preferred_zoneref;
3744 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist,
3745 ac->highest_zoneidx, ac->nodemask) {
3746 struct page *page;
3747 unsigned long mark;
3748
3749 if (cpusets_enabled() &&
3750 (alloc_flags & ALLOC_CPUSET) &&
3751 !__cpuset_zone_allowed(zone, gfp_mask))
3752 continue;
3753 /*
3754 * When allocating a page cache page for writing, we
3755 * want to get it from a node that is within its dirty
3756 * limit, such that no single node holds more than its
3757 * proportional share of globally allowed dirty pages.
3758 * The dirty limits take into account the node's
3759 * lowmem reserves and high watermark so that kswapd
3760 * should be able to balance it without having to
3761 * write pages from its LRU list.
3762 *
3763 * XXX: For now, allow allocations to potentially
3764 * exceed the per-node dirty limit in the slowpath
3765 * (spread_dirty_pages unset) before going into reclaim,
3766 * which is important when on a NUMA setup the allowed
3767 * nodes are together not big enough to reach the
3768 * global limit. The proper fix for these situations
3769 * will require awareness of nodes in the
3770 * dirty-throttling and the flusher threads.
3771 */
3772 if (ac->spread_dirty_pages) {
3773 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3774 continue;
3775
3776 if (!node_dirty_ok(zone->zone_pgdat)) {
3777 last_pgdat_dirty_limit = zone->zone_pgdat;
3778 continue;
3779 }
3780 }
3781
3782 if (no_fallback && nr_online_nodes > 1 &&
3783 zone != ac->preferred_zoneref->zone) {
3784 int local_nid;
3785
3786 /*
3787 * If moving to a remote node, retry but allow
3788 * fragmenting fallbacks. Locality is more important
3789 * than fragmentation avoidance.
3790 */
3791 local_nid = zone_to_nid(ac->preferred_zoneref->zone);
3792 if (zone_to_nid(zone) != local_nid) {
3793 alloc_flags &= ~ALLOC_NOFRAGMENT;
3794 goto retry;
3795 }
3796 }
3797
3798 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3799 if (!zone_watermark_fast(zone, order, mark,
3800 ac->highest_zoneidx, alloc_flags,
3801 gfp_mask)) {
3802 int ret;
3803
3804#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3805 /*
3806 * Watermark failed for this zone, but see if we can
3807 * grow this zone if it contains deferred pages.
3808 */
3809 if (static_branch_unlikely(&deferred_pages)) {
3810 if (_deferred_grow_zone(zone, order))
3811 goto try_this_zone;
3812 }
3813#endif
3814 /* Checked here to keep the fast path fast */
3815 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3816 if (alloc_flags & ALLOC_NO_WATERMARKS)
3817 goto try_this_zone;
3818
3819 if (node_reclaim_mode == 0 ||
3820 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3821 continue;
3822
3823 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3824 switch (ret) {
3825 case NODE_RECLAIM_NOSCAN:
3826 /* did not scan */
3827 continue;
3828 case NODE_RECLAIM_FULL:
3829 /* scanned but unreclaimable */
3830 continue;
3831 default:
3832 /* did we reclaim enough */
3833 if (zone_watermark_ok(zone, order, mark,
3834 ac->highest_zoneidx, alloc_flags))
3835 goto try_this_zone;
3836
3837 continue;
3838 }
3839 }
3840
3841try_this_zone:
3842 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3843 gfp_mask, alloc_flags, ac->migratetype);
3844 if (page) {
3845 prep_new_page(page, order, gfp_mask, alloc_flags);
3846
3847 /*
3848 * If this is a high-order atomic allocation then check
3849 * if the pageblock should be reserved for the future
3850 */
3851 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3852 reserve_highatomic_pageblock(page, zone, order);
3853
3854 return page;
3855 } else {
3856#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3857 /* Try again if zone has deferred pages */
3858 if (static_branch_unlikely(&deferred_pages)) {
3859 if (_deferred_grow_zone(zone, order))
3860 goto try_this_zone;
3861 }
3862#endif
3863 }
3864 }
3865
3866 /*
3867 * It's possible on a UMA machine to get through all zones that are
3868 * fragmented. If avoiding fragmentation, reset and try again.
3869 */
3870 if (no_fallback) {
3871 alloc_flags &= ~ALLOC_NOFRAGMENT;
3872 goto retry;
3873 }
3874
3875 return NULL;
3876}
3877
3878static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3879{
3880 unsigned int filter = SHOW_MEM_FILTER_NODES;
3881
3882 /*
3883 * This documents exceptions given to allocations in certain
3884 * contexts that are allowed to allocate outside current's set
3885 * of allowed nodes.
3886 */
3887 if (!(gfp_mask & __GFP_NOMEMALLOC))
3888 if (tsk_is_oom_victim(current) ||
3889 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3890 filter &= ~SHOW_MEM_FILTER_NODES;
3891 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3892 filter &= ~SHOW_MEM_FILTER_NODES;
3893
3894 show_mem(filter, nodemask);
3895}
3896
3897void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3898{
3899 struct va_format vaf;
3900 va_list args;
3901 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3902
3903 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3904 return;
3905
3906 va_start(args, fmt);
3907 vaf.fmt = fmt;
3908 vaf.va = &args;
3909 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3910 current->comm, &vaf, gfp_mask, &gfp_mask,
3911 nodemask_pr_args(nodemask));
3912 va_end(args);
3913
3914 cpuset_print_current_mems_allowed();
3915 pr_cont("\n");
3916 dump_stack();
3917 warn_alloc_show_mem(gfp_mask, nodemask);
3918}
3919
3920static inline struct page *
3921__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3922 unsigned int alloc_flags,
3923 const struct alloc_context *ac)
3924{
3925 struct page *page;
3926
3927 page = get_page_from_freelist(gfp_mask, order,
3928 alloc_flags|ALLOC_CPUSET, ac);
3929 /*
3930 * fallback to ignore cpuset restriction if our nodes
3931 * are depleted
3932 */
3933 if (!page)
3934 page = get_page_from_freelist(gfp_mask, order,
3935 alloc_flags, ac);
3936
3937 return page;
3938}
3939
3940static inline struct page *
3941__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3942 const struct alloc_context *ac, unsigned long *did_some_progress)
3943{
3944 struct oom_control oc = {
3945 .zonelist = ac->zonelist,
3946 .nodemask = ac->nodemask,
3947 .memcg = NULL,
3948 .gfp_mask = gfp_mask,
3949 .order = order,
3950 };
3951 struct page *page;
3952
3953 *did_some_progress = 0;
3954
3955 /*
3956 * Acquire the oom lock. If that fails, somebody else is
3957 * making progress for us.
3958 */
3959 if (!mutex_trylock(&oom_lock)) {
3960 *did_some_progress = 1;
3961 schedule_timeout_uninterruptible(1);
3962 return NULL;
3963 }
3964
3965 /*
3966 * Go through the zonelist yet one more time, keep very high watermark
3967 * here, this is only to catch a parallel oom killing, we must fail if
3968 * we're still under heavy pressure. But make sure that this reclaim
3969 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3970 * allocation which will never fail due to oom_lock already held.
3971 */
3972 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3973 ~__GFP_DIRECT_RECLAIM, order,
3974 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3975 if (page)
3976 goto out;
3977
3978 /* Coredumps can quickly deplete all memory reserves */
3979 if (current->flags & PF_DUMPCORE)
3980 goto out;
3981 /* The OOM killer will not help higher order allocs */
3982 if (order > PAGE_ALLOC_COSTLY_ORDER)
3983 goto out;
3984 /*
3985 * We have already exhausted all our reclaim opportunities without any
3986 * success so it is time to admit defeat. We will skip the OOM killer
3987 * because it is very likely that the caller has a more reasonable
3988 * fallback than shooting a random task.
3989 */
3990 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3991 goto out;
3992 /* The OOM killer does not needlessly kill tasks for lowmem */
3993 if (ac->highest_zoneidx < ZONE_NORMAL)
3994 goto out;
3995 if (pm_suspended_storage())
3996 goto out;
3997 /*
3998 * XXX: GFP_NOFS allocations should rather fail than rely on
3999 * other request to make a forward progress.
4000 * We are in an unfortunate situation where out_of_memory cannot
4001 * do much for this context but let's try it to at least get
4002 * access to memory reserved if the current task is killed (see
4003 * out_of_memory). Once filesystems are ready to handle allocation
4004 * failures more gracefully we should just bail out here.
4005 */
4006
4007 /* The OOM killer may not free memory on a specific node */
4008 if (gfp_mask & __GFP_THISNODE)
4009 goto out;
4010
4011 /* Exhausted what can be done so it's blame time */
4012 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
4013 *did_some_progress = 1;
4014
4015 /*
4016 * Help non-failing allocations by giving them access to memory
4017 * reserves
4018 */
4019 if (gfp_mask & __GFP_NOFAIL)
4020 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
4021 ALLOC_NO_WATERMARKS, ac);
4022 }
4023out:
4024 mutex_unlock(&oom_lock);
4025 return page;
4026}
4027
4028/*
4029 * Maximum number of compaction retries wit a progress before OOM
4030 * killer is consider as the only way to move forward.
4031 */
4032#define MAX_COMPACT_RETRIES 16
4033
4034#ifdef CONFIG_COMPACTION
4035/* Try memory compaction for high-order allocations before reclaim */
4036static struct page *
4037__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4038 unsigned int alloc_flags, const struct alloc_context *ac,
4039 enum compact_priority prio, enum compact_result *compact_result)
4040{
4041 struct page *page = NULL;
4042 unsigned long pflags;
4043 unsigned int noreclaim_flag;
4044
4045 if (!order)
4046 return NULL;
4047
4048 psi_memstall_enter(&pflags);
4049 noreclaim_flag = memalloc_noreclaim_save();
4050
4051 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
4052 prio, &page);
4053
4054 memalloc_noreclaim_restore(noreclaim_flag);
4055 psi_memstall_leave(&pflags);
4056
4057 /*
4058 * At least in one zone compaction wasn't deferred or skipped, so let's
4059 * count a compaction stall
4060 */
4061 count_vm_event(COMPACTSTALL);
4062
4063 /* Prep a captured page if available */
4064 if (page)
4065 prep_new_page(page, order, gfp_mask, alloc_flags);
4066
4067 /* Try get a page from the freelist if available */
4068 if (!page)
4069 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4070
4071 if (page) {
4072 struct zone *zone = page_zone(page);
4073
4074 zone->compact_blockskip_flush = false;
4075 compaction_defer_reset(zone, order, true);
4076 count_vm_event(COMPACTSUCCESS);
4077 return page;
4078 }
4079
4080 /*
4081 * It's bad if compaction run occurs and fails. The most likely reason
4082 * is that pages exist, but not enough to satisfy watermarks.
4083 */
4084 count_vm_event(COMPACTFAIL);
4085
4086 cond_resched();
4087
4088 return NULL;
4089}
4090
4091static inline bool
4092should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
4093 enum compact_result compact_result,
4094 enum compact_priority *compact_priority,
4095 int *compaction_retries)
4096{
4097 int max_retries = MAX_COMPACT_RETRIES;
4098 int min_priority;
4099 bool ret = false;
4100 int retries = *compaction_retries;
4101 enum compact_priority priority = *compact_priority;
4102
4103 if (!order)
4104 return false;
4105
4106 if (compaction_made_progress(compact_result))
4107 (*compaction_retries)++;
4108
4109 /*
4110 * compaction considers all the zone as desperately out of memory
4111 * so it doesn't really make much sense to retry except when the
4112 * failure could be caused by insufficient priority
4113 */
4114 if (compaction_failed(compact_result))
4115 goto check_priority;
4116
4117 /*
4118 * compaction was skipped because there are not enough order-0 pages
4119 * to work with, so we retry only if it looks like reclaim can help.
4120 */
4121 if (compaction_needs_reclaim(compact_result)) {
4122 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
4123 goto out;
4124 }
4125
4126 /*
4127 * make sure the compaction wasn't deferred or didn't bail out early
4128 * due to locks contention before we declare that we should give up.
4129 * But the next retry should use a higher priority if allowed, so
4130 * we don't just keep bailing out endlessly.
4131 */
4132 if (compaction_withdrawn(compact_result)) {
4133 goto check_priority;
4134 }
4135
4136 /*
4137 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
4138 * costly ones because they are de facto nofail and invoke OOM
4139 * killer to move on while costly can fail and users are ready
4140 * to cope with that. 1/4 retries is rather arbitrary but we
4141 * would need much more detailed feedback from compaction to
4142 * make a better decision.
4143 */
4144 if (order > PAGE_ALLOC_COSTLY_ORDER)
4145 max_retries /= 4;
4146 if (*compaction_retries <= max_retries) {
4147 ret = true;
4148 goto out;
4149 }
4150
4151 /*
4152 * Make sure there are attempts at the highest priority if we exhausted
4153 * all retries or failed at the lower priorities.
4154 */
4155check_priority:
4156 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
4157 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
4158
4159 if (*compact_priority > min_priority) {
4160 (*compact_priority)--;
4161 *compaction_retries = 0;
4162 ret = true;
4163 }
4164out:
4165 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
4166 return ret;
4167}
4168#else
4169static inline struct page *
4170__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
4171 unsigned int alloc_flags, const struct alloc_context *ac,
4172 enum compact_priority prio, enum compact_result *compact_result)
4173{
4174 *compact_result = COMPACT_SKIPPED;
4175 return NULL;
4176}
4177
4178static inline bool
4179should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
4180 enum compact_result compact_result,
4181 enum compact_priority *compact_priority,
4182 int *compaction_retries)
4183{
4184 struct zone *zone;
4185 struct zoneref *z;
4186
4187 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
4188 return false;
4189
4190 /*
4191 * There are setups with compaction disabled which would prefer to loop
4192 * inside the allocator rather than hit the oom killer prematurely.
4193 * Let's give them a good hope and keep retrying while the order-0
4194 * watermarks are OK.
4195 */
4196 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4197 ac->highest_zoneidx, ac->nodemask) {
4198 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
4199 ac->highest_zoneidx, alloc_flags))
4200 return true;
4201 }
4202 return false;
4203}
4204#endif /* CONFIG_COMPACTION */
4205
4206#ifdef CONFIG_LOCKDEP
4207static struct lockdep_map __fs_reclaim_map =
4208 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
4209
4210static bool __need_fs_reclaim(gfp_t gfp_mask)
4211{
4212 gfp_mask = current_gfp_context(gfp_mask);
4213
4214 /* no reclaim without waiting on it */
4215 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
4216 return false;
4217
4218 /* this guy won't enter reclaim */
4219 if (current->flags & PF_MEMALLOC)
4220 return false;
4221
4222 /* We're only interested __GFP_FS allocations for now */
4223 if (!(gfp_mask & __GFP_FS))
4224 return false;
4225
4226 if (gfp_mask & __GFP_NOLOCKDEP)
4227 return false;
4228
4229 return true;
4230}
4231
4232void __fs_reclaim_acquire(void)
4233{
4234 lock_map_acquire(&__fs_reclaim_map);
4235}
4236
4237void __fs_reclaim_release(void)
4238{
4239 lock_map_release(&__fs_reclaim_map);
4240}
4241
4242void fs_reclaim_acquire(gfp_t gfp_mask)
4243{
4244 if (__need_fs_reclaim(gfp_mask))
4245 __fs_reclaim_acquire();
4246}
4247EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
4248
4249void fs_reclaim_release(gfp_t gfp_mask)
4250{
4251 if (__need_fs_reclaim(gfp_mask))
4252 __fs_reclaim_release();
4253}
4254EXPORT_SYMBOL_GPL(fs_reclaim_release);
4255#endif
4256
4257/* Perform direct synchronous page reclaim */
4258static int
4259__perform_reclaim(gfp_t gfp_mask, unsigned int order,
4260 const struct alloc_context *ac)
4261{
4262 int progress;
4263 unsigned int noreclaim_flag;
4264 unsigned long pflags;
4265
4266 cond_resched();
4267
4268 /* We now go into synchronous reclaim */
4269 cpuset_memory_pressure_bump();
4270 psi_memstall_enter(&pflags);
4271 fs_reclaim_acquire(gfp_mask);
4272 noreclaim_flag = memalloc_noreclaim_save();
4273
4274 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
4275 ac->nodemask);
4276
4277 memalloc_noreclaim_restore(noreclaim_flag);
4278 fs_reclaim_release(gfp_mask);
4279 psi_memstall_leave(&pflags);
4280
4281 cond_resched();
4282
4283 return progress;
4284}
4285
4286/* The really slow allocator path where we enter direct reclaim */
4287static inline struct page *
4288__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
4289 unsigned int alloc_flags, const struct alloc_context *ac,
4290 unsigned long *did_some_progress)
4291{
4292 struct page *page = NULL;
4293 bool drained = false;
4294
4295 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
4296 if (unlikely(!(*did_some_progress)))
4297 return NULL;
4298
4299retry:
4300 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4301
4302 /*
4303 * If an allocation failed after direct reclaim, it could be because
4304 * pages are pinned on the per-cpu lists or in high alloc reserves.
4305 * Shrink them and try again
4306 */
4307 if (!page && !drained) {
4308 unreserve_highatomic_pageblock(ac, false);
4309 drain_all_pages(NULL);
4310 drained = true;
4311 goto retry;
4312 }
4313
4314 return page;
4315}
4316
4317static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
4318 const struct alloc_context *ac)
4319{
4320 struct zoneref *z;
4321 struct zone *zone;
4322 pg_data_t *last_pgdat = NULL;
4323 enum zone_type highest_zoneidx = ac->highest_zoneidx;
4324
4325 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
4326 ac->nodemask) {
4327 if (last_pgdat != zone->zone_pgdat)
4328 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
4329 last_pgdat = zone->zone_pgdat;
4330 }
4331}
4332
4333static inline unsigned int
4334gfp_to_alloc_flags(gfp_t gfp_mask)
4335{
4336 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
4337
4338 /*
4339 * __GFP_HIGH is assumed to be the same as ALLOC_HIGH
4340 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4341 * to save two branches.
4342 */
4343 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
4344 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4345
4346 /*
4347 * The caller may dip into page reserves a bit more if the caller
4348 * cannot run direct reclaim, or if the caller has realtime scheduling
4349 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4350 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
4351 */
4352 alloc_flags |= (__force int)
4353 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4354
4355 if (gfp_mask & __GFP_ATOMIC) {
4356 /*
4357 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4358 * if it can't schedule.
4359 */
4360 if (!(gfp_mask & __GFP_NOMEMALLOC))
4361 alloc_flags |= ALLOC_HARDER;
4362 /*
4363 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
4364 * comment for __cpuset_node_allowed().
4365 */
4366 alloc_flags &= ~ALLOC_CPUSET;
4367 } else if (unlikely(rt_task(current)) && !in_interrupt())
4368 alloc_flags |= ALLOC_HARDER;
4369
4370 alloc_flags = current_alloc_flags(gfp_mask, alloc_flags);
4371
4372 return alloc_flags;
4373}
4374
4375static bool oom_reserves_allowed(struct task_struct *tsk)
4376{
4377 if (!tsk_is_oom_victim(tsk))
4378 return false;
4379
4380 /*
4381 * !MMU doesn't have oom reaper so give access to memory reserves
4382 * only to the thread with TIF_MEMDIE set
4383 */
4384 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4385 return false;
4386
4387 return true;
4388}
4389
4390/*
4391 * Distinguish requests which really need access to full memory
4392 * reserves from oom victims which can live with a portion of it
4393 */
4394static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4395{
4396 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4397 return 0;
4398 if (gfp_mask & __GFP_MEMALLOC)
4399 return ALLOC_NO_WATERMARKS;
4400 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4401 return ALLOC_NO_WATERMARKS;
4402 if (!in_interrupt()) {
4403 if (current->flags & PF_MEMALLOC)
4404 return ALLOC_NO_WATERMARKS;
4405 else if (oom_reserves_allowed(current))
4406 return ALLOC_OOM;
4407 }
4408
4409 return 0;
4410}
4411
4412bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4413{
4414 return !!__gfp_pfmemalloc_flags(gfp_mask);
4415}
4416
4417/*
4418 * Checks whether it makes sense to retry the reclaim to make a forward progress
4419 * for the given allocation request.
4420 *
4421 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4422 * without success, or when we couldn't even meet the watermark if we
4423 * reclaimed all remaining pages on the LRU lists.
