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/*
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/kasan.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/memremap.h>
46#include <linux/stop_machine.h>
47#include <linux/sort.h>
48#include <linux/pfn.h>
49#include <linux/backing-dev.h>
50#include <linux/fault-inject.h>
51#include <linux/page-isolation.h>
52#include <linux/page_ext.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
71#include <asm/sections.h>
72#include <asm/tlbflush.h>
73#include <asm/div64.h>
74#include "internal.h"
75
76/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
77static DEFINE_MUTEX(pcp_batch_high_lock);
78#define MIN_PERCPU_PAGELIST_FRACTION (8)
79
80#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
81DEFINE_PER_CPU(int, numa_node);
82EXPORT_PER_CPU_SYMBOL(numa_node);
83#endif
84
85DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
86
87#ifdef CONFIG_HAVE_MEMORYLESS_NODES
88/*
89 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
90 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
91 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
92 * defined in <linux/topology.h>.
93 */
94DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
95EXPORT_PER_CPU_SYMBOL(_numa_mem_);
96int _node_numa_mem_[MAX_NUMNODES];
97#endif
98
99/* work_structs for global per-cpu drains */
100DEFINE_MUTEX(pcpu_drain_mutex);
101DEFINE_PER_CPU(struct work_struct, pcpu_drain);
102
103#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
104volatile unsigned long latent_entropy __latent_entropy;
105EXPORT_SYMBOL(latent_entropy);
106#endif
107
108/*
109 * Array of node states.
110 */
111nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
112 [N_POSSIBLE] = NODE_MASK_ALL,
113 [N_ONLINE] = { { [0] = 1UL } },
114#ifndef CONFIG_NUMA
115 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
116#ifdef CONFIG_HIGHMEM
117 [N_HIGH_MEMORY] = { { [0] = 1UL } },
118#endif
119 [N_MEMORY] = { { [0] = 1UL } },
120 [N_CPU] = { { [0] = 1UL } },
121#endif /* NUMA */
122};
123EXPORT_SYMBOL(node_states);
124
125/* Protect totalram_pages and zone->managed_pages */
126static DEFINE_SPINLOCK(managed_page_count_lock);
127
128unsigned long totalram_pages __read_mostly;
129unsigned long totalreserve_pages __read_mostly;
130unsigned long totalcma_pages __read_mostly;
131
132int percpu_pagelist_fraction;
133gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
134
135/*
136 * A cached value of the page's pageblock's migratetype, used when the page is
137 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
138 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
139 * Also the migratetype set in the page does not necessarily match the pcplist
140 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
141 * other index - this ensures that it will be put on the correct CMA freelist.
142 */
143static inline int get_pcppage_migratetype(struct page *page)
144{
145 return page->index;
146}
147
148static inline void set_pcppage_migratetype(struct page *page, int migratetype)
149{
150 page->index = migratetype;
151}
152
153#ifdef CONFIG_PM_SLEEP
154/*
155 * The following functions are used by the suspend/hibernate code to temporarily
156 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
157 * while devices are suspended. To avoid races with the suspend/hibernate code,
158 * they should always be called with pm_mutex held (gfp_allowed_mask also should
159 * only be modified with pm_mutex held, unless the suspend/hibernate code is
160 * guaranteed not to run in parallel with that modification).
161 */
162
163static gfp_t saved_gfp_mask;
164
165void pm_restore_gfp_mask(void)
166{
167 WARN_ON(!mutex_is_locked(&pm_mutex));
168 if (saved_gfp_mask) {
169 gfp_allowed_mask = saved_gfp_mask;
170 saved_gfp_mask = 0;
171 }
172}
173
174void pm_restrict_gfp_mask(void)
175{
176 WARN_ON(!mutex_is_locked(&pm_mutex));
177 WARN_ON(saved_gfp_mask);
178 saved_gfp_mask = gfp_allowed_mask;
179 gfp_allowed_mask &= ~(__GFP_IO | __GFP_FS);
180}
181
182bool pm_suspended_storage(void)
183{
184 if ((gfp_allowed_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
185 return false;
186 return true;
187}
188#endif /* CONFIG_PM_SLEEP */
189
190#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
191unsigned int pageblock_order __read_mostly;
192#endif
193
194static void __free_pages_ok(struct page *page, unsigned int order);
195
196/*
197 * results with 256, 32 in the lowmem_reserve sysctl:
198 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
199 * 1G machine -> (16M dma, 784M normal, 224M high)
200 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
201 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
202 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
203 *
204 * TBD: should special case ZONE_DMA32 machines here - in those we normally
205 * don't need any ZONE_NORMAL reservation
206 */
207int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
208#ifdef CONFIG_ZONE_DMA
209 [ZONE_DMA] = 256,
210#endif
211#ifdef CONFIG_ZONE_DMA32
212 [ZONE_DMA32] = 256,
213#endif
214 [ZONE_NORMAL] = 32,
215#ifdef CONFIG_HIGHMEM
216 [ZONE_HIGHMEM] = 0,
217#endif
218 [ZONE_MOVABLE] = 0,
219};
220
221EXPORT_SYMBOL(totalram_pages);
222
223static char * const zone_names[MAX_NR_ZONES] = {
224#ifdef CONFIG_ZONE_DMA
225 "DMA",
226#endif
227#ifdef CONFIG_ZONE_DMA32
228 "DMA32",
229#endif
230 "Normal",
231#ifdef CONFIG_HIGHMEM
232 "HighMem",
233#endif
234 "Movable",
235#ifdef CONFIG_ZONE_DEVICE
236 "Device",
237#endif
238};
239
240char * const migratetype_names[MIGRATE_TYPES] = {
241 "Unmovable",
242 "Movable",
243 "Reclaimable",
244 "HighAtomic",
245#ifdef CONFIG_CMA
246 "CMA",
247#endif
248#ifdef CONFIG_MEMORY_ISOLATION
249 "Isolate",
250#endif
251};
252
253compound_page_dtor * const compound_page_dtors[] = {
254 NULL,
255 free_compound_page,
256#ifdef CONFIG_HUGETLB_PAGE
257 free_huge_page,
258#endif
259#ifdef CONFIG_TRANSPARENT_HUGEPAGE
260 free_transhuge_page,
261#endif
262};
263
264int min_free_kbytes = 1024;
265int user_min_free_kbytes = -1;
266int watermark_scale_factor = 10;
267
268static unsigned long nr_kernel_pages __meminitdata;
269static unsigned long nr_all_pages __meminitdata;
270static unsigned long dma_reserve __meminitdata;
271
272#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
273static unsigned long arch_zone_lowest_possible_pfn[MAX_NR_ZONES] __meminitdata;
274static unsigned long arch_zone_highest_possible_pfn[MAX_NR_ZONES] __meminitdata;
275static unsigned long required_kernelcore __initdata;
276static unsigned long required_kernelcore_percent __initdata;
277static unsigned long required_movablecore __initdata;
278static unsigned long required_movablecore_percent __initdata;
279static unsigned long zone_movable_pfn[MAX_NUMNODES] __meminitdata;
280static bool mirrored_kernelcore __meminitdata;
281
282/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
283int movable_zone;
284EXPORT_SYMBOL(movable_zone);
285#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
286
287#if MAX_NUMNODES > 1
288int nr_node_ids __read_mostly = MAX_NUMNODES;
289int nr_online_nodes __read_mostly = 1;
290EXPORT_SYMBOL(nr_node_ids);
291EXPORT_SYMBOL(nr_online_nodes);
292#endif
293
294int page_group_by_mobility_disabled __read_mostly;
295
296#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
297/* Returns true if the struct page for the pfn is uninitialised */
298static inline bool __meminit early_page_uninitialised(unsigned long pfn)
299{
300 int nid = early_pfn_to_nid(pfn);
301
302 if (node_online(nid) && pfn >= NODE_DATA(nid)->first_deferred_pfn)
303 return true;
304
305 return false;
306}
307
308/*
309 * Returns false when the remaining initialisation should be deferred until
310 * later in the boot cycle when it can be parallelised.
311 */
312static inline bool update_defer_init(pg_data_t *pgdat,
313 unsigned long pfn, unsigned long zone_end,
314 unsigned long *nr_initialised)
315{
316 /* Always populate low zones for address-constrained allocations */
317 if (zone_end < pgdat_end_pfn(pgdat))
318 return true;
319 (*nr_initialised)++;
320 if ((*nr_initialised > pgdat->static_init_pgcnt) &&
321 (pfn & (PAGES_PER_SECTION - 1)) == 0) {
322 pgdat->first_deferred_pfn = pfn;
323 return false;
324 }
325
326 return true;
327}
328#else
329static inline bool early_page_uninitialised(unsigned long pfn)
330{
331 return false;
332}
333
334static inline bool update_defer_init(pg_data_t *pgdat,
335 unsigned long pfn, unsigned long zone_end,
336 unsigned long *nr_initialised)
337{
338 return true;
339}
340#endif
341
342/* Return a pointer to the bitmap storing bits affecting a block of pages */
343static inline unsigned long *get_pageblock_bitmap(struct page *page,
344 unsigned long pfn)
345{
346#ifdef CONFIG_SPARSEMEM
347 return __pfn_to_section(pfn)->pageblock_flags;
348#else
349 return page_zone(page)->pageblock_flags;
350#endif /* CONFIG_SPARSEMEM */
351}
352
353static inline int pfn_to_bitidx(struct page *page, unsigned long pfn)
354{
355#ifdef CONFIG_SPARSEMEM
356 pfn &= (PAGES_PER_SECTION-1);
357 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
358#else
359 pfn = pfn - round_down(page_zone(page)->zone_start_pfn, pageblock_nr_pages);
360 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
361#endif /* CONFIG_SPARSEMEM */
362}
363
364/**
365 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
366 * @page: The page within the block of interest
367 * @pfn: The target page frame number
368 * @end_bitidx: The last bit of interest to retrieve
369 * @mask: mask of bits that the caller is interested in
370 *
371 * Return: pageblock_bits flags
372 */
373static __always_inline unsigned long __get_pfnblock_flags_mask(struct page *page,
374 unsigned long pfn,
375 unsigned long end_bitidx,
376 unsigned long mask)
377{
378 unsigned long *bitmap;
379 unsigned long bitidx, word_bitidx;
380 unsigned long word;
381
382 bitmap = get_pageblock_bitmap(page, pfn);
383 bitidx = pfn_to_bitidx(page, pfn);
384 word_bitidx = bitidx / BITS_PER_LONG;
385 bitidx &= (BITS_PER_LONG-1);
386
387 word = bitmap[word_bitidx];
388 bitidx += end_bitidx;
389 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
390}
391
392unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
393 unsigned long end_bitidx,
394 unsigned long mask)
395{
396 return __get_pfnblock_flags_mask(page, pfn, end_bitidx, mask);
397}
398
399static __always_inline int get_pfnblock_migratetype(struct page *page, unsigned long pfn)
400{
401 return __get_pfnblock_flags_mask(page, pfn, PB_migrate_end, MIGRATETYPE_MASK);
402}
403
404/**
405 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
406 * @page: The page within the block of interest
407 * @flags: The flags to set
408 * @pfn: The target page frame number
409 * @end_bitidx: The last bit of interest
410 * @mask: mask of bits that the caller is interested in
411 */
412void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
413 unsigned long pfn,
414 unsigned long end_bitidx,
415 unsigned long mask)
416{
417 unsigned long *bitmap;
418 unsigned long bitidx, word_bitidx;
419 unsigned long old_word, word;
420
421 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
422
423 bitmap = get_pageblock_bitmap(page, pfn);
424 bitidx = pfn_to_bitidx(page, pfn);
425 word_bitidx = bitidx / BITS_PER_LONG;
426 bitidx &= (BITS_PER_LONG-1);
427
428 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
429
430 bitidx += end_bitidx;
431 mask <<= (BITS_PER_LONG - bitidx - 1);
432 flags <<= (BITS_PER_LONG - bitidx - 1);
433
434 word = READ_ONCE(bitmap[word_bitidx]);
435 for (;;) {
436 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
437 if (word == old_word)
438 break;
439 word = old_word;
440 }
441}
442
443void set_pageblock_migratetype(struct page *page, int migratetype)
444{
445 if (unlikely(page_group_by_mobility_disabled &&
446 migratetype < MIGRATE_PCPTYPES))
447 migratetype = MIGRATE_UNMOVABLE;
448
449 set_pageblock_flags_group(page, (unsigned long)migratetype,
450 PB_migrate, PB_migrate_end);
451}
452
453#ifdef CONFIG_DEBUG_VM
454static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
455{
456 int ret = 0;
457 unsigned seq;
458 unsigned long pfn = page_to_pfn(page);
459 unsigned long sp, start_pfn;
460
461 do {
462 seq = zone_span_seqbegin(zone);
463 start_pfn = zone->zone_start_pfn;
464 sp = zone->spanned_pages;
465 if (!zone_spans_pfn(zone, pfn))
466 ret = 1;
467 } while (zone_span_seqretry(zone, seq));
468
469 if (ret)
470 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
471 pfn, zone_to_nid(zone), zone->name,
472 start_pfn, start_pfn + sp);
473
474 return ret;
475}
476
477static int page_is_consistent(struct zone *zone, struct page *page)
478{
479 if (!pfn_valid_within(page_to_pfn(page)))
480 return 0;
481 if (zone != page_zone(page))
482 return 0;
483
484 return 1;
485}
486/*
487 * Temporary debugging check for pages not lying within a given zone.
488 */
489static int __maybe_unused bad_range(struct zone *zone, struct page *page)
490{
491 if (page_outside_zone_boundaries(zone, page))
492 return 1;
493 if (!page_is_consistent(zone, page))
494 return 1;
495
496 return 0;
497}
498#else
499static inline int __maybe_unused bad_range(struct zone *zone, struct page *page)
500{
501 return 0;
502}
503#endif
504
505static void bad_page(struct page *page, const char *reason,
506 unsigned long bad_flags)
507{
508 static unsigned long resume;
509 static unsigned long nr_shown;
510 static unsigned long nr_unshown;
511
512 /*
513 * Allow a burst of 60 reports, then keep quiet for that minute;
514 * or allow a steady drip of one report per second.
515 */
516 if (nr_shown == 60) {
517 if (time_before(jiffies, resume)) {
518 nr_unshown++;
519 goto out;
520 }
521 if (nr_unshown) {
522 pr_alert(
523 "BUG: Bad page state: %lu messages suppressed\n",
524 nr_unshown);
525 nr_unshown = 0;
526 }
527 nr_shown = 0;
528 }
529 if (nr_shown++ == 0)
530 resume = jiffies + 60 * HZ;
531
532 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
533 current->comm, page_to_pfn(page));
534 __dump_page(page, reason);
535 bad_flags &= page->flags;
536 if (bad_flags)
537 pr_alert("bad because of flags: %#lx(%pGp)\n",
538 bad_flags, &bad_flags);
539 dump_page_owner(page);
540
541 print_modules();
542 dump_stack();
543out:
544 /* Leave bad fields for debug, except PageBuddy could make trouble */
545 page_mapcount_reset(page); /* remove PageBuddy */
546 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
547}
548
549/*
550 * Higher-order pages are called "compound pages". They are structured thusly:
551 *
552 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
553 *
554 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
555 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
556 *
557 * The first tail page's ->compound_dtor holds the offset in array of compound
558 * page destructors. See compound_page_dtors.
559 *
560 * The first tail page's ->compound_order holds the order of allocation.
561 * This usage means that zero-order pages may not be compound.
562 */
563
564void free_compound_page(struct page *page)
565{
566 __free_pages_ok(page, compound_order(page));
567}
568
569void prep_compound_page(struct page *page, unsigned int order)
570{
571 int i;
572 int nr_pages = 1 << order;
573
574 set_compound_page_dtor(page, COMPOUND_PAGE_DTOR);
575 set_compound_order(page, order);
576 __SetPageHead(page);
577 for (i = 1; i < nr_pages; i++) {
578 struct page *p = page + i;
579 set_page_count(p, 0);
580 p->mapping = TAIL_MAPPING;
581 set_compound_head(p, page);
582 }
583 atomic_set(compound_mapcount_ptr(page), -1);
584}
585
586#ifdef CONFIG_DEBUG_PAGEALLOC
587unsigned int _debug_guardpage_minorder;
588bool _debug_pagealloc_enabled __read_mostly
589 = IS_ENABLED(CONFIG_DEBUG_PAGEALLOC_ENABLE_DEFAULT);
590EXPORT_SYMBOL(_debug_pagealloc_enabled);
591bool _debug_guardpage_enabled __read_mostly;
592
593static int __init early_debug_pagealloc(char *buf)
594{
595 if (!buf)
596 return -EINVAL;
597 return kstrtobool(buf, &_debug_pagealloc_enabled);
598}
599early_param("debug_pagealloc", early_debug_pagealloc);
600
601static bool need_debug_guardpage(void)
602{
603 /* If we don't use debug_pagealloc, we don't need guard page */
604 if (!debug_pagealloc_enabled())
605 return false;
606
607 if (!debug_guardpage_minorder())
608 return false;
609
610 return true;
611}
612
613static void init_debug_guardpage(void)
614{
615 if (!debug_pagealloc_enabled())
616 return;
617
618 if (!debug_guardpage_minorder())
619 return;
620
621 _debug_guardpage_enabled = true;
622}
623
624struct page_ext_operations debug_guardpage_ops = {
625 .need = need_debug_guardpage,
626 .init = init_debug_guardpage,
627};
628
629static int __init debug_guardpage_minorder_setup(char *buf)
630{
631 unsigned long res;
632
633 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
634 pr_err("Bad debug_guardpage_minorder value\n");
635 return 0;
636 }
637 _debug_guardpage_minorder = res;
638 pr_info("Setting debug_guardpage_minorder to %lu\n", res);
639 return 0;
640}
641early_param("debug_guardpage_minorder", debug_guardpage_minorder_setup);
642
643static inline bool set_page_guard(struct zone *zone, struct page *page,
644 unsigned int order, int migratetype)
645{
646 struct page_ext *page_ext;
647
648 if (!debug_guardpage_enabled())
649 return false;
650
651 if (order >= debug_guardpage_minorder())
652 return false;
653
654 page_ext = lookup_page_ext(page);
655 if (unlikely(!page_ext))
656 return false;
657
658 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
659
660 INIT_LIST_HEAD(&page->lru);
661 set_page_private(page, order);
662 /* Guard pages are not available for any usage */
663 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
664
665 return true;
666}
667
668static inline void clear_page_guard(struct zone *zone, struct page *page,
669 unsigned int order, int migratetype)
670{
671 struct page_ext *page_ext;
672
673 if (!debug_guardpage_enabled())
674 return;
675
676 page_ext = lookup_page_ext(page);
677 if (unlikely(!page_ext))
678 return;
679
680 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
681
682 set_page_private(page, 0);
683 if (!is_migrate_isolate(migratetype))
684 __mod_zone_freepage_state(zone, (1 << order), migratetype);
685}
686#else
687struct page_ext_operations debug_guardpage_ops;
688static inline bool set_page_guard(struct zone *zone, struct page *page,
689 unsigned int order, int migratetype) { return false; }
690static inline void clear_page_guard(struct zone *zone, struct page *page,
691 unsigned int order, int migratetype) {}
692#endif
693
694static inline void set_page_order(struct page *page, unsigned int order)
695{
696 set_page_private(page, order);
697 __SetPageBuddy(page);
698}
699
700static inline void rmv_page_order(struct page *page)
701{
702 __ClearPageBuddy(page);
703 set_page_private(page, 0);
704}
705
706/*
707 * This function checks whether a page is free && is the buddy
708 * we can do coalesce a page and its buddy if
709 * (a) the buddy is not in a hole (check before calling!) &&
710 * (b) the buddy is in the buddy system &&
711 * (c) a page and its buddy have the same order &&
712 * (d) a page and its buddy are in the same zone.
713 *
714 * For recording whether a page is in the buddy system, we set ->_mapcount
715 * PAGE_BUDDY_MAPCOUNT_VALUE.
716 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
717 * serialized by zone->lock.
718 *
719 * For recording page's order, we use page_private(page).
720 */
721static inline int page_is_buddy(struct page *page, struct page *buddy,
722 unsigned int order)
723{
724 if (page_is_guard(buddy) && page_order(buddy) == order) {
725 if (page_zone_id(page) != page_zone_id(buddy))
726 return 0;
727
728 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
729
730 return 1;
731 }
732
733 if (PageBuddy(buddy) && page_order(buddy) == order) {
734 /*
735 * zone check is done late to avoid uselessly
736 * calculating zone/node ids for pages that could
737 * never merge.
738 */
739 if (page_zone_id(page) != page_zone_id(buddy))
740 return 0;
741
742 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
743
744 return 1;
745 }
746 return 0;
747}
748
749/*
750 * Freeing function for a buddy system allocator.
751 *
752 * The concept of a buddy system is to maintain direct-mapped table
753 * (containing bit values) for memory blocks of various "orders".
754 * The bottom level table contains the map for the smallest allocatable
755 * units of memory (here, pages), and each level above it describes
756 * pairs of units from the levels below, hence, "buddies".
757 * At a high level, all that happens here is marking the table entry
758 * at the bottom level available, and propagating the changes upward
759 * as necessary, plus some accounting needed to play nicely with other
760 * parts of the VM system.
761 * At each level, we keep a list of pages, which are heads of continuous
762 * free pages of length of (1 << order) and marked with _mapcount
763 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
764 * field.
765 * So when we are allocating or freeing one, we can derive the state of the
766 * other. That is, if we allocate a small block, and both were
767 * free, the remainder of the region must be split into blocks.
768 * If a block is freed, and its buddy is also free, then this
769 * triggers coalescing into a block of larger size.
770 *
771 * -- nyc
772 */
773
774static inline void __free_one_page(struct page *page,
775 unsigned long pfn,
776 struct zone *zone, unsigned int order,
777 int migratetype)
778{
779 unsigned long combined_pfn;
780 unsigned long uninitialized_var(buddy_pfn);
781 struct page *buddy;
782 unsigned int max_order;
783
784 max_order = min_t(unsigned int, MAX_ORDER, pageblock_order + 1);
785
786 VM_BUG_ON(!zone_is_initialized(zone));
787 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
788
789 VM_BUG_ON(migratetype == -1);
790 if (likely(!is_migrate_isolate(migratetype)))
791 __mod_zone_freepage_state(zone, 1 << order, migratetype);
792
793 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
794 VM_BUG_ON_PAGE(bad_range(zone, page), page);
795
796continue_merging:
797 while (order < max_order - 1) {
798 buddy_pfn = __find_buddy_pfn(pfn, order);
799 buddy = page + (buddy_pfn - pfn);
800
801 if (!pfn_valid_within(buddy_pfn))
802 goto done_merging;
803 if (!page_is_buddy(page, buddy, order))
804 goto done_merging;
805 /*
806 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
807 * merge with it and move up one order.
808 */
809 if (page_is_guard(buddy)) {
810 clear_page_guard(zone, buddy, order, migratetype);
811 } else {
812 list_del(&buddy->lru);
813 zone->free_area[order].nr_free--;
814 rmv_page_order(buddy);
815 }
816 combined_pfn = buddy_pfn & pfn;
817 page = page + (combined_pfn - pfn);
818 pfn = combined_pfn;
819 order++;
820 }
821 if (max_order < MAX_ORDER) {
822 /* If we are here, it means order is >= pageblock_order.
823 * We want to prevent merge between freepages on isolate
824 * pageblock and normal pageblock. Without this, pageblock
825 * isolation could cause incorrect freepage or CMA accounting.
826 *
827 * We don't want to hit this code for the more frequent
828 * low-order merging.
829 */
830 if (unlikely(has_isolate_pageblock(zone))) {
831 int buddy_mt;
832
833 buddy_pfn = __find_buddy_pfn(pfn, order);
834 buddy = page + (buddy_pfn - pfn);
835 buddy_mt = get_pageblock_migratetype(buddy);
836
837 if (migratetype != buddy_mt
838 && (is_migrate_isolate(migratetype) ||
839 is_migrate_isolate(buddy_mt)))
840 goto done_merging;
841 }
842 max_order++;
843 goto continue_merging;
844 }
845
846done_merging:
847 set_page_order(page, order);
848
849 /*
850 * If this is not the largest possible page, check if the buddy
851 * of the next-highest order is free. If it is, it's possible
852 * that pages are being freed that will coalesce soon. In case,
853 * that is happening, add the free page to the tail of the list
854 * so it's less likely to be used soon and more likely to be merged
855 * as a higher order page
856 */
857 if ((order < MAX_ORDER-2) && pfn_valid_within(buddy_pfn)) {
858 struct page *higher_page, *higher_buddy;
859 combined_pfn = buddy_pfn & pfn;
860 higher_page = page + (combined_pfn - pfn);
861 buddy_pfn = __find_buddy_pfn(combined_pfn, order + 1);
862 higher_buddy = higher_page + (buddy_pfn - combined_pfn);
863 if (pfn_valid_within(buddy_pfn) &&
864 page_is_buddy(higher_page, higher_buddy, order + 1)) {
865 list_add_tail(&page->lru,
866 &zone->free_area[order].free_list[migratetype]);
867 goto out;
868 }
869 }
870
871 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
872out:
873 zone->free_area[order].nr_free++;
874}
875
876/*
877 * A bad page could be due to a number of fields. Instead of multiple branches,
878 * try and check multiple fields with one check. The caller must do a detailed
879 * check if necessary.
880 */
881static inline bool page_expected_state(struct page *page,
882 unsigned long check_flags)
883{
884 if (unlikely(atomic_read(&page->_mapcount) != -1))
885 return false;
886
887 if (unlikely((unsigned long)page->mapping |
888 page_ref_count(page) |
889#ifdef CONFIG_MEMCG
890 (unsigned long)page->mem_cgroup |
891#endif
892 (page->flags & check_flags)))
893 return false;
894
895 return true;
896}
897
898static void free_pages_check_bad(struct page *page)
899{
900 const char *bad_reason;
901 unsigned long bad_flags;
902
903 bad_reason = NULL;
904 bad_flags = 0;
905
906 if (unlikely(atomic_read(&page->_mapcount) != -1))
907 bad_reason = "nonzero mapcount";
908 if (unlikely(page->mapping != NULL))
909 bad_reason = "non-NULL mapping";
910 if (unlikely(page_ref_count(page) != 0))
911 bad_reason = "nonzero _refcount";
912 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
913 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
914 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
915 }
916#ifdef CONFIG_MEMCG
917 if (unlikely(page->mem_cgroup))
918 bad_reason = "page still charged to cgroup";
919#endif
920 bad_page(page, bad_reason, bad_flags);
921}
922
923static inline int free_pages_check(struct page *page)
924{
925 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
926 return 0;
927
928 /* Something has gone sideways, find it */
929 free_pages_check_bad(page);
930 return 1;
931}
932
933static int free_tail_pages_check(struct page *head_page, struct page *page)
934{
935 int ret = 1;
936
937 /*
938 * We rely page->lru.next never has bit 0 set, unless the page
939 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
940 */
941 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
942
943 if (!IS_ENABLED(CONFIG_DEBUG_VM)) {
944 ret = 0;
945 goto out;
946 }
947 switch (page - head_page) {
948 case 1:
949 /* the first tail page: ->mapping is compound_mapcount() */
950 if (unlikely(compound_mapcount(page))) {
951 bad_page(page, "nonzero compound_mapcount", 0);
952 goto out;
953 }
954 break;
955 case 2:
956 /*
957 * the second tail page: ->mapping is
958 * page_deferred_list().next -- ignore value.
959 */
960 break;
961 default:
962 if (page->mapping != TAIL_MAPPING) {
963 bad_page(page, "corrupted mapping in tail page", 0);
964 goto out;
965 }
966 break;
967 }
968 if (unlikely(!PageTail(page))) {
969 bad_page(page, "PageTail not set", 0);
970 goto out;
971 }
972 if (unlikely(compound_head(page) != head_page)) {
973 bad_page(page, "compound_head not consistent", 0);
974 goto out;
975 }
976 ret = 0;
977out:
978 page->mapping = NULL;
979 clear_compound_head(page);
980 return ret;
981}
982
983static __always_inline bool free_pages_prepare(struct page *page,
984 unsigned int order, bool check_free)
985{
986 int bad = 0;
987
988 VM_BUG_ON_PAGE(PageTail(page), page);
989
990 trace_mm_page_free(page, order);
991
992 /*
993 * Check tail pages before head page information is cleared to
994 * avoid checking PageCompound for order-0 pages.
995 */
996 if (unlikely(order)) {
997 bool compound = PageCompound(page);
998 int i;
999
1000 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1001
1002 if (compound)
1003 ClearPageDoubleMap(page);
1004 for (i = 1; i < (1 << order); i++) {
1005 if (compound)
1006 bad += free_tail_pages_check(page, page + i);
1007 if (unlikely(free_pages_check(page + i))) {
1008 bad++;
1009 continue;
1010 }
1011 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1012 }
1013 }
1014 if (PageMappingFlags(page))
1015 page->mapping = NULL;
1016 if (memcg_kmem_enabled() && PageKmemcg(page))
1017 memcg_kmem_uncharge(page, order);
1018 if (check_free)
1019 bad += free_pages_check(page);
1020 if (bad)
1021 return false;
1022
1023 page_cpupid_reset_last(page);
1024 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1025 reset_page_owner(page, order);
1026
1027 if (!PageHighMem(page)) {
1028 debug_check_no_locks_freed(page_address(page),
1029 PAGE_SIZE << order);
1030 debug_check_no_obj_freed(page_address(page),
1031 PAGE_SIZE << order);
1032 }
1033 arch_free_page(page, order);
1034 kernel_poison_pages(page, 1 << order, 0);
1035 kernel_map_pages(page, 1 << order, 0);
1036 kasan_free_pages(page, order);
1037
1038 return true;
1039}
1040
1041#ifdef CONFIG_DEBUG_VM
1042static inline bool free_pcp_prepare(struct page *page)
1043{
1044 return free_pages_prepare(page, 0, true);
1045}
1046
1047static inline bool bulkfree_pcp_prepare(struct page *page)
1048{
1049 return false;
1050}
1051#else
1052static bool free_pcp_prepare(struct page *page)
1053{
1054 return free_pages_prepare(page, 0, false);
1055}
1056
1057static bool bulkfree_pcp_prepare(struct page *page)
1058{
1059 return free_pages_check(page);
1060}
1061#endif /* CONFIG_DEBUG_VM */
1062
1063static inline void prefetch_buddy(struct page *page)
1064{
1065 unsigned long pfn = page_to_pfn(page);
1066 unsigned long buddy_pfn = __find_buddy_pfn(pfn, 0);
1067 struct page *buddy = page + (buddy_pfn - pfn);
1068
1069 prefetch(buddy);
1070}
1071
1072/*
1073 * Frees a number of pages from the PCP lists
1074 * Assumes all pages on list are in same zone, and of same order.
1075 * count is the number of pages to free.
1076 *
1077 * If the zone was previously in an "all pages pinned" state then look to
1078 * see if this freeing clears that state.
1079 *
1080 * And clear the zone's pages_scanned counter, to hold off the "all pages are
1081 * pinned" detection logic.
1082 */
1083static void free_pcppages_bulk(struct zone *zone, int count,
1084 struct per_cpu_pages *pcp)
1085{
1086 int migratetype = 0;
1087 int batch_free = 0;
1088 int prefetch_nr = 0;
1089 bool isolated_pageblocks;
1090 struct page *page, *tmp;
1091 LIST_HEAD(head);
1092
1093 while (count) {
1094 struct list_head *list;
1095
1096 /*
1097 * Remove pages from lists in a round-robin fashion. A
1098 * batch_free count is maintained that is incremented when an
1099 * empty list is encountered. This is so more pages are freed
1100 * off fuller lists instead of spinning excessively around empty
1101 * lists
1102 */
1103 do {
1104 batch_free++;
1105 if (++migratetype == MIGRATE_PCPTYPES)
1106 migratetype = 0;
1107 list = &pcp->lists[migratetype];
1108 } while (list_empty(list));
1109
1110 /* This is the only non-empty list. Free them all. */
1111 if (batch_free == MIGRATE_PCPTYPES)
1112 batch_free = count;
1113
1114 do {
1115 page = list_last_entry(list, struct page, lru);
1116 /* must delete to avoid corrupting pcp list */
1117 list_del(&page->lru);
1118 pcp->count--;
1119
1120 if (bulkfree_pcp_prepare(page))
1121 continue;
1122
1123 list_add_tail(&page->lru, &head);
1124
1125 /*
1126 * We are going to put the page back to the global
1127 * pool, prefetch its buddy to speed up later access
1128 * under zone->lock. It is believed the overhead of
1129 * an additional test and calculating buddy_pfn here
1130 * can be offset by reduced memory latency later. To
1131 * avoid excessive prefetching due to large count, only
1132 * prefetch buddy for the first pcp->batch nr of pages.
1133 */
1134 if (prefetch_nr++ < pcp->batch)
1135 prefetch_buddy(page);
1136 } while (--count && --batch_free && !list_empty(list));
1137 }
1138
1139 spin_lock(&zone->lock);
1140 isolated_pageblocks = has_isolate_pageblock(zone);
1141
1142 /*
1143 * Use safe version since after __free_one_page(),
1144 * page->lru.next will not point to original list.
1145 */
1146 list_for_each_entry_safe(page, tmp, &head, lru) {
1147 int mt = get_pcppage_migratetype(page);
1148 /* MIGRATE_ISOLATE page should not go to pcplists */
1149 VM_BUG_ON_PAGE(is_migrate_isolate(mt), page);
1150 /* Pageblock could have been isolated meanwhile */
1151 if (unlikely(isolated_pageblocks))
1152 mt = get_pageblock_migratetype(page);
1153
1154 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
1155 trace_mm_page_pcpu_drain(page, 0, mt);
1156 }
1157 spin_unlock(&zone->lock);
1158}
1159
1160static void free_one_page(struct zone *zone,
1161 struct page *page, unsigned long pfn,
1162 unsigned int order,
1163 int migratetype)
1164{
1165 spin_lock(&zone->lock);
1166 if (unlikely(has_isolate_pageblock(zone) ||
1167 is_migrate_isolate(migratetype))) {
1168 migratetype = get_pfnblock_migratetype(page, pfn);
1169 }
1170 __free_one_page(page, pfn, zone, order, migratetype);
1171 spin_unlock(&zone->lock);
1172}
1173
1174static void __meminit __init_single_page(struct page *page, unsigned long pfn,
1175 unsigned long zone, int nid)
1176{
1177 mm_zero_struct_page(page);
1178 set_page_links(page, zone, nid, pfn);
1179 init_page_count(page);
1180 page_mapcount_reset(page);
1181 page_cpupid_reset_last(page);
1182
1183 INIT_LIST_HEAD(&page->lru);
1184#ifdef WANT_PAGE_VIRTUAL
1185 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
1186 if (!is_highmem_idx(zone))
1187 set_page_address(page, __va(pfn << PAGE_SHIFT));
1188#endif
1189}
1190
1191#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1192static void __meminit init_reserved_page(unsigned long pfn)
1193{
1194 pg_data_t *pgdat;
1195 int nid, zid;
1196
1197 if (!early_page_uninitialised(pfn))
1198 return;
1199
1200 nid = early_pfn_to_nid(pfn);
1201 pgdat = NODE_DATA(nid);
1202
1203 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1204 struct zone *zone = &pgdat->node_zones[zid];
1205
1206 if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
1207 break;
1208 }
1209 __init_single_page(pfn_to_page(pfn), pfn, zid, nid);
1210}
1211#else
1212static inline void init_reserved_page(unsigned long pfn)
1213{
1214}
1215#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1216
1217/*
1218 * Initialised pages do not have PageReserved set. This function is
1219 * called for each range allocated by the bootmem allocator and
1220 * marks the pages PageReserved. The remaining valid pages are later
1221 * sent to the buddy page allocator.
1222 */
1223void __meminit reserve_bootmem_region(phys_addr_t start, phys_addr_t end)
1224{
1225 unsigned long start_pfn = PFN_DOWN(start);
1226 unsigned long end_pfn = PFN_UP(end);
1227
1228 for (; start_pfn < end_pfn; start_pfn++) {
1229 if (pfn_valid(start_pfn)) {
1230 struct page *page = pfn_to_page(start_pfn);
1231
1232 init_reserved_page(start_pfn);
1233
1234 /* Avoid false-positive PageTail() */
1235 INIT_LIST_HEAD(&page->lru);
1236
1237 SetPageReserved(page);
1238 }
1239 }
1240}
1241
1242static void __free_pages_ok(struct page *page, unsigned int order)
1243{
1244 unsigned long flags;
1245 int migratetype;
1246 unsigned long pfn = page_to_pfn(page);
1247
1248 if (!free_pages_prepare(page, order, true))
1249 return;
1250
1251 migratetype = get_pfnblock_migratetype(page, pfn);
1252 local_irq_save(flags);
1253 __count_vm_events(PGFREE, 1 << order);
1254 free_one_page(page_zone(page), page, pfn, order, migratetype);
1255 local_irq_restore(flags);
1256}
1257
1258static void __init __free_pages_boot_core(struct page *page, unsigned int order)
1259{
1260 unsigned int nr_pages = 1 << order;
1261 struct page *p = page;
1262 unsigned int loop;
1263
1264 prefetchw(p);
1265 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
1266 prefetchw(p + 1);
1267 __ClearPageReserved(p);
1268 set_page_count(p, 0);
1269 }
1270 __ClearPageReserved(p);
1271 set_page_count(p, 0);
1272
1273 page_zone(page)->managed_pages += nr_pages;
1274 set_page_refcounted(page);
1275 __free_pages(page, order);
1276}
1277
1278#if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1279 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1280
1281static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata;
1282
1283int __meminit early_pfn_to_nid(unsigned long pfn)
1284{
1285 static DEFINE_SPINLOCK(early_pfn_lock);
1286 int nid;
1287
1288 spin_lock(&early_pfn_lock);
1289 nid = __early_pfn_to_nid(pfn, &early_pfnnid_cache);
1290 if (nid < 0)
1291 nid = first_online_node;
1292 spin_unlock(&early_pfn_lock);
1293
1294 return nid;
1295}
1296#endif
1297
1298#ifdef CONFIG_NODES_SPAN_OTHER_NODES
1299static inline bool __meminit __maybe_unused
1300meminit_pfn_in_nid(unsigned long pfn, int node,
1301 struct mminit_pfnnid_cache *state)
1302{
1303 int nid;
1304
1305 nid = __early_pfn_to_nid(pfn, state);
1306 if (nid >= 0 && nid != node)
1307 return false;
1308 return true;
1309}
1310
1311/* Only safe to use early in boot when initialisation is single-threaded */
1312static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1313{
1314 return meminit_pfn_in_nid(pfn, node, &early_pfnnid_cache);
1315}
1316
1317#else
1318
1319static inline bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
1320{
1321 return true;
1322}
1323static inline bool __meminit __maybe_unused
1324meminit_pfn_in_nid(unsigned long pfn, int node,
1325 struct mminit_pfnnid_cache *state)
1326{
1327 return true;
1328}
1329#endif
1330
1331
1332void __init __free_pages_bootmem(struct page *page, unsigned long pfn,
1333 unsigned int order)
1334{
1335 if (early_page_uninitialised(pfn))
1336 return;
1337 return __free_pages_boot_core(page, order);
1338}
1339
1340/*
1341 * Check that the whole (or subset of) a pageblock given by the interval of
1342 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1343 * with the migration of free compaction scanner. The scanners then need to
1344 * use only pfn_valid_within() check for arches that allow holes within
1345 * pageblocks.