4424 *
4425 * Returns true if a retry is viable or false to enter the oom path.
4426 */
4427static inline bool
4428should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4429 struct alloc_context *ac, int alloc_flags,
4430 bool did_some_progress, int *no_progress_loops)
4431{
4432 struct zone *zone;
4433 struct zoneref *z;
4434 bool ret = false;
4435
4436 /*
4437 * Costly allocations might have made a progress but this doesn't mean
4438 * their order will become available due to high fragmentation so
4439 * always increment the no progress counter for them
4440 */
4441 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4442 *no_progress_loops = 0;
4443 else
4444 (*no_progress_loops)++;
4445
4446 /*
4447 * Make sure we converge to OOM if we cannot make any progress
4448 * several times in the row.
4449 */
4450 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
4451 /* Before OOM, exhaust highatomic_reserve */
4452 return unreserve_highatomic_pageblock(ac, true);
4453 }
4454
4455 /*
4456 * Keep reclaiming pages while there is a chance this will lead
4457 * somewhere. If none of the target zones can satisfy our allocation
4458 * request even if all reclaimable pages are considered then we are
4459 * screwed and have to go OOM.
4460 */
4461 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4462 ac->highest_zoneidx, ac->nodemask) {
4463 unsigned long available;
4464 unsigned long reclaimable;
4465 unsigned long min_wmark = min_wmark_pages(zone);
4466 bool wmark;
4467
4468 available = reclaimable = zone_reclaimable_pages(zone);
4469 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4470
4471 /*
4472 * Would the allocation succeed if we reclaimed all
4473 * reclaimable pages?
4474 */
4475 wmark = __zone_watermark_ok(zone, order, min_wmark,
4476 ac->highest_zoneidx, alloc_flags, available);
4477 trace_reclaim_retry_zone(z, order, reclaimable,
4478 available, min_wmark, *no_progress_loops, wmark);
4479 if (wmark) {
4480 /*
4481 * If we didn't make any progress and have a lot of
4482 * dirty + writeback pages then we should wait for
4483 * an IO to complete to slow down the reclaim and
4484 * prevent from pre mature OOM
4485 */
4486 if (!did_some_progress) {
4487 unsigned long write_pending;
4488
4489 write_pending = zone_page_state_snapshot(zone,
4490 NR_ZONE_WRITE_PENDING);
4491
4492 if (2 * write_pending > reclaimable) {
4493 congestion_wait(BLK_RW_ASYNC, HZ/10);
4494 return true;
4495 }
4496 }
4497
4498 ret = true;
4499 goto out;
4500 }
4501 }
4502
4503out:
4504 /*
4505 * Memory allocation/reclaim might be called from a WQ context and the
4506 * current implementation of the WQ concurrency control doesn't
4507 * recognize that a particular WQ is congested if the worker thread is
4508 * looping without ever sleeping. Therefore we have to do a short sleep
4509 * here rather than calling cond_resched().
4510 */
4511 if (current->flags & PF_WQ_WORKER)
4512 schedule_timeout_uninterruptible(1);
4513 else
4514 cond_resched();
4515 return ret;
4516}
4517
4518static inline bool
4519check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4520{
4521 /*
4522 * It's possible that cpuset's mems_allowed and the nodemask from
4523 * mempolicy don't intersect. This should be normally dealt with by
4524 * policy_nodemask(), but it's possible to race with cpuset update in
4525 * such a way the check therein was true, and then it became false
4526 * before we got our cpuset_mems_cookie here.
4527 * This assumes that for all allocations, ac->nodemask can come only
4528 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4529 * when it does not intersect with the cpuset restrictions) or the
4530 * caller can deal with a violated nodemask.
4531 */
4532 if (cpusets_enabled() && ac->nodemask &&
4533 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4534 ac->nodemask = NULL;
4535 return true;
4536 }
4537
4538 /*
4539 * When updating a task's mems_allowed or mempolicy nodemask, it is
4540 * possible to race with parallel threads in such a way that our
4541 * allocation can fail while the mask is being updated. If we are about
4542 * to fail, check if the cpuset changed during allocation and if so,
4543 * retry.
4544 */
4545 if (read_mems_allowed_retry(cpuset_mems_cookie))
4546 return true;
4547
4548 return false;
4549}
4550
4551static inline struct page *
4552__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4553 struct alloc_context *ac)
4554{
4555 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4556 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4557 struct page *page = NULL;
4558 unsigned int alloc_flags;
4559 unsigned long did_some_progress;
4560 enum compact_priority compact_priority;
4561 enum compact_result compact_result;
4562 int compaction_retries;
4563 int no_progress_loops;
4564 unsigned int cpuset_mems_cookie;
4565 int reserve_flags;
4566
4567 /*
4568 * We also sanity check to catch abuse of atomic reserves being used by
4569 * callers that are not in atomic context.
4570 */
4571 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4572 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4573 gfp_mask &= ~__GFP_ATOMIC;
4574
4575retry_cpuset:
4576 compaction_retries = 0;
4577 no_progress_loops = 0;
4578 compact_priority = DEF_COMPACT_PRIORITY;
4579 cpuset_mems_cookie = read_mems_allowed_begin();
4580
4581 /*
4582 * The fast path uses conservative alloc_flags to succeed only until
4583 * kswapd needs to be woken up, and to avoid the cost of setting up
4584 * alloc_flags precisely. So we do that now.
4585 */
4586 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4587
4588 /*
4589 * We need to recalculate the starting point for the zonelist iterator
4590 * because we might have used different nodemask in the fast path, or
4591 * there was a cpuset modification and we are retrying - otherwise we
4592 * could end up iterating over non-eligible zones endlessly.
4593 */
4594 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4595 ac->highest_zoneidx, ac->nodemask);
4596 if (!ac->preferred_zoneref->zone)
4597 goto nopage;
4598
4599 if (alloc_flags & ALLOC_KSWAPD)
4600 wake_all_kswapds(order, gfp_mask, ac);
4601
4602 /*
4603 * The adjusted alloc_flags might result in immediate success, so try
4604 * that first
4605 */
4606 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4607 if (page)
4608 goto got_pg;
4609
4610 /*
4611 * For costly allocations, try direct compaction first, as it's likely
4612 * that we have enough base pages and don't need to reclaim. For non-
4613 * movable high-order allocations, do that as well, as compaction will
4614 * try prevent permanent fragmentation by migrating from blocks of the
4615 * same migratetype.
4616 * Don't try this for allocations that are allowed to ignore
4617 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4618 */
4619 if (can_direct_reclaim &&
4620 (costly_order ||
4621 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4622 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4623 page = __alloc_pages_direct_compact(gfp_mask, order,
4624 alloc_flags, ac,
4625 INIT_COMPACT_PRIORITY,
4626 &compact_result);
4627 if (page)
4628 goto got_pg;
4629
4630 /*
4631 * Checks for costly allocations with __GFP_NORETRY, which
4632 * includes some THP page fault allocations
4633 */
4634 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4635 /*
4636 * If allocating entire pageblock(s) and compaction
4637 * failed because all zones are below low watermarks
4638 * or is prohibited because it recently failed at this
4639 * order, fail immediately unless the allocator has
4640 * requested compaction and reclaim retry.
4641 *
4642 * Reclaim is
4643 * - potentially very expensive because zones are far
4644 * below their low watermarks or this is part of very
4645 * bursty high order allocations,
4646 * - not guaranteed to help because isolate_freepages()
4647 * may not iterate over freed pages as part of its
4648 * linear scan, and
4649 * - unlikely to make entire pageblocks free on its
4650 * own.
4651 */
4652 if (compact_result == COMPACT_SKIPPED ||
4653 compact_result == COMPACT_DEFERRED)
4654 goto nopage;
4655
4656 /*
4657 * Looks like reclaim/compaction is worth trying, but
4658 * sync compaction could be very expensive, so keep
4659 * using async compaction.
4660 */
4661 compact_priority = INIT_COMPACT_PRIORITY;
4662 }
4663 }
4664
4665retry:
4666 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4667 if (alloc_flags & ALLOC_KSWAPD)
4668 wake_all_kswapds(order, gfp_mask, ac);
4669
4670 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4671 if (reserve_flags)
4672 alloc_flags = current_alloc_flags(gfp_mask, reserve_flags);
4673
4674 /*
4675 * Reset the nodemask and zonelist iterators if memory policies can be
4676 * ignored. These allocations are high priority and system rather than
4677 * user oriented.
4678 */
4679 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4680 ac->nodemask = NULL;
4681 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4682 ac->highest_zoneidx, ac->nodemask);
4683 }
4684
4685 /* Attempt with potentially adjusted zonelist and alloc_flags */
4686 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4687 if (page)
4688 goto got_pg;
4689
4690 /* Caller is not willing to reclaim, we can't balance anything */
4691 if (!can_direct_reclaim)
4692 goto nopage;
4693
4694 /* Avoid recursion of direct reclaim */
4695 if (current->flags & PF_MEMALLOC)
4696 goto nopage;
4697
4698 /* Try direct reclaim and then allocating */
4699 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4700 &did_some_progress);
4701 if (page)
4702 goto got_pg;
4703
4704 /* Try direct compaction and then allocating */
4705 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4706 compact_priority, &compact_result);
4707 if (page)
4708 goto got_pg;
4709
4710 /* Do not loop if specifically requested */
4711 if (gfp_mask & __GFP_NORETRY)
4712 goto nopage;
4713
4714 /*
4715 * Do not retry costly high order allocations unless they are
4716 * __GFP_RETRY_MAYFAIL
4717 */
4718 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4719 goto nopage;
4720
4721 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4722 did_some_progress > 0, &no_progress_loops))
4723 goto retry;
4724
4725 /*
4726 * It doesn't make any sense to retry for the compaction if the order-0
4727 * reclaim is not able to make any progress because the current
4728 * implementation of the compaction depends on the sufficient amount
4729 * of free memory (see __compaction_suitable)
4730 */
4731 if (did_some_progress > 0 &&
4732 should_compact_retry(ac, order, alloc_flags,
4733 compact_result, &compact_priority,
4734 &compaction_retries))
4735 goto retry;
4736
4737
4738 /* Deal with possible cpuset update races before we start OOM killing */
4739 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4740 goto retry_cpuset;
4741
4742 /* Reclaim has failed us, start killing things */
4743 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4744 if (page)
4745 goto got_pg;
4746
4747 /* Avoid allocations with no watermarks from looping endlessly */
4748 if (tsk_is_oom_victim(current) &&
4749 (alloc_flags & ALLOC_OOM ||
4750 (gfp_mask & __GFP_NOMEMALLOC)))
4751 goto nopage;
4752
4753 /* Retry as long as the OOM killer is making progress */
4754 if (did_some_progress) {
4755 no_progress_loops = 0;
4756 goto retry;
4757 }
4758
4759nopage:
4760 /* Deal with possible cpuset update races before we fail */
4761 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4762 goto retry_cpuset;
4763
4764 /*
4765 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4766 * we always retry
4767 */
4768 if (gfp_mask & __GFP_NOFAIL) {
4769 /*
4770 * All existing users of the __GFP_NOFAIL are blockable, so warn
4771 * of any new users that actually require GFP_NOWAIT
4772 */
4773 if (WARN_ON_ONCE(!can_direct_reclaim))
4774 goto fail;
4775
4776 /*
4777 * PF_MEMALLOC request from this context is rather bizarre
4778 * because we cannot reclaim anything and only can loop waiting
4779 * for somebody to do a work for us
4780 */
4781 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4782
4783 /*
4784 * non failing costly orders are a hard requirement which we
4785 * are not prepared for much so let's warn about these users
4786 * so that we can identify them and convert them to something
4787 * else.
4788 */
4789 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4790
4791 /*
4792 * Help non-failing allocations by giving them access to memory
4793 * reserves but do not use ALLOC_NO_WATERMARKS because this
4794 * could deplete whole memory reserves which would just make
4795 * the situation worse
4796 */
4797 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4798 if (page)
4799 goto got_pg;
4800
4801 cond_resched();
4802 goto retry;
4803 }
4804fail:
4805 warn_alloc(gfp_mask, ac->nodemask,
4806 "page allocation failure: order:%u", order);
4807got_pg:
4808 return page;
4809}
4810
4811static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4812 int preferred_nid, nodemask_t *nodemask,
4813 struct alloc_context *ac, gfp_t *alloc_mask,
4814 unsigned int *alloc_flags)
4815{
4816 ac->highest_zoneidx = gfp_zone(gfp_mask);
4817 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4818 ac->nodemask = nodemask;
4819 ac->migratetype = gfp_migratetype(gfp_mask);
4820
4821 if (cpusets_enabled()) {
4822 *alloc_mask |= __GFP_HARDWALL;
4823 /*
4824 * When we are in the interrupt context, it is irrelevant
4825 * to the current task context. It means that any node ok.
4826 */
4827 if (!in_interrupt() && !ac->nodemask)
4828 ac->nodemask = &cpuset_current_mems_allowed;
4829 else
4830 *alloc_flags |= ALLOC_CPUSET;
4831 }
4832
4833 fs_reclaim_acquire(gfp_mask);
4834 fs_reclaim_release(gfp_mask);
4835
4836 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4837
4838 if (should_fail_alloc_page(gfp_mask, order))
4839 return false;
4840
4841 *alloc_flags = current_alloc_flags(gfp_mask, *alloc_flags);
4842
4843 return true;
4844}
4845
4846/* Determine whether to spread dirty pages and what the first usable zone */
4847static inline void finalise_ac(gfp_t gfp_mask, struct alloc_context *ac)
4848{
4849 /* Dirty zone balancing only done in the fast path */
4850 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4851
4852 /*
4853 * The preferred zone is used for statistics but crucially it is
4854 * also used as the starting point for the zonelist iterator. It
4855 * may get reset for allocations that ignore memory policies.
4856 */
4857 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4858 ac->highest_zoneidx, ac->nodemask);
4859}
4860
4861/*
4862 * This is the 'heart' of the zoned buddy allocator.
4863 */
4864struct page *
4865__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4866 nodemask_t *nodemask)
4867{
4868 struct page *page;
4869 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4870 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4871 struct alloc_context ac = { };
4872
4873 /*
4874 * There are several places where we assume that the order value is sane
4875 * so bail out early if the request is out of bound.
4876 */
4877 if (unlikely(order >= MAX_ORDER)) {
4878 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4879 return NULL;
4880 }
4881
4882 gfp_mask &= gfp_allowed_mask;
4883 alloc_mask = gfp_mask;
4884 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4885 return NULL;
4886
4887 finalise_ac(gfp_mask, &ac);
4888
4889 /*
4890 * Forbid the first pass from falling back to types that fragment
4891 * memory until all local zones are considered.
4892 */
4893 alloc_flags |= alloc_flags_nofragment(ac.preferred_zoneref->zone, gfp_mask);
4894
4895 /* First allocation attempt */
4896 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4897 if (likely(page))
4898 goto out;
4899
4900 /*
4901 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4902 * resp. GFP_NOIO which has to be inherited for all allocation requests
4903 * from a particular context which has been marked by
4904 * memalloc_no{fs,io}_{save,restore}.
4905 */
4906 alloc_mask = current_gfp_context(gfp_mask);
4907 ac.spread_dirty_pages = false;
4908
4909 /*
4910 * Restore the original nodemask if it was potentially replaced with
4911 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4912 */
4913 ac.nodemask = nodemask;
4914
4915 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4916
4917out:
4918 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4919 unlikely(__memcg_kmem_charge_page(page, gfp_mask, order) != 0)) {
4920 __free_pages(page, order);
4921 page = NULL;
4922 }
4923
4924 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4925
4926 return page;
4927}
4928EXPORT_SYMBOL(__alloc_pages_nodemask);
4929
4930/*
4931 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4932 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4933 * you need to access high mem.
4934 */
4935unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4936{
4937 struct page *page;
4938
4939 page = alloc_pages(gfp_mask & ~__GFP_HIGHMEM, order);
4940 if (!page)
4941 return 0;
4942 return (unsigned long) page_address(page);
4943}
4944EXPORT_SYMBOL(__get_free_pages);
4945
4946unsigned long get_zeroed_page(gfp_t gfp_mask)
4947{
4948 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4949}
4950EXPORT_SYMBOL(get_zeroed_page);
4951
4952static inline void free_the_page(struct page *page, unsigned int order)
4953{
4954 if (order == 0) /* Via pcp? */
4955 free_unref_page(page);
4956 else
4957 __free_pages_ok(page, order);
4958}
4959
4960void __free_pages(struct page *page, unsigned int order)
4961{
4962 if (put_page_testzero(page))
4963 free_the_page(page, order);
4964}
4965EXPORT_SYMBOL(__free_pages);
4966
4967void free_pages(unsigned long addr, unsigned int order)
4968{
4969 if (addr != 0) {
4970 VM_BUG_ON(!virt_addr_valid((void *)addr));
4971 __free_pages(virt_to_page((void *)addr), order);
4972 }
4973}
4974
4975EXPORT_SYMBOL(free_pages);
4976
4977/*
4978 * Page Fragment:
4979 * An arbitrary-length arbitrary-offset area of memory which resides
4980 * within a 0 or higher order page. Multiple fragments within that page
4981 * are individually refcounted, in the page's reference counter.
4982 *
4983 * The page_frag functions below provide a simple allocation framework for
4984 * page fragments. This is used by the network stack and network device
4985 * drivers to provide a backing region of memory for use as either an
4986 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4987 */
4988static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4989 gfp_t gfp_mask)
4990{
4991 struct page *page = NULL;
4992 gfp_t gfp = gfp_mask;
4993
4994#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4995 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4996 __GFP_NOMEMALLOC;
4997 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4998 PAGE_FRAG_CACHE_MAX_ORDER);
4999 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
5000#endif
5001 if (unlikely(!page))
5002 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
5003
5004 nc->va = page ? page_address(page) : NULL;
5005
5006 return page;
5007}
5008
5009void __page_frag_cache_drain(struct page *page, unsigned int count)
5010{
5011 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
5012
5013 if (page_ref_sub_and_test(page, count))
5014 free_the_page(page, compound_order(page));
5015}
5016EXPORT_SYMBOL(__page_frag_cache_drain);
5017
5018void *page_frag_alloc(struct page_frag_cache *nc,
5019 unsigned int fragsz, gfp_t gfp_mask)
5020{
5021 unsigned int size = PAGE_SIZE;
5022 struct page *page;
5023 int offset;
5024
5025 if (unlikely(!nc->va)) {
5026refill:
5027 page = __page_frag_cache_refill(nc, gfp_mask);
5028 if (!page)
5029 return NULL;
5030
5031#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5032 /* if size can vary use size else just use PAGE_SIZE */
5033 size = nc->size;
5034#endif
5035 /* Even if we own the page, we do not use atomic_set().
5036 * This would break get_page_unless_zero() users.
5037 */
5038 page_ref_add(page, PAGE_FRAG_CACHE_MAX_SIZE);
5039
5040 /* reset page count bias and offset to start of new frag */
5041 nc->pfmemalloc = page_is_pfmemalloc(page);
5042 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5043 nc->offset = size;
5044 }
5045
5046 offset = nc->offset - fragsz;
5047 if (unlikely(offset < 0)) {
5048 page = virt_to_page(nc->va);
5049
5050 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
5051 goto refill;
5052
5053#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
5054 /* if size can vary use size else just use PAGE_SIZE */
5055 size = nc->size;
5056#endif
5057 /* OK, page count is 0, we can safely set it */
5058 set_page_count(page, PAGE_FRAG_CACHE_MAX_SIZE + 1);
5059
5060 /* reset page count bias and offset to start of new frag */
5061 nc->pagecnt_bias = PAGE_FRAG_CACHE_MAX_SIZE + 1;
5062 offset = size - fragsz;
5063 }
5064
5065 nc->pagecnt_bias--;
5066 nc->offset = offset;
5067
5068 return nc->va + offset;
5069}
5070EXPORT_SYMBOL(page_frag_alloc);
5071
5072/*
5073 * Frees a page fragment allocated out of either a compound or order 0 page.