1346 *
1347 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1348 *
1349 * It's possible on some configurations to have a setup like node0 node1 node0
1350 * i.e. it's possible that all pages within a zones range of pages do not
1351 * belong to a single zone. We assume that a border between node0 and node1
1352 * can occur within a single pageblock, but not a node0 node1 node0
1353 * interleaving within a single pageblock. It is therefore sufficient to check
1354 * the first and last page of a pageblock and avoid checking each individual
1355 * page in a pageblock.
1356 */
1357struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1358 unsigned long end_pfn, struct zone *zone)
1359{
1360 struct page *start_page;
1361 struct page *end_page;
1362
1363 /* end_pfn is one past the range we are checking */
1364 end_pfn--;
1365
1366 if (!pfn_valid(start_pfn) || !pfn_valid(end_pfn))
1367 return NULL;
1368
1369 start_page = pfn_to_online_page(start_pfn);
1370 if (!start_page)
1371 return NULL;
1372
1373 if (page_zone(start_page) != zone)
1374 return NULL;
1375
1376 end_page = pfn_to_page(end_pfn);
1377
1378 /* This gives a shorter code than deriving page_zone(end_page) */
1379 if (page_zone_id(start_page) != page_zone_id(end_page))
1380 return NULL;
1381
1382 return start_page;
1383}
1384
1385void set_zone_contiguous(struct zone *zone)
1386{
1387 unsigned long block_start_pfn = zone->zone_start_pfn;
1388 unsigned long block_end_pfn;
1389
1390 block_end_pfn = ALIGN(block_start_pfn + 1, pageblock_nr_pages);
1391 for (; block_start_pfn < zone_end_pfn(zone);
1392 block_start_pfn = block_end_pfn,
1393 block_end_pfn += pageblock_nr_pages) {
1394
1395 block_end_pfn = min(block_end_pfn, zone_end_pfn(zone));
1396
1397 if (!__pageblock_pfn_to_page(block_start_pfn,
1398 block_end_pfn, zone))
1399 return;
1400 }
1401
1402 /* We confirm that there is no hole */
1403 zone->contiguous = true;
1404}
1405
1406void clear_zone_contiguous(struct zone *zone)
1407{
1408 zone->contiguous = false;
1409}
1410
1411#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1412static void __init deferred_free_range(unsigned long pfn,
1413 unsigned long nr_pages)
1414{
1415 struct page *page;
1416 unsigned long i;
1417
1418 if (!nr_pages)
1419 return;
1420
1421 page = pfn_to_page(pfn);
1422
1423 /* Free a large naturally-aligned chunk if possible */
1424 if (nr_pages == pageblock_nr_pages &&
1425 (pfn & (pageblock_nr_pages - 1)) == 0) {
1426 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1427 __free_pages_boot_core(page, pageblock_order);
1428 return;
1429 }
1430
1431 for (i = 0; i < nr_pages; i++, page++, pfn++) {
1432 if ((pfn & (pageblock_nr_pages - 1)) == 0)
1433 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1434 __free_pages_boot_core(page, 0);
1435 }
1436}
1437
1438/* Completion tracking for deferred_init_memmap() threads */
1439static atomic_t pgdat_init_n_undone __initdata;
1440static __initdata DECLARE_COMPLETION(pgdat_init_all_done_comp);
1441
1442static inline void __init pgdat_init_report_one_done(void)
1443{
1444 if (atomic_dec_and_test(&pgdat_init_n_undone))
1445 complete(&pgdat_init_all_done_comp);
1446}
1447
1448/*
1449 * Returns true if page needs to be initialized or freed to buddy allocator.
1450 *
1451 * First we check if pfn is valid on architectures where it is possible to have
1452 * holes within pageblock_nr_pages. On systems where it is not possible, this
1453 * function is optimized out.
1454 *
1455 * Then, we check if a current large page is valid by only checking the validity
1456 * of the head pfn.
1457 *
1458 * Finally, meminit_pfn_in_nid is checked on systems where pfns can interleave
1459 * within a node: a pfn is between start and end of a node, but does not belong
1460 * to this memory node.
1461 */
1462static inline bool __init
1463deferred_pfn_valid(int nid, unsigned long pfn,
1464 struct mminit_pfnnid_cache *nid_init_state)
1465{
1466 if (!pfn_valid_within(pfn))
1467 return false;
1468 if (!(pfn & (pageblock_nr_pages - 1)) && !pfn_valid(pfn))
1469 return false;
1470 if (!meminit_pfn_in_nid(pfn, nid, nid_init_state))
1471 return false;
1472 return true;
1473}
1474
1475/*
1476 * Free pages to buddy allocator. Try to free aligned pages in
1477 * pageblock_nr_pages sizes.
1478 */
1479static void __init deferred_free_pages(int nid, int zid, unsigned long pfn,
1480 unsigned long end_pfn)
1481{
1482 struct mminit_pfnnid_cache nid_init_state = { };
1483 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1484 unsigned long nr_free = 0;
1485
1486 for (; pfn < end_pfn; pfn++) {
1487 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1488 deferred_free_range(pfn - nr_free, nr_free);
1489 nr_free = 0;
1490 } else if (!(pfn & nr_pgmask)) {
1491 deferred_free_range(pfn - nr_free, nr_free);
1492 nr_free = 1;
1493 touch_nmi_watchdog();
1494 } else {
1495 nr_free++;
1496 }
1497 }
1498 /* Free the last block of pages to allocator */
1499 deferred_free_range(pfn - nr_free, nr_free);
1500}
1501
1502/*
1503 * Initialize struct pages. We minimize pfn page lookups and scheduler checks
1504 * by performing it only once every pageblock_nr_pages.
1505 * Return number of pages initialized.
1506 */
1507static unsigned long __init deferred_init_pages(int nid, int zid,
1508 unsigned long pfn,
1509 unsigned long end_pfn)
1510{
1511 struct mminit_pfnnid_cache nid_init_state = { };
1512 unsigned long nr_pgmask = pageblock_nr_pages - 1;
1513 unsigned long nr_pages = 0;
1514 struct page *page = NULL;
1515
1516 for (; pfn < end_pfn; pfn++) {
1517 if (!deferred_pfn_valid(nid, pfn, &nid_init_state)) {
1518 page = NULL;
1519 continue;
1520 } else if (!page || !(pfn & nr_pgmask)) {
1521 page = pfn_to_page(pfn);
1522 touch_nmi_watchdog();
1523 } else {
1524 page++;
1525 }
1526 __init_single_page(page, pfn, zid, nid);
1527 nr_pages++;
1528 }
1529 return (nr_pages);
1530}
1531
1532/* Initialise remaining memory on a node */
1533static int __init deferred_init_memmap(void *data)
1534{
1535 pg_data_t *pgdat = data;
1536 int nid = pgdat->node_id;
1537 unsigned long start = jiffies;
1538 unsigned long nr_pages = 0;
1539 unsigned long spfn, epfn, first_init_pfn, flags;
1540 phys_addr_t spa, epa;
1541 int zid;
1542 struct zone *zone;
1543 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1544 u64 i;
1545
1546 /* Bind memory initialisation thread to a local node if possible */
1547 if (!cpumask_empty(cpumask))
1548 set_cpus_allowed_ptr(current, cpumask);
1549
1550 pgdat_resize_lock(pgdat, &flags);
1551 first_init_pfn = pgdat->first_deferred_pfn;
1552 if (first_init_pfn == ULONG_MAX) {
1553 pgdat_resize_unlock(pgdat, &flags);
1554 pgdat_init_report_one_done();
1555 return 0;
1556 }
1557
1558 /* Sanity check boundaries */
1559 BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
1560 BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
1561 pgdat->first_deferred_pfn = ULONG_MAX;
1562
1563 /* Only the highest zone is deferred so find it */
1564 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1565 zone = pgdat->node_zones + zid;
1566 if (first_init_pfn < zone_end_pfn(zone))
1567 break;
1568 }
1569 first_init_pfn = max(zone->zone_start_pfn, first_init_pfn);
1570
1571 /*
1572 * Initialize and free pages. We do it in two loops: first we initialize
1573 * struct page, than free to buddy allocator, because while we are
1574 * freeing pages we can access pages that are ahead (computing buddy
1575 * page in __free_one_page()).
1576 */
1577 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1578 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1579 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1580 nr_pages += deferred_init_pages(nid, zid, spfn, epfn);
1581 }
1582 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1583 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1584 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1585 deferred_free_pages(nid, zid, spfn, epfn);
1586 }
1587 pgdat_resize_unlock(pgdat, &flags);
1588
1589 /* Sanity check that the next zone really is unpopulated */
1590 WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
1591
1592 pr_info("node %d initialised, %lu pages in %ums\n", nid, nr_pages,
1593 jiffies_to_msecs(jiffies - start));
1594
1595 pgdat_init_report_one_done();
1596 return 0;
1597}
1598
1599/*
1600 * During boot we initialize deferred pages on-demand, as needed, but once
1601 * page_alloc_init_late() has finished, the deferred pages are all initialized,
1602 * and we can permanently disable that path.
1603 */
1604static DEFINE_STATIC_KEY_TRUE(deferred_pages);
1605
1606/*
1607 * If this zone has deferred pages, try to grow it by initializing enough
1608 * deferred pages to satisfy the allocation specified by order, rounded up to
1609 * the nearest PAGES_PER_SECTION boundary. So we're adding memory in increments
1610 * of SECTION_SIZE bytes by initializing struct pages in increments of
1611 * PAGES_PER_SECTION * sizeof(struct page) bytes.
1612 *
1613 * Return true when zone was grown, otherwise return false. We return true even
1614 * when we grow less than requested, to let the caller decide if there are
1615 * enough pages to satisfy the allocation.
1616 *
1617 * Note: We use noinline because this function is needed only during boot, and
1618 * it is called from a __ref function _deferred_grow_zone. This way we are
1619 * making sure that it is not inlined into permanent text section.
1620 */
1621static noinline bool __init
1622deferred_grow_zone(struct zone *zone, unsigned int order)
1623{
1624 int zid = zone_idx(zone);
1625 int nid = zone_to_nid(zone);
1626 pg_data_t *pgdat = NODE_DATA(nid);
1627 unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
1628 unsigned long nr_pages = 0;
1629 unsigned long first_init_pfn, spfn, epfn, t, flags;
1630 unsigned long first_deferred_pfn = pgdat->first_deferred_pfn;
1631 phys_addr_t spa, epa;
1632 u64 i;
1633
1634 /* Only the last zone may have deferred pages */
1635 if (zone_end_pfn(zone) != pgdat_end_pfn(pgdat))
1636 return false;
1637
1638 pgdat_resize_lock(pgdat, &flags);
1639
1640 /*
1641 * If deferred pages have been initialized while we were waiting for
1642 * the lock, return true, as the zone was grown. The caller will retry
1643 * this zone. We won't return to this function since the caller also
1644 * has this static branch.
1645 */
1646 if (!static_branch_unlikely(&deferred_pages)) {
1647 pgdat_resize_unlock(pgdat, &flags);
1648 return true;
1649 }
1650
1651 /*
1652 * If someone grew this zone while we were waiting for spinlock, return
1653 * true, as there might be enough pages already.
1654 */
1655 if (first_deferred_pfn != pgdat->first_deferred_pfn) {
1656 pgdat_resize_unlock(pgdat, &flags);
1657 return true;
1658 }
1659
1660 first_init_pfn = max(zone->zone_start_pfn, first_deferred_pfn);
1661
1662 if (first_init_pfn >= pgdat_end_pfn(pgdat)) {
1663 pgdat_resize_unlock(pgdat, &flags);
1664 return false;
1665 }
1666
1667 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1668 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1669 epfn = min_t(unsigned long, zone_end_pfn(zone), PFN_DOWN(epa));
1670
1671 while (spfn < epfn && nr_pages < nr_pages_needed) {
1672 t = ALIGN(spfn + PAGES_PER_SECTION, PAGES_PER_SECTION);
1673 first_deferred_pfn = min(t, epfn);
1674 nr_pages += deferred_init_pages(nid, zid, spfn,
1675 first_deferred_pfn);
1676 spfn = first_deferred_pfn;
1677 }
1678
1679 if (nr_pages >= nr_pages_needed)
1680 break;
1681 }
1682
1683 for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &spa, &epa, NULL) {
1684 spfn = max_t(unsigned long, first_init_pfn, PFN_UP(spa));
1685 epfn = min_t(unsigned long, first_deferred_pfn, PFN_DOWN(epa));
1686 deferred_free_pages(nid, zid, spfn, epfn);
1687
1688 if (first_deferred_pfn == epfn)
1689 break;
1690 }
1691 pgdat->first_deferred_pfn = first_deferred_pfn;
1692 pgdat_resize_unlock(pgdat, &flags);
1693
1694 return nr_pages > 0;
1695}
1696
1697/*
1698 * deferred_grow_zone() is __init, but it is called from
1699 * get_page_from_freelist() during early boot until deferred_pages permanently
1700 * disables this call. This is why we have refdata wrapper to avoid warning,
1701 * and to ensure that the function body gets unloaded.
1702 */
1703static bool __ref
1704_deferred_grow_zone(struct zone *zone, unsigned int order)
1705{
1706 return deferred_grow_zone(zone, order);
1707}
1708
1709#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1710
1711void __init page_alloc_init_late(void)
1712{
1713 struct zone *zone;
1714
1715#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1716 int nid;
1717
1718 /* There will be num_node_state(N_MEMORY) threads */
1719 atomic_set(&pgdat_init_n_undone, num_node_state(N_MEMORY));
1720 for_each_node_state(nid, N_MEMORY) {
1721 kthread_run(deferred_init_memmap, NODE_DATA(nid), "pgdatinit%d", nid);
1722 }
1723
1724 /* Block until all are initialised */
1725 wait_for_completion(&pgdat_init_all_done_comp);
1726
1727 /*
1728 * We initialized the rest of the deferred pages. Permanently disable
1729 * on-demand struct page initialization.
1730 */
1731 static_branch_disable(&deferred_pages);
1732
1733 /* Reinit limits that are based on free pages after the kernel is up */
1734 files_maxfiles_init();
1735#endif
1736#ifdef CONFIG_ARCH_DISCARD_MEMBLOCK
1737 /* Discard memblock private memory */
1738 memblock_discard();
1739#endif
1740
1741 for_each_populated_zone(zone)
1742 set_zone_contiguous(zone);
1743}
1744
1745#ifdef CONFIG_CMA
1746/* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1747void __init init_cma_reserved_pageblock(struct page *page)
1748{
1749 unsigned i = pageblock_nr_pages;
1750 struct page *p = page;
1751
1752 do {
1753 __ClearPageReserved(p);
1754 set_page_count(p, 0);
1755 } while (++p, --i);
1756
1757 set_pageblock_migratetype(page, MIGRATE_CMA);
1758
1759 if (pageblock_order >= MAX_ORDER) {
1760 i = pageblock_nr_pages;
1761 p = page;
1762 do {
1763 set_page_refcounted(p);
1764 __free_pages(p, MAX_ORDER - 1);
1765 p += MAX_ORDER_NR_PAGES;
1766 } while (i -= MAX_ORDER_NR_PAGES);
1767 } else {
1768 set_page_refcounted(page);
1769 __free_pages(page, pageblock_order);
1770 }
1771
1772 adjust_managed_page_count(page, pageblock_nr_pages);
1773}
1774#endif
1775
1776/*
1777 * The order of subdivision here is critical for the IO subsystem.
1778 * Please do not alter this order without good reasons and regression
1779 * testing. Specifically, as large blocks of memory are subdivided,
1780 * the order in which smaller blocks are delivered depends on the order
1781 * they're subdivided in this function. This is the primary factor
1782 * influencing the order in which pages are delivered to the IO
1783 * subsystem according to empirical testing, and this is also justified
1784 * by considering the behavior of a buddy system containing a single
1785 * large block of memory acted on by a series of small allocations.
1786 * This behavior is a critical factor in sglist merging's success.
1787 *
1788 * -- nyc
1789 */
1790static inline void expand(struct zone *zone, struct page *page,
1791 int low, int high, struct free_area *area,
1792 int migratetype)
1793{
1794 unsigned long size = 1 << high;
1795
1796 while (high > low) {
1797 area--;
1798 high--;
1799 size >>= 1;
1800 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1801
1802 /*
1803 * Mark as guard pages (or page), that will allow to
1804 * merge back to allocator when buddy will be freed.
1805 * Corresponding page table entries will not be touched,
1806 * pages will stay not present in virtual address space
1807 */
1808 if (set_page_guard(zone, &page[size], high, migratetype))
1809 continue;
1810
1811 list_add(&page[size].lru, &area->free_list[migratetype]);
1812 area->nr_free++;
1813 set_page_order(&page[size], high);
1814 }
1815}
1816
1817static void check_new_page_bad(struct page *page)
1818{
1819 const char *bad_reason = NULL;
1820 unsigned long bad_flags = 0;
1821
1822 if (unlikely(atomic_read(&page->_mapcount) != -1))
1823 bad_reason = "nonzero mapcount";
1824 if (unlikely(page->mapping != NULL))
1825 bad_reason = "non-NULL mapping";
1826 if (unlikely(page_ref_count(page) != 0))
1827 bad_reason = "nonzero _count";
1828 if (unlikely(page->flags & __PG_HWPOISON)) {
1829 bad_reason = "HWPoisoned (hardware-corrupted)";
1830 bad_flags = __PG_HWPOISON;
1831 /* Don't complain about hwpoisoned pages */
1832 page_mapcount_reset(page); /* remove PageBuddy */
1833 return;
1834 }
1835 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
1836 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
1837 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
1838 }
1839#ifdef CONFIG_MEMCG
1840 if (unlikely(page->mem_cgroup))
1841 bad_reason = "page still charged to cgroup";
1842#endif
1843 bad_page(page, bad_reason, bad_flags);
1844}
1845
1846/*
1847 * This page is about to be returned from the page allocator
1848 */
1849static inline int check_new_page(struct page *page)
1850{
1851 if (likely(page_expected_state(page,
1852 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1853 return 0;
1854
1855 check_new_page_bad(page);
1856 return 1;
1857}
1858
1859static inline bool free_pages_prezeroed(void)
1860{
1861 return IS_ENABLED(CONFIG_PAGE_POISONING_ZERO) &&
1862 page_poisoning_enabled();
1863}
1864
1865#ifdef CONFIG_DEBUG_VM
1866static bool check_pcp_refill(struct page *page)
1867{
1868 return false;
1869}
1870
1871static bool check_new_pcp(struct page *page)
1872{
1873 return check_new_page(page);
1874}
1875#else
1876static bool check_pcp_refill(struct page *page)
1877{
1878 return check_new_page(page);
1879}
1880static bool check_new_pcp(struct page *page)
1881{
1882 return false;
1883}
1884#endif /* CONFIG_DEBUG_VM */
1885
1886static bool check_new_pages(struct page *page, unsigned int order)
1887{
1888 int i;
1889 for (i = 0; i < (1 << order); i++) {
1890 struct page *p = page + i;
1891
1892 if (unlikely(check_new_page(p)))
1893 return true;
1894 }
1895
1896 return false;
1897}
1898
1899inline void post_alloc_hook(struct page *page, unsigned int order,
1900 gfp_t gfp_flags)
1901{
1902 set_page_private(page, 0);
1903 set_page_refcounted(page);
1904
1905 arch_alloc_page(page, order);
1906 kernel_map_pages(page, 1 << order, 1);
1907 kernel_poison_pages(page, 1 << order, 1);
1908 kasan_alloc_pages(page, order);
1909 set_page_owner(page, order, gfp_flags);
1910}
1911
1912static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1913 unsigned int alloc_flags)
1914{
1915 int i;
1916
1917 post_alloc_hook(page, order, gfp_flags);
1918
1919 if (!free_pages_prezeroed() && (gfp_flags & __GFP_ZERO))
1920 for (i = 0; i < (1 << order); i++)
1921 clear_highpage(page + i);
1922
1923 if (order && (gfp_flags & __GFP_COMP))
1924 prep_compound_page(page, order);
1925
1926 /*
1927 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1928 * allocate the page. The expectation is that the caller is taking
1929 * steps that will free more memory. The caller should avoid the page
1930 * being used for !PFMEMALLOC purposes.