5074 */
5075void page_frag_free(void *addr)
5076{
5077 struct page *page = virt_to_head_page(addr);
5078
5079 if (unlikely(put_page_testzero(page)))
5080 free_the_page(page, compound_order(page));
5081}
5082EXPORT_SYMBOL(page_frag_free);
5083
5084static void *make_alloc_exact(unsigned long addr, unsigned int order,
5085 size_t size)
5086{
5087 if (addr) {
5088 unsigned long alloc_end = addr + (PAGE_SIZE << order);
5089 unsigned long used = addr + PAGE_ALIGN(size);
5090
5091 split_page(virt_to_page((void *)addr), order);
5092 while (used < alloc_end) {
5093 free_page(used);
5094 used += PAGE_SIZE;
5095 }
5096 }
5097 return (void *)addr;
5098}
5099
5100/**
5101 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
5102 * @size: the number of bytes to allocate
5103 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5104 *
5105 * This function is similar to alloc_pages(), except that it allocates the
5106 * minimum number of pages to satisfy the request. alloc_pages() can only
5107 * allocate memory in power-of-two pages.
5108 *
5109 * This function is also limited by MAX_ORDER.
5110 *
5111 * Memory allocated by this function must be released by free_pages_exact().
5112 *
5113 * Return: pointer to the allocated area or %NULL in case of error.
5114 */
5115void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
5116{
5117 unsigned int order = get_order(size);
5118 unsigned long addr;
5119
5120 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5121 gfp_mask &= ~__GFP_COMP;
5122
5123 addr = __get_free_pages(gfp_mask, order);
5124 return make_alloc_exact(addr, order, size);
5125}
5126EXPORT_SYMBOL(alloc_pages_exact);
5127
5128/**
5129 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
5130 * pages on a node.
5131 * @nid: the preferred node ID where memory should be allocated
5132 * @size: the number of bytes to allocate
5133 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
5134 *
5135 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
5136 * back.
5137 *
5138 * Return: pointer to the allocated area or %NULL in case of error.
5139 */
5140void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
5141{
5142 unsigned int order = get_order(size);
5143 struct page *p;
5144
5145 if (WARN_ON_ONCE(gfp_mask & __GFP_COMP))
5146 gfp_mask &= ~__GFP_COMP;
5147
5148 p = alloc_pages_node(nid, gfp_mask, order);
5149 if (!p)
5150 return NULL;
5151 return make_alloc_exact((unsigned long)page_address(p), order, size);
5152}
5153
5154/**
5155 * free_pages_exact - release memory allocated via alloc_pages_exact()
5156 * @virt: the value returned by alloc_pages_exact.
5157 * @size: size of allocation, same value as passed to alloc_pages_exact().
5158 *
5159 * Release the memory allocated by a previous call to alloc_pages_exact.
5160 */
5161void free_pages_exact(void *virt, size_t size)
5162{
5163 unsigned long addr = (unsigned long)virt;
5164 unsigned long end = addr + PAGE_ALIGN(size);
5165
5166 while (addr < end) {
5167 free_page(addr);
5168 addr += PAGE_SIZE;
5169 }
5170}
5171EXPORT_SYMBOL(free_pages_exact);
5172
5173/**
5174 * nr_free_zone_pages - count number of pages beyond high watermark
5175 * @offset: The zone index of the highest zone
5176 *
5177 * nr_free_zone_pages() counts the number of pages which are beyond the
5178 * high watermark within all zones at or below a given zone index. For each
5179 * zone, the number of pages is calculated as:
5180 *
5181 * nr_free_zone_pages = managed_pages - high_pages
5182 *
5183 * Return: number of pages beyond high watermark.
5184 */
5185static unsigned long nr_free_zone_pages(int offset)
5186{
5187 struct zoneref *z;
5188 struct zone *zone;
5189
5190 /* Just pick one node, since fallback list is circular */
5191 unsigned long sum = 0;
5192
5193 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
5194
5195 for_each_zone_zonelist(zone, z, zonelist, offset) {
5196 unsigned long size = zone_managed_pages(zone);
5197 unsigned long high = high_wmark_pages(zone);
5198 if (size > high)
5199 sum += size - high;
5200 }
5201
5202 return sum;
5203}
5204
5205/**
5206 * nr_free_buffer_pages - count number of pages beyond high watermark
5207 *
5208 * nr_free_buffer_pages() counts the number of pages which are beyond the high
5209 * watermark within ZONE_DMA and ZONE_NORMAL.
5210 *
5211 * Return: number of pages beyond high watermark within ZONE_DMA and
5212 * ZONE_NORMAL.
5213 */
5214unsigned long nr_free_buffer_pages(void)
5215{
5216 return nr_free_zone_pages(gfp_zone(GFP_USER));
5217}
5218EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5219
5220static inline void show_node(struct zone *zone)
5221{
5222 if (IS_ENABLED(CONFIG_NUMA))
5223 printk("Node %d ", zone_to_nid(zone));
5224}
5225
5226long si_mem_available(void)
5227{
5228 long available;
5229 unsigned long pagecache;
5230 unsigned long wmark_low = 0;
5231 unsigned long pages[NR_LRU_LISTS];
5232 unsigned long reclaimable;
5233 struct zone *zone;
5234 int lru;
5235
5236 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
5237 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
5238
5239 for_each_zone(zone)
5240 wmark_low += low_wmark_pages(zone);
5241
5242 /*
5243 * Estimate the amount of memory available for userspace allocations,
5244 * without causing swapping.
5245 */
5246 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
5247
5248 /*
5249 * Not all the page cache can be freed, otherwise the system will
5250 * start swapping. Assume at least half of the page cache, or the
5251 * low watermark worth of cache, needs to stay.
5252 */
5253 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
5254 pagecache -= min(pagecache / 2, wmark_low);
5255 available += pagecache;
5256
5257 /*
5258 * Part of the reclaimable slab and other kernel memory consists of
5259 * items that are in use, and cannot be freed. Cap this estimate at the
5260 * low watermark.
5261 */
5262 reclaimable = global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B) +
5263 global_node_page_state(NR_KERNEL_MISC_RECLAIMABLE);
5264 available += reclaimable - min(reclaimable / 2, wmark_low);
5265
5266 if (available < 0)
5267 available = 0;
5268 return available;
5269}
5270EXPORT_SYMBOL_GPL(si_mem_available);
5271
5272void si_meminfo(struct sysinfo *val)
5273{
5274 val->totalram = totalram_pages();
5275 val->sharedram = global_node_page_state(NR_SHMEM);
5276 val->freeram = global_zone_page_state(NR_FREE_PAGES);
5277 val->bufferram = nr_blockdev_pages();
5278 val->totalhigh = totalhigh_pages();
5279 val->freehigh = nr_free_highpages();
5280 val->mem_unit = PAGE_SIZE;
5281}
5282
5283EXPORT_SYMBOL(si_meminfo);
5284
5285#ifdef CONFIG_NUMA
5286void si_meminfo_node(struct sysinfo *val, int nid)
5287{
5288 int zone_type; /* needs to be signed */
5289 unsigned long managed_pages = 0;
5290 unsigned long managed_highpages = 0;
5291 unsigned long free_highpages = 0;
5292 pg_data_t *pgdat = NODE_DATA(nid);
5293
5294 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
5295 managed_pages += zone_managed_pages(&pgdat->node_zones[zone_type]);
5296 val->totalram = managed_pages;
5297 val->sharedram = node_page_state(pgdat, NR_SHMEM);
5298 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
5299#ifdef CONFIG_HIGHMEM
5300 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
5301 struct zone *zone = &pgdat->node_zones[zone_type];
5302
5303 if (is_highmem(zone)) {
5304 managed_highpages += zone_managed_pages(zone);
5305 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
5306 }
5307 }
5308 val->totalhigh = managed_highpages;
5309 val->freehigh = free_highpages;
5310#else
5311 val->totalhigh = managed_highpages;
5312 val->freehigh = free_highpages;
5313#endif
5314 val->mem_unit = PAGE_SIZE;
5315}
5316#endif
5317
5318/*
5319 * Determine whether the node should be displayed or not, depending on whether
5320 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
5321 */
5322static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
5323{
5324 if (!(flags & SHOW_MEM_FILTER_NODES))
5325 return false;
5326
5327 /*
5328 * no node mask - aka implicit memory numa policy. Do not bother with
5329 * the synchronization - read_mems_allowed_begin - because we do not
5330 * have to be precise here.
5331 */
5332 if (!nodemask)
5333 nodemask = &cpuset_current_mems_allowed;
5334
5335 return !node_isset(nid, *nodemask);
5336}
5337
5338#define K(x) ((x) << (PAGE_SHIFT-10))
5339
5340static void show_migration_types(unsigned char type)
5341{
5342 static const char types[MIGRATE_TYPES] = {
5343 [MIGRATE_UNMOVABLE] = 'U',
5344 [MIGRATE_MOVABLE] = 'M',
5345 [MIGRATE_RECLAIMABLE] = 'E',
5346 [MIGRATE_HIGHATOMIC] = 'H',
5347#ifdef CONFIG_CMA
5348 [MIGRATE_CMA] = 'C',
5349#endif
5350#ifdef CONFIG_MEMORY_ISOLATION
5351 [MIGRATE_ISOLATE] = 'I',
5352#endif
5353 };
5354 char tmp[MIGRATE_TYPES + 1];
5355 char *p = tmp;
5356 int i;
5357
5358 for (i = 0; i < MIGRATE_TYPES; i++) {
5359 if (type & (1 << i))
5360 *p++ = types[i];
5361 }
5362
5363 *p = '\0';
5364 printk(KERN_CONT "(%s) ", tmp);
5365}
5366
5367/*
5368 * Show free area list (used inside shift_scroll-lock stuff)
5369 * We also calculate the percentage fragmentation. We do this by counting the
5370 * memory on each free list with the exception of the first item on the list.
5371 *
5372 * Bits in @filter:
5373 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
5374 * cpuset.
5375 */
5376void show_free_areas(unsigned int filter, nodemask_t *nodemask)
5377{
5378 unsigned long free_pcp = 0;
5379 int cpu;
5380 struct zone *zone;
5381 pg_data_t *pgdat;
5382
5383 for_each_populated_zone(zone) {
5384 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5385 continue;
5386
5387 for_each_online_cpu(cpu)
5388 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5389 }
5390
5391 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
5392 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
5393 " unevictable:%lu dirty:%lu writeback:%lu\n"
5394 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
5395 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
5396 " free:%lu free_pcp:%lu free_cma:%lu\n",
5397 global_node_page_state(NR_ACTIVE_ANON),
5398 global_node_page_state(NR_INACTIVE_ANON),
5399 global_node_page_state(NR_ISOLATED_ANON),
5400 global_node_page_state(NR_ACTIVE_FILE),
5401 global_node_page_state(NR_INACTIVE_FILE),
5402 global_node_page_state(NR_ISOLATED_FILE),
5403 global_node_page_state(NR_UNEVICTABLE),
5404 global_node_page_state(NR_FILE_DIRTY),
5405 global_node_page_state(NR_WRITEBACK),
5406 global_node_page_state_pages(NR_SLAB_RECLAIMABLE_B),
5407 global_node_page_state_pages(NR_SLAB_UNRECLAIMABLE_B),
5408 global_node_page_state(NR_FILE_MAPPED),
5409 global_node_page_state(NR_SHMEM),
5410 global_zone_page_state(NR_PAGETABLE),
5411 global_zone_page_state(NR_BOUNCE),
5412 global_zone_page_state(NR_FREE_PAGES),
5413 free_pcp,
5414 global_zone_page_state(NR_FREE_CMA_PAGES));
5415
5416 for_each_online_pgdat(pgdat) {
5417 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
5418 continue;
5419
5420 printk("Node %d"
5421 " active_anon:%lukB"
5422 " inactive_anon:%lukB"
5423 " active_file:%lukB"
5424 " inactive_file:%lukB"
5425 " unevictable:%lukB"
5426 " isolated(anon):%lukB"
5427 " isolated(file):%lukB"
5428 " mapped:%lukB"
5429 " dirty:%lukB"
5430 " writeback:%lukB"
5431 " shmem:%lukB"
5432#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5433 " shmem_thp: %lukB"
5434 " shmem_pmdmapped: %lukB"
5435 " anon_thp: %lukB"
5436#endif
5437 " writeback_tmp:%lukB"
5438 " kernel_stack:%lukB"
5439#ifdef CONFIG_SHADOW_CALL_STACK
5440 " shadow_call_stack:%lukB"
5441#endif
5442 " all_unreclaimable? %s"
5443 "\n",
5444 pgdat->node_id,
5445 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
5446 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
5447 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
5448 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
5449 K(node_page_state(pgdat, NR_UNEVICTABLE)),
5450 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
5451 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
5452 K(node_page_state(pgdat, NR_FILE_MAPPED)),
5453 K(node_page_state(pgdat, NR_FILE_DIRTY)),
5454 K(node_page_state(pgdat, NR_WRITEBACK)),
5455 K(node_page_state(pgdat, NR_SHMEM)),
5456#ifdef CONFIG_TRANSPARENT_HUGEPAGE
5457 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
5458 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
5459 * HPAGE_PMD_NR),
5460 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
5461#endif
5462 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
5463 node_page_state(pgdat, NR_KERNEL_STACK_KB),
5464#ifdef CONFIG_SHADOW_CALL_STACK
5465 node_page_state(pgdat, NR_KERNEL_SCS_KB),
5466#endif
5467 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
5468 "yes" : "no");
5469 }
5470
5471 for_each_populated_zone(zone) {
5472 int i;
5473
5474 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5475 continue;
5476
5477 free_pcp = 0;
5478 for_each_online_cpu(cpu)
5479 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
5480
5481 show_node(zone);
5482 printk(KERN_CONT
5483 "%s"
5484 " free:%lukB"
5485 " min:%lukB"
5486 " low:%lukB"
5487 " high:%lukB"
5488 " reserved_highatomic:%luKB"
5489 " active_anon:%lukB"
5490 " inactive_anon:%lukB"
5491 " active_file:%lukB"
5492 " inactive_file:%lukB"
5493 " unevictable:%lukB"
5494 " writepending:%lukB"
5495 " present:%lukB"
5496 " managed:%lukB"
5497 " mlocked:%lukB"
5498 " pagetables:%lukB"
5499 " bounce:%lukB"
5500 " free_pcp:%lukB"
5501 " local_pcp:%ukB"
5502 " free_cma:%lukB"
5503 "\n",
5504 zone->name,
5505 K(zone_page_state(zone, NR_FREE_PAGES)),
5506 K(min_wmark_pages(zone)),
5507 K(low_wmark_pages(zone)),
5508 K(high_wmark_pages(zone)),
5509 K(zone->nr_reserved_highatomic),
5510 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
5511 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
5512 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
5513 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
5514 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
5515 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
5516 K(zone->present_pages),
5517 K(zone_managed_pages(zone)),
5518 K(zone_page_state(zone, NR_MLOCK)),
5519 K(zone_page_state(zone, NR_PAGETABLE)),
5520 K(zone_page_state(zone, NR_BOUNCE)),
5521 K(free_pcp),
5522 K(this_cpu_read(zone->pageset->pcp.count)),
5523 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
5524 printk("lowmem_reserve[]:");
5525 for (i = 0; i < MAX_NR_ZONES; i++)
5526 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
5527 printk(KERN_CONT "\n");
5528 }
5529
5530 for_each_populated_zone(zone) {
5531 unsigned int order;
5532 unsigned long nr[MAX_ORDER], flags, total = 0;
5533 unsigned char types[MAX_ORDER];
5534
5535 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5536 continue;
5537 show_node(zone);
5538 printk(KERN_CONT "%s: ", zone->name);
5539
5540 spin_lock_irqsave(&zone->lock, flags);
5541 for (order = 0; order < MAX_ORDER; order++) {
5542 struct free_area *area = &zone->free_area[order];
5543 int type;
5544
5545 nr[order] = area->nr_free;
5546 total += nr[order] << order;
5547
5548 types[order] = 0;
5549 for (type = 0; type < MIGRATE_TYPES; type++) {
5550 if (!free_area_empty(area, type))
5551 types[order] |= 1 << type;
5552 }
5553 }
5554 spin_unlock_irqrestore(&zone->lock, flags);
5555 for (order = 0; order < MAX_ORDER; order++) {
5556 printk(KERN_CONT "%lu*%lukB ",
5557 nr[order], K(1UL) << order);
5558 if (nr[order])
5559 show_migration_types(types[order]);
5560 }
5561 printk(KERN_CONT "= %lukB\n", K(total));
5562 }
5563
5564 hugetlb_show_meminfo();
5565
5566 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5567
5568 show_swap_cache_info();
5569}
5570
5571static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5572{
5573 zoneref->zone = zone;
5574 zoneref->zone_idx = zone_idx(zone);
5575}
5576
5577/*
5578 * Builds allocation fallback zone lists.
5579 *
5580 * Add all populated zones of a node to the zonelist.
5581 */
5582static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5583{
5584 struct zone *zone;
5585 enum zone_type zone_type = MAX_NR_ZONES;
5586 int nr_zones = 0;
5587
5588 do {
5589 zone_type--;
5590 zone = pgdat->node_zones + zone_type;
5591 if (managed_zone(zone)) {
5592 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5593 check_highest_zone(zone_type);
5594 }
5595 } while (zone_type);
5596
5597 return nr_zones;
5598}
5599
5600#ifdef CONFIG_NUMA
5601
5602static int __parse_numa_zonelist_order(char *s)
5603{
5604 /*
5605 * We used to support different zonlists modes but they turned
5606 * out to be just not useful. Let's keep the warning in place
5607 * if somebody still use the cmd line parameter so that we do
5608 * not fail it silently
5609 */
5610 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5611 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5612 return -EINVAL;
5613 }
5614 return 0;
5615}
5616
5617char numa_zonelist_order[] = "Node";
5618
5619/*
5620 * sysctl handler for numa_zonelist_order
5621 */
5622int numa_zonelist_order_handler(struct ctl_table *table, int write,
5623 void *buffer, size_t *length, loff_t *ppos)
5624{
5625 if (write)
5626 return __parse_numa_zonelist_order(buffer);
5627 return proc_dostring(table, write, buffer, length, ppos);
5628}
5629
5630
5631#define MAX_NODE_LOAD (nr_online_nodes)
5632static int node_load[MAX_NUMNODES];
5633
5634/**
5635 * find_next_best_node - find the next node that should appear in a given node's fallback list
5636 * @node: node whose fallback list we're appending
5637 * @used_node_mask: nodemask_t of already used nodes
5638 *
5639 * We use a number of factors to determine which is the next node that should
5640 * appear on a given node's fallback list. The node should not have appeared
5641 * already in @node's fallback list, and it should be the next closest node
5642 * according to the distance array (which contains arbitrary distance values
5643 * from each node to each node in the system), and should also prefer nodes
5644 * with no CPUs, since presumably they'll have very little allocation pressure
5645 * on them otherwise.
5646 *
5647 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5648 */
5649static int find_next_best_node(int node, nodemask_t *used_node_mask)
5650{
5651 int n, val;
5652 int min_val = INT_MAX;
5653 int best_node = NUMA_NO_NODE;
5654 const struct cpumask *tmp = cpumask_of_node(0);
5655
5656 /* Use the local node if we haven't already */
5657 if (!node_isset(node, *used_node_mask)) {
5658 node_set(node, *used_node_mask);
5659 return node;
5660 }
5661
5662 for_each_node_state(n, N_MEMORY) {
5663
5664 /* Don't want a node to appear more than once */
5665 if (node_isset(n, *used_node_mask))
5666 continue;
5667
5668 /* Use the distance array to find the distance */
5669 val = node_distance(node, n);
5670
5671 /* Penalize nodes under us ("prefer the next node") */
5672 val += (n < node);
5673
5674 /* Give preference to headless and unused nodes */
5675 tmp = cpumask_of_node(n);
5676 if (!cpumask_empty(tmp))
5677 val += PENALTY_FOR_NODE_WITH_CPUS;
5678
5679 /* Slight preference for less loaded node */
5680 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5681 val += node_load[n];
5682
5683 if (val < min_val) {
5684 min_val = val;
5685 best_node = n;
5686 }
5687 }
5688
5689 if (best_node >= 0)
5690 node_set(best_node, *used_node_mask);
5691
5692 return best_node;
5693}
5694
5695
5696/*
5697 * Build zonelists ordered by node and zones within node.
5698 * This results in maximum locality--normal zone overflows into local
5699 * DMA zone, if any--but risks exhausting DMA zone.