1931 */
1932 if (alloc_flags & ALLOC_NO_WATERMARKS)
1933 set_page_pfmemalloc(page);
1934 else
1935 clear_page_pfmemalloc(page);
1936}
1937
1938/*
1939 * Go through the free lists for the given migratetype and remove
1940 * the smallest available page from the freelists
1941 */
1942static __always_inline
1943struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1944 int migratetype)
1945{
1946 unsigned int current_order;
1947 struct free_area *area;
1948 struct page *page;
1949
1950 /* Find a page of the appropriate size in the preferred list */
1951 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1952 area = &(zone->free_area[current_order]);
1953 page = list_first_entry_or_null(&area->free_list[migratetype],
1954 struct page, lru);
1955 if (!page)
1956 continue;
1957 list_del(&page->lru);
1958 rmv_page_order(page);
1959 area->nr_free--;
1960 expand(zone, page, order, current_order, area, migratetype);
1961 set_pcppage_migratetype(page, migratetype);
1962 return page;
1963 }
1964
1965 return NULL;
1966}
1967
1968
1969/*
1970 * This array describes the order lists are fallen back to when
1971 * the free lists for the desirable migrate type are depleted
1972 */
1973static int fallbacks[MIGRATE_TYPES][4] = {
1974 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1975 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_TYPES },
1976 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_TYPES },
1977#ifdef CONFIG_CMA
1978 [MIGRATE_CMA] = { MIGRATE_TYPES }, /* Never used */
1979#endif
1980#ifdef CONFIG_MEMORY_ISOLATION
1981 [MIGRATE_ISOLATE] = { MIGRATE_TYPES }, /* Never used */
1982#endif
1983};
1984
1985#ifdef CONFIG_CMA
1986static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1987 unsigned int order)
1988{
1989 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1990}
1991#else
1992static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1993 unsigned int order) { return NULL; }
1994#endif
1995
1996/*
1997 * Move the free pages in a range to the free lists of the requested type.
1998 * Note that start_page and end_pages are not aligned on a pageblock
1999 * boundary. If alignment is required, use move_freepages_block()
2000 */
2001static int move_freepages(struct zone *zone,
2002 struct page *start_page, struct page *end_page,
2003 int migratetype, int *num_movable)
2004{
2005 struct page *page;
2006 unsigned int order;
2007 int pages_moved = 0;
2008
2009#ifndef CONFIG_HOLES_IN_ZONE
2010 /*
2011 * page_zone is not safe to call in this context when
2012 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
2013 * anyway as we check zone boundaries in move_freepages_block().
2014 * Remove at a later date when no bug reports exist related to
2015 * grouping pages by mobility
2016 */
2017 VM_BUG_ON(pfn_valid(page_to_pfn(start_page)) &&
2018 pfn_valid(page_to_pfn(end_page)) &&
2019 page_zone(start_page) != page_zone(end_page));
2020#endif
2021
2022 if (num_movable)
2023 *num_movable = 0;
2024
2025 for (page = start_page; page <= end_page;) {
2026 if (!pfn_valid_within(page_to_pfn(page))) {
2027 page++;
2028 continue;
2029 }
2030
2031 /* Make sure we are not inadvertently changing nodes */
2032 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
2033
2034 if (!PageBuddy(page)) {
2035 /*
2036 * We assume that pages that could be isolated for
2037 * migration are movable. But we don't actually try
2038 * isolating, as that would be expensive.
2039 */
2040 if (num_movable &&
2041 (PageLRU(page) || __PageMovable(page)))
2042 (*num_movable)++;
2043
2044 page++;
2045 continue;
2046 }
2047
2048 order = page_order(page);
2049 list_move(&page->lru,
2050 &zone->free_area[order].free_list[migratetype]);
2051 page += 1 << order;
2052 pages_moved += 1 << order;
2053 }
2054
2055 return pages_moved;
2056}
2057
2058int move_freepages_block(struct zone *zone, struct page *page,
2059 int migratetype, int *num_movable)
2060{
2061 unsigned long start_pfn, end_pfn;
2062 struct page *start_page, *end_page;
2063
2064 start_pfn = page_to_pfn(page);
2065 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
2066 start_page = pfn_to_page(start_pfn);
2067 end_page = start_page + pageblock_nr_pages - 1;
2068 end_pfn = start_pfn + pageblock_nr_pages - 1;
2069
2070 /* Do not cross zone boundaries */
2071 if (!zone_spans_pfn(zone, start_pfn))
2072 start_page = page;
2073 if (!zone_spans_pfn(zone, end_pfn))
2074 return 0;
2075
2076 return move_freepages(zone, start_page, end_page, migratetype,
2077 num_movable);
2078}
2079
2080static void change_pageblock_range(struct page *pageblock_page,
2081 int start_order, int migratetype)
2082{
2083 int nr_pageblocks = 1 << (start_order - pageblock_order);
2084
2085 while (nr_pageblocks--) {
2086 set_pageblock_migratetype(pageblock_page, migratetype);
2087 pageblock_page += pageblock_nr_pages;
2088 }
2089}
2090
2091/*
2092 * When we are falling back to another migratetype during allocation, try to
2093 * steal extra free pages from the same pageblocks to satisfy further
2094 * allocations, instead of polluting multiple pageblocks.
2095 *
2096 * If we are stealing a relatively large buddy page, it is likely there will
2097 * be more free pages in the pageblock, so try to steal them all. For
2098 * reclaimable and unmovable allocations, we steal regardless of page size,
2099 * as fragmentation caused by those allocations polluting movable pageblocks
2100 * is worse than movable allocations stealing from unmovable and reclaimable
2101 * pageblocks.
2102 */
2103static bool can_steal_fallback(unsigned int order, int start_mt)
2104{
2105 /*
2106 * Leaving this order check is intended, although there is
2107 * relaxed order check in next check. The reason is that
2108 * we can actually steal whole pageblock if this condition met,
2109 * but, below check doesn't guarantee it and that is just heuristic
2110 * so could be changed anytime.
2111 */
2112 if (order >= pageblock_order)
2113 return true;
2114
2115 if (order >= pageblock_order / 2 ||
2116 start_mt == MIGRATE_RECLAIMABLE ||
2117 start_mt == MIGRATE_UNMOVABLE ||
2118 page_group_by_mobility_disabled)
2119 return true;
2120
2121 return false;
2122}
2123
2124/*
2125 * This function implements actual steal behaviour. If order is large enough,
2126 * we can steal whole pageblock. If not, we first move freepages in this
2127 * pageblock to our migratetype and determine how many already-allocated pages
2128 * are there in the pageblock with a compatible migratetype. If at least half
2129 * of pages are free or compatible, we can change migratetype of the pageblock
2130 * itself, so pages freed in the future will be put on the correct free list.
2131 */
2132static void steal_suitable_fallback(struct zone *zone, struct page *page,
2133 int start_type, bool whole_block)
2134{
2135 unsigned int current_order = page_order(page);
2136 struct free_area *area;
2137 int free_pages, movable_pages, alike_pages;
2138 int old_block_type;
2139
2140 old_block_type = get_pageblock_migratetype(page);
2141
2142 /*
2143 * This can happen due to races and we want to prevent broken
2144 * highatomic accounting.
2145 */
2146 if (is_migrate_highatomic(old_block_type))
2147 goto single_page;
2148
2149 /* Take ownership for orders >= pageblock_order */
2150 if (current_order >= pageblock_order) {
2151 change_pageblock_range(page, current_order, start_type);
2152 goto single_page;
2153 }
2154
2155 /* We are not allowed to try stealing from the whole block */
2156 if (!whole_block)
2157 goto single_page;
2158
2159 free_pages = move_freepages_block(zone, page, start_type,
2160 &movable_pages);
2161 /*
2162 * Determine how many pages are compatible with our allocation.
2163 * For movable allocation, it's the number of movable pages which
2164 * we just obtained. For other types it's a bit more tricky.
2165 */
2166 if (start_type == MIGRATE_MOVABLE) {
2167 alike_pages = movable_pages;
2168 } else {
2169 /*
2170 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
2171 * to MOVABLE pageblock, consider all non-movable pages as
2172 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
2173 * vice versa, be conservative since we can't distinguish the
2174 * exact migratetype of non-movable pages.
2175 */
2176 if (old_block_type == MIGRATE_MOVABLE)
2177 alike_pages = pageblock_nr_pages
2178 - (free_pages + movable_pages);
2179 else
2180 alike_pages = 0;
2181 }
2182
2183 /* moving whole block can fail due to zone boundary conditions */
2184 if (!free_pages)
2185 goto single_page;
2186
2187 /*
2188 * If a sufficient number of pages in the block are either free or of
2189 * comparable migratability as our allocation, claim the whole block.
2190 */
2191 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
2192 page_group_by_mobility_disabled)
2193 set_pageblock_migratetype(page, start_type);
2194
2195 return;
2196
2197single_page:
2198 area = &zone->free_area[current_order];
2199 list_move(&page->lru, &area->free_list[start_type]);
2200}
2201
2202/*
2203 * Check whether there is a suitable fallback freepage with requested order.
2204 * If only_stealable is true, this function returns fallback_mt only if
2205 * we can steal other freepages all together. This would help to reduce
2206 * fragmentation due to mixed migratetype pages in one pageblock.
2207 */
2208int find_suitable_fallback(struct free_area *area, unsigned int order,
2209 int migratetype, bool only_stealable, bool *can_steal)
2210{
2211 int i;
2212 int fallback_mt;
2213
2214 if (area->nr_free == 0)
2215 return -1;
2216
2217 *can_steal = false;
2218 for (i = 0;; i++) {
2219 fallback_mt = fallbacks[migratetype][i];
2220 if (fallback_mt == MIGRATE_TYPES)
2221 break;
2222
2223 if (list_empty(&area->free_list[fallback_mt]))
2224 continue;
2225
2226 if (can_steal_fallback(order, migratetype))
2227 *can_steal = true;
2228
2229 if (!only_stealable)
2230 return fallback_mt;
2231
2232 if (*can_steal)
2233 return fallback_mt;
2234 }
2235
2236 return -1;
2237}
2238
2239/*
2240 * Reserve a pageblock for exclusive use of high-order atomic allocations if
2241 * there are no empty page blocks that contain a page with a suitable order
2242 */
2243static void reserve_highatomic_pageblock(struct page *page, struct zone *zone,
2244 unsigned int alloc_order)
2245{
2246 int mt;
2247 unsigned long max_managed, flags;
2248
2249 /*
2250 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
2251 * Check is race-prone but harmless.
2252 */
2253 max_managed = (zone->managed_pages / 100) + pageblock_nr_pages;
2254 if (zone->nr_reserved_highatomic >= max_managed)
2255 return;
2256
2257 spin_lock_irqsave(&zone->lock, flags);
2258
2259 /* Recheck the nr_reserved_highatomic limit under the lock */
2260 if (zone->nr_reserved_highatomic >= max_managed)
2261 goto out_unlock;
2262
2263 /* Yoink! */
2264 mt = get_pageblock_migratetype(page);
2265 if (!is_migrate_highatomic(mt) && !is_migrate_isolate(mt)
2266 && !is_migrate_cma(mt)) {
2267 zone->nr_reserved_highatomic += pageblock_nr_pages;
2268 set_pageblock_migratetype(page, MIGRATE_HIGHATOMIC);
2269 move_freepages_block(zone, page, MIGRATE_HIGHATOMIC, NULL);
2270 }
2271
2272out_unlock:
2273 spin_unlock_irqrestore(&zone->lock, flags);
2274}
2275
2276/*
2277 * Used when an allocation is about to fail under memory pressure. This
2278 * potentially hurts the reliability of high-order allocations when under
2279 * intense memory pressure but failed atomic allocations should be easier
2280 * to recover from than an OOM.
2281 *
2282 * If @force is true, try to unreserve a pageblock even though highatomic
2283 * pageblock is exhausted.
2284 */
2285static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2286 bool force)
2287{
2288 struct zonelist *zonelist = ac->zonelist;
2289 unsigned long flags;
2290 struct zoneref *z;
2291 struct zone *zone;
2292 struct page *page;
2293 int order;
2294 bool ret;
2295
2296 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2297 ac->nodemask) {
2298 /*
2299 * Preserve at least one pageblock unless memory pressure
2300 * is really high.
2301 */
2302 if (!force && zone->nr_reserved_highatomic <=
2303 pageblock_nr_pages)
2304 continue;
2305
2306 spin_lock_irqsave(&zone->lock, flags);
2307 for (order = 0; order < MAX_ORDER; order++) {
2308 struct free_area *area = &(zone->free_area[order]);
2309
2310 page = list_first_entry_or_null(
2311 &area->free_list[MIGRATE_HIGHATOMIC],
2312 struct page, lru);
2313 if (!page)
2314 continue;
2315
2316 /*
2317 * In page freeing path, migratetype change is racy so
2318 * we can counter several free pages in a pageblock
2319 * in this loop althoug we changed the pageblock type
2320 * from highatomic to ac->migratetype. So we should
2321 * adjust the count once.
2322 */
2323 if (is_migrate_highatomic_page(page)) {
2324 /*
2325 * It should never happen but changes to
2326 * locking could inadvertently allow a per-cpu
2327 * drain to add pages to MIGRATE_HIGHATOMIC
2328 * while unreserving so be safe and watch for
2329 * underflows.
2330 */
2331 zone->nr_reserved_highatomic -= min(
2332 pageblock_nr_pages,
2333 zone->nr_reserved_highatomic);
2334 }
2335
2336 /*
2337 * Convert to ac->migratetype and avoid the normal
2338 * pageblock stealing heuristics. Minimally, the caller
2339 * is doing the work and needs the pages. More
2340 * importantly, if the block was always converted to
2341 * MIGRATE_UNMOVABLE or another type then the number
2342 * of pageblocks that cannot be completely freed
2343 * may increase.
2344 */
2345 set_pageblock_migratetype(page, ac->migratetype);
2346 ret = move_freepages_block(zone, page, ac->migratetype,
2347 NULL);
2348 if (ret) {
2349 spin_unlock_irqrestore(&zone->lock, flags);
2350 return ret;
2351 }
2352 }
2353 spin_unlock_irqrestore(&zone->lock, flags);
2354 }
2355
2356 return false;
2357}
2358
2359/*
2360 * Try finding a free buddy page on the fallback list and put it on the free
2361 * list of requested migratetype, possibly along with other pages from the same
2362 * block, depending on fragmentation avoidance heuristics. Returns true if
2363 * fallback was found so that __rmqueue_smallest() can grab it.
2364 *
2365 * The use of signed ints for order and current_order is a deliberate
2366 * deviation from the rest of this file, to make the for loop
2367 * condition simpler.
2368 */
2369static __always_inline bool
2370__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
2371{
2372 struct free_area *area;
2373 int current_order;
2374 struct page *page;
2375 int fallback_mt;
2376 bool can_steal;
2377
2378 /*
2379 * Find the largest available free page in the other list. This roughly
2380 * approximates finding the pageblock with the most free pages, which
2381 * would be too costly to do exactly.
2382 */
2383 for (current_order = MAX_ORDER - 1; current_order >= order;
2384 --current_order) {
2385 area = &(zone->free_area[current_order]);
2386 fallback_mt = find_suitable_fallback(area, current_order,
2387 start_migratetype, false, &can_steal);
2388 if (fallback_mt == -1)
2389 continue;
2390
2391 /*
2392 * We cannot steal all free pages from the pageblock and the
2393 * requested migratetype is movable. In that case it's better to
2394 * steal and split the smallest available page instead of the
2395 * largest available page, because even if the next movable
2396 * allocation falls back into a different pageblock than this
2397 * one, it won't cause permanent fragmentation.
2398 */
2399 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2400 && current_order > order)
2401 goto find_smallest;
2402
2403 goto do_steal;
2404 }
2405
2406 return false;
2407
2408find_smallest:
2409 for (current_order = order; current_order < MAX_ORDER;
2410 current_order++) {
2411 area = &(zone->free_area[current_order]);
2412 fallback_mt = find_suitable_fallback(area, current_order,
2413 start_migratetype, false, &can_steal);
2414 if (fallback_mt != -1)
2415 break;
2416 }
2417
2418 /*
2419 * This should not happen - we already found a suitable fallback
2420 * when looking for the largest page.
2421 */
2422 VM_BUG_ON(current_order == MAX_ORDER);
2423
2424do_steal:
2425 page = list_first_entry(&area->free_list[fallback_mt],
2426 struct page, lru);
2427
2428 steal_suitable_fallback(zone, page, start_migratetype, can_steal);
2429
2430 trace_mm_page_alloc_extfrag(page, order, current_order,
2431 start_migratetype, fallback_mt);
2432
2433 return true;
2434
2435}
2436
2437/*
2438 * Do the hard work of removing an element from the buddy allocator.
2439 * Call me with the zone->lock already held.
2440 */
2441static __always_inline struct page *
2442__rmqueue(struct zone *zone, unsigned int order, int migratetype)
2443{
2444 struct page *page;
2445
2446retry:
2447 page = __rmqueue_smallest(zone, order, migratetype);
2448 if (unlikely(!page)) {
2449 if (migratetype == MIGRATE_MOVABLE)
2450 page = __rmqueue_cma_fallback(zone, order);
2451
2452 if (!page && __rmqueue_fallback(zone, order, migratetype))
2453 goto retry;
2454 }
2455
2456 trace_mm_page_alloc_zone_locked(page, order, migratetype);
2457 return page;
2458}
2459
2460/*
2461 * Obtain a specified number of elements from the buddy allocator, all under
2462 * a single hold of the lock, for efficiency. Add them to the supplied list.
2463 * Returns the number of new pages which were placed at *list.
2464 */
2465static int rmqueue_bulk(struct zone *zone, unsigned int order,
2466 unsigned long count, struct list_head *list,
2467 int migratetype)
2468{
2469 int i, alloced = 0;
2470
2471 spin_lock(&zone->lock);
2472 for (i = 0; i < count; ++i) {
2473 struct page *page = __rmqueue(zone, order, migratetype);
2474 if (unlikely(page == NULL))
2475 break;
2476
2477 if (unlikely(check_pcp_refill(page)))
2478 continue;
2479
2480 /*
2481 * Split buddy pages returned by expand() are received here in
2482 * physical page order. The page is added to the tail of
2483 * caller's list. From the callers perspective, the linked list
2484 * is ordered by page number under some conditions. This is
2485 * useful for IO devices that can forward direction from the
2486 * head, thus also in the physical page order. This is useful
2487 * for IO devices that can merge IO requests if the physical
2488 * pages are ordered properly.
2489 */
2490 list_add_tail(&page->lru, list);
2491 alloced++;
2492 if (is_migrate_cma(get_pcppage_migratetype(page)))
2493 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
2494 -(1 << order));
2495 }
2496
2497 /*
2498 * i pages were removed from the buddy list even if some leak due
2499 * to check_pcp_refill failing so adjust NR_FREE_PAGES based
2500 * on i. Do not confuse with 'alloced' which is the number of
2501 * pages added to the pcp list.
2502 */
2503 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
2504 spin_unlock(&zone->lock);
2505 return alloced;
2506}
2507
2508#ifdef CONFIG_NUMA
2509/*
2510 * Called from the vmstat counter updater to drain pagesets of this
2511 * currently executing processor on remote nodes after they have
2512 * expired.
2513 *
2514 * Note that this function must be called with the thread pinned to
2515 * a single processor.
2516 */
2517void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2518{
2519 unsigned long flags;
2520 int to_drain, batch;
2521
2522 local_irq_save(flags);
2523 batch = READ_ONCE(pcp->batch);
2524 to_drain = min(pcp->count, batch);
2525 if (to_drain > 0)
2526 free_pcppages_bulk(zone, to_drain, pcp);
2527 local_irq_restore(flags);
2528}
2529#endif
2530
2531/*
2532 * Drain pcplists of the indicated processor and zone.
2533 *
2534 * The processor must either be the current processor and the
2535 * thread pinned to the current processor or a processor that
2536 * is not online.
2537 */
2538static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2539{
2540 unsigned long flags;
2541 struct per_cpu_pageset *pset;
2542 struct per_cpu_pages *pcp;
2543
2544 local_irq_save(flags);
2545 pset = per_cpu_ptr(zone->pageset, cpu);
2546
2547 pcp = &pset->pcp;
2548 if (pcp->count)
2549 free_pcppages_bulk(zone, pcp->count, pcp);
2550 local_irq_restore(flags);
2551}
2552
2553/*
2554 * Drain pcplists of all zones on the indicated processor.
2555 *
2556 * The processor must either be the current processor and the
2557 * thread pinned to the current processor or a processor that
2558 * is not online.
2559 */
2560static void drain_pages(unsigned int cpu)
2561{
2562 struct zone *zone;
2563
2564 for_each_populated_zone(zone) {
2565 drain_pages_zone(cpu, zone);
2566 }
2567}
2568
2569/*
2570 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2571 *
2572 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
2573 * the single zone's pages.
2574 */
2575void drain_local_pages(struct zone *zone)
2576{
2577 int cpu = smp_processor_id();
2578
2579 if (zone)
2580 drain_pages_zone(cpu, zone);
2581 else
2582 drain_pages(cpu);
2583}
2584
2585static void drain_local_pages_wq(struct work_struct *work)
2586{
2587 /*
2588 * drain_all_pages doesn't use proper cpu hotplug protection so
2589 * we can race with cpu offline when the WQ can move this from
2590 * a cpu pinned worker to an unbound one. We can operate on a different
2591 * cpu which is allright but we also have to make sure to not move to
2592 * a different one.
2593 */
2594 preempt_disable();
2595 drain_local_pages(NULL);
2596 preempt_enable();
2597}
2598
2599/*
2600 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2601 *
2602 * When zone parameter is non-NULL, spill just the single zone's pages.
2603 *
2604 * Note that this can be extremely slow as the draining happens in a workqueue.
2605 */
2606void drain_all_pages(struct zone *zone)
2607{
2608 int cpu;
2609
2610 /*
2611 * Allocate in the BSS so we wont require allocation in
2612 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2613 */
2614 static cpumask_t cpus_with_pcps;
2615
2616 /*
2617 * Make sure nobody triggers this path before mm_percpu_wq is fully
2618 * initialized.
2619 */
2620 if (WARN_ON_ONCE(!mm_percpu_wq))
2621 return;
2622
2623 /*
2624 * Do not drain if one is already in progress unless it's specific to
2625 * a zone. Such callers are primarily CMA and memory hotplug and need
2626 * the drain to be complete when the call returns.
2627 */
2628 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2629 if (!zone)
2630 return;
2631 mutex_lock(&pcpu_drain_mutex);
2632 }
2633
2634 /*
2635 * We don't care about racing with CPU hotplug event
2636 * as offline notification will cause the notified
2637 * cpu to drain that CPU pcps and on_each_cpu_mask
2638 * disables preemption as part of its processing
2639 */
2640 for_each_online_cpu(cpu) {
2641 struct per_cpu_pageset *pcp;
2642 struct zone *z;
2643 bool has_pcps = false;
2644
2645 if (zone) {
2646 pcp = per_cpu_ptr(zone->pageset, cpu);
2647 if (pcp->pcp.count)
2648 has_pcps = true;
2649 } else {
2650 for_each_populated_zone(z) {
2651 pcp = per_cpu_ptr(z->pageset, cpu);
2652 if (pcp->pcp.count) {
2653 has_pcps = true;
2654 break;
2655 }
2656 }
2657 }
2658
2659 if (has_pcps)
2660 cpumask_set_cpu(cpu, &cpus_with_pcps);
2661 else
2662 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2663 }
2664
2665 for_each_cpu(cpu, &cpus_with_pcps) {
2666 struct work_struct *work = per_cpu_ptr(&pcpu_drain, cpu);
2667 INIT_WORK(work, drain_local_pages_wq);
2668 queue_work_on(cpu, mm_percpu_wq, work);
2669 }
2670 for_each_cpu(cpu, &cpus_with_pcps)
2671 flush_work(per_cpu_ptr(&pcpu_drain, cpu));
2672
2673 mutex_unlock(&pcpu_drain_mutex);
2674}
2675
2676#ifdef CONFIG_HIBERNATION
2677
2678/*
2679 * Touch the watchdog for every WD_PAGE_COUNT pages.
2680 */
2681#define WD_PAGE_COUNT (128*1024)
2682
2683void mark_free_pages(struct zone *zone)
2684{
2685 unsigned long pfn, max_zone_pfn, page_count = WD_PAGE_COUNT;
2686 unsigned long flags;
2687 unsigned int order, t;
2688 struct page *page;
2689
2690 if (zone_is_empty(zone))
2691 return;
2692
2693 spin_lock_irqsave(&zone->lock, flags);
2694
2695 max_zone_pfn = zone_end_pfn(zone);
2696 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
2697 if (pfn_valid(pfn)) {
2698 page = pfn_to_page(pfn);
2699
2700 if (!--page_count) {
2701 touch_nmi_watchdog();
2702 page_count = WD_PAGE_COUNT;
2703 }
2704
2705 if (page_zone(page) != zone)
2706 continue;
2707
2708 if (!swsusp_page_is_forbidden(page))
2709 swsusp_unset_page_free(page);
2710 }
2711
2712 for_each_migratetype_order(order, t) {
2713 list_for_each_entry(page,
2714 &zone->free_area[order].free_list[t], lru) {
2715 unsigned long i;
2716
2717 pfn = page_to_pfn(page);
2718 for (i = 0; i < (1UL << order); i++) {
2719 if (!--page_count) {
2720 touch_nmi_watchdog();
2721 page_count = WD_PAGE_COUNT;
2722 }
2723 swsusp_set_page_free(pfn_to_page(pfn + i));
2724 }
2725 }
2726 }
2727 spin_unlock_irqrestore(&zone->lock, flags);
2728}
2729#endif /* CONFIG_PM */
2730
2731static bool free_unref_page_prepare(struct page *page, unsigned long pfn)
2732{
2733 int migratetype;
2734
2735 if (!free_pcp_prepare(page))
2736 return false;
2737
2738 migratetype = get_pfnblock_migratetype(page, pfn);
2739 set_pcppage_migratetype(page, migratetype);
2740 return true;
2741}
2742
2743static void free_unref_page_commit(struct page *page, unsigned long pfn)
2744{
2745 struct zone *zone = page_zone(page);
2746 struct per_cpu_pages *pcp;
2747 int migratetype;
2748
2749 migratetype = get_pcppage_migratetype(page);
2750 __count_vm_event(PGFREE);
2751
2752 /*
2753 * We only track unmovable, reclaimable and movable on pcp lists.
2754 * Free ISOLATE pages back to the allocator because they are being
2755 * offlined but treat HIGHATOMIC as movable pages so we can get those
2756 * areas back if necessary. Otherwise, we may have to free
2757 * excessively into the page allocator
2758 */
2759 if (migratetype >= MIGRATE_PCPTYPES) {
2760 if (unlikely(is_migrate_isolate(migratetype))) {
2761 free_one_page(zone, page, pfn, 0, migratetype);
2762 return;
2763 }
2764 migratetype = MIGRATE_MOVABLE;
2765 }
2766
2767 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2768 list_add(&page->lru, &pcp->lists[migratetype]);
2769 pcp->count++;
2770 if (pcp->count >= pcp->high) {
2771 unsigned long batch = READ_ONCE(pcp->batch);
2772 free_pcppages_bulk(zone, batch, pcp);
2773 }
2774}
2775
2776/*
2777 * Free a 0-order page
2778 */
2779void free_unref_page(struct page *page)
2780{
2781 unsigned long flags;
2782 unsigned long pfn = page_to_pfn(page);
2783
2784 if (!free_unref_page_prepare(page, pfn))
2785 return;
2786
2787 local_irq_save(flags);
2788 free_unref_page_commit(page, pfn);
2789 local_irq_restore(flags);
2790}
2791
2792/*
2793 * Free a list of 0-order pages
2794 */
2795void free_unref_page_list(struct list_head *list)
2796{
2797 struct page *page, *next;
2798 unsigned long flags, pfn;
2799 int batch_count = 0;
2800
2801 /* Prepare pages for freeing */
2802 list_for_each_entry_safe(page, next, list, lru) {
2803 pfn = page_to_pfn(page);
2804 if (!free_unref_page_prepare(page, pfn))
2805 list_del(&page->lru);
2806 set_page_private(page, pfn);
2807 }
2808
2809 local_irq_save(flags);
2810 list_for_each_entry_safe(page, next, list, lru) {
2811 unsigned long pfn = page_private(page);
2812
2813 set_page_private(page, 0);
2814 trace_mm_page_free_batched(page);
2815 free_unref_page_commit(page, pfn);
2816
2817 /*
2818 * Guard against excessive IRQ disabled times when we get
2819 * a large list of pages to free.
2820 */
2821 if (++batch_count == SWAP_CLUSTER_MAX) {
2822 local_irq_restore(flags);
2823 batch_count = 0;
2824 local_irq_save(flags);
2825 }
2826 }
2827 local_irq_restore(flags);
2828}
2829
2830/*
2831 * split_page takes a non-compound higher-order page, and splits it into
2832 * n (1<<order) sub-pages: page[0..n]
2833 * Each sub-page must be freed individually.
2834 *
2835 * Note: this is probably too low level an operation for use in drivers.
2836 * Please consult with lkml before using this in your driver.
2837 */
2838void split_page(struct page *page, unsigned int order)
2839{
2840 int i;
2841
2842 VM_BUG_ON_PAGE(PageCompound(page), page);
2843 VM_BUG_ON_PAGE(!page_count(page), page);
2844
2845 for (i = 1; i < (1 << order); i++)
2846 set_page_refcounted(page + i);
2847 split_page_owner(page, order);
2848}
2849EXPORT_SYMBOL_GPL(split_page);
2850
2851int __isolate_free_page(struct page *page, unsigned int order)
2852{
2853 unsigned long watermark;
2854 struct zone *zone;
2855 int mt;
2856
2857 BUG_ON(!PageBuddy(page));
2858
2859 zone = page_zone(page);
2860 mt = get_pageblock_migratetype(page);
2861
2862 if (!is_migrate_isolate(mt)) {
2863 /*
2864 * Obey watermarks as if the page was being allocated. We can
2865 * emulate a high-order watermark check with a raised order-0
2866 * watermark, because we already know our high-order page
2867 * exists.
2868 */
2869 watermark = min_wmark_pages(zone) + (1UL << order);
2870 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2871 return 0;
2872
2873 __mod_zone_freepage_state(zone, -(1UL << order), mt);
2874 }
2875
2876 /* Remove page from free list */
2877 list_del(&page->lru);
2878 zone->free_area[order].nr_free--;
2879 rmv_page_order(page);
2880
2881 /*
2882 * Set the pageblock if the isolated page is at least half of a
2883 * pageblock
2884 */
2885 if (order >= pageblock_order - 1) {
2886 struct page *endpage = page + (1 << order) - 1;
2887 for (; page < endpage; page += pageblock_nr_pages) {
2888 int mt = get_pageblock_migratetype(page);
2889 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt)
2890 && !is_migrate_highatomic(mt))
2891 set_pageblock_migratetype(page,
2892 MIGRATE_MOVABLE);
2893 }
2894 }
2895
2896
2897 return 1UL << order;
2898}
2899
2900/*
2901 * Update NUMA hit/miss statistics
2902 *
2903 * Must be called with interrupts disabled.
2904 */
2905static inline void zone_statistics(struct zone *preferred_zone, struct zone *z)
2906{
2907#ifdef CONFIG_NUMA
2908 enum numa_stat_item local_stat = NUMA_LOCAL;
2909
2910 /* skip numa counters update if numa stats is disabled */
2911 if (!static_branch_likely(&vm_numa_stat_key))
2912 return;
2913
2914 if (z->node != numa_node_id())
2915 local_stat = NUMA_OTHER;
2916
2917 if (z->node == preferred_zone->node)
2918 __inc_numa_state(z, NUMA_HIT);
2919 else {
2920 __inc_numa_state(z, NUMA_MISS);
2921 __inc_numa_state(preferred_zone, NUMA_FOREIGN);
2922 }
2923 __inc_numa_state(z, local_stat);
2924#endif
2925}
2926
2927/* Remove page from the per-cpu list, caller must protect the list */
2928static struct page *__rmqueue_pcplist(struct zone *zone, int migratetype,
2929 struct per_cpu_pages *pcp,
2930 struct list_head *list)
2931{
2932 struct page *page;
2933
2934 do {
2935 if (list_empty(list)) {
2936 pcp->count += rmqueue_bulk(zone, 0,
2937 pcp->batch, list,
2938 migratetype);
2939 if (unlikely(list_empty(list)))
2940 return NULL;
2941 }
2942
2943 page = list_first_entry(list, struct page, lru);
2944 list_del(&page->lru);
2945 pcp->count--;
2946 } while (check_new_pcp(page));
2947
2948 return page;
2949}
2950
2951/* Lock and remove page from the per-cpu list */
2952static struct page *rmqueue_pcplist(struct zone *preferred_zone,
2953 struct zone *zone, unsigned int order,
2954 gfp_t gfp_flags, int migratetype)
2955{
2956 struct per_cpu_pages *pcp;
2957 struct list_head *list;
2958 struct page *page;
2959 unsigned long flags;
2960
2961 local_irq_save(flags);
2962 pcp = &this_cpu_ptr(zone->pageset)->pcp;
2963 list = &pcp->lists[migratetype];
2964 page = __rmqueue_pcplist(zone, migratetype, pcp, list);
2965 if (page) {
2966 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2967 zone_statistics(preferred_zone, zone);
2968 }
2969 local_irq_restore(flags);
2970 return page;
2971}
2972
2973/*
2974 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2975 */
2976static inline
2977struct page *rmqueue(struct zone *preferred_zone,
2978 struct zone *zone, unsigned int order,
2979 gfp_t gfp_flags, unsigned int alloc_flags,
2980 int migratetype)
2981{
2982 unsigned long flags;
2983 struct page *page;
2984
2985 if (likely(order == 0)) {
2986 page = rmqueue_pcplist(preferred_zone, zone, order,
2987 gfp_flags, migratetype);
2988 goto out;
2989 }
2990
2991 /*
2992 * We most definitely don't want callers attempting to
2993 * allocate greater than order-1 page units with __GFP_NOFAIL.
2994 */
2995 WARN_ON_ONCE((gfp_flags & __GFP_NOFAIL) && (order > 1));
2996 spin_lock_irqsave(&zone->lock, flags);
2997
2998 do {
2999 page = NULL;
3000 if (alloc_flags & ALLOC_HARDER) {
3001 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
3002 if (page)
3003 trace_mm_page_alloc_zone_locked(page, order, migratetype);
3004 }
3005 if (!page)
3006 page = __rmqueue(zone, order, migratetype);
3007 } while (page && check_new_pages(page, order));
3008 spin_unlock(&zone->lock);
3009 if (!page)
3010 goto failed;
3011 __mod_zone_freepage_state(zone, -(1 << order),
3012 get_pcppage_migratetype(page));
3013
3014 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3015 zone_statistics(preferred_zone, zone);
3016 local_irq_restore(flags);
3017
3018out:
3019 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3020 return page;
3021
3022failed:
3023 local_irq_restore(flags);
3024 return NULL;
3025}
3026
3027#ifdef CONFIG_FAIL_PAGE_ALLOC
3028
3029static struct {
3030 struct fault_attr attr;
3031
3032 bool ignore_gfp_highmem;
3033 bool ignore_gfp_reclaim;
3034 u32 min_order;
3035} fail_page_alloc = {
3036 .attr = FAULT_ATTR_INITIALIZER,
3037 .ignore_gfp_reclaim = true,
3038 .ignore_gfp_highmem = true,
3039 .min_order = 1,
3040};
3041
3042static int __init setup_fail_page_alloc(char *str)
3043{
3044 return setup_fault_attr(&fail_page_alloc.attr, str);
3045}
3046__setup("fail_page_alloc=", setup_fail_page_alloc);
3047
3048static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3049{
3050 if (order < fail_page_alloc.min_order)
3051 return false;
3052 if (gfp_mask & __GFP_NOFAIL)
3053 return false;
3054 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
3055 return false;
3056 if (fail_page_alloc.ignore_gfp_reclaim &&
3057 (gfp_mask & __GFP_DIRECT_RECLAIM))
3058 return false;
3059
3060 return should_fail(&fail_page_alloc.attr, 1 << order);
3061}
3062
3063#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3064
3065static int __init fail_page_alloc_debugfs(void)
3066{
3067 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3068 struct dentry *dir;
3069
3070 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
3071 &fail_page_alloc.attr);
3072 if (IS_ERR(dir))
3073 return PTR_ERR(dir);
3074
3075 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
3076 &fail_page_alloc.ignore_gfp_reclaim))
3077 goto fail;
3078 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
3079 &fail_page_alloc.ignore_gfp_highmem))
3080 goto fail;
3081 if (!debugfs_create_u32("min-order", mode, dir,
3082 &fail_page_alloc.min_order))
3083 goto fail;
3084
3085 return 0;
3086fail:
3087 debugfs_remove_recursive(dir);
3088
3089 return -ENOMEM;
3090}
3091
3092late_initcall(fail_page_alloc_debugfs);
3093
3094#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3095
3096#else /* CONFIG_FAIL_PAGE_ALLOC */
3097
3098static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
3099{
3100 return false;
3101}
3102
3103#endif /* CONFIG_FAIL_PAGE_ALLOC */
3104
3105/*
3106 * Return true if free base pages are above 'mark'. For high-order checks it
3107 * will return true of the order-0 watermark is reached and there is at least
3108 * one free page of a suitable size. Checking now avoids taking the zone lock
3109 * to check in the allocation paths if no pages are free.