5700 */
5701static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5702 unsigned nr_nodes)
5703{
5704 struct zoneref *zonerefs;
5705 int i;
5706
5707 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5708
5709 for (i = 0; i < nr_nodes; i++) {
5710 int nr_zones;
5711
5712 pg_data_t *node = NODE_DATA(node_order[i]);
5713
5714 nr_zones = build_zonerefs_node(node, zonerefs);
5715 zonerefs += nr_zones;
5716 }
5717 zonerefs->zone = NULL;
5718 zonerefs->zone_idx = 0;
5719}
5720
5721/*
5722 * Build gfp_thisnode zonelists
5723 */
5724static void build_thisnode_zonelists(pg_data_t *pgdat)
5725{
5726 struct zoneref *zonerefs;
5727 int nr_zones;
5728
5729 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5730 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5731 zonerefs += nr_zones;
5732 zonerefs->zone = NULL;
5733 zonerefs->zone_idx = 0;
5734}
5735
5736/*
5737 * Build zonelists ordered by zone and nodes within zones.
5738 * This results in conserving DMA zone[s] until all Normal memory is
5739 * exhausted, but results in overflowing to remote node while memory
5740 * may still exist in local DMA zone.
5741 */
5742
5743static void build_zonelists(pg_data_t *pgdat)
5744{
5745 static int node_order[MAX_NUMNODES];
5746 int node, load, nr_nodes = 0;
5747 nodemask_t used_mask = NODE_MASK_NONE;
5748 int local_node, prev_node;
5749
5750 /* NUMA-aware ordering of nodes */
5751 local_node = pgdat->node_id;
5752 load = nr_online_nodes;
5753 prev_node = local_node;
5754
5755 memset(node_order, 0, sizeof(node_order));
5756 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5757 /*
5758 * We don't want to pressure a particular node.
5759 * So adding penalty to the first node in same
5760 * distance group to make it round-robin.
5761 */
5762 if (node_distance(local_node, node) !=
5763 node_distance(local_node, prev_node))
5764 node_load[node] = load;
5765
5766 node_order[nr_nodes++] = node;
5767 prev_node = node;
5768 load--;
5769 }
5770
5771 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5772 build_thisnode_zonelists(pgdat);
5773}
5774
5775#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5776/*
5777 * Return node id of node used for "local" allocations.
5778 * I.e., first node id of first zone in arg node's generic zonelist.
5779 * Used for initializing percpu 'numa_mem', which is used primarily
5780 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5781 */
5782int local_memory_node(int node)
5783{
5784 struct zoneref *z;
5785
5786 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5787 gfp_zone(GFP_KERNEL),
5788 NULL);
5789 return zone_to_nid(z->zone);
5790}
5791#endif
5792
5793static void setup_min_unmapped_ratio(void);
5794static void setup_min_slab_ratio(void);
5795#else /* CONFIG_NUMA */
5796
5797static void build_zonelists(pg_data_t *pgdat)
5798{
5799 int node, local_node;
5800 struct zoneref *zonerefs;
5801 int nr_zones;
5802
5803 local_node = pgdat->node_id;
5804
5805 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5806 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5807 zonerefs += nr_zones;
5808
5809 /*
5810 * Now we build the zonelist so that it contains the zones
5811 * of all the other nodes.
5812 * We don't want to pressure a particular node, so when
5813 * building the zones for node N, we make sure that the
5814 * zones coming right after the local ones are those from
5815 * node N+1 (modulo N)
5816 */
5817 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5818 if (!node_online(node))
5819 continue;
5820 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5821 zonerefs += nr_zones;
5822 }
5823 for (node = 0; node < local_node; node++) {
5824 if (!node_online(node))
5825 continue;
5826 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5827 zonerefs += nr_zones;
5828 }
5829
5830 zonerefs->zone = NULL;
5831 zonerefs->zone_idx = 0;
5832}
5833
5834#endif /* CONFIG_NUMA */
5835
5836/*
5837 * Boot pageset table. One per cpu which is going to be used for all
5838 * zones and all nodes. The parameters will be set in such a way
5839 * that an item put on a list will immediately be handed over to
5840 * the buddy list. This is safe since pageset manipulation is done
5841 * with interrupts disabled.
5842 *
5843 * The boot_pagesets must be kept even after bootup is complete for
5844 * unused processors and/or zones. They do play a role for bootstrapping
5845 * hotplugged processors.
5846 *
5847 * zoneinfo_show() and maybe other functions do
5848 * not check if the processor is online before following the pageset pointer.
5849 * Other parts of the kernel may not check if the zone is available.
5850 */
5851static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5852static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5853static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5854
5855static void __build_all_zonelists(void *data)
5856{
5857 int nid;
5858 int __maybe_unused cpu;
5859 pg_data_t *self = data;
5860 static DEFINE_SPINLOCK(lock);
5861
5862 spin_lock(&lock);
5863
5864#ifdef CONFIG_NUMA
5865 memset(node_load, 0, sizeof(node_load));
5866#endif
5867
5868 /*
5869 * This node is hotadded and no memory is yet present. So just
5870 * building zonelists is fine - no need to touch other nodes.
5871 */
5872 if (self && !node_online(self->node_id)) {
5873 build_zonelists(self);
5874 } else {
5875 for_each_online_node(nid) {
5876 pg_data_t *pgdat = NODE_DATA(nid);
5877
5878 build_zonelists(pgdat);
5879 }
5880
5881#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5882 /*
5883 * We now know the "local memory node" for each node--
5884 * i.e., the node of the first zone in the generic zonelist.
5885 * Set up numa_mem percpu variable for on-line cpus. During
5886 * boot, only the boot cpu should be on-line; we'll init the
5887 * secondary cpus' numa_mem as they come on-line. During
5888 * node/memory hotplug, we'll fixup all on-line cpus.
5889 */
5890 for_each_online_cpu(cpu)
5891 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5892#endif
5893 }
5894
5895 spin_unlock(&lock);
5896}
5897
5898static noinline void __init
5899build_all_zonelists_init(void)
5900{
5901 int cpu;
5902
5903 __build_all_zonelists(NULL);
5904
5905 /*
5906 * Initialize the boot_pagesets that are going to be used
5907 * for bootstrapping processors. The real pagesets for
5908 * each zone will be allocated later when the per cpu
5909 * allocator is available.
5910 *
5911 * boot_pagesets are used also for bootstrapping offline
5912 * cpus if the system is already booted because the pagesets
5913 * are needed to initialize allocators on a specific cpu too.
5914 * F.e. the percpu allocator needs the page allocator which
5915 * needs the percpu allocator in order to allocate its pagesets
5916 * (a chicken-egg dilemma).
5917 */
5918 for_each_possible_cpu(cpu)
5919 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5920
5921 mminit_verify_zonelist();
5922 cpuset_init_current_mems_allowed();
5923}
5924
5925/*
5926 * unless system_state == SYSTEM_BOOTING.
5927 *
5928 * __ref due to call of __init annotated helper build_all_zonelists_init
5929 * [protected by SYSTEM_BOOTING].
5930 */
5931void __ref build_all_zonelists(pg_data_t *pgdat)
5932{
5933 unsigned long vm_total_pages;
5934
5935 if (system_state == SYSTEM_BOOTING) {
5936 build_all_zonelists_init();
5937 } else {
5938 __build_all_zonelists(pgdat);
5939 /* cpuset refresh routine should be here */
5940 }
5941 /* Get the number of free pages beyond high watermark in all zones. */
5942 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5943 /*
5944 * Disable grouping by mobility if the number of pages in the
5945 * system is too low to allow the mechanism to work. It would be
5946 * more accurate, but expensive to check per-zone. This check is
5947 * made on memory-hotadd so a system can start with mobility
5948 * disabled and enable it later
5949 */
5950 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5951 page_group_by_mobility_disabled = 1;
5952 else
5953 page_group_by_mobility_disabled = 0;
5954
5955 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5956 nr_online_nodes,
5957 page_group_by_mobility_disabled ? "off" : "on",
5958 vm_total_pages);
5959#ifdef CONFIG_NUMA
5960 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5961#endif
5962}
5963
5964/* If zone is ZONE_MOVABLE but memory is mirrored, it is an overlapped init */
5965static bool __meminit
5966overlap_memmap_init(unsigned long zone, unsigned long *pfn)
5967{
5968 static struct memblock_region *r;
5969
5970 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5971 if (!r || *pfn >= memblock_region_memory_end_pfn(r)) {
5972 for_each_memblock(memory, r) {
5973 if (*pfn < memblock_region_memory_end_pfn(r))
5974 break;
5975 }
5976 }
5977 if (*pfn >= memblock_region_memory_base_pfn(r) &&
5978 memblock_is_mirror(r)) {
5979 *pfn = memblock_region_memory_end_pfn(r);
5980 return true;
5981 }
5982 }
5983 return false;
5984}
5985
5986/*
5987 * Initially all pages are reserved - free ones are freed
5988 * up by memblock_free_all() once the early boot process is
5989 * done. Non-atomic initialization, single-pass.
5990 */
5991void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5992 unsigned long start_pfn, enum meminit_context context,
5993 struct vmem_altmap *altmap)
5994{
5995 unsigned long pfn, end_pfn = start_pfn + size;
5996 struct page *page;
5997
5998 if (highest_memmap_pfn < end_pfn - 1)
5999 highest_memmap_pfn = end_pfn - 1;
6000
6001#ifdef CONFIG_ZONE_DEVICE
6002 /*
6003 * Honor reservation requested by the driver for this ZONE_DEVICE
6004 * memory. We limit the total number of pages to initialize to just
6005 * those that might contain the memory mapping. We will defer the
6006 * ZONE_DEVICE page initialization until after we have released
6007 * the hotplug lock.
6008 */
6009 if (zone == ZONE_DEVICE) {
6010 if (!altmap)
6011 return;
6012
6013 if (start_pfn == altmap->base_pfn)
6014 start_pfn += altmap->reserve;
6015 end_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6016 }
6017#endif
6018
6019 for (pfn = start_pfn; pfn < end_pfn; ) {
6020 /*
6021 * There can be holes in boot-time mem_map[]s handed to this
6022 * function. They do not exist on hotplugged memory.
6023 */
6024 if (context == MEMINIT_EARLY) {
6025 if (overlap_memmap_init(zone, &pfn))
6026 continue;
6027 if (defer_init(nid, pfn, end_pfn))
6028 break;
6029 }
6030
6031 page = pfn_to_page(pfn);
6032 __init_single_page(page, pfn, zone, nid);
6033 if (context == MEMINIT_HOTPLUG)
6034 __SetPageReserved(page);
6035
6036 /*
6037 * Mark the block movable so that blocks are reserved for
6038 * movable at startup. This will force kernel allocations
6039 * to reserve their blocks rather than leaking throughout
6040 * the address space during boot when many long-lived
6041 * kernel allocations are made.
6042 *
6043 * bitmap is created for zone's valid pfn range. but memmap
6044 * can be created for invalid pages (for alignment)
6045 * check here not to call set_pageblock_migratetype() against
6046 * pfn out of zone.
6047 */
6048 if (!(pfn & (pageblock_nr_pages - 1))) {
6049 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6050 cond_resched();
6051 }
6052 pfn++;
6053 }
6054}
6055
6056#ifdef CONFIG_ZONE_DEVICE
6057void __ref memmap_init_zone_device(struct zone *zone,
6058 unsigned long start_pfn,
6059 unsigned long nr_pages,
6060 struct dev_pagemap *pgmap)
6061{
6062 unsigned long pfn, end_pfn = start_pfn + nr_pages;
6063 struct pglist_data *pgdat = zone->zone_pgdat;
6064 struct vmem_altmap *altmap = pgmap_altmap(pgmap);
6065 unsigned long zone_idx = zone_idx(zone);
6066 unsigned long start = jiffies;
6067 int nid = pgdat->node_id;
6068
6069 if (WARN_ON_ONCE(!pgmap || zone_idx(zone) != ZONE_DEVICE))
6070 return;
6071
6072 /*
6073 * The call to memmap_init_zone should have already taken care
6074 * of the pages reserved for the memmap, so we can just jump to
6075 * the end of that region and start processing the device pages.
6076 */
6077 if (altmap) {
6078 start_pfn = altmap->base_pfn + vmem_altmap_offset(altmap);
6079 nr_pages = end_pfn - start_pfn;
6080 }
6081
6082 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
6083 struct page *page = pfn_to_page(pfn);
6084
6085 __init_single_page(page, pfn, zone_idx, nid);
6086
6087 /*
6088 * Mark page reserved as it will need to wait for onlining
6089 * phase for it to be fully associated with a zone.
6090 *
6091 * We can use the non-atomic __set_bit operation for setting
6092 * the flag as we are still initializing the pages.
6093 */
6094 __SetPageReserved(page);
6095
6096 /*
6097 * ZONE_DEVICE pages union ->lru with a ->pgmap back pointer
6098 * and zone_device_data. It is a bug if a ZONE_DEVICE page is
6099 * ever freed or placed on a driver-private list.
6100 */
6101 page->pgmap = pgmap;
6102 page->zone_device_data = NULL;
6103
6104 /*
6105 * Mark the block movable so that blocks are reserved for
6106 * movable at startup. This will force kernel allocations
6107 * to reserve their blocks rather than leaking throughout
6108 * the address space during boot when many long-lived
6109 * kernel allocations are made.
6110 *
6111 * bitmap is created for zone's valid pfn range. but memmap
6112 * can be created for invalid pages (for alignment)
6113 * check here not to call set_pageblock_migratetype() against
6114 * pfn out of zone.
6115 *
6116 * Please note that MEMINIT_HOTPLUG path doesn't clear memmap
6117 * because this is done early in section_activate()
6118 */
6119 if (!(pfn & (pageblock_nr_pages - 1))) {
6120 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
6121 cond_resched();
6122 }
6123 }
6124
6125 pr_info("%s initialised %lu pages in %ums\n", __func__,
6126 nr_pages, jiffies_to_msecs(jiffies - start));
6127}
6128
6129#endif
6130static void __meminit zone_init_free_lists(struct zone *zone)
6131{
6132 unsigned int order, t;
6133 for_each_migratetype_order(order, t) {
6134 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
6135 zone->free_area[order].nr_free = 0;
6136 }
6137}
6138
6139void __meminit __weak memmap_init(unsigned long size, int nid,
6140 unsigned long zone,
6141 unsigned long range_start_pfn)
6142{
6143 unsigned long start_pfn, end_pfn;
6144 unsigned long range_end_pfn = range_start_pfn + size;
6145 int i;
6146
6147 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6148 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6149 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6150
6151 if (end_pfn > start_pfn) {
6152 size = end_pfn - start_pfn;
6153 memmap_init_zone(size, nid, zone, start_pfn,
6154 MEMINIT_EARLY, NULL);
6155 }
6156 }
6157}
6158
6159static int zone_batchsize(struct zone *zone)
6160{
6161#ifdef CONFIG_MMU
6162 int batch;
6163
6164 /*
6165 * The per-cpu-pages pools are set to around 1000th of the
6166 * size of the zone.
6167 */
6168 batch = zone_managed_pages(zone) / 1024;
6169 /* But no more than a meg. */
6170 if (batch * PAGE_SIZE > 1024 * 1024)
6171 batch = (1024 * 1024) / PAGE_SIZE;
6172 batch /= 4; /* We effectively *= 4 below */
6173 if (batch < 1)
6174 batch = 1;
6175
6176 /*
6177 * Clamp the batch to a 2^n - 1 value. Having a power
6178 * of 2 value was found to be more likely to have
6179 * suboptimal cache aliasing properties in some cases.
6180 *
6181 * For example if 2 tasks are alternately allocating
6182 * batches of pages, one task can end up with a lot
6183 * of pages of one half of the possible page colors
6184 * and the other with pages of the other colors.
6185 */
6186 batch = rounddown_pow_of_two(batch + batch/2) - 1;
6187
6188 return batch;
6189
6190#else
6191 /* The deferral and batching of frees should be suppressed under NOMMU
6192 * conditions.
6193 *
6194 * The problem is that NOMMU needs to be able to allocate large chunks
6195 * of contiguous memory as there's no hardware page translation to
6196 * assemble apparent contiguous memory from discontiguous pages.
6197 *
6198 * Queueing large contiguous runs of pages for batching, however,
6199 * causes the pages to actually be freed in smaller chunks. As there
6200 * can be a significant delay between the individual batches being
6201 * recycled, this leads to the once large chunks of space being
6202 * fragmented and becoming unavailable for high-order allocations.
6203 */
6204 return 0;
6205#endif
6206}
6207
6208/*
6209 * pcp->high and pcp->batch values are related and dependent on one another:
6210 * ->batch must never be higher then ->high.
6211 * The following function updates them in a safe manner without read side
6212 * locking.
6213 *
6214 * Any new users of pcp->batch and pcp->high should ensure they can cope with
6215 * those fields changing asynchronously (acording to the above rule).
6216 *
6217 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
6218 * outside of boot time (or some other assurance that no concurrent updaters
6219 * exist).
6220 */
6221static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
6222 unsigned long batch)
6223{
6224 /* start with a fail safe value for batch */
6225 pcp->batch = 1;
6226 smp_wmb();
6227
6228 /* Update high, then batch, in order */
6229 pcp->high = high;
6230 smp_wmb();
6231
6232 pcp->batch = batch;
6233}
6234
6235/* a companion to pageset_set_high() */
6236static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
6237{
6238 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
6239}
6240
6241static void pageset_init(struct per_cpu_pageset *p)
6242{
6243 struct per_cpu_pages *pcp;
6244 int migratetype;
6245
6246 memset(p, 0, sizeof(*p));
6247
6248 pcp = &p->pcp;
6249 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
6250 INIT_LIST_HEAD(&pcp->lists[migratetype]);
6251}
6252
6253static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
6254{
6255 pageset_init(p);
6256 pageset_set_batch(p, batch);
6257}
6258
6259/*
6260 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
6261 * to the value high for the pageset p.
6262 */
6263static void pageset_set_high(struct per_cpu_pageset *p,
6264 unsigned long high)
6265{
6266 unsigned long batch = max(1UL, high / 4);
6267 if ((high / 4) > (PAGE_SHIFT * 8))
6268 batch = PAGE_SHIFT * 8;
6269
6270 pageset_update(&p->pcp, high, batch);
6271}
6272
6273static void pageset_set_high_and_batch(struct zone *zone,
6274 struct per_cpu_pageset *pcp)
6275{
6276 if (percpu_pagelist_fraction)
6277 pageset_set_high(pcp,
6278 (zone_managed_pages(zone) /
6279 percpu_pagelist_fraction));
6280 else
6281 pageset_set_batch(pcp, zone_batchsize(zone));
6282}
6283
6284static void __meminit zone_pageset_init(struct zone *zone, int cpu)
6285{
6286 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
6287
6288 pageset_init(pcp);
6289 pageset_set_high_and_batch(zone, pcp);
6290}
6291
6292void __meminit setup_zone_pageset(struct zone *zone)
6293{
6294 int cpu;
6295 zone->pageset = alloc_percpu(struct per_cpu_pageset);
6296 for_each_possible_cpu(cpu)
6297 zone_pageset_init(zone, cpu);
6298}
6299
6300/*
6301 * Allocate per cpu pagesets and initialize them.
6302 * Before this call only boot pagesets were available.
6303 */
6304void __init setup_per_cpu_pageset(void)
6305{
6306 struct pglist_data *pgdat;
6307 struct zone *zone;
6308 int __maybe_unused cpu;
6309
6310 for_each_populated_zone(zone)
6311 setup_zone_pageset(zone);
6312
6313#ifdef CONFIG_NUMA
6314 /*
6315 * Unpopulated zones continue using the boot pagesets.
6316 * The numa stats for these pagesets need to be reset.
6317 * Otherwise, they will end up skewing the stats of
6318 * the nodes these zones are associated with.
6319 */
6320 for_each_possible_cpu(cpu) {
6321 struct per_cpu_pageset *pcp = &per_cpu(boot_pageset, cpu);
6322 memset(pcp->vm_numa_stat_diff, 0,
6323 sizeof(pcp->vm_numa_stat_diff));
6324 }
6325#endif
6326
6327 for_each_online_pgdat(pgdat)
6328 pgdat->per_cpu_nodestats =
6329 alloc_percpu(struct per_cpu_nodestat);
6330}
6331
6332static __meminit void zone_pcp_init(struct zone *zone)
6333{
6334 /*
6335 * per cpu subsystem is not up at this point. The following code
6336 * relies on the ability of the linker to provide the
6337 * offset of a (static) per cpu variable into the per cpu area.