3110 */
3111bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3112 int classzone_idx, unsigned int alloc_flags,
3113 long free_pages)
3114{
3115 long min = mark;
3116 int o;
3117 const bool alloc_harder = (alloc_flags & (ALLOC_HARDER|ALLOC_OOM));
3118
3119 /* free_pages may go negative - that's OK */
3120 free_pages -= (1 << order) - 1;
3121
3122 if (alloc_flags & ALLOC_HIGH)
3123 min -= min / 2;
3124
3125 /*
3126 * If the caller does not have rights to ALLOC_HARDER then subtract
3127 * the high-atomic reserves. This will over-estimate the size of the
3128 * atomic reserve but it avoids a search.
3129 */
3130 if (likely(!alloc_harder)) {
3131 free_pages -= z->nr_reserved_highatomic;
3132 } else {
3133 /*
3134 * OOM victims can try even harder than normal ALLOC_HARDER
3135 * users on the grounds that it's definitely going to be in
3136 * the exit path shortly and free memory. Any allocation it
3137 * makes during the free path will be small and short-lived.
3138 */
3139 if (alloc_flags & ALLOC_OOM)
3140 min -= min / 2;
3141 else
3142 min -= min / 4;
3143 }
3144
3145
3146#ifdef CONFIG_CMA
3147 /* If allocation can't use CMA areas don't use free CMA pages */
3148 if (!(alloc_flags & ALLOC_CMA))
3149 free_pages -= zone_page_state(z, NR_FREE_CMA_PAGES);
3150#endif
3151
3152 /*
3153 * Check watermarks for an order-0 allocation request. If these
3154 * are not met, then a high-order request also cannot go ahead
3155 * even if a suitable page happened to be free.
3156 */
3157 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
3158 return false;
3159
3160 /* If this is an order-0 request then the watermark is fine */
3161 if (!order)
3162 return true;
3163
3164 /* For a high-order request, check at least one suitable page is free */
3165 for (o = order; o < MAX_ORDER; o++) {
3166 struct free_area *area = &z->free_area[o];
3167 int mt;
3168
3169 if (!area->nr_free)
3170 continue;
3171
3172 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3173 if (!list_empty(&area->free_list[mt]))
3174 return true;
3175 }
3176
3177#ifdef CONFIG_CMA
3178 if ((alloc_flags & ALLOC_CMA) &&
3179 !list_empty(&area->free_list[MIGRATE_CMA])) {
3180 return true;
3181 }
3182#endif
3183 if (alloc_harder &&
3184 !list_empty(&area->free_list[MIGRATE_HIGHATOMIC]))
3185 return true;
3186 }
3187 return false;
3188}
3189
3190bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3191 int classzone_idx, unsigned int alloc_flags)
3192{
3193 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3194 zone_page_state(z, NR_FREE_PAGES));
3195}
3196
3197static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3198 unsigned long mark, int classzone_idx, unsigned int alloc_flags)
3199{
3200 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3201 long cma_pages = 0;
3202
3203#ifdef CONFIG_CMA
3204 /* If allocation can't use CMA areas don't use free CMA pages */
3205 if (!(alloc_flags & ALLOC_CMA))
3206 cma_pages = zone_page_state(z, NR_FREE_CMA_PAGES);
3207#endif
3208
3209 /*
3210 * Fast check for order-0 only. If this fails then the reserves
3211 * need to be calculated. There is a corner case where the check
3212 * passes but only the high-order atomic reserve are free. If
3213 * the caller is !atomic then it'll uselessly search the free
3214 * list. That corner case is then slower but it is harmless.
3215 */
3216 if (!order && (free_pages - cma_pages) > mark + z->lowmem_reserve[classzone_idx])
3217 return true;
3218
3219 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
3220 free_pages);
3221}
3222
3223bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3224 unsigned long mark, int classzone_idx)
3225{
3226 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3227
3228 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3229 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3230
3231 return __zone_watermark_ok(z, order, mark, classzone_idx, 0,
3232 free_pages);
3233}
3234
3235#ifdef CONFIG_NUMA
3236static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3237{
3238 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3239 RECLAIM_DISTANCE;
3240}
3241#else /* CONFIG_NUMA */
3242static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3243{
3244 return true;
3245}
3246#endif /* CONFIG_NUMA */
3247
3248/*
3249 * get_page_from_freelist goes through the zonelist trying to allocate
3250 * a page.
3251 */
3252static struct page *
3253get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3254 const struct alloc_context *ac)
3255{
3256 struct zoneref *z = ac->preferred_zoneref;
3257 struct zone *zone;
3258 struct pglist_data *last_pgdat_dirty_limit = NULL;
3259
3260 /*
3261 * Scan zonelist, looking for a zone with enough free.
3262 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
3263 */
3264 for_next_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3265 ac->nodemask) {
3266 struct page *page;
3267 unsigned long mark;
3268
3269 if (cpusets_enabled() &&
3270 (alloc_flags & ALLOC_CPUSET) &&
3271 !__cpuset_zone_allowed(zone, gfp_mask))
3272 continue;
3273 /*
3274 * When allocating a page cache page for writing, we
3275 * want to get it from a node that is within its dirty
3276 * limit, such that no single node holds more than its
3277 * proportional share of globally allowed dirty pages.
3278 * The dirty limits take into account the node's
3279 * lowmem reserves and high watermark so that kswapd
3280 * should be able to balance it without having to
3281 * write pages from its LRU list.
3282 *
3283 * XXX: For now, allow allocations to potentially
3284 * exceed the per-node dirty limit in the slowpath
3285 * (spread_dirty_pages unset) before going into reclaim,
3286 * which is important when on a NUMA setup the allowed
3287 * nodes are together not big enough to reach the
3288 * global limit. The proper fix for these situations
3289 * will require awareness of nodes in the
3290 * dirty-throttling and the flusher threads.
3291 */
3292 if (ac->spread_dirty_pages) {
3293 if (last_pgdat_dirty_limit == zone->zone_pgdat)
3294 continue;
3295
3296 if (!node_dirty_ok(zone->zone_pgdat)) {
3297 last_pgdat_dirty_limit = zone->zone_pgdat;
3298 continue;
3299 }
3300 }
3301
3302 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
3303 if (!zone_watermark_fast(zone, order, mark,
3304 ac_classzone_idx(ac), alloc_flags)) {
3305 int ret;
3306
3307#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3308 /*
3309 * Watermark failed for this zone, but see if we can
3310 * grow this zone if it contains deferred pages.
3311 */
3312 if (static_branch_unlikely(&deferred_pages)) {
3313 if (_deferred_grow_zone(zone, order))
3314 goto try_this_zone;
3315 }
3316#endif
3317 /* Checked here to keep the fast path fast */
3318 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3319 if (alloc_flags & ALLOC_NO_WATERMARKS)
3320 goto try_this_zone;
3321
3322 if (node_reclaim_mode == 0 ||
3323 !zone_allows_reclaim(ac->preferred_zoneref->zone, zone))
3324 continue;
3325
3326 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3327 switch (ret) {
3328 case NODE_RECLAIM_NOSCAN:
3329 /* did not scan */
3330 continue;
3331 case NODE_RECLAIM_FULL:
3332 /* scanned but unreclaimable */
3333 continue;
3334 default:
3335 /* did we reclaim enough */
3336 if (zone_watermark_ok(zone, order, mark,
3337 ac_classzone_idx(ac), alloc_flags))
3338 goto try_this_zone;
3339
3340 continue;
3341 }
3342 }
3343
3344try_this_zone:
3345 page = rmqueue(ac->preferred_zoneref->zone, zone, order,
3346 gfp_mask, alloc_flags, ac->migratetype);
3347 if (page) {
3348 prep_new_page(page, order, gfp_mask, alloc_flags);
3349
3350 /*
3351 * If this is a high-order atomic allocation then check
3352 * if the pageblock should be reserved for the future
3353 */
3354 if (unlikely(order && (alloc_flags & ALLOC_HARDER)))
3355 reserve_highatomic_pageblock(page, zone, order);
3356
3357 return page;
3358 } else {
3359#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
3360 /* Try again if zone has deferred pages */
3361 if (static_branch_unlikely(&deferred_pages)) {
3362 if (_deferred_grow_zone(zone, order))
3363 goto try_this_zone;
3364 }
3365#endif
3366 }
3367 }
3368
3369 return NULL;
3370}
3371
3372/*
3373 * Large machines with many possible nodes should not always dump per-node
3374 * meminfo in irq context.
3375 */
3376static inline bool should_suppress_show_mem(void)
3377{
3378 bool ret = false;
3379
3380#if NODES_SHIFT > 8
3381 ret = in_interrupt();
3382#endif
3383 return ret;
3384}
3385
3386static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3387{
3388 unsigned int filter = SHOW_MEM_FILTER_NODES;
3389 static DEFINE_RATELIMIT_STATE(show_mem_rs, HZ, 1);
3390
3391 if (should_suppress_show_mem() || !__ratelimit(&show_mem_rs))
3392 return;
3393
3394 /*
3395 * This documents exceptions given to allocations in certain
3396 * contexts that are allowed to allocate outside current's set
3397 * of allowed nodes.
3398 */
3399 if (!(gfp_mask & __GFP_NOMEMALLOC))
3400 if (tsk_is_oom_victim(current) ||
3401 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3402 filter &= ~SHOW_MEM_FILTER_NODES;
3403 if (in_interrupt() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3404 filter &= ~SHOW_MEM_FILTER_NODES;
3405
3406 show_mem(filter, nodemask);
3407}
3408
3409void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3410{
3411 struct va_format vaf;
3412 va_list args;
3413 static DEFINE_RATELIMIT_STATE(nopage_rs, DEFAULT_RATELIMIT_INTERVAL,
3414 DEFAULT_RATELIMIT_BURST);
3415
3416 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
3417 return;
3418
3419 va_start(args, fmt);
3420 vaf.fmt = fmt;
3421 vaf.va = &args;
3422 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl\n",
3423 current->comm, &vaf, gfp_mask, &gfp_mask,
3424 nodemask_pr_args(nodemask));
3425 va_end(args);
3426
3427 cpuset_print_current_mems_allowed();
3428
3429 dump_stack();
3430 warn_alloc_show_mem(gfp_mask, nodemask);
3431}
3432
3433static inline struct page *
3434__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3435 unsigned int alloc_flags,
3436 const struct alloc_context *ac)
3437{
3438 struct page *page;
3439
3440 page = get_page_from_freelist(gfp_mask, order,
3441 alloc_flags|ALLOC_CPUSET, ac);
3442 /*
3443 * fallback to ignore cpuset restriction if our nodes
3444 * are depleted
3445 */
3446 if (!page)
3447 page = get_page_from_freelist(gfp_mask, order,
3448 alloc_flags, ac);
3449
3450 return page;
3451}
3452
3453static inline struct page *
3454__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3455 const struct alloc_context *ac, unsigned long *did_some_progress)
3456{
3457 struct oom_control oc = {
3458 .zonelist = ac->zonelist,
3459 .nodemask = ac->nodemask,
3460 .memcg = NULL,
3461 .gfp_mask = gfp_mask,
3462 .order = order,
3463 };
3464 struct page *page;
3465
3466 *did_some_progress = 0;
3467
3468 /*
3469 * Acquire the oom lock. If that fails, somebody else is
3470 * making progress for us.
3471 */
3472 if (!mutex_trylock(&oom_lock)) {
3473 *did_some_progress = 1;
3474 schedule_timeout_uninterruptible(1);
3475 return NULL;
3476 }
3477
3478 /*
3479 * Go through the zonelist yet one more time, keep very high watermark
3480 * here, this is only to catch a parallel oom killing, we must fail if
3481 * we're still under heavy pressure. But make sure that this reclaim
3482 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3483 * allocation which will never fail due to oom_lock already held.
3484 */
3485 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3486 ~__GFP_DIRECT_RECLAIM, order,
3487 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3488 if (page)
3489 goto out;
3490
3491 /* Coredumps can quickly deplete all memory reserves */
3492 if (current->flags & PF_DUMPCORE)
3493 goto out;
3494 /* The OOM killer will not help higher order allocs */
3495 if (order > PAGE_ALLOC_COSTLY_ORDER)
3496 goto out;
3497 /*
3498 * We have already exhausted all our reclaim opportunities without any
3499 * success so it is time to admit defeat. We will skip the OOM killer
3500 * because it is very likely that the caller has a more reasonable
3501 * fallback than shooting a random task.
3502 */
3503 if (gfp_mask & __GFP_RETRY_MAYFAIL)
3504 goto out;
3505 /* The OOM killer does not needlessly kill tasks for lowmem */
3506 if (ac->high_zoneidx < ZONE_NORMAL)
3507 goto out;
3508 if (pm_suspended_storage())
3509 goto out;
3510 /*
3511 * XXX: GFP_NOFS allocations should rather fail than rely on
3512 * other request to make a forward progress.
3513 * We are in an unfortunate situation where out_of_memory cannot
3514 * do much for this context but let's try it to at least get
3515 * access to memory reserved if the current task is killed (see
3516 * out_of_memory). Once filesystems are ready to handle allocation
3517 * failures more gracefully we should just bail out here.
3518 */
3519
3520 /* The OOM killer may not free memory on a specific node */
3521 if (gfp_mask & __GFP_THISNODE)
3522 goto out;
3523
3524 /* Exhausted what can be done so it's blame time */
3525 if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
3526 *did_some_progress = 1;
3527
3528 /*
3529 * Help non-failing allocations by giving them access to memory
3530 * reserves
3531 */
3532 if (gfp_mask & __GFP_NOFAIL)
3533 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3534 ALLOC_NO_WATERMARKS, ac);
3535 }
3536out:
3537 mutex_unlock(&oom_lock);
3538 return page;
3539}
3540
3541/*
3542 * Maximum number of compaction retries wit a progress before OOM
3543 * killer is consider as the only way to move forward.
3544 */
3545#define MAX_COMPACT_RETRIES 16
3546
3547#ifdef CONFIG_COMPACTION
3548/* Try memory compaction for high-order allocations before reclaim */
3549static struct page *
3550__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3551 unsigned int alloc_flags, const struct alloc_context *ac,
3552 enum compact_priority prio, enum compact_result *compact_result)
3553{
3554 struct page *page;
3555 unsigned int noreclaim_flag;
3556
3557 if (!order)
3558 return NULL;
3559
3560 noreclaim_flag = memalloc_noreclaim_save();
3561 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3562 prio);
3563 memalloc_noreclaim_restore(noreclaim_flag);
3564
3565 if (*compact_result <= COMPACT_INACTIVE)
3566 return NULL;
3567
3568 /*
3569 * At least in one zone compaction wasn't deferred or skipped, so let's
3570 * count a compaction stall
3571 */
3572 count_vm_event(COMPACTSTALL);
3573
3574 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3575
3576 if (page) {
3577 struct zone *zone = page_zone(page);
3578
3579 zone->compact_blockskip_flush = false;
3580 compaction_defer_reset(zone, order, true);
3581 count_vm_event(COMPACTSUCCESS);
3582 return page;
3583 }
3584
3585 /*
3586 * It's bad if compaction run occurs and fails. The most likely reason
3587 * is that pages exist, but not enough to satisfy watermarks.
3588 */
3589 count_vm_event(COMPACTFAIL);
3590
3591 cond_resched();
3592
3593 return NULL;
3594}
3595
3596static inline bool
3597should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3598 enum compact_result compact_result,
3599 enum compact_priority *compact_priority,
3600 int *compaction_retries)
3601{
3602 int max_retries = MAX_COMPACT_RETRIES;
3603 int min_priority;
3604 bool ret = false;
3605 int retries = *compaction_retries;
3606 enum compact_priority priority = *compact_priority;
3607
3608 if (!order)
3609 return false;
3610
3611 if (compaction_made_progress(compact_result))
3612 (*compaction_retries)++;
3613
3614 /*
3615 * compaction considers all the zone as desperately out of memory
3616 * so it doesn't really make much sense to retry except when the
3617 * failure could be caused by insufficient priority
3618 */
3619 if (compaction_failed(compact_result))
3620 goto check_priority;
3621
3622 /*
3623 * make sure the compaction wasn't deferred or didn't bail out early
3624 * due to locks contention before we declare that we should give up.
3625 * But do not retry if the given zonelist is not suitable for
3626 * compaction.
3627 */
3628 if (compaction_withdrawn(compact_result)) {
3629 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3630 goto out;
3631 }
3632
3633 /*
3634 * !costly requests are much more important than __GFP_RETRY_MAYFAIL
3635 * costly ones because they are de facto nofail and invoke OOM
3636 * killer to move on while costly can fail and users are ready
3637 * to cope with that. 1/4 retries is rather arbitrary but we
3638 * would need much more detailed feedback from compaction to
3639 * make a better decision.
3640 */
3641 if (order > PAGE_ALLOC_COSTLY_ORDER)
3642 max_retries /= 4;
3643 if (*compaction_retries <= max_retries) {
3644 ret = true;
3645 goto out;
3646 }
3647
3648 /*
3649 * Make sure there are attempts at the highest priority if we exhausted
3650 * all retries or failed at the lower priorities.
3651 */
3652check_priority:
3653 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3654 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3655
3656 if (*compact_priority > min_priority) {
3657 (*compact_priority)--;
3658 *compaction_retries = 0;
3659 ret = true;
3660 }
3661out:
3662 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3663 return ret;
3664}
3665#else
3666static inline struct page *
3667__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3668 unsigned int alloc_flags, const struct alloc_context *ac,
3669 enum compact_priority prio, enum compact_result *compact_result)
3670{
3671 *compact_result = COMPACT_SKIPPED;
3672 return NULL;
3673}
3674
3675static inline bool
3676should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3677 enum compact_result compact_result,
3678 enum compact_priority *compact_priority,
3679 int *compaction_retries)
3680{
3681 struct zone *zone;
3682 struct zoneref *z;
3683
3684 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3685 return false;
3686
3687 /*
3688 * There are setups with compaction disabled which would prefer to loop
3689 * inside the allocator rather than hit the oom killer prematurely.
3690 * Let's give them a good hope and keep retrying while the order-0
3691 * watermarks are OK.