6338 */
6339 zone->pageset = &boot_pageset;
6340
6341 if (populated_zone(zone))
6342 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
6343 zone->name, zone->present_pages,
6344 zone_batchsize(zone));
6345}
6346
6347void __meminit init_currently_empty_zone(struct zone *zone,
6348 unsigned long zone_start_pfn,
6349 unsigned long size)
6350{
6351 struct pglist_data *pgdat = zone->zone_pgdat;
6352 int zone_idx = zone_idx(zone) + 1;
6353
6354 if (zone_idx > pgdat->nr_zones)
6355 pgdat->nr_zones = zone_idx;
6356
6357 zone->zone_start_pfn = zone_start_pfn;
6358
6359 mminit_dprintk(MMINIT_TRACE, "memmap_init",
6360 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
6361 pgdat->node_id,
6362 (unsigned long)zone_idx(zone),
6363 zone_start_pfn, (zone_start_pfn + size));
6364
6365 zone_init_free_lists(zone);
6366 zone->initialized = 1;
6367}
6368
6369/**
6370 * get_pfn_range_for_nid - Return the start and end page frames for a node
6371 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
6372 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
6373 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
6374 *
6375 * It returns the start and end page frame of a node based on information
6376 * provided by memblock_set_node(). If called for a node
6377 * with no available memory, a warning is printed and the start and end
6378 * PFNs will be 0.
6379 */
6380void __init get_pfn_range_for_nid(unsigned int nid,
6381 unsigned long *start_pfn, unsigned long *end_pfn)
6382{
6383 unsigned long this_start_pfn, this_end_pfn;
6384 int i;
6385
6386 *start_pfn = -1UL;
6387 *end_pfn = 0;
6388
6389 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
6390 *start_pfn = min(*start_pfn, this_start_pfn);
6391 *end_pfn = max(*end_pfn, this_end_pfn);
6392 }
6393
6394 if (*start_pfn == -1UL)
6395 *start_pfn = 0;
6396}
6397
6398/*
6399 * This finds a zone that can be used for ZONE_MOVABLE pages. The
6400 * assumption is made that zones within a node are ordered in monotonic
6401 * increasing memory addresses so that the "highest" populated zone is used
6402 */
6403static void __init find_usable_zone_for_movable(void)
6404{
6405 int zone_index;
6406 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
6407 if (zone_index == ZONE_MOVABLE)
6408 continue;
6409
6410 if (arch_zone_highest_possible_pfn[zone_index] >
6411 arch_zone_lowest_possible_pfn[zone_index])
6412 break;
6413 }
6414
6415 VM_BUG_ON(zone_index == -1);
6416 movable_zone = zone_index;
6417}
6418
6419/*
6420 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
6421 * because it is sized independent of architecture. Unlike the other zones,
6422 * the starting point for ZONE_MOVABLE is not fixed. It may be different
6423 * in each node depending on the size of each node and how evenly kernelcore
6424 * is distributed. This helper function adjusts the zone ranges
6425 * provided by the architecture for a given node by using the end of the
6426 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
6427 * zones within a node are in order of monotonic increases memory addresses
6428 */
6429static void __init adjust_zone_range_for_zone_movable(int nid,
6430 unsigned long zone_type,
6431 unsigned long node_start_pfn,
6432 unsigned long node_end_pfn,
6433 unsigned long *zone_start_pfn,
6434 unsigned long *zone_end_pfn)
6435{
6436 /* Only adjust if ZONE_MOVABLE is on this node */
6437 if (zone_movable_pfn[nid]) {
6438 /* Size ZONE_MOVABLE */
6439 if (zone_type == ZONE_MOVABLE) {
6440 *zone_start_pfn = zone_movable_pfn[nid];
6441 *zone_end_pfn = min(node_end_pfn,
6442 arch_zone_highest_possible_pfn[movable_zone]);
6443
6444 /* Adjust for ZONE_MOVABLE starting within this range */
6445 } else if (!mirrored_kernelcore &&
6446 *zone_start_pfn < zone_movable_pfn[nid] &&
6447 *zone_end_pfn > zone_movable_pfn[nid]) {
6448 *zone_end_pfn = zone_movable_pfn[nid];
6449
6450 /* Check if this whole range is within ZONE_MOVABLE */
6451 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
6452 *zone_start_pfn = *zone_end_pfn;
6453 }
6454}
6455
6456/*
6457 * Return the number of pages a zone spans in a node, including holes
6458 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
6459 */
6460static unsigned long __init zone_spanned_pages_in_node(int nid,
6461 unsigned long zone_type,
6462 unsigned long node_start_pfn,
6463 unsigned long node_end_pfn,
6464 unsigned long *zone_start_pfn,
6465 unsigned long *zone_end_pfn)
6466{
6467 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6468 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6469 /* When hotadd a new node from cpu_up(), the node should be empty */
6470 if (!node_start_pfn && !node_end_pfn)
6471 return 0;
6472
6473 /* Get the start and end of the zone */
6474 *zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6475 *zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6476 adjust_zone_range_for_zone_movable(nid, zone_type,
6477 node_start_pfn, node_end_pfn,
6478 zone_start_pfn, zone_end_pfn);
6479
6480 /* Check that this node has pages within the zone's required range */
6481 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
6482 return 0;
6483
6484 /* Move the zone boundaries inside the node if necessary */
6485 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
6486 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
6487
6488 /* Return the spanned pages */
6489 return *zone_end_pfn - *zone_start_pfn;
6490}
6491
6492/*
6493 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
6494 * then all holes in the requested range will be accounted for.
6495 */
6496unsigned long __init __absent_pages_in_range(int nid,
6497 unsigned long range_start_pfn,
6498 unsigned long range_end_pfn)
6499{
6500 unsigned long nr_absent = range_end_pfn - range_start_pfn;
6501 unsigned long start_pfn, end_pfn;
6502 int i;
6503
6504 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6505 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
6506 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
6507 nr_absent -= end_pfn - start_pfn;
6508 }
6509 return nr_absent;
6510}
6511
6512/**
6513 * absent_pages_in_range - Return number of page frames in holes within a range
6514 * @start_pfn: The start PFN to start searching for holes
6515 * @end_pfn: The end PFN to stop searching for holes
6516 *
6517 * Return: the number of pages frames in memory holes within a range.
6518 */
6519unsigned long __init absent_pages_in_range(unsigned long start_pfn,
6520 unsigned long end_pfn)
6521{
6522 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
6523}
6524
6525/* Return the number of page frames in holes in a zone on a node */
6526static unsigned long __init zone_absent_pages_in_node(int nid,
6527 unsigned long zone_type,
6528 unsigned long node_start_pfn,
6529 unsigned long node_end_pfn)
6530{
6531 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
6532 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
6533 unsigned long zone_start_pfn, zone_end_pfn;
6534 unsigned long nr_absent;
6535
6536 /* When hotadd a new node from cpu_up(), the node should be empty */
6537 if (!node_start_pfn && !node_end_pfn)
6538 return 0;
6539
6540 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
6541 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
6542
6543 adjust_zone_range_for_zone_movable(nid, zone_type,
6544 node_start_pfn, node_end_pfn,
6545 &zone_start_pfn, &zone_end_pfn);
6546 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
6547
6548 /*
6549 * ZONE_MOVABLE handling.
6550 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6551 * and vice versa.
6552 */
6553 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6554 unsigned long start_pfn, end_pfn;
6555 struct memblock_region *r;
6556
6557 for_each_memblock(memory, r) {
6558 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6559 zone_start_pfn, zone_end_pfn);
6560 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6561 zone_start_pfn, zone_end_pfn);
6562
6563 if (zone_type == ZONE_MOVABLE &&
6564 memblock_is_mirror(r))
6565 nr_absent += end_pfn - start_pfn;
6566
6567 if (zone_type == ZONE_NORMAL &&
6568 !memblock_is_mirror(r))
6569 nr_absent += end_pfn - start_pfn;
6570 }
6571 }
6572
6573 return nr_absent;
6574}
6575
6576static void __init calculate_node_totalpages(struct pglist_data *pgdat,
6577 unsigned long node_start_pfn,
6578 unsigned long node_end_pfn)
6579{
6580 unsigned long realtotalpages = 0, totalpages = 0;
6581 enum zone_type i;
6582
6583 for (i = 0; i < MAX_NR_ZONES; i++) {
6584 struct zone *zone = pgdat->node_zones + i;
6585 unsigned long zone_start_pfn, zone_end_pfn;
6586 unsigned long spanned, absent;
6587 unsigned long size, real_size;
6588
6589 spanned = zone_spanned_pages_in_node(pgdat->node_id, i,
6590 node_start_pfn,
6591 node_end_pfn,
6592 &zone_start_pfn,
6593 &zone_end_pfn);
6594 absent = zone_absent_pages_in_node(pgdat->node_id, i,
6595 node_start_pfn,
6596 node_end_pfn);
6597
6598 size = spanned;
6599 real_size = size - absent;
6600
6601 if (size)
6602 zone->zone_start_pfn = zone_start_pfn;
6603 else
6604 zone->zone_start_pfn = 0;
6605 zone->spanned_pages = size;
6606 zone->present_pages = real_size;
6607
6608 totalpages += size;
6609 realtotalpages += real_size;
6610 }
6611
6612 pgdat->node_spanned_pages = totalpages;
6613 pgdat->node_present_pages = realtotalpages;
6614 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6615 realtotalpages);
6616}
6617
6618#ifndef CONFIG_SPARSEMEM
6619/*
6620 * Calculate the size of the zone->blockflags rounded to an unsigned long
6621 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6622 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6623 * round what is now in bits to nearest long in bits, then return it in
6624 * bytes.
6625 */
6626static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6627{
6628 unsigned long usemapsize;
6629
6630 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6631 usemapsize = roundup(zonesize, pageblock_nr_pages);
6632 usemapsize = usemapsize >> pageblock_order;
6633 usemapsize *= NR_PAGEBLOCK_BITS;
6634 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6635
6636 return usemapsize / 8;
6637}
6638
6639static void __ref setup_usemap(struct pglist_data *pgdat,
6640 struct zone *zone,
6641 unsigned long zone_start_pfn,
6642 unsigned long zonesize)
6643{
6644 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6645 zone->pageblock_flags = NULL;
6646 if (usemapsize) {
6647 zone->pageblock_flags =
6648 memblock_alloc_node(usemapsize, SMP_CACHE_BYTES,
6649 pgdat->node_id);
6650 if (!zone->pageblock_flags)
6651 panic("Failed to allocate %ld bytes for zone %s pageblock flags on node %d\n",
6652 usemapsize, zone->name, pgdat->node_id);
6653 }
6654}
6655#else
6656static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6657 unsigned long zone_start_pfn, unsigned long zonesize) {}
6658#endif /* CONFIG_SPARSEMEM */
6659
6660#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6661
6662/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6663void __init set_pageblock_order(void)
6664{
6665 unsigned int order;
6666
6667 /* Check that pageblock_nr_pages has not already been setup */
6668 if (pageblock_order)
6669 return;
6670
6671 if (HPAGE_SHIFT > PAGE_SHIFT)
6672 order = HUGETLB_PAGE_ORDER;
6673 else
6674 order = MAX_ORDER - 1;
6675
6676 /*
6677 * Assume the largest contiguous order of interest is a huge page.
6678 * This value may be variable depending on boot parameters on IA64 and
6679 * powerpc.
6680 */
6681 pageblock_order = order;
6682}
6683#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6684
6685/*
6686 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6687 * is unused as pageblock_order is set at compile-time. See
6688 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6689 * the kernel config
6690 */
6691void __init set_pageblock_order(void)
6692{
6693}
6694
6695#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6696
6697static unsigned long __init calc_memmap_size(unsigned long spanned_pages,
6698 unsigned long present_pages)
6699{
6700 unsigned long pages = spanned_pages;
6701
6702 /*
6703 * Provide a more accurate estimation if there are holes within
6704 * the zone and SPARSEMEM is in use. If there are holes within the
6705 * zone, each populated memory region may cost us one or two extra
6706 * memmap pages due to alignment because memmap pages for each
6707 * populated regions may not be naturally aligned on page boundary.
6708 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6709 */
6710 if (spanned_pages > present_pages + (present_pages >> 4) &&
6711 IS_ENABLED(CONFIG_SPARSEMEM))
6712 pages = present_pages;
6713
6714 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6715}
6716
6717#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6718static void pgdat_init_split_queue(struct pglist_data *pgdat)
6719{
6720 struct deferred_split *ds_queue = &pgdat->deferred_split_queue;
6721
6722 spin_lock_init(&ds_queue->split_queue_lock);
6723 INIT_LIST_HEAD(&ds_queue->split_queue);
6724 ds_queue->split_queue_len = 0;
6725}
6726#else
6727static void pgdat_init_split_queue(struct pglist_data *pgdat) {}
6728#endif
6729
6730#ifdef CONFIG_COMPACTION
6731static void pgdat_init_kcompactd(struct pglist_data *pgdat)
6732{
6733 init_waitqueue_head(&pgdat->kcompactd_wait);
6734}
6735#else
6736static void pgdat_init_kcompactd(struct pglist_data *pgdat) {}
6737#endif
6738
6739static void __meminit pgdat_init_internals(struct pglist_data *pgdat)
6740{
6741 pgdat_resize_init(pgdat);
6742
6743 pgdat_init_split_queue(pgdat);
6744 pgdat_init_kcompactd(pgdat);
6745
6746 init_waitqueue_head(&pgdat->kswapd_wait);
6747 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6748
6749 pgdat_page_ext_init(pgdat);
6750 spin_lock_init(&pgdat->lru_lock);
6751 lruvec_init(&pgdat->__lruvec);
6752}
6753
6754static void __meminit zone_init_internals(struct zone *zone, enum zone_type idx, int nid,
6755 unsigned long remaining_pages)
6756{
6757 atomic_long_set(&zone->managed_pages, remaining_pages);
6758 zone_set_nid(zone, nid);
6759 zone->name = zone_names[idx];
6760 zone->zone_pgdat = NODE_DATA(nid);
6761 spin_lock_init(&zone->lock);
6762 zone_seqlock_init(zone);
6763 zone_pcp_init(zone);
6764}
6765
6766/*
6767 * Set up the zone data structures
6768 * - init pgdat internals
6769 * - init all zones belonging to this node
6770 *
6771 * NOTE: this function is only called during memory hotplug
6772 */
6773#ifdef CONFIG_MEMORY_HOTPLUG
6774void __ref free_area_init_core_hotplug(int nid)
6775{
6776 enum zone_type z;
6777 pg_data_t *pgdat = NODE_DATA(nid);
6778
6779 pgdat_init_internals(pgdat);
6780 for (z = 0; z < MAX_NR_ZONES; z++)
6781 zone_init_internals(&pgdat->node_zones[z], z, nid, 0);
6782}
6783#endif
6784
6785/*
6786 * Set up the zone data structures:
6787 * - mark all pages reserved
6788 * - mark all memory queues empty
6789 * - clear the memory bitmaps
6790 *
6791 * NOTE: pgdat should get zeroed by caller.
6792 * NOTE: this function is only called during early init.
6793 */
6794static void __init free_area_init_core(struct pglist_data *pgdat)
6795{
6796 enum zone_type j;
6797 int nid = pgdat->node_id;
6798
6799 pgdat_init_internals(pgdat);
6800 pgdat->per_cpu_nodestats = &boot_nodestats;
6801
6802 for (j = 0; j < MAX_NR_ZONES; j++) {
6803 struct zone *zone = pgdat->node_zones + j;
6804 unsigned long size, freesize, memmap_pages;
6805 unsigned long zone_start_pfn = zone->zone_start_pfn;
6806
6807 size = zone->spanned_pages;
6808 freesize = zone->present_pages;
6809
6810 /*
6811 * Adjust freesize so that it accounts for how much memory
6812 * is used by this zone for memmap. This affects the watermark
6813 * and per-cpu initialisations
6814 */
6815 memmap_pages = calc_memmap_size(size, freesize);
6816 if (!is_highmem_idx(j)) {
6817 if (freesize >= memmap_pages) {
6818 freesize -= memmap_pages;
6819 if (memmap_pages)
6820 printk(KERN_DEBUG
6821 " %s zone: %lu pages used for memmap\n",
6822 zone_names[j], memmap_pages);
6823 } else
6824 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6825 zone_names[j], memmap_pages, freesize);
6826 }
6827
6828 /* Account for reserved pages */
6829 if (j == 0 && freesize > dma_reserve) {
6830 freesize -= dma_reserve;
6831 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6832 zone_names[0], dma_reserve);
6833 }
6834
6835 if (!is_highmem_idx(j))
6836 nr_kernel_pages += freesize;
6837 /* Charge for highmem memmap if there are enough kernel pages */
6838 else if (nr_kernel_pages > memmap_pages * 2)
6839 nr_kernel_pages -= memmap_pages;
6840 nr_all_pages += freesize;
6841
6842 /*
6843 * Set an approximate value for lowmem here, it will be adjusted
6844 * when the bootmem allocator frees pages into the buddy system.
6845 * And all highmem pages will be managed by the buddy system.
6846 */
6847 zone_init_internals(zone, j, nid, freesize);
6848
6849 if (!size)
6850 continue;
6851
6852 set_pageblock_order();
6853 setup_usemap(pgdat, zone, zone_start_pfn, size);
6854 init_currently_empty_zone(zone, zone_start_pfn, size);
6855 memmap_init(size, nid, j, zone_start_pfn);
6856 }
6857}
6858
6859#ifdef CONFIG_FLAT_NODE_MEM_MAP
6860static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6861{
6862 unsigned long __maybe_unused start = 0;
6863 unsigned long __maybe_unused offset = 0;
6864
6865 /* Skip empty nodes */
6866 if (!pgdat->node_spanned_pages)
6867 return;
6868
6869 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6870 offset = pgdat->node_start_pfn - start;
6871 /* ia64 gets its own node_mem_map, before this, without bootmem */
6872 if (!pgdat->node_mem_map) {
6873 unsigned long size, end;
6874 struct page *map;
6875
6876 /*
6877 * The zone's endpoints aren't required to be MAX_ORDER
6878 * aligned but the node_mem_map endpoints must be in order
6879 * for the buddy allocator to function correctly.
6880 */
6881 end = pgdat_end_pfn(pgdat);
6882 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6883 size = (end - start) * sizeof(struct page);
6884 map = memblock_alloc_node(size, SMP_CACHE_BYTES,
6885 pgdat->node_id);
6886 if (!map)
6887 panic("Failed to allocate %ld bytes for node %d memory map\n",
6888 size, pgdat->node_id);
6889 pgdat->node_mem_map = map + offset;
6890 }
6891 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6892 __func__, pgdat->node_id, (unsigned long)pgdat,
6893 (unsigned long)pgdat->node_mem_map);
6894#ifndef CONFIG_NEED_MULTIPLE_NODES
6895 /*
6896 * With no DISCONTIG, the global mem_map is just set as node 0's
6897 */
6898 if (pgdat == NODE_DATA(0)) {
6899 mem_map = NODE_DATA(0)->node_mem_map;
6900 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6901 mem_map -= offset;
6902 }
6903#endif
6904}
6905#else
6906static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6907#endif /* CONFIG_FLAT_NODE_MEM_MAP */
6908
6909#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6910static inline void pgdat_set_deferred_range(pg_data_t *pgdat)
6911{
6912 pgdat->first_deferred_pfn = ULONG_MAX;
6913}
6914#else
6915static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {}
6916#endif
6917
6918static void __init free_area_init_node(int nid)
6919{
6920 pg_data_t *pgdat = NODE_DATA(nid);
6921 unsigned long start_pfn = 0;
6922 unsigned long end_pfn = 0;
6923
6924 /* pg_data_t should be reset to zero when it's allocated */
6925 WARN_ON(pgdat->nr_zones || pgdat->kswapd_highest_zoneidx);
6926
6927 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6928
6929 pgdat->node_id = nid;
6930 pgdat->node_start_pfn = start_pfn;
6931 pgdat->per_cpu_nodestats = NULL;
6932
6933 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6934 (u64)start_pfn << PAGE_SHIFT,
6935 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6936 calculate_node_totalpages(pgdat, start_pfn, end_pfn);
6937
6938 alloc_node_mem_map(pgdat);
6939 pgdat_set_deferred_range(pgdat);
6940
6941 free_area_init_core(pgdat);
6942}
6943
6944void __init free_area_init_memoryless_node(int nid)
6945{
6946 free_area_init_node(nid);
6947}
6948
6949#if !defined(CONFIG_FLAT_NODE_MEM_MAP)
6950/*
6951 * Initialize all valid struct pages in the range [spfn, epfn) and mark them
6952 * PageReserved(). Return the number of struct pages that were initialized.
6953 */
6954static u64 __init init_unavailable_range(unsigned long spfn, unsigned long epfn)
6955{
6956 unsigned long pfn;
6957 u64 pgcnt = 0;
6958
6959 for (pfn = spfn; pfn < epfn; pfn++) {
6960 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages))) {
6961 pfn = ALIGN_DOWN(pfn, pageblock_nr_pages)
6962 + pageblock_nr_pages - 1;
6963 continue;
6964 }
6965 /*
6966 * Use a fake node/zone (0) for now. Some of these pages
6967 * (in memblock.reserved but not in memblock.memory) will
6968 * get re-initialized via reserve_bootmem_region() later.
6969 */
6970 __init_single_page(pfn_to_page(pfn), pfn, 0, 0);
6971 __SetPageReserved(pfn_to_page(pfn));
6972 pgcnt++;
6973 }
6974
6975 return pgcnt;
6976}
6977
6978/*
6979 * Only struct pages that are backed by physical memory are zeroed and
6980 * initialized by going through __init_single_page(). But, there are some
6981 * struct pages which are reserved in memblock allocator and their fields
6982 * may be accessed (for example page_to_pfn() on some configuration accesses
6983 * flags). We must explicitly initialize those struct pages.
6984 *
6985 * This function also addresses a similar issue where struct pages are left
6986 * uninitialized because the physical address range is not covered by
6987 * memblock.memory or memblock.reserved. That could happen when memblock
6988 * layout is manually configured via memmap=, or when the highest physical
6989 * address (max_pfn) does not end on a section boundary.
6990 */
6991static void __init init_unavailable_mem(void)
6992{
6993 phys_addr_t start, end;
6994 u64 i, pgcnt;
6995 phys_addr_t next = 0;
6996
6997 /*
6998 * Loop through unavailable ranges not covered by memblock.memory.
6999 */
7000 pgcnt = 0;
7001 for_each_mem_range(i, &memblock.memory, NULL,
7002 NUMA_NO_NODE, MEMBLOCK_NONE, &start, &end, NULL) {
7003 if (next < start)
7004 pgcnt += init_unavailable_range(PFN_DOWN(next),
7005 PFN_UP(start));
7006 next = end;
7007 }
7008
7009 /*
7010 * Early sections always have a fully populated memmap for the whole
7011 * section - see pfn_valid(). If the last section has holes at the
7012 * end and that section is marked "online", the memmap will be
7013 * considered initialized. Make sure that memmap has a well defined
7014 * state.
7015 */
7016 pgcnt += init_unavailable_range(PFN_DOWN(next),
7017 round_up(max_pfn, PAGES_PER_SECTION));
7018
7019 /*
7020 * Struct pages that do not have backing memory. This could be because
7021 * firmware is using some of this memory, or for some other reasons.
7022 */
7023 if (pgcnt)
7024 pr_info("Zeroed struct page in unavailable ranges: %lld pages", pgcnt);
7025}
7026#else
7027static inline void __init init_unavailable_mem(void)
7028{
7029}
7030#endif /* !CONFIG_FLAT_NODE_MEM_MAP */
7031
7032#if MAX_NUMNODES > 1
7033/*
7034 * Figure out the number of possible node ids.
7035 */
7036void __init setup_nr_node_ids(void)
7037{
7038 unsigned int highest;
7039
7040 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
7041 nr_node_ids = highest + 1;
7042}
7043#endif
7044
7045/**
7046 * node_map_pfn_alignment - determine the maximum internode alignment
7047 *
7048 * This function should be called after node map is populated and sorted.
7049 * It calculates the maximum power of two alignment which can distinguish
7050 * all the nodes.
7051 *
7052 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
7053 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
7054 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
7055 * shifted, 1GiB is enough and this function will indicate so.
7056 *
7057 * This is used to test whether pfn -> nid mapping of the chosen memory
7058 * model has fine enough granularity to avoid incorrect mapping for the
7059 * populated node map.
7060 *
7061 * Return: the determined alignment in pfn's. 0 if there is no alignment
7062 * requirement (single node).
7063 */
7064unsigned long __init node_map_pfn_alignment(void)
7065{
7066 unsigned long accl_mask = 0, last_end = 0;
7067 unsigned long start, end, mask;
7068 int last_nid = NUMA_NO_NODE;
7069 int i, nid;
7070
7071 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
7072 if (!start || last_nid < 0 || last_nid == nid) {
7073 last_nid = nid;
7074 last_end = end;
7075 continue;
7076 }
7077
7078 /*
7079 * Start with a mask granular enough to pin-point to the
7080 * start pfn and tick off bits one-by-one until it becomes
7081 * too coarse to separate the current node from the last.
7082 */
7083 mask = ~((1 << __ffs(start)) - 1);
7084 while (mask && last_end <= (start & (mask << 1)))
7085 mask <<= 1;
7086
7087 /* accumulate all internode masks */
7088 accl_mask |= mask;
7089 }
7090
7091 /* convert mask to number of pages */
7092 return ~accl_mask + 1;
7093}
7094
7095/**
7096 * find_min_pfn_with_active_regions - Find the minimum PFN registered
7097 *
7098 * Return: the minimum PFN based on information provided via
7099 * memblock_set_node().
7100 */
7101unsigned long __init find_min_pfn_with_active_regions(void)
7102{
7103 return PHYS_PFN(memblock_start_of_DRAM());
7104}
7105
7106/*
7107 * early_calculate_totalpages()
7108 * Sum pages in active regions for movable zone.
7109 * Populate N_MEMORY for calculating usable_nodes.
7110 */
7111static unsigned long __init early_calculate_totalpages(void)
7112{
7113 unsigned long totalpages = 0;
7114 unsigned long start_pfn, end_pfn;
7115 int i, nid;
7116
7117 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7118 unsigned long pages = end_pfn - start_pfn;
7119
7120 totalpages += pages;
7121 if (pages)
7122 node_set_state(nid, N_MEMORY);
7123 }
7124 return totalpages;
7125}
7126
7127/*
7128 * Find the PFN the Movable zone begins in each node. Kernel memory
7129 * is spread evenly between nodes as long as the nodes have enough
7130 * memory. When they don't, some nodes will have more kernelcore than
7131 * others
7132 */
7133static void __init find_zone_movable_pfns_for_nodes(void)
7134{
7135 int i, nid;
7136 unsigned long usable_startpfn;
7137 unsigned long kernelcore_node, kernelcore_remaining;
7138 /* save the state before borrow the nodemask */
7139 nodemask_t saved_node_state = node_states[N_MEMORY];
7140 unsigned long totalpages = early_calculate_totalpages();
7141 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
7142 struct memblock_region *r;
7143
7144 /* Need to find movable_zone earlier when movable_node is specified. */
7145 find_usable_zone_for_movable();
7146
7147 /*
7148 * If movable_node is specified, ignore kernelcore and movablecore
7149 * options.
7150 */
7151 if (movable_node_is_enabled()) {
7152 for_each_memblock(memory, r) {
7153 if (!memblock_is_hotpluggable(r))
7154 continue;
7155
7156 nid = memblock_get_region_node(r);
7157
7158 usable_startpfn = PFN_DOWN(r->base);
7159 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7160 min(usable_startpfn, zone_movable_pfn[nid]) :
7161 usable_startpfn;
7162 }
7163
7164 goto out2;
7165 }
7166
7167 /*
7168 * If kernelcore=mirror is specified, ignore movablecore option
7169 */
7170 if (mirrored_kernelcore) {
7171 bool mem_below_4gb_not_mirrored = false;
7172
7173 for_each_memblock(memory, r) {
7174 if (memblock_is_mirror(r))
7175 continue;
7176
7177 nid = memblock_get_region_node(r);
7178
7179 usable_startpfn = memblock_region_memory_base_pfn(r);
7180
7181 if (usable_startpfn < 0x100000) {
7182 mem_below_4gb_not_mirrored = true;
7183 continue;
7184 }
7185
7186 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
7187 min(usable_startpfn, zone_movable_pfn[nid]) :
7188 usable_startpfn;
7189 }
7190
7191 if (mem_below_4gb_not_mirrored)
7192 pr_warn("This configuration results in unmirrored kernel memory.\n");
7193
7194 goto out2;
7195 }
7196
7197 /*
7198 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
7199 * amount of necessary memory.
7200 */
7201 if (required_kernelcore_percent)
7202 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
7203 10000UL;
7204 if (required_movablecore_percent)
7205 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
7206 10000UL;
7207
7208 /*
7209 * If movablecore= was specified, calculate what size of
7210 * kernelcore that corresponds so that memory usable for
7211 * any allocation type is evenly spread. If both kernelcore
7212 * and movablecore are specified, then the value of kernelcore
7213 * will be used for required_kernelcore if it's greater than
7214 * what movablecore would have allowed.
7215 */
7216 if (required_movablecore) {
7217 unsigned long corepages;
7218
7219 /*
7220 * Round-up so that ZONE_MOVABLE is at least as large as what
7221 * was requested by the user
7222 */
7223 required_movablecore =
7224 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
7225 required_movablecore = min(totalpages, required_movablecore);
7226 corepages = totalpages - required_movablecore;
7227
7228 required_kernelcore = max(required_kernelcore, corepages);
7229 }
7230
7231 /*
7232 * If kernelcore was not specified or kernelcore size is larger
7233 * than totalpages, there is no ZONE_MOVABLE.
7234 */
7235 if (!required_kernelcore || required_kernelcore >= totalpages)
7236 goto out;
7237
7238 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
7239 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
7240
7241restart:
7242 /* Spread kernelcore memory as evenly as possible throughout nodes */
7243 kernelcore_node = required_kernelcore / usable_nodes;
7244 for_each_node_state(nid, N_MEMORY) {
7245 unsigned long start_pfn, end_pfn;
7246
7247 /*
7248 * Recalculate kernelcore_node if the division per node
7249 * now exceeds what is necessary to satisfy the requested
7250 * amount of memory for the kernel
7251 */
7252 if (required_kernelcore < kernelcore_node)
7253 kernelcore_node = required_kernelcore / usable_nodes;
7254
7255 /*
7256 * As the map is walked, we track how much memory is usable
7257 * by the kernel using kernelcore_remaining. When it is
7258 * 0, the rest of the node is usable by ZONE_MOVABLE
7259 */
7260 kernelcore_remaining = kernelcore_node;
7261
7262 /* Go through each range of PFNs within this node */
7263 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
7264 unsigned long size_pages;
7265
7266 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
7267 if (start_pfn >= end_pfn)
7268 continue;
7269
7270 /* Account for what is only usable for kernelcore */
7271 if (start_pfn < usable_startpfn) {
7272 unsigned long kernel_pages;
7273 kernel_pages = min(end_pfn, usable_startpfn)
7274 - start_pfn;
7275
7276 kernelcore_remaining -= min(kernel_pages,
7277 kernelcore_remaining);
7278 required_kernelcore -= min(kernel_pages,
7279 required_kernelcore);
7280
7281 /* Continue if range is now fully accounted */
7282 if (end_pfn <= usable_startpfn) {
7283
7284 /*
7285 * Push zone_movable_pfn to the end so
7286 * that if we have to rebalance
7287 * kernelcore across nodes, we will
7288 * not double account here
7289 */
7290 zone_movable_pfn[nid] = end_pfn;
7291 continue;
7292 }
7293 start_pfn = usable_startpfn;
7294 }
7295
7296 /*
7297 * The usable PFN range for ZONE_MOVABLE is from
7298 * start_pfn->end_pfn. Calculate size_pages as the
7299 * number of pages used as kernelcore
7300 */
7301 size_pages = end_pfn - start_pfn;
7302 if (size_pages > kernelcore_remaining)
7303 size_pages = kernelcore_remaining;
7304 zone_movable_pfn[nid] = start_pfn + size_pages;
7305
7306 /*
7307 * Some kernelcore has been met, update counts and
7308 * break if the kernelcore for this node has been
7309 * satisfied
7310 */
7311 required_kernelcore -= min(required_kernelcore,
7312 size_pages);
7313 kernelcore_remaining -= size_pages;
7314 if (!kernelcore_remaining)
7315 break;
7316 }
7317 }
7318
7319 /*
7320 * If there is still required_kernelcore, we do another pass with one
7321 * less node in the count. This will push zone_movable_pfn[nid] further
7322 * along on the nodes that still have memory until kernelcore is
7323 * satisfied
7324 */
7325 usable_nodes--;
7326 if (usable_nodes && required_kernelcore > usable_nodes)
7327 goto restart;
7328
7329out2:
7330 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
7331 for (nid = 0; nid < MAX_NUMNODES; nid++)
7332 zone_movable_pfn[nid] =
7333 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
7334
7335out:
7336 /* restore the node_state */
7337 node_states[N_MEMORY] = saved_node_state;
7338}
7339
7340/* Any regular or high memory on that node ? */
7341static void check_for_memory(pg_data_t *pgdat, int nid)
7342{
7343 enum zone_type zone_type;
7344
7345 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
7346 struct zone *zone = &pgdat->node_zones[zone_type];
7347 if (populated_zone(zone)) {
7348 if (IS_ENABLED(CONFIG_HIGHMEM))
7349 node_set_state(nid, N_HIGH_MEMORY);
7350 if (zone_type <= ZONE_NORMAL)
7351 node_set_state(nid, N_NORMAL_MEMORY);
7352 break;
7353 }
7354 }
7355}
7356
7357/*
7358 * Some architecturs, e.g. ARC may have ZONE_HIGHMEM below ZONE_NORMAL. For
7359 * such cases we allow max_zone_pfn sorted in the descending order
7360 */
7361bool __weak arch_has_descending_max_zone_pfns(void)
7362{
7363 return false;
7364}
7365
7366/**
7367 * free_area_init - Initialise all pg_data_t and zone data
7368 * @max_zone_pfn: an array of max PFNs for each zone
7369 *
7370 * This will call free_area_init_node() for each active node in the system.
7371 * Using the page ranges provided by memblock_set_node(), the size of each
7372 * zone in each node and their holes is calculated. If the maximum PFN
7373 * between two adjacent zones match, it is assumed that the zone is empty.
7374 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
7375 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
7376 * starts where the previous one ended. For example, ZONE_DMA32 starts
7377 * at arch_max_dma_pfn.
7378 */
7379void __init free_area_init(unsigned long *max_zone_pfn)
7380{
7381 unsigned long start_pfn, end_pfn;
7382 int i, nid, zone;
7383 bool descending;
7384
7385 /* Record where the zone boundaries are */
7386 memset(arch_zone_lowest_possible_pfn, 0,
7387 sizeof(arch_zone_lowest_possible_pfn));
7388 memset(arch_zone_highest_possible_pfn, 0,
7389 sizeof(arch_zone_highest_possible_pfn));
7390
7391 start_pfn = find_min_pfn_with_active_regions();
7392 descending = arch_has_descending_max_zone_pfns();
7393
7394 for (i = 0; i < MAX_NR_ZONES; i++) {
7395 if (descending)
7396 zone = MAX_NR_ZONES - i - 1;
7397 else
7398 zone = i;
7399
7400 if (zone == ZONE_MOVABLE)
7401 continue;
7402
7403 end_pfn = max(max_zone_pfn[zone], start_pfn);
7404 arch_zone_lowest_possible_pfn[zone] = start_pfn;
7405 arch_zone_highest_possible_pfn[zone] = end_pfn;
7406
7407 start_pfn = end_pfn;
7408 }
7409
7410 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
7411 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
7412 find_zone_movable_pfns_for_nodes();
7413
7414 /* Print out the zone ranges */
7415 pr_info("Zone ranges:\n");
7416 for (i = 0; i < MAX_NR_ZONES; i++) {
7417 if (i == ZONE_MOVABLE)
7418 continue;
7419 pr_info(" %-8s ", zone_names[i]);
7420 if (arch_zone_lowest_possible_pfn[i] ==
7421 arch_zone_highest_possible_pfn[i])
7422 pr_cont("empty\n");
7423 else
7424 pr_cont("[mem %#018Lx-%#018Lx]\n",
7425 (u64)arch_zone_lowest_possible_pfn[i]
7426 << PAGE_SHIFT,
7427 ((u64)arch_zone_highest_possible_pfn[i]
7428 << PAGE_SHIFT) - 1);
7429 }
7430
7431 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
7432 pr_info("Movable zone start for each node\n");
7433 for (i = 0; i < MAX_NUMNODES; i++) {
7434 if (zone_movable_pfn[i])
7435 pr_info(" Node %d: %#018Lx\n", i,
7436 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
7437 }
7438
7439 /*
7440 * Print out the early node map, and initialize the
7441 * subsection-map relative to active online memory ranges to
7442 * enable future "sub-section" extensions of the memory map.
7443 */
7444 pr_info("Early memory node ranges\n");
7445 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
7446 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
7447 (u64)start_pfn << PAGE_SHIFT,
7448 ((u64)end_pfn << PAGE_SHIFT) - 1);
7449 subsection_map_init(start_pfn, end_pfn - start_pfn);
7450 }
7451
7452 /* Initialise every node */
7453 mminit_verify_pageflags_layout();
7454 setup_nr_node_ids();
7455 init_unavailable_mem();
7456 for_each_online_node(nid) {
7457 pg_data_t *pgdat = NODE_DATA(nid);
7458 free_area_init_node(nid);
7459
7460 /* Any memory on that node */
7461 if (pgdat->node_present_pages)
7462 node_set_state(nid, N_MEMORY);
7463 check_for_memory(pgdat, nid);
7464 }
7465}
7466
7467static int __init cmdline_parse_core(char *p, unsigned long *core,
7468 unsigned long *percent)
7469{
7470 unsigned long long coremem;
7471 char *endptr;
7472
7473 if (!p)
7474 return -EINVAL;
7475
7476 /* Value may be a percentage of total memory, otherwise bytes */
7477 coremem = simple_strtoull(p, &endptr, 0);
7478 if (*endptr == '%') {
7479 /* Paranoid check for percent values greater than 100 */
7480 WARN_ON(coremem > 100);
7481
7482 *percent = coremem;
7483 } else {
7484 coremem = memparse(p, &p);
7485 /* Paranoid check that UL is enough for the coremem value */
7486 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
7487
7488 *core = coremem >> PAGE_SHIFT;
7489 *percent = 0UL;
7490 }
7491 return 0;
7492}
7493
7494/*
7495 * kernelcore=size sets the amount of memory for use for allocations that
7496 * cannot be reclaimed or migrated.
7497 */
7498static int __init cmdline_parse_kernelcore(char *p)
7499{
7500 /* parse kernelcore=mirror */
7501 if (parse_option_str(p, "mirror")) {
7502 mirrored_kernelcore = true;
7503 return 0;
7504 }
7505
7506 return cmdline_parse_core(p, &required_kernelcore,
7507 &required_kernelcore_percent);
7508}
7509
7510/*
7511 * movablecore=size sets the amount of memory for use for allocations that
7512 * can be reclaimed or migrated.
7513 */
7514static int __init cmdline_parse_movablecore(char *p)
7515{
7516 return cmdline_parse_core(p, &required_movablecore,
7517 &required_movablecore_percent);
7518}
7519
7520early_param("kernelcore", cmdline_parse_kernelcore);
7521early_param("movablecore", cmdline_parse_movablecore);
7522
7523void adjust_managed_page_count(struct page *page, long count)
7524{
7525 atomic_long_add(count, &page_zone(page)->managed_pages);
7526 totalram_pages_add(count);
7527#ifdef CONFIG_HIGHMEM
7528 if (PageHighMem(page))
7529 totalhigh_pages_add(count);
7530#endif
7531}
7532EXPORT_SYMBOL(adjust_managed_page_count);
7533
7534unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
7535{
7536 void *pos;
7537 unsigned long pages = 0;
7538
7539 start = (void *)PAGE_ALIGN((unsigned long)start);
7540 end = (void *)((unsigned long)end & PAGE_MASK);
7541 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
7542 struct page *page = virt_to_page(pos);
7543 void *direct_map_addr;
7544
7545 /*
7546 * 'direct_map_addr' might be different from 'pos'
7547 * because some architectures' virt_to_page()
7548 * work with aliases. Getting the direct map
7549 * address ensures that we get a _writeable_
7550 * alias for the memset().