3692 */
3693 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3694 ac->nodemask) {
3695 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3696 ac_classzone_idx(ac), alloc_flags))
3697 return true;
3698 }
3699 return false;
3700}
3701#endif /* CONFIG_COMPACTION */
3702
3703#ifdef CONFIG_LOCKDEP
3704struct lockdep_map __fs_reclaim_map =
3705 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3706
3707static bool __need_fs_reclaim(gfp_t gfp_mask)
3708{
3709 gfp_mask = current_gfp_context(gfp_mask);
3710
3711 /* no reclaim without waiting on it */
3712 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3713 return false;
3714
3715 /* this guy won't enter reclaim */
3716 if (current->flags & PF_MEMALLOC)
3717 return false;
3718
3719 /* We're only interested __GFP_FS allocations for now */
3720 if (!(gfp_mask & __GFP_FS))
3721 return false;
3722
3723 if (gfp_mask & __GFP_NOLOCKDEP)
3724 return false;
3725
3726 return true;
3727}
3728
3729void fs_reclaim_acquire(gfp_t gfp_mask)
3730{
3731 if (__need_fs_reclaim(gfp_mask))
3732 lock_map_acquire(&__fs_reclaim_map);
3733}
3734EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3735
3736void fs_reclaim_release(gfp_t gfp_mask)
3737{
3738 if (__need_fs_reclaim(gfp_mask))
3739 lock_map_release(&__fs_reclaim_map);
3740}
3741EXPORT_SYMBOL_GPL(fs_reclaim_release);
3742#endif
3743
3744/* Perform direct synchronous page reclaim */
3745static int
3746__perform_reclaim(gfp_t gfp_mask, unsigned int order,
3747 const struct alloc_context *ac)
3748{
3749 struct reclaim_state reclaim_state;
3750 int progress;
3751 unsigned int noreclaim_flag;
3752
3753 cond_resched();
3754
3755 /* We now go into synchronous reclaim */
3756 cpuset_memory_pressure_bump();
3757 noreclaim_flag = memalloc_noreclaim_save();
3758 fs_reclaim_acquire(gfp_mask);
3759 reclaim_state.reclaimed_slab = 0;
3760 current->reclaim_state = &reclaim_state;
3761
3762 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3763 ac->nodemask);
3764
3765 current->reclaim_state = NULL;
3766 fs_reclaim_release(gfp_mask);
3767 memalloc_noreclaim_restore(noreclaim_flag);
3768
3769 cond_resched();
3770
3771 return progress;
3772}
3773
3774/* The really slow allocator path where we enter direct reclaim */
3775static inline struct page *
3776__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3777 unsigned int alloc_flags, const struct alloc_context *ac,
3778 unsigned long *did_some_progress)
3779{
3780 struct page *page = NULL;
3781 bool drained = false;
3782
3783 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3784 if (unlikely(!(*did_some_progress)))
3785 return NULL;
3786
3787retry:
3788 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3789
3790 /*
3791 * If an allocation failed after direct reclaim, it could be because
3792 * pages are pinned on the per-cpu lists or in high alloc reserves.
3793 * Shrink them them and try again
3794 */
3795 if (!page && !drained) {
3796 unreserve_highatomic_pageblock(ac, false);
3797 drain_all_pages(NULL);
3798 drained = true;
3799 goto retry;
3800 }
3801
3802 return page;
3803}
3804
3805static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3806 const struct alloc_context *ac)
3807{
3808 struct zoneref *z;
3809 struct zone *zone;
3810 pg_data_t *last_pgdat = NULL;
3811 enum zone_type high_zoneidx = ac->high_zoneidx;
3812
3813 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, high_zoneidx,
3814 ac->nodemask) {
3815 if (last_pgdat != zone->zone_pgdat)
3816 wakeup_kswapd(zone, gfp_mask, order, high_zoneidx);
3817 last_pgdat = zone->zone_pgdat;
3818 }
3819}
3820
3821static inline unsigned int
3822gfp_to_alloc_flags(gfp_t gfp_mask)
3823{
3824 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3825
3826 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
3827 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
3828
3829 /*
3830 * The caller may dip into page reserves a bit more if the caller
3831 * cannot run direct reclaim, or if the caller has realtime scheduling
3832 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
3833 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
3834 */
3835 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
3836
3837 if (gfp_mask & __GFP_ATOMIC) {
3838 /*
3839 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
3840 * if it can't schedule.
3841 */
3842 if (!(gfp_mask & __GFP_NOMEMALLOC))
3843 alloc_flags |= ALLOC_HARDER;
3844 /*
3845 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
3846 * comment for __cpuset_node_allowed().
3847 */
3848 alloc_flags &= ~ALLOC_CPUSET;
3849 } else if (unlikely(rt_task(current)) && !in_interrupt())
3850 alloc_flags |= ALLOC_HARDER;
3851
3852#ifdef CONFIG_CMA
3853 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3854 alloc_flags |= ALLOC_CMA;
3855#endif
3856 return alloc_flags;
3857}
3858
3859static bool oom_reserves_allowed(struct task_struct *tsk)
3860{
3861 if (!tsk_is_oom_victim(tsk))
3862 return false;
3863
3864 /*
3865 * !MMU doesn't have oom reaper so give access to memory reserves
3866 * only to the thread with TIF_MEMDIE set
3867 */
3868 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
3869 return false;
3870
3871 return true;
3872}
3873
3874/*
3875 * Distinguish requests which really need access to full memory
3876 * reserves from oom victims which can live with a portion of it
3877 */
3878static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
3879{
3880 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
3881 return 0;
3882 if (gfp_mask & __GFP_MEMALLOC)
3883 return ALLOC_NO_WATERMARKS;
3884 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
3885 return ALLOC_NO_WATERMARKS;
3886 if (!in_interrupt()) {
3887 if (current->flags & PF_MEMALLOC)
3888 return ALLOC_NO_WATERMARKS;
3889 else if (oom_reserves_allowed(current))
3890 return ALLOC_OOM;
3891 }
3892
3893 return 0;
3894}
3895
3896bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
3897{
3898 return !!__gfp_pfmemalloc_flags(gfp_mask);
3899}
3900
3901/*
3902 * Checks whether it makes sense to retry the reclaim to make a forward progress
3903 * for the given allocation request.
3904 *
3905 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
3906 * without success, or when we couldn't even meet the watermark if we
3907 * reclaimed all remaining pages on the LRU lists.
3908 *
3909 * Returns true if a retry is viable or false to enter the oom path.
3910 */
3911static inline bool
3912should_reclaim_retry(gfp_t gfp_mask, unsigned order,
3913 struct alloc_context *ac, int alloc_flags,
3914 bool did_some_progress, int *no_progress_loops)
3915{
3916 struct zone *zone;
3917 struct zoneref *z;
3918
3919 /*
3920 * Costly allocations might have made a progress but this doesn't mean
3921 * their order will become available due to high fragmentation so
3922 * always increment the no progress counter for them
3923 */
3924 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
3925 *no_progress_loops = 0;
3926 else
3927 (*no_progress_loops)++;
3928
3929 /*
3930 * Make sure we converge to OOM if we cannot make any progress
3931 * several times in the row.
3932 */
3933 if (*no_progress_loops > MAX_RECLAIM_RETRIES) {
3934 /* Before OOM, exhaust highatomic_reserve */
3935 return unreserve_highatomic_pageblock(ac, true);
3936 }
3937
3938 /*
3939 * Keep reclaiming pages while there is a chance this will lead
3940 * somewhere. If none of the target zones can satisfy our allocation
3941 * request even if all reclaimable pages are considered then we are
3942 * screwed and have to go OOM.
3943 */
3944 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, ac->high_zoneidx,
3945 ac->nodemask) {
3946 unsigned long available;
3947 unsigned long reclaimable;
3948 unsigned long min_wmark = min_wmark_pages(zone);
3949 bool wmark;
3950
3951 available = reclaimable = zone_reclaimable_pages(zone);
3952 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
3953
3954 /*
3955 * Would the allocation succeed if we reclaimed all
3956 * reclaimable pages?
3957 */
3958 wmark = __zone_watermark_ok(zone, order, min_wmark,
3959 ac_classzone_idx(ac), alloc_flags, available);
3960 trace_reclaim_retry_zone(z, order, reclaimable,
3961 available, min_wmark, *no_progress_loops, wmark);
3962 if (wmark) {
3963 /*
3964 * If we didn't make any progress and have a lot of
3965 * dirty + writeback pages then we should wait for
3966 * an IO to complete to slow down the reclaim and
3967 * prevent from pre mature OOM
3968 */
3969 if (!did_some_progress) {
3970 unsigned long write_pending;
3971
3972 write_pending = zone_page_state_snapshot(zone,
3973 NR_ZONE_WRITE_PENDING);
3974
3975 if (2 * write_pending > reclaimable) {
3976 congestion_wait(BLK_RW_ASYNC, HZ/10);
3977 return true;
3978 }
3979 }
3980
3981 /*
3982 * Memory allocation/reclaim might be called from a WQ
3983 * context and the current implementation of the WQ
3984 * concurrency control doesn't recognize that
3985 * a particular WQ is congested if the worker thread is
3986 * looping without ever sleeping. Therefore we have to
3987 * do a short sleep here rather than calling
3988 * cond_resched().
3989 */
3990 if (current->flags & PF_WQ_WORKER)
3991 schedule_timeout_uninterruptible(1);
3992 else
3993 cond_resched();
3994
3995 return true;
3996 }
3997 }
3998
3999 return false;
4000}
4001
4002static inline bool
4003check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4004{
4005 /*
4006 * It's possible that cpuset's mems_allowed and the nodemask from
4007 * mempolicy don't intersect. This should be normally dealt with by
4008 * policy_nodemask(), but it's possible to race with cpuset update in
4009 * such a way the check therein was true, and then it became false
4010 * before we got our cpuset_mems_cookie here.
4011 * This assumes that for all allocations, ac->nodemask can come only
4012 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4013 * when it does not intersect with the cpuset restrictions) or the
4014 * caller can deal with a violated nodemask.
4015 */
4016 if (cpusets_enabled() && ac->nodemask &&
4017 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4018 ac->nodemask = NULL;
4019 return true;
4020 }
4021
4022 /*
4023 * When updating a task's mems_allowed or mempolicy nodemask, it is
4024 * possible to race with parallel threads in such a way that our
4025 * allocation can fail while the mask is being updated. If we are about
4026 * to fail, check if the cpuset changed during allocation and if so,
4027 * retry.
4028 */
4029 if (read_mems_allowed_retry(cpuset_mems_cookie))
4030 return true;
4031
4032 return false;
4033}
4034
4035static inline struct page *
4036__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4037 struct alloc_context *ac)
4038{
4039 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4040 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4041 struct page *page = NULL;
4042 unsigned int alloc_flags;
4043 unsigned long did_some_progress;
4044 enum compact_priority compact_priority;
4045 enum compact_result compact_result;
4046 int compaction_retries;
4047 int no_progress_loops;
4048 unsigned int cpuset_mems_cookie;
4049 int reserve_flags;
4050
4051 /*
4052 * In the slowpath, we sanity check order to avoid ever trying to
4053 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
4054 * be using allocators in order of preference for an area that is
4055 * too large.
4056 */
4057 if (order >= MAX_ORDER) {
4058 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
4059 return NULL;
4060 }
4061
4062 /*
4063 * We also sanity check to catch abuse of atomic reserves being used by
4064 * callers that are not in atomic context.
4065 */
4066 if (WARN_ON_ONCE((gfp_mask & (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)) ==
4067 (__GFP_ATOMIC|__GFP_DIRECT_RECLAIM)))
4068 gfp_mask &= ~__GFP_ATOMIC;
4069
4070retry_cpuset:
4071 compaction_retries = 0;
4072 no_progress_loops = 0;
4073 compact_priority = DEF_COMPACT_PRIORITY;
4074 cpuset_mems_cookie = read_mems_allowed_begin();
4075
4076 /*
4077 * The fast path uses conservative alloc_flags to succeed only until
4078 * kswapd needs to be woken up, and to avoid the cost of setting up
4079 * alloc_flags precisely. So we do that now.
4080 */
4081 alloc_flags = gfp_to_alloc_flags(gfp_mask);
4082
4083 /*
4084 * We need to recalculate the starting point for the zonelist iterator
4085 * because we might have used different nodemask in the fast path, or
4086 * there was a cpuset modification and we are retrying - otherwise we
4087 * could end up iterating over non-eligible zones endlessly.
4088 */
4089 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4090 ac->high_zoneidx, ac->nodemask);
4091 if (!ac->preferred_zoneref->zone)
4092 goto nopage;
4093
4094 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4095 wake_all_kswapds(order, gfp_mask, ac);
4096
4097 /*
4098 * The adjusted alloc_flags might result in immediate success, so try
4099 * that first
4100 */
4101 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4102 if (page)
4103 goto got_pg;
4104
4105 /*
4106 * For costly allocations, try direct compaction first, as it's likely
4107 * that we have enough base pages and don't need to reclaim. For non-
4108 * movable high-order allocations, do that as well, as compaction will
4109 * try prevent permanent fragmentation by migrating from blocks of the
4110 * same migratetype.
4111 * Don't try this for allocations that are allowed to ignore
4112 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4113 */
4114 if (can_direct_reclaim &&
4115 (costly_order ||
4116 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4117 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4118 page = __alloc_pages_direct_compact(gfp_mask, order,
4119 alloc_flags, ac,
4120 INIT_COMPACT_PRIORITY,
4121 &compact_result);
4122 if (page)
4123 goto got_pg;
4124
4125 /*
4126 * Checks for costly allocations with __GFP_NORETRY, which
4127 * includes THP page fault allocations
4128 */
4129 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4130 /*
4131 * If compaction is deferred for high-order allocations,
4132 * it is because sync compaction recently failed. If
4133 * this is the case and the caller requested a THP
4134 * allocation, we do not want to heavily disrupt the
4135 * system, so we fail the allocation instead of entering
4136 * direct reclaim.
4137 */
4138 if (compact_result == COMPACT_DEFERRED)
4139 goto nopage;
4140
4141 /*
4142 * Looks like reclaim/compaction is worth trying, but
4143 * sync compaction could be very expensive, so keep
4144 * using async compaction.
4145 */
4146 compact_priority = INIT_COMPACT_PRIORITY;
4147 }
4148 }
4149
4150retry:
4151 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4152 if (gfp_mask & __GFP_KSWAPD_RECLAIM)
4153 wake_all_kswapds(order, gfp_mask, ac);
4154
4155 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4156 if (reserve_flags)
4157 alloc_flags = reserve_flags;
4158
4159 /*
4160 * Reset the zonelist iterators if memory policies can be ignored.
4161 * These allocations are high priority and system rather than user
4162 * orientated.
4163 */
4164 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4165 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
4166 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4167 ac->high_zoneidx, ac->nodemask);
4168 }
4169
4170 /* Attempt with potentially adjusted zonelist and alloc_flags */
4171 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4172 if (page)
4173 goto got_pg;
4174
4175 /* Caller is not willing to reclaim, we can't balance anything */
4176 if (!can_direct_reclaim)
4177 goto nopage;
4178
4179 /* Avoid recursion of direct reclaim */
4180 if (current->flags & PF_MEMALLOC)
4181 goto nopage;
4182
4183 /* Try direct reclaim and then allocating */
4184 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4185 &did_some_progress);
4186 if (page)
4187 goto got_pg;
4188
4189 /* Try direct compaction and then allocating */
4190 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4191 compact_priority, &compact_result);
4192 if (page)
4193 goto got_pg;
4194
4195 /* Do not loop if specifically requested */
4196 if (gfp_mask & __GFP_NORETRY)
4197 goto nopage;
4198
4199 /*
4200 * Do not retry costly high order allocations unless they are
4201 * __GFP_RETRY_MAYFAIL
4202 */
4203 if (costly_order && !(gfp_mask & __GFP_RETRY_MAYFAIL))
4204 goto nopage;
4205
4206 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4207 did_some_progress > 0, &no_progress_loops))
4208 goto retry;
4209
4210 /*
4211 * It doesn't make any sense to retry for the compaction if the order-0
4212 * reclaim is not able to make any progress because the current
4213 * implementation of the compaction depends on the sufficient amount
4214 * of free memory (see __compaction_suitable)
4215 */
4216 if (did_some_progress > 0 &&
4217 should_compact_retry(ac, order, alloc_flags,
4218 compact_result, &compact_priority,
4219 &compaction_retries))
4220 goto retry;
4221
4222
4223 /* Deal with possible cpuset update races before we start OOM killing */
4224 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4225 goto retry_cpuset;
4226
4227 /* Reclaim has failed us, start killing things */
4228 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4229 if (page)
4230 goto got_pg;
4231
4232 /* Avoid allocations with no watermarks from looping endlessly */
4233 if (tsk_is_oom_victim(current) &&
4234 (alloc_flags == ALLOC_OOM ||
4235 (gfp_mask & __GFP_NOMEMALLOC)))
4236 goto nopage;
4237
4238 /* Retry as long as the OOM killer is making progress */
4239 if (did_some_progress) {
4240 no_progress_loops = 0;
4241 goto retry;
4242 }
4243
4244nopage:
4245 /* Deal with possible cpuset update races before we fail */
4246 if (check_retry_cpuset(cpuset_mems_cookie, ac))
4247 goto retry_cpuset;
4248
4249 /*
4250 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4251 * we always retry
4252 */
4253 if (gfp_mask & __GFP_NOFAIL) {
4254 /*
4255 * All existing users of the __GFP_NOFAIL are blockable, so warn
4256 * of any new users that actually require GFP_NOWAIT
4257 */
4258 if (WARN_ON_ONCE(!can_direct_reclaim))
4259 goto fail;
4260
4261 /*
4262 * PF_MEMALLOC request from this context is rather bizarre
4263 * because we cannot reclaim anything and only can loop waiting
4264 * for somebody to do a work for us
4265 */
4266 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4267
4268 /*
4269 * non failing costly orders are a hard requirement which we
4270 * are not prepared for much so let's warn about these users
4271 * so that we can identify them and convert them to something
4272 * else.
4273 */
4274 WARN_ON_ONCE(order > PAGE_ALLOC_COSTLY_ORDER);
4275
4276 /*
4277 * Help non-failing allocations by giving them access to memory
4278 * reserves but do not use ALLOC_NO_WATERMARKS because this
4279 * could deplete whole memory reserves which would just make
4280 * the situation worse
4281 */
4282 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_HARDER, ac);
4283 if (page)
4284 goto got_pg;
4285
4286 cond_resched();
4287 goto retry;
4288 }
4289fail:
4290 warn_alloc(gfp_mask, ac->nodemask,
4291 "page allocation failure: order:%u", order);
4292got_pg:
4293 return page;
4294}
4295
4296static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4297 int preferred_nid, nodemask_t *nodemask,
4298 struct alloc_context *ac, gfp_t *alloc_mask,
4299 unsigned int *alloc_flags)
4300{
4301 ac->high_zoneidx = gfp_zone(gfp_mask);
4302 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4303 ac->nodemask = nodemask;
4304 ac->migratetype = gfpflags_to_migratetype(gfp_mask);
4305
4306 if (cpusets_enabled()) {
4307 *alloc_mask |= __GFP_HARDWALL;
4308 if (!ac->nodemask)
4309 ac->nodemask = &cpuset_current_mems_allowed;
4310 else
4311 *alloc_flags |= ALLOC_CPUSET;
4312 }
4313
4314 fs_reclaim_acquire(gfp_mask);
4315 fs_reclaim_release(gfp_mask);
4316
4317 might_sleep_if(gfp_mask & __GFP_DIRECT_RECLAIM);
4318
4319 if (should_fail_alloc_page(gfp_mask, order))
4320 return false;
4321
4322 if (IS_ENABLED(CONFIG_CMA) && ac->migratetype == MIGRATE_MOVABLE)
4323 *alloc_flags |= ALLOC_CMA;
4324
4325 return true;
4326}
4327
4328/* Determine whether to spread dirty pages and what the first usable zone */
4329static inline void finalise_ac(gfp_t gfp_mask,
4330 unsigned int order, struct alloc_context *ac)
4331{
4332 /* Dirty zone balancing only done in the fast path */
4333 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4334
4335 /*
4336 * The preferred zone is used for statistics but crucially it is
4337 * also used as the starting point for the zonelist iterator. It
4338 * may get reset for allocations that ignore memory policies.
4339 */
4340 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4341 ac->high_zoneidx, ac->nodemask);
4342}
4343
4344/*
4345 * This is the 'heart' of the zoned buddy allocator.