7551 */
7552 direct_map_addr = page_address(page);
7553 if ((unsigned int)poison <= 0xFF)
7554 memset(direct_map_addr, poison, PAGE_SIZE);
7555
7556 free_reserved_page(page);
7557 }
7558
7559 if (pages && s)
7560 pr_info("Freeing %s memory: %ldK\n",
7561 s, pages << (PAGE_SHIFT - 10));
7562
7563 return pages;
7564}
7565
7566#ifdef CONFIG_HIGHMEM
7567void free_highmem_page(struct page *page)
7568{
7569 __free_reserved_page(page);
7570 totalram_pages_inc();
7571 atomic_long_inc(&page_zone(page)->managed_pages);
7572 totalhigh_pages_inc();
7573}
7574#endif
7575
7576
7577void __init mem_init_print_info(const char *str)
7578{
7579 unsigned long physpages, codesize, datasize, rosize, bss_size;
7580 unsigned long init_code_size, init_data_size;
7581
7582 physpages = get_num_physpages();
7583 codesize = _etext - _stext;
7584 datasize = _edata - _sdata;
7585 rosize = __end_rodata - __start_rodata;
7586 bss_size = __bss_stop - __bss_start;
7587 init_data_size = __init_end - __init_begin;
7588 init_code_size = _einittext - _sinittext;
7589
7590 /*
7591 * Detect special cases and adjust section sizes accordingly:
7592 * 1) .init.* may be embedded into .data sections
7593 * 2) .init.text.* may be out of [__init_begin, __init_end],
7594 * please refer to arch/tile/kernel/vmlinux.lds.S.
7595 * 3) .rodata.* may be embedded into .text or .data sections.
7596 */
7597#define adj_init_size(start, end, size, pos, adj) \
7598 do { \
7599 if (start <= pos && pos < end && size > adj) \
7600 size -= adj; \
7601 } while (0)
7602
7603 adj_init_size(__init_begin, __init_end, init_data_size,
7604 _sinittext, init_code_size);
7605 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
7606 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
7607 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
7608 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
7609
7610#undef adj_init_size
7611
7612 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
7613#ifdef CONFIG_HIGHMEM
7614 ", %luK highmem"
7615#endif
7616 "%s%s)\n",
7617 nr_free_pages() << (PAGE_SHIFT - 10),
7618 physpages << (PAGE_SHIFT - 10),
7619 codesize >> 10, datasize >> 10, rosize >> 10,
7620 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7621 (physpages - totalram_pages() - totalcma_pages) << (PAGE_SHIFT - 10),
7622 totalcma_pages << (PAGE_SHIFT - 10),
7623#ifdef CONFIG_HIGHMEM
7624 totalhigh_pages() << (PAGE_SHIFT - 10),
7625#endif
7626 str ? ", " : "", str ? str : "");
7627}
7628
7629/**
7630 * set_dma_reserve - set the specified number of pages reserved in the first zone
7631 * @new_dma_reserve: The number of pages to mark reserved
7632 *
7633 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7634 * In the DMA zone, a significant percentage may be consumed by kernel image
7635 * and other unfreeable allocations which can skew the watermarks badly. This
7636 * function may optionally be used to account for unfreeable pages in the
7637 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7638 * smaller per-cpu batchsize.
7639 */
7640void __init set_dma_reserve(unsigned long new_dma_reserve)
7641{
7642 dma_reserve = new_dma_reserve;
7643}
7644
7645static int page_alloc_cpu_dead(unsigned int cpu)
7646{
7647
7648 lru_add_drain_cpu(cpu);
7649 drain_pages(cpu);
7650
7651 /*
7652 * Spill the event counters of the dead processor
7653 * into the current processors event counters.
7654 * This artificially elevates the count of the current
7655 * processor.
7656 */
7657 vm_events_fold_cpu(cpu);
7658
7659 /*
7660 * Zero the differential counters of the dead processor
7661 * so that the vm statistics are consistent.
7662 *
7663 * This is only okay since the processor is dead and cannot
7664 * race with what we are doing.
7665 */
7666 cpu_vm_stats_fold(cpu);
7667 return 0;
7668}
7669
7670#ifdef CONFIG_NUMA
7671int hashdist = HASHDIST_DEFAULT;
7672
7673static int __init set_hashdist(char *str)
7674{
7675 if (!str)
7676 return 0;
7677 hashdist = simple_strtoul(str, &str, 0);
7678 return 1;
7679}
7680__setup("hashdist=", set_hashdist);
7681#endif
7682
7683void __init page_alloc_init(void)
7684{
7685 int ret;
7686
7687#ifdef CONFIG_NUMA
7688 if (num_node_state(N_MEMORY) == 1)
7689 hashdist = 0;
7690#endif
7691
7692 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7693 "mm/page_alloc:dead", NULL,
7694 page_alloc_cpu_dead);
7695 WARN_ON(ret < 0);
7696}
7697
7698/*
7699 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7700 * or min_free_kbytes changes.
7701 */
7702static void calculate_totalreserve_pages(void)
7703{
7704 struct pglist_data *pgdat;
7705 unsigned long reserve_pages = 0;
7706 enum zone_type i, j;
7707
7708 for_each_online_pgdat(pgdat) {
7709
7710 pgdat->totalreserve_pages = 0;
7711
7712 for (i = 0; i < MAX_NR_ZONES; i++) {
7713 struct zone *zone = pgdat->node_zones + i;
7714 long max = 0;
7715 unsigned long managed_pages = zone_managed_pages(zone);
7716
7717 /* Find valid and maximum lowmem_reserve in the zone */
7718 for (j = i; j < MAX_NR_ZONES; j++) {
7719 if (zone->lowmem_reserve[j] > max)
7720 max = zone->lowmem_reserve[j];
7721 }
7722
7723 /* we treat the high watermark as reserved pages. */
7724 max += high_wmark_pages(zone);
7725
7726 if (max > managed_pages)
7727 max = managed_pages;
7728
7729 pgdat->totalreserve_pages += max;
7730
7731 reserve_pages += max;
7732 }
7733 }
7734 totalreserve_pages = reserve_pages;
7735}
7736
7737/*
7738 * setup_per_zone_lowmem_reserve - called whenever
7739 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7740 * has a correct pages reserved value, so an adequate number of
7741 * pages are left in the zone after a successful __alloc_pages().
7742 */
7743static void setup_per_zone_lowmem_reserve(void)
7744{
7745 struct pglist_data *pgdat;
7746 enum zone_type j, idx;
7747
7748 for_each_online_pgdat(pgdat) {
7749 for (j = 0; j < MAX_NR_ZONES; j++) {
7750 struct zone *zone = pgdat->node_zones + j;
7751 unsigned long managed_pages = zone_managed_pages(zone);
7752
7753 zone->lowmem_reserve[j] = 0;
7754
7755 idx = j;
7756 while (idx) {
7757 struct zone *lower_zone;
7758
7759 idx--;
7760 lower_zone = pgdat->node_zones + idx;
7761
7762 if (!sysctl_lowmem_reserve_ratio[idx] ||
7763 !zone_managed_pages(lower_zone)) {
7764 lower_zone->lowmem_reserve[j] = 0;
7765 continue;
7766 } else {
7767 lower_zone->lowmem_reserve[j] =
7768 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7769 }
7770 managed_pages += zone_managed_pages(lower_zone);
7771 }
7772 }
7773 }
7774
7775 /* update totalreserve_pages */
7776 calculate_totalreserve_pages();
7777}
7778
7779static void __setup_per_zone_wmarks(void)
7780{
7781 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7782 unsigned long lowmem_pages = 0;
7783 struct zone *zone;
7784 unsigned long flags;
7785
7786 /* Calculate total number of !ZONE_HIGHMEM pages */
7787 for_each_zone(zone) {
7788 if (!is_highmem(zone))
7789 lowmem_pages += zone_managed_pages(zone);
7790 }
7791
7792 for_each_zone(zone) {
7793 u64 tmp;
7794
7795 spin_lock_irqsave(&zone->lock, flags);
7796 tmp = (u64)pages_min * zone_managed_pages(zone);
7797 do_div(tmp, lowmem_pages);
7798 if (is_highmem(zone)) {
7799 /*
7800 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7801 * need highmem pages, so cap pages_min to a small
7802 * value here.
7803 *
7804 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7805 * deltas control async page reclaim, and so should
7806 * not be capped for highmem.
7807 */
7808 unsigned long min_pages;
7809
7810 min_pages = zone_managed_pages(zone) / 1024;
7811 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7812 zone->_watermark[WMARK_MIN] = min_pages;
7813 } else {
7814 /*
7815 * If it's a lowmem zone, reserve a number of pages
7816 * proportionate to the zone's size.
7817 */
7818 zone->_watermark[WMARK_MIN] = tmp;
7819 }
7820
7821 /*
7822 * Set the kswapd watermarks distance according to the
7823 * scale factor in proportion to available memory, but
7824 * ensure a minimum size on small systems.
7825 */
7826 tmp = max_t(u64, tmp >> 2,
7827 mult_frac(zone_managed_pages(zone),
7828 watermark_scale_factor, 10000));
7829
7830 zone->watermark_boost = 0;
7831 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7832 zone->_watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7833
7834 spin_unlock_irqrestore(&zone->lock, flags);
7835 }
7836
7837 /* update totalreserve_pages */
7838 calculate_totalreserve_pages();
7839}
7840
7841/**
7842 * setup_per_zone_wmarks - called when min_free_kbytes changes
7843 * or when memory is hot-{added|removed}
7844 *
7845 * Ensures that the watermark[min,low,high] values for each zone are set
7846 * correctly with respect to min_free_kbytes.
7847 */
7848void setup_per_zone_wmarks(void)
7849{
7850 static DEFINE_SPINLOCK(lock);
7851
7852 spin_lock(&lock);
7853 __setup_per_zone_wmarks();
7854 spin_unlock(&lock);
7855}
7856
7857/*
7858 * Initialise min_free_kbytes.
7859 *
7860 * For small machines we want it small (128k min). For large machines
7861 * we want it large (256MB max). But it is not linear, because network
7862 * bandwidth does not increase linearly with machine size. We use
7863 *
7864 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7865 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7866 *
7867 * which yields
7868 *
7869 * 16MB: 512k
7870 * 32MB: 724k
7871 * 64MB: 1024k
7872 * 128MB: 1448k
7873 * 256MB: 2048k
7874 * 512MB: 2896k
7875 * 1024MB: 4096k
7876 * 2048MB: 5792k
7877 * 4096MB: 8192k
7878 * 8192MB: 11584k
7879 * 16384MB: 16384k
7880 */
7881int __meminit init_per_zone_wmark_min(void)
7882{
7883 unsigned long lowmem_kbytes;
7884 int new_min_free_kbytes;
7885
7886 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7887 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7888
7889 if (new_min_free_kbytes > user_min_free_kbytes) {
7890 min_free_kbytes = new_min_free_kbytes;
7891 if (min_free_kbytes < 128)
7892 min_free_kbytes = 128;
7893 if (min_free_kbytes > 262144)
7894 min_free_kbytes = 262144;
7895 } else {
7896 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7897 new_min_free_kbytes, user_min_free_kbytes);
7898 }
7899 setup_per_zone_wmarks();
7900 refresh_zone_stat_thresholds();
7901 setup_per_zone_lowmem_reserve();
7902
7903#ifdef CONFIG_NUMA
7904 setup_min_unmapped_ratio();
7905 setup_min_slab_ratio();
7906#endif
7907
7908 khugepaged_min_free_kbytes_update();
7909
7910 return 0;
7911}
7912postcore_initcall(init_per_zone_wmark_min)
7913
7914/*
7915 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7916 * that we can call two helper functions whenever min_free_kbytes
7917 * changes.
7918 */
7919int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7920 void *buffer, size_t *length, loff_t *ppos)
7921{
7922 int rc;
7923
7924 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7925 if (rc)
7926 return rc;
7927
7928 if (write) {
7929 user_min_free_kbytes = min_free_kbytes;
7930 setup_per_zone_wmarks();
7931 }
7932 return 0;
7933}
7934
7935int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7936 void *buffer, size_t *length, loff_t *ppos)
7937{
7938 int rc;
7939
7940 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7941 if (rc)
7942 return rc;
7943
7944 if (write)
7945 setup_per_zone_wmarks();
7946
7947 return 0;
7948}
7949
7950#ifdef CONFIG_NUMA
7951static void setup_min_unmapped_ratio(void)
7952{
7953 pg_data_t *pgdat;
7954 struct zone *zone;
7955
7956 for_each_online_pgdat(pgdat)
7957 pgdat->min_unmapped_pages = 0;
7958
7959 for_each_zone(zone)
7960 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
7961 sysctl_min_unmapped_ratio) / 100;
7962}
7963
7964
7965int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7966 void *buffer, size_t *length, loff_t *ppos)
7967{
7968 int rc;
7969
7970 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7971 if (rc)
7972 return rc;
7973
7974 setup_min_unmapped_ratio();
7975
7976 return 0;
7977}
7978
7979static void setup_min_slab_ratio(void)
7980{
7981 pg_data_t *pgdat;
7982 struct zone *zone;
7983
7984 for_each_online_pgdat(pgdat)
7985 pgdat->min_slab_pages = 0;
7986
7987 for_each_zone(zone)
7988 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
7989 sysctl_min_slab_ratio) / 100;
7990}
7991
7992int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7993 void *buffer, size_t *length, loff_t *ppos)
7994{
7995 int rc;
7996
7997 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7998 if (rc)
7999 return rc;
8000
8001 setup_min_slab_ratio();
8002
8003 return 0;
8004}
8005#endif
8006
8007/*
8008 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
8009 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
8010 * whenever sysctl_lowmem_reserve_ratio changes.
8011 *
8012 * The reserve ratio obviously has absolutely no relation with the
8013 * minimum watermarks. The lowmem reserve ratio can only make sense
8014 * if in function of the boot time zone sizes.
8015 */
8016int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
8017 void *buffer, size_t *length, loff_t *ppos)
8018{
8019 int i;
8020
8021 proc_dointvec_minmax(table, write, buffer, length, ppos);
8022
8023 for (i = 0; i < MAX_NR_ZONES; i++) {
8024 if (sysctl_lowmem_reserve_ratio[i] < 1)
8025 sysctl_lowmem_reserve_ratio[i] = 0;
8026 }
8027
8028 setup_per_zone_lowmem_reserve();
8029 return 0;
8030}
8031
8032static void __zone_pcp_update(struct zone *zone)
8033{
8034 unsigned int cpu;
8035
8036 for_each_possible_cpu(cpu)
8037 pageset_set_high_and_batch(zone,
8038 per_cpu_ptr(zone->pageset, cpu));
8039}
8040
8041/*
8042 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
8043 * cpu. It is the fraction of total pages in each zone that a hot per cpu
8044 * pagelist can have before it gets flushed back to buddy allocator.
8045 */
8046int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
8047 void *buffer, size_t *length, loff_t *ppos)
8048{
8049 struct zone *zone;
8050 int old_percpu_pagelist_fraction;
8051 int ret;
8052
8053 mutex_lock(&pcp_batch_high_lock);
8054 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
8055
8056 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
8057 if (!write || ret < 0)
8058 goto out;
8059
8060 /* Sanity checking to avoid pcp imbalance */
8061 if (percpu_pagelist_fraction &&
8062 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
8063 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
8064 ret = -EINVAL;
8065 goto out;
8066 }
8067
8068 /* No change? */
8069 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
8070 goto out;
8071
8072 for_each_populated_zone(zone)
8073 __zone_pcp_update(zone);
8074out:
8075 mutex_unlock(&pcp_batch_high_lock);
8076 return ret;
8077}
8078
8079#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
8080/*
8081 * Returns the number of pages that arch has reserved but
8082 * is not known to alloc_large_system_hash().
8083 */
8084static unsigned long __init arch_reserved_kernel_pages(void)
8085{
8086 return 0;
8087}
8088#endif
8089
8090/*
8091 * Adaptive scale is meant to reduce sizes of hash tables on large memory
8092 * machines. As memory size is increased the scale is also increased but at
8093 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
8094 * quadruples the scale is increased by one, which means the size of hash table
8095 * only doubles, instead of quadrupling as well.
8096 * Because 32-bit systems cannot have large physical memory, where this scaling
8097 * makes sense, it is disabled on such platforms.
8098 */
8099#if __BITS_PER_LONG > 32
8100#define ADAPT_SCALE_BASE (64ul << 30)
8101#define ADAPT_SCALE_SHIFT 2
8102#define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
8103#endif
8104
8105/*
8106 * allocate a large system hash table from bootmem
8107 * - it is assumed that the hash table must contain an exact power-of-2
8108 * quantity of entries
8109 * - limit is the number of hash buckets, not the total allocation size
8110 */
8111void *__init alloc_large_system_hash(const char *tablename,
8112 unsigned long bucketsize,
8113 unsigned long numentries,
8114 int scale,
8115 int flags,
8116 unsigned int *_hash_shift,
8117 unsigned int *_hash_mask,
8118 unsigned long low_limit,
8119 unsigned long high_limit)
8120{
8121 unsigned long long max = high_limit;
8122 unsigned long log2qty, size;
8123 void *table = NULL;
8124 gfp_t gfp_flags;
8125 bool virt;
8126
8127 /* allow the kernel cmdline to have a say */
8128 if (!numentries) {
8129 /* round applicable memory size up to nearest megabyte */
8130 numentries = nr_kernel_pages;
8131 numentries -= arch_reserved_kernel_pages();
8132
8133 /* It isn't necessary when PAGE_SIZE >= 1MB */
8134 if (PAGE_SHIFT < 20)
8135 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
8136
8137#if __BITS_PER_LONG > 32
8138 if (!high_limit) {
8139 unsigned long adapt;
8140
8141 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
8142 adapt <<= ADAPT_SCALE_SHIFT)
8143 scale++;
8144 }
8145#endif
8146
8147 /* limit to 1 bucket per 2^scale bytes of low memory */
8148 if (scale > PAGE_SHIFT)
8149 numentries >>= (scale - PAGE_SHIFT);
8150 else
8151 numentries <<= (PAGE_SHIFT - scale);
8152
8153 /* Make sure we've got at least a 0-order allocation.. */
8154 if (unlikely(flags & HASH_SMALL)) {
8155 /* Makes no sense without HASH_EARLY */
8156 WARN_ON(!(flags & HASH_EARLY));
8157 if (!(numentries >> *_hash_shift)) {
8158 numentries = 1UL << *_hash_shift;
8159 BUG_ON(!numentries);
8160 }
8161 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
8162 numentries = PAGE_SIZE / bucketsize;
8163 }
8164 numentries = roundup_pow_of_two(numentries);
8165
8166 /* limit allocation size to 1/16 total memory by default */
8167 if (max == 0) {
8168 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
8169 do_div(max, bucketsize);
8170 }
8171 max = min(max, 0x80000000ULL);
8172
8173 if (numentries < low_limit)
8174 numentries = low_limit;
8175 if (numentries > max)
8176 numentries = max;
8177
8178 log2qty = ilog2(numentries);
8179
8180 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
8181 do {
8182 virt = false;
8183 size = bucketsize << log2qty;
8184 if (flags & HASH_EARLY) {
8185 if (flags & HASH_ZERO)
8186 table = memblock_alloc(size, SMP_CACHE_BYTES);
8187 else
8188 table = memblock_alloc_raw(size,
8189 SMP_CACHE_BYTES);
8190 } else if (get_order(size) >= MAX_ORDER || hashdist) {
8191 table = __vmalloc(size, gfp_flags);
8192 virt = true;
8193 } else {
8194 /*
8195 * If bucketsize is not a power-of-two, we may free
8196 * some pages at the end of hash table which
8197 * alloc_pages_exact() automatically does
8198 */
8199 table = alloc_pages_exact(size, gfp_flags);
8200 kmemleak_alloc(table, size, 1, gfp_flags);
8201 }
8202 } while (!table && size > PAGE_SIZE && --log2qty);
8203
8204 if (!table)
8205 panic("Failed to allocate %s hash table\n", tablename);
8206
8207 pr_info("%s hash table entries: %ld (order: %d, %lu bytes, %s)\n",
8208 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size,
8209 virt ? "vmalloc" : "linear");
8210
8211 if (_hash_shift)
8212 *_hash_shift = log2qty;
8213 if (_hash_mask)
8214 *_hash_mask = (1 << log2qty) - 1;
8215
8216 return table;
8217}
8218
8219/*
8220 * This function checks whether pageblock includes unmovable pages or not.
8221 *
8222 * PageLRU check without isolation or lru_lock could race so that
8223 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
8224 * check without lock_page also may miss some movable non-lru pages at
8225 * race condition. So you can't expect this function should be exact.
8226 *
8227 * Returns a page without holding a reference. If the caller wants to
8228 * dereference that page (e.g., dumping), it has to make sure that it
8229 * cannot get removed (e.g., via memory unplug) concurrently.
8230 *
8231 */
8232struct page *has_unmovable_pages(struct zone *zone, struct page *page,
8233 int migratetype, int flags)
8234{
8235 unsigned long iter = 0;
8236 unsigned long pfn = page_to_pfn(page);
8237
8238 /*
8239 * TODO we could make this much more efficient by not checking every
8240 * page in the range if we know all of them are in MOVABLE_ZONE and
8241 * that the movable zone guarantees that pages are migratable but
8242 * the later is not the case right now unfortunatelly. E.g. movablecore
8243 * can still lead to having bootmem allocations in zone_movable.
8244 */
8245
8246 if (is_migrate_cma_page(page)) {
8247 /*
8248 * CMA allocations (alloc_contig_range) really need to mark
8249 * isolate CMA pageblocks even when they are not movable in fact
8250 * so consider them movable here.
8251 */
8252 if (is_migrate_cma(migratetype))
8253 return NULL;
8254
8255 return page;
8256 }
8257
8258 for (; iter < pageblock_nr_pages; iter++) {
8259 if (!pfn_valid_within(pfn + iter))
8260 continue;
8261
8262 page = pfn_to_page(pfn + iter);
8263
8264 if (PageReserved(page))
8265 return page;
8266
8267 /*
8268 * If the zone is movable and we have ruled out all reserved
8269 * pages then it should be reasonably safe to assume the rest
8270 * is movable.
8271 */
8272 if (zone_idx(zone) == ZONE_MOVABLE)
8273 continue;
8274
8275 /*
8276 * Hugepages are not in LRU lists, but they're movable.
8277 * THPs are on the LRU, but need to be counted as #small pages.
8278 * We need not scan over tail pages because we don't
8279 * handle each tail page individually in migration.
8280 */
8281 if (PageHuge(page) || PageTransCompound(page)) {
8282 struct page *head = compound_head(page);
8283 unsigned int skip_pages;
8284
8285 if (PageHuge(page)) {
8286 if (!hugepage_migration_supported(page_hstate(head)))
8287 return page;
8288 } else if (!PageLRU(head) && !__PageMovable(head)) {
8289 return page;
8290 }
8291
8292 skip_pages = compound_nr(head) - (page - head);
8293 iter += skip_pages - 1;
8294 continue;
8295 }
8296
8297 /*
8298 * We can't use page_count without pin a page
8299 * because another CPU can free compound page.
8300 * This check already skips compound tails of THP
8301 * because their page->_refcount is zero at all time.
8302 */
8303 if (!page_ref_count(page)) {
8304 if (PageBuddy(page))
8305 iter += (1 << page_order(page)) - 1;
8306 continue;
8307 }
8308
8309 /*
8310 * The HWPoisoned page may be not in buddy system, and
8311 * page_count() is not 0.
8312 */
8313 if ((flags & MEMORY_OFFLINE) && PageHWPoison(page))
8314 continue;
8315
8316 /*
8317 * We treat all PageOffline() pages as movable when offlining
8318 * to give drivers a chance to decrement their reference count
8319 * in MEM_GOING_OFFLINE in order to indicate that these pages
8320 * can be offlined as there are no direct references anymore.
8321 * For actually unmovable PageOffline() where the driver does
8322 * not support this, we will fail later when trying to actually
8323 * move these pages that still have a reference count > 0.
8324 * (false negatives in this function only)
8325 */
8326 if ((flags & MEMORY_OFFLINE) && PageOffline(page))
8327 continue;
8328
8329 if (__PageMovable(page) || PageLRU(page))
8330 continue;
8331
8332 /*
8333 * If there are RECLAIMABLE pages, we need to check
8334 * it. But now, memory offline itself doesn't call
8335 * shrink_node_slabs() and it still to be fixed.
8336 */
8337 /*
8338 * If the page is not RAM, page_count()should be 0.
8339 * we don't need more check. This is an _used_ not-movable page.
8340 *
8341 * The problematic thing here is PG_reserved pages. PG_reserved
8342 * is set to both of a memory hole page and a _used_ kernel
8343 * page at boot.
8344 */
8345 return page;
8346 }
8347 return NULL;
8348}
8349
8350#ifdef CONFIG_CONTIG_ALLOC
8351static unsigned long pfn_max_align_down(unsigned long pfn)
8352{
8353 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
8354 pageblock_nr_pages) - 1);
8355}
8356
8357static unsigned long pfn_max_align_up(unsigned long pfn)
8358{
8359 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
8360 pageblock_nr_pages));
8361}
8362
8363/* [start, end) must belong to a single zone. */
8364static int __alloc_contig_migrate_range(struct compact_control *cc,
8365 unsigned long start, unsigned long end)
8366{
8367 /* This function is based on compact_zone() from compaction.c. */
8368 unsigned int nr_reclaimed;
8369 unsigned long pfn = start;
8370 unsigned int tries = 0;
8371 int ret = 0;
8372 struct migration_target_control mtc = {
8373 .nid = zone_to_nid(cc->zone),
8374 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
8375 };
8376
8377 migrate_prep();
8378
8379 while (pfn < end || !list_empty(&cc->migratepages)) {
8380 if (fatal_signal_pending(current)) {
8381 ret = -EINTR;
8382 break;
8383 }
8384
8385 if (list_empty(&cc->migratepages)) {
8386 cc->nr_migratepages = 0;
8387 pfn = isolate_migratepages_range(cc, pfn, end);
8388 if (!pfn) {
8389 ret = -EINTR;
8390 break;
8391 }
8392 tries = 0;
8393 } else if (++tries == 5) {
8394 ret = ret < 0 ? ret : -EBUSY;
8395 break;
8396 }
8397
8398 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
8399 &cc->migratepages);
8400 cc->nr_migratepages -= nr_reclaimed;
8401
8402 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
8403 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE);
8404 }
8405 if (ret < 0) {
8406 putback_movable_pages(&cc->migratepages);
8407 return ret;
8408 }
8409 return 0;
8410}
8411
8412/**
8413 * alloc_contig_range() -- tries to allocate given range of pages
8414 * @start: start PFN to allocate
8415 * @end: one-past-the-last PFN to allocate
8416 * @migratetype: migratetype of the underlaying pageblocks (either
8417 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
8418 * in range must have the same migratetype and it must
8419 * be either of the two.
8420 * @gfp_mask: GFP mask to use during compaction
8421 *
8422 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
8423 * aligned. The PFN range must belong to a single zone.
8424 *
8425 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
8426 * pageblocks in the range. Once isolated, the pageblocks should not
8427 * be modified by others.
8428 *
8429 * Return: zero on success or negative error code. On success all
8430 * pages which PFN is in [start, end) are allocated for the caller and
8431 * need to be freed with free_contig_range().
8432 */
8433int alloc_contig_range(unsigned long start, unsigned long end,
8434 unsigned migratetype, gfp_t gfp_mask)
8435{
8436 unsigned long outer_start, outer_end;
8437 unsigned int order;
8438 int ret = 0;
8439
8440 struct compact_control cc = {
8441 .nr_migratepages = 0,
8442 .order = -1,
8443 .zone = page_zone(pfn_to_page(start)),
8444 .mode = MIGRATE_SYNC,
8445 .ignore_skip_hint = true,
8446 .no_set_skip_hint = true,
8447 .gfp_mask = current_gfp_context(gfp_mask),
8448 .alloc_contig = true,
8449 };
8450 INIT_LIST_HEAD(&cc.migratepages);
8451
8452 /*
8453 * What we do here is we mark all pageblocks in range as
8454 * MIGRATE_ISOLATE. Because pageblock and max order pages may
8455 * have different sizes, and due to the way page allocator
8456 * work, we align the range to biggest of the two pages so
8457 * that page allocator won't try to merge buddies from
8458 * different pageblocks and change MIGRATE_ISOLATE to some
8459 * other migration type.
8460 *
8461 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
8462 * migrate the pages from an unaligned range (ie. pages that
8463 * we are interested in). This will put all the pages in
8464 * range back to page allocator as MIGRATE_ISOLATE.
8465 *
8466 * When this is done, we take the pages in range from page
8467 * allocator removing them from the buddy system. This way
8468 * page allocator will never consider using them.
8469 *
8470 * This lets us mark the pageblocks back as
8471 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
8472 * aligned range but not in the unaligned, original range are
8473 * put back to page allocator so that buddy can use them.
8474 */
8475
8476 ret = start_isolate_page_range(pfn_max_align_down(start),
8477 pfn_max_align_up(end), migratetype, 0);
8478 if (ret < 0)
8479 return ret;
8480
8481 /*
8482 * In case of -EBUSY, we'd like to know which page causes problem.
8483 * So, just fall through. test_pages_isolated() has a tracepoint
8484 * which will report the busy page.
8485 *
8486 * It is possible that busy pages could become available before
8487 * the call to test_pages_isolated, and the range will actually be
8488 * allocated. So, if we fall through be sure to clear ret so that
8489 * -EBUSY is not accidentally used or returned to caller.
8490 */
8491 ret = __alloc_contig_migrate_range(&cc, start, end);
8492 if (ret && ret != -EBUSY)
8493 goto done;
8494 ret =0;
8495
8496 /*
8497 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
8498 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
8499 * more, all pages in [start, end) are free in page allocator.
8500 * What we are going to do is to allocate all pages from
8501 * [start, end) (that is remove them from page allocator).
8502 *
8503 * The only problem is that pages at the beginning and at the
8504 * end of interesting range may be not aligned with pages that
8505 * page allocator holds, ie. they can be part of higher order
8506 * pages. Because of this, we reserve the bigger range and
8507 * once this is done free the pages we are not interested in.
8508 *
8509 * We don't have to hold zone->lock here because the pages are
8510 * isolated thus they won't get removed from buddy.
8511 */
8512
8513 lru_add_drain_all();
8514
8515 order = 0;
8516 outer_start = start;
8517 while (!PageBuddy(pfn_to_page(outer_start))) {
8518 if (++order >= MAX_ORDER) {
8519 outer_start = start;
8520 break;
8521 }
8522 outer_start &= ~0UL << order;
8523 }
8524
8525 if (outer_start != start) {
8526 order = page_order(pfn_to_page(outer_start));
8527
8528 /*
8529 * outer_start page could be small order buddy page and
8530 * it doesn't include start page. Adjust outer_start
8531 * in this case to report failed page properly
8532 * on tracepoint in test_pages_isolated()
8533 */
8534 if (outer_start + (1UL << order) <= start)
8535 outer_start = start;
8536 }
8537
8538 /* Make sure the range is really isolated. */
8539 if (test_pages_isolated(outer_start, end, 0)) {
8540 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
8541 __func__, outer_start, end);
8542 ret = -EBUSY;
8543 goto done;
8544 }
8545
8546 /* Grab isolated pages from freelists. */
8547 outer_end = isolate_freepages_range(&cc, outer_start, end);
8548 if (!outer_end) {
8549 ret = -EBUSY;
8550 goto done;
8551 }
8552
8553 /* Free head and tail (if any) */
8554 if (start != outer_start)
8555 free_contig_range(outer_start, start - outer_start);
8556 if (end != outer_end)
8557 free_contig_range(end, outer_end - end);
8558
8559done:
8560 undo_isolate_page_range(pfn_max_align_down(start),
8561 pfn_max_align_up(end), migratetype);
8562 return ret;
8563}
8564EXPORT_SYMBOL(alloc_contig_range);
8565
8566static int __alloc_contig_pages(unsigned long start_pfn,
8567 unsigned long nr_pages, gfp_t gfp_mask)
8568{
8569 unsigned long end_pfn = start_pfn + nr_pages;
8570
8571 return alloc_contig_range(start_pfn, end_pfn, MIGRATE_MOVABLE,
8572 gfp_mask);
8573}
8574
8575static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
8576 unsigned long nr_pages)
8577{
8578 unsigned long i, end_pfn = start_pfn + nr_pages;
8579 struct page *page;
8580
8581 for (i = start_pfn; i < end_pfn; i++) {
8582 page = pfn_to_online_page(i);
8583 if (!page)
8584 return false;
8585
8586 if (page_zone(page) != z)
8587 return false;
8588
8589 if (PageReserved(page))
8590 return false;
8591
8592 if (page_count(page) > 0)
8593 return false;
8594
8595 if (PageHuge(page))
8596 return false;
8597 }
8598 return true;
8599}
8600
8601static bool zone_spans_last_pfn(const struct zone *zone,
8602 unsigned long start_pfn, unsigned long nr_pages)
8603{
8604 unsigned long last_pfn = start_pfn + nr_pages - 1;
8605
8606 return zone_spans_pfn(zone, last_pfn);
8607}
8608
8609/**
8610 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
8611 * @nr_pages: Number of contiguous pages to allocate
8612 * @gfp_mask: GFP mask to limit search and used during compaction
8613 * @nid: Target node
8614 * @nodemask: Mask for other possible nodes
8615 *
8616 * This routine is a wrapper around alloc_contig_range(). It scans over zones
8617 * on an applicable zonelist to find a contiguous pfn range which can then be
8618 * tried for allocation with alloc_contig_range(). This routine is intended
8619 * for allocation requests which can not be fulfilled with the buddy allocator.
8620 *
8621 * The allocated memory is always aligned to a page boundary. If nr_pages is a
8622 * power of two then the alignment is guaranteed to be to the given nr_pages
8623 * (e.g. 1GB request would be aligned to 1GB).
8624 *
8625 * Allocated pages can be freed with free_contig_range() or by manually calling
8626 * __free_page() on each allocated page.
8627 *
8628 * Return: pointer to contiguous pages on success, or NULL if not successful.
8629 */
8630struct page *alloc_contig_pages(unsigned long nr_pages, gfp_t gfp_mask,
8631 int nid, nodemask_t *nodemask)
8632{
8633 unsigned long ret, pfn, flags;
8634 struct zonelist *zonelist;
8635 struct zone *zone;
8636 struct zoneref *z;
8637
8638 zonelist = node_zonelist(nid, gfp_mask);
8639 for_each_zone_zonelist_nodemask(zone, z, zonelist,
8640 gfp_zone(gfp_mask), nodemask) {
8641 spin_lock_irqsave(&zone->lock, flags);
8642
8643 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
8644 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
8645 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
8646 /*
8647 * We release the zone lock here because
8648 * alloc_contig_range() will also lock the zone
8649 * at some point. If there's an allocation
8650 * spinning on this lock, it may win the race
8651 * and cause alloc_contig_range() to fail...
8652 */
8653 spin_unlock_irqrestore(&zone->lock, flags);
8654 ret = __alloc_contig_pages(pfn, nr_pages,
8655 gfp_mask);
8656 if (!ret)
8657 return pfn_to_page(pfn);
8658 spin_lock_irqsave(&zone->lock, flags);
8659 }
8660 pfn += nr_pages;
8661 }
8662 spin_unlock_irqrestore(&zone->lock, flags);
8663 }
8664 return NULL;
8665}
8666#endif /* CONFIG_CONTIG_ALLOC */
8667
8668void free_contig_range(unsigned long pfn, unsigned int nr_pages)
8669{
8670 unsigned int count = 0;
8671
8672 for (; nr_pages--; pfn++) {
8673 struct page *page = pfn_to_page(pfn);
8674
8675 count += page_count(page) != 1;
8676 __free_page(page);
8677 }
8678 WARN(count != 0, "%d pages are still in use!\n", count);
8679}
8680EXPORT_SYMBOL(free_contig_range);
8681
8682/*
8683 * The zone indicated has a new number of managed_pages; batch sizes and percpu
8684 * page high values need to be recalulated.
8685 */
8686void __meminit zone_pcp_update(struct zone *zone)
8687{
8688 mutex_lock(&pcp_batch_high_lock);
8689 __zone_pcp_update(zone);
8690 mutex_unlock(&pcp_batch_high_lock);
8691}
8692
8693void zone_pcp_reset(struct zone *zone)
8694{
8695 unsigned long flags;
8696 int cpu;
8697 struct per_cpu_pageset *pset;
8698
8699 /* avoid races with drain_pages() */
8700 local_irq_save(flags);
8701 if (zone->pageset != &boot_pageset) {
8702 for_each_online_cpu(cpu) {
8703 pset = per_cpu_ptr(zone->pageset, cpu);
8704 drain_zonestat(zone, pset);
8705 }
8706 free_percpu(zone->pageset);
8707 zone->pageset = &boot_pageset;
8708 }
8709 local_irq_restore(flags);
8710}
8711
8712#ifdef CONFIG_MEMORY_HOTREMOVE
8713/*
8714 * All pages in the range must be in a single zone and isolated
8715 * before calling this.
8716 */
8717unsigned long
8718__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
8719{
8720 struct page *page;
8721 struct zone *zone;
8722 unsigned int order;
8723 unsigned long pfn;
8724 unsigned long flags;
8725 unsigned long offlined_pages = 0;
8726
8727 /* find the first valid pfn */
8728 for (pfn = start_pfn; pfn < end_pfn; pfn++)
8729 if (pfn_valid(pfn))
8730 break;
8731 if (pfn == end_pfn)
8732 return offlined_pages;
8733
8734 offline_mem_sections(pfn, end_pfn);
8735 zone = page_zone(pfn_to_page(pfn));
8736 spin_lock_irqsave(&zone->lock, flags);
8737 pfn = start_pfn;
8738 while (pfn < end_pfn) {
8739 if (!pfn_valid(pfn)) {
8740 pfn++;
8741 continue;
8742 }
8743 page = pfn_to_page(pfn);
8744 /*
8745 * The HWPoisoned page may be not in buddy system, and
8746 * page_count() is not 0.
8747 */
8748 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8749 pfn++;
8750 offlined_pages++;
8751 continue;
8752 }
8753 /*
8754 * At this point all remaining PageOffline() pages have a
8755 * reference count of 0 and can simply be skipped.
8756 */
8757 if (PageOffline(page)) {
8758 BUG_ON(page_count(page));
8759 BUG_ON(PageBuddy(page));
8760 pfn++;
8761 offlined_pages++;
8762 continue;
8763 }
8764
8765 BUG_ON(page_count(page));
8766 BUG_ON(!PageBuddy(page));
8767 order = page_order(page);
8768 offlined_pages += 1 << order;
8769 del_page_from_free_list(page, zone, order);
8770 pfn += (1 << order);
8771 }
8772 spin_unlock_irqrestore(&zone->lock, flags);
8773
8774 return offlined_pages;
8775}
8776#endif
8777
8778bool is_free_buddy_page(struct page *page)
8779{
8780 struct zone *zone = page_zone(page);
8781 unsigned long pfn = page_to_pfn(page);
8782 unsigned long flags;
8783 unsigned int order;
8784
8785 spin_lock_irqsave(&zone->lock, flags);
8786 for (order = 0; order < MAX_ORDER; order++) {
8787 struct page *page_head = page - (pfn & ((1 << order) - 1));
8788
8789 if (PageBuddy(page_head) && page_order(page_head) >= order)
8790 break;
8791 }
8792 spin_unlock_irqrestore(&zone->lock, flags);
8793
8794 return order < MAX_ORDER;
8795}
8796
8797#ifdef CONFIG_MEMORY_FAILURE
8798/*
8799 * Set PG_hwpoison flag if a given page is confirmed to be a free page. This
8800 * test is performed under the zone lock to prevent a race against page
8801 * allocation.
8802 */
8803bool set_hwpoison_free_buddy_page(struct page *page)
8804{
8805 struct zone *zone = page_zone(page);
8806 unsigned long pfn = page_to_pfn(page);
8807 unsigned long flags;
8808 unsigned int order;
8809 bool hwpoisoned = false;
8810
8811 spin_lock_irqsave(&zone->lock, flags);
8812 for (order = 0; order < MAX_ORDER; order++) {
8813 struct page *page_head = page - (pfn & ((1 << order) - 1));
8814
8815 if (PageBuddy(page_head) && page_order(page_head) >= order) {
8816 if (!TestSetPageHWPoison(page))
8817 hwpoisoned = true;
8818 break;
8819 }
8820 }
8821 spin_unlock_irqrestore(&zone->lock, flags);
8822
8823 return hwpoisoned;
8824}
8825#endif