4346 */
4347struct page *
4348__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order, int preferred_nid,
4349 nodemask_t *nodemask)
4350{
4351 struct page *page;
4352 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4353 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
4354 struct alloc_context ac = { };
4355
4356 gfp_mask &= gfp_allowed_mask;
4357 alloc_mask = gfp_mask;
4358 if (!prepare_alloc_pages(gfp_mask, order, preferred_nid, nodemask, &ac, &alloc_mask, &alloc_flags))
4359 return NULL;
4360
4361 finalise_ac(gfp_mask, order, &ac);
4362
4363 /* First allocation attempt */
4364 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
4365 if (likely(page))
4366 goto out;
4367
4368 /*
4369 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4370 * resp. GFP_NOIO which has to be inherited for all allocation requests
4371 * from a particular context which has been marked by
4372 * memalloc_no{fs,io}_{save,restore}.
4373 */
4374 alloc_mask = current_gfp_context(gfp_mask);
4375 ac.spread_dirty_pages = false;
4376
4377 /*
4378 * Restore the original nodemask if it was potentially replaced with
4379 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4380 */
4381 if (unlikely(ac.nodemask != nodemask))
4382 ac.nodemask = nodemask;
4383
4384 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
4385
4386out:
4387 if (memcg_kmem_enabled() && (gfp_mask & __GFP_ACCOUNT) && page &&
4388 unlikely(memcg_kmem_charge(page, gfp_mask, order) != 0)) {
4389 __free_pages(page, order);
4390 page = NULL;
4391 }
4392
4393 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
4394
4395 return page;
4396}
4397EXPORT_SYMBOL(__alloc_pages_nodemask);
4398
4399/*
4400 * Common helper functions.
4401 */
4402unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
4403{
4404 struct page *page;
4405
4406 /*
4407 * __get_free_pages() returns a virtual address, which cannot represent
4408 * a highmem page
4409 */
4410 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
4411
4412 page = alloc_pages(gfp_mask, order);
4413 if (!page)
4414 return 0;
4415 return (unsigned long) page_address(page);
4416}
4417EXPORT_SYMBOL(__get_free_pages);
4418
4419unsigned long get_zeroed_page(gfp_t gfp_mask)
4420{
4421 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
4422}
4423EXPORT_SYMBOL(get_zeroed_page);
4424
4425void __free_pages(struct page *page, unsigned int order)
4426{
4427 if (put_page_testzero(page)) {
4428 if (order == 0)
4429 free_unref_page(page);
4430 else
4431 __free_pages_ok(page, order);
4432 }
4433}
4434
4435EXPORT_SYMBOL(__free_pages);
4436
4437void free_pages(unsigned long addr, unsigned int order)
4438{
4439 if (addr != 0) {
4440 VM_BUG_ON(!virt_addr_valid((void *)addr));
4441 __free_pages(virt_to_page((void *)addr), order);
4442 }
4443}
4444
4445EXPORT_SYMBOL(free_pages);
4446
4447/*
4448 * Page Fragment:
4449 * An arbitrary-length arbitrary-offset area of memory which resides
4450 * within a 0 or higher order page. Multiple fragments within that page
4451 * are individually refcounted, in the page's reference counter.
4452 *
4453 * The page_frag functions below provide a simple allocation framework for
4454 * page fragments. This is used by the network stack and network device
4455 * drivers to provide a backing region of memory for use as either an
4456 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
4457 */
4458static struct page *__page_frag_cache_refill(struct page_frag_cache *nc,
4459 gfp_t gfp_mask)
4460{
4461 struct page *page = NULL;
4462 gfp_t gfp = gfp_mask;
4463
4464#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4465 gfp_mask |= __GFP_COMP | __GFP_NOWARN | __GFP_NORETRY |
4466 __GFP_NOMEMALLOC;
4467 page = alloc_pages_node(NUMA_NO_NODE, gfp_mask,
4468 PAGE_FRAG_CACHE_MAX_ORDER);
4469 nc->size = page ? PAGE_FRAG_CACHE_MAX_SIZE : PAGE_SIZE;
4470#endif
4471 if (unlikely(!page))
4472 page = alloc_pages_node(NUMA_NO_NODE, gfp, 0);
4473
4474 nc->va = page ? page_address(page) : NULL;
4475
4476 return page;
4477}
4478
4479void __page_frag_cache_drain(struct page *page, unsigned int count)
4480{
4481 VM_BUG_ON_PAGE(page_ref_count(page) == 0, page);
4482
4483 if (page_ref_sub_and_test(page, count)) {
4484 unsigned int order = compound_order(page);
4485
4486 if (order == 0)
4487 free_unref_page(page);
4488 else
4489 __free_pages_ok(page, order);
4490 }
4491}
4492EXPORT_SYMBOL(__page_frag_cache_drain);
4493
4494void *page_frag_alloc(struct page_frag_cache *nc,
4495 unsigned int fragsz, gfp_t gfp_mask)
4496{
4497 unsigned int size = PAGE_SIZE;
4498 struct page *page;
4499 int offset;
4500
4501 if (unlikely(!nc->va)) {
4502refill:
4503 page = __page_frag_cache_refill(nc, gfp_mask);
4504 if (!page)
4505 return NULL;
4506
4507#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4508 /* if size can vary use size else just use PAGE_SIZE */
4509 size = nc->size;
4510#endif
4511 /* Even if we own the page, we do not use atomic_set().
4512 * This would break get_page_unless_zero() users.
4513 */
4514 page_ref_add(page, size - 1);
4515
4516 /* reset page count bias and offset to start of new frag */
4517 nc->pfmemalloc = page_is_pfmemalloc(page);
4518 nc->pagecnt_bias = size;
4519 nc->offset = size;
4520 }
4521
4522 offset = nc->offset - fragsz;
4523 if (unlikely(offset < 0)) {
4524 page = virt_to_page(nc->va);
4525
4526 if (!page_ref_sub_and_test(page, nc->pagecnt_bias))
4527 goto refill;
4528
4529#if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
4530 /* if size can vary use size else just use PAGE_SIZE */
4531 size = nc->size;
4532#endif
4533 /* OK, page count is 0, we can safely set it */
4534 set_page_count(page, size);
4535
4536 /* reset page count bias and offset to start of new frag */
4537 nc->pagecnt_bias = size;
4538 offset = size - fragsz;
4539 }
4540
4541 nc->pagecnt_bias--;
4542 nc->offset = offset;
4543
4544 return nc->va + offset;
4545}
4546EXPORT_SYMBOL(page_frag_alloc);
4547
4548/*
4549 * Frees a page fragment allocated out of either a compound or order 0 page.
4550 */
4551void page_frag_free(void *addr)
4552{
4553 struct page *page = virt_to_head_page(addr);
4554
4555 if (unlikely(put_page_testzero(page)))
4556 __free_pages_ok(page, compound_order(page));
4557}
4558EXPORT_SYMBOL(page_frag_free);
4559
4560static void *make_alloc_exact(unsigned long addr, unsigned int order,
4561 size_t size)
4562{
4563 if (addr) {
4564 unsigned long alloc_end = addr + (PAGE_SIZE << order);
4565 unsigned long used = addr + PAGE_ALIGN(size);
4566
4567 split_page(virt_to_page((void *)addr), order);
4568 while (used < alloc_end) {
4569 free_page(used);
4570 used += PAGE_SIZE;
4571 }
4572 }
4573 return (void *)addr;
4574}
4575
4576/**
4577 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4578 * @size: the number of bytes to allocate
4579 * @gfp_mask: GFP flags for the allocation
4580 *
4581 * This function is similar to alloc_pages(), except that it allocates the
4582 * minimum number of pages to satisfy the request. alloc_pages() can only
4583 * allocate memory in power-of-two pages.
4584 *
4585 * This function is also limited by MAX_ORDER.
4586 *
4587 * Memory allocated by this function must be released by free_pages_exact().
4588 */
4589void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
4590{
4591 unsigned int order = get_order(size);
4592 unsigned long addr;
4593
4594 addr = __get_free_pages(gfp_mask, order);
4595 return make_alloc_exact(addr, order, size);
4596}
4597EXPORT_SYMBOL(alloc_pages_exact);
4598
4599/**
4600 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4601 * pages on a node.
4602 * @nid: the preferred node ID where memory should be allocated
4603 * @size: the number of bytes to allocate
4604 * @gfp_mask: GFP flags for the allocation
4605 *
4606 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4607 * back.
4608 */
4609void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
4610{
4611 unsigned int order = get_order(size);
4612 struct page *p = alloc_pages_node(nid, gfp_mask, order);
4613 if (!p)
4614 return NULL;
4615 return make_alloc_exact((unsigned long)page_address(p), order, size);
4616}
4617
4618/**
4619 * free_pages_exact - release memory allocated via alloc_pages_exact()
4620 * @virt: the value returned by alloc_pages_exact.
4621 * @size: size of allocation, same value as passed to alloc_pages_exact().
4622 *
4623 * Release the memory allocated by a previous call to alloc_pages_exact.
4624 */
4625void free_pages_exact(void *virt, size_t size)
4626{
4627 unsigned long addr = (unsigned long)virt;
4628 unsigned long end = addr + PAGE_ALIGN(size);
4629
4630 while (addr < end) {
4631 free_page(addr);
4632 addr += PAGE_SIZE;
4633 }
4634}
4635EXPORT_SYMBOL(free_pages_exact);
4636
4637/**
4638 * nr_free_zone_pages - count number of pages beyond high watermark
4639 * @offset: The zone index of the highest zone
4640 *
4641 * nr_free_zone_pages() counts the number of counts pages which are beyond the
4642 * high watermark within all zones at or below a given zone index. For each
4643 * zone, the number of pages is calculated as:
4644 *
4645 * nr_free_zone_pages = managed_pages - high_pages
4646 */
4647static unsigned long nr_free_zone_pages(int offset)
4648{
4649 struct zoneref *z;
4650 struct zone *zone;
4651
4652 /* Just pick one node, since fallback list is circular */
4653 unsigned long sum = 0;
4654
4655 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4656
4657 for_each_zone_zonelist(zone, z, zonelist, offset) {
4658 unsigned long size = zone->managed_pages;
4659 unsigned long high = high_wmark_pages(zone);
4660 if (size > high)
4661 sum += size - high;
4662 }
4663
4664 return sum;
4665}
4666
4667/**
4668 * nr_free_buffer_pages - count number of pages beyond high watermark
4669 *
4670 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4671 * watermark within ZONE_DMA and ZONE_NORMAL.
4672 */
4673unsigned long nr_free_buffer_pages(void)
4674{
4675 return nr_free_zone_pages(gfp_zone(GFP_USER));
4676}
4677EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
4678
4679/**
4680 * nr_free_pagecache_pages - count number of pages beyond high watermark
4681 *
4682 * nr_free_pagecache_pages() counts the number of pages which are beyond the
4683 * high watermark within all zones.
4684 */
4685unsigned long nr_free_pagecache_pages(void)
4686{
4687 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
4688}
4689
4690static inline void show_node(struct zone *zone)
4691{
4692 if (IS_ENABLED(CONFIG_NUMA))
4693 printk("Node %d ", zone_to_nid(zone));
4694}
4695
4696long si_mem_available(void)
4697{
4698 long available;
4699 unsigned long pagecache;
4700 unsigned long wmark_low = 0;
4701 unsigned long pages[NR_LRU_LISTS];
4702 struct zone *zone;
4703 int lru;
4704
4705 for (lru = LRU_BASE; lru < NR_LRU_LISTS; lru++)
4706 pages[lru] = global_node_page_state(NR_LRU_BASE + lru);
4707
4708 for_each_zone(zone)
4709 wmark_low += zone->watermark[WMARK_LOW];
4710
4711 /*
4712 * Estimate the amount of memory available for userspace allocations,
4713 * without causing swapping.
4714 */
4715 available = global_zone_page_state(NR_FREE_PAGES) - totalreserve_pages;
4716
4717 /*
4718 * Not all the page cache can be freed, otherwise the system will
4719 * start swapping. Assume at least half of the page cache, or the
4720 * low watermark worth of cache, needs to stay.
4721 */
4722 pagecache = pages[LRU_ACTIVE_FILE] + pages[LRU_INACTIVE_FILE];
4723 pagecache -= min(pagecache / 2, wmark_low);
4724 available += pagecache;
4725
4726 /*
4727 * Part of the reclaimable slab consists of items that are in use,
4728 * and cannot be freed. Cap this estimate at the low watermark.
4729 */
4730 available += global_node_page_state(NR_SLAB_RECLAIMABLE) -
4731 min(global_node_page_state(NR_SLAB_RECLAIMABLE) / 2,
4732 wmark_low);
4733
4734 /*
4735 * Part of the kernel memory, which can be released under memory
4736 * pressure.
4737 */
4738 available += global_node_page_state(NR_INDIRECTLY_RECLAIMABLE_BYTES) >>
4739 PAGE_SHIFT;
4740
4741 if (available < 0)
4742 available = 0;
4743 return available;
4744}
4745EXPORT_SYMBOL_GPL(si_mem_available);
4746
4747void si_meminfo(struct sysinfo *val)
4748{
4749 val->totalram = totalram_pages;
4750 val->sharedram = global_node_page_state(NR_SHMEM);
4751 val->freeram = global_zone_page_state(NR_FREE_PAGES);
4752 val->bufferram = nr_blockdev_pages();
4753 val->totalhigh = totalhigh_pages;
4754 val->freehigh = nr_free_highpages();
4755 val->mem_unit = PAGE_SIZE;
4756}
4757
4758EXPORT_SYMBOL(si_meminfo);
4759
4760#ifdef CONFIG_NUMA
4761void si_meminfo_node(struct sysinfo *val, int nid)
4762{
4763 int zone_type; /* needs to be signed */
4764 unsigned long managed_pages = 0;
4765 unsigned long managed_highpages = 0;
4766 unsigned long free_highpages = 0;
4767 pg_data_t *pgdat = NODE_DATA(nid);
4768
4769 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
4770 managed_pages += pgdat->node_zones[zone_type].managed_pages;
4771 val->totalram = managed_pages;
4772 val->sharedram = node_page_state(pgdat, NR_SHMEM);
4773 val->freeram = sum_zone_node_page_state(nid, NR_FREE_PAGES);
4774#ifdef CONFIG_HIGHMEM
4775 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
4776 struct zone *zone = &pgdat->node_zones[zone_type];
4777
4778 if (is_highmem(zone)) {
4779 managed_highpages += zone->managed_pages;
4780 free_highpages += zone_page_state(zone, NR_FREE_PAGES);
4781 }
4782 }
4783 val->totalhigh = managed_highpages;
4784 val->freehigh = free_highpages;
4785#else
4786 val->totalhigh = managed_highpages;
4787 val->freehigh = free_highpages;
4788#endif
4789 val->mem_unit = PAGE_SIZE;
4790}
4791#endif
4792
4793/*
4794 * Determine whether the node should be displayed or not, depending on whether
4795 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
4796 */
4797static bool show_mem_node_skip(unsigned int flags, int nid, nodemask_t *nodemask)
4798{
4799 if (!(flags & SHOW_MEM_FILTER_NODES))
4800 return false;
4801
4802 /*
4803 * no node mask - aka implicit memory numa policy. Do not bother with
4804 * the synchronization - read_mems_allowed_begin - because we do not
4805 * have to be precise here.
4806 */
4807 if (!nodemask)
4808 nodemask = &cpuset_current_mems_allowed;
4809
4810 return !node_isset(nid, *nodemask);
4811}
4812
4813#define K(x) ((x) << (PAGE_SHIFT-10))
4814
4815static void show_migration_types(unsigned char type)
4816{
4817 static const char types[MIGRATE_TYPES] = {
4818 [MIGRATE_UNMOVABLE] = 'U',
4819 [MIGRATE_MOVABLE] = 'M',
4820 [MIGRATE_RECLAIMABLE] = 'E',
4821 [MIGRATE_HIGHATOMIC] = 'H',
4822#ifdef CONFIG_CMA
4823 [MIGRATE_CMA] = 'C',
4824#endif
4825#ifdef CONFIG_MEMORY_ISOLATION
4826 [MIGRATE_ISOLATE] = 'I',
4827#endif
4828 };
4829 char tmp[MIGRATE_TYPES + 1];
4830 char *p = tmp;
4831 int i;
4832
4833 for (i = 0; i < MIGRATE_TYPES; i++) {
4834 if (type & (1 << i))
4835 *p++ = types[i];
4836 }
4837
4838 *p = '\0';
4839 printk(KERN_CONT "(%s) ", tmp);
4840}
4841
4842/*
4843 * Show free area list (used inside shift_scroll-lock stuff)
4844 * We also calculate the percentage fragmentation. We do this by counting the
4845 * memory on each free list with the exception of the first item on the list.
4846 *
4847 * Bits in @filter:
4848 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
4849 * cpuset.
4850 */
4851void show_free_areas(unsigned int filter, nodemask_t *nodemask)
4852{
4853 unsigned long free_pcp = 0;
4854 int cpu;
4855 struct zone *zone;
4856 pg_data_t *pgdat;
4857
4858 for_each_populated_zone(zone) {
4859 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4860 continue;
4861
4862 for_each_online_cpu(cpu)
4863 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4864 }
4865
4866 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
4867 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
4868 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
4869 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
4870 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
4871 " free:%lu free_pcp:%lu free_cma:%lu\n",
4872 global_node_page_state(NR_ACTIVE_ANON),
4873 global_node_page_state(NR_INACTIVE_ANON),
4874 global_node_page_state(NR_ISOLATED_ANON),
4875 global_node_page_state(NR_ACTIVE_FILE),
4876 global_node_page_state(NR_INACTIVE_FILE),
4877 global_node_page_state(NR_ISOLATED_FILE),
4878 global_node_page_state(NR_UNEVICTABLE),
4879 global_node_page_state(NR_FILE_DIRTY),
4880 global_node_page_state(NR_WRITEBACK),
4881 global_node_page_state(NR_UNSTABLE_NFS),
4882 global_node_page_state(NR_SLAB_RECLAIMABLE),
4883 global_node_page_state(NR_SLAB_UNRECLAIMABLE),
4884 global_node_page_state(NR_FILE_MAPPED),
4885 global_node_page_state(NR_SHMEM),
4886 global_zone_page_state(NR_PAGETABLE),
4887 global_zone_page_state(NR_BOUNCE),
4888 global_zone_page_state(NR_FREE_PAGES),
4889 free_pcp,
4890 global_zone_page_state(NR_FREE_CMA_PAGES));
4891
4892 for_each_online_pgdat(pgdat) {
4893 if (show_mem_node_skip(filter, pgdat->node_id, nodemask))
4894 continue;
4895
4896 printk("Node %d"
4897 " active_anon:%lukB"
4898 " inactive_anon:%lukB"
4899 " active_file:%lukB"
4900 " inactive_file:%lukB"
4901 " unevictable:%lukB"
4902 " isolated(anon):%lukB"
4903 " isolated(file):%lukB"
4904 " mapped:%lukB"
4905 " dirty:%lukB"
4906 " writeback:%lukB"
4907 " shmem:%lukB"
4908#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4909 " shmem_thp: %lukB"
4910 " shmem_pmdmapped: %lukB"
4911 " anon_thp: %lukB"
4912#endif
4913 " writeback_tmp:%lukB"
4914 " unstable:%lukB"
4915 " all_unreclaimable? %s"
4916 "\n",
4917 pgdat->node_id,
4918 K(node_page_state(pgdat, NR_ACTIVE_ANON)),
4919 K(node_page_state(pgdat, NR_INACTIVE_ANON)),
4920 K(node_page_state(pgdat, NR_ACTIVE_FILE)),
4921 K(node_page_state(pgdat, NR_INACTIVE_FILE)),
4922 K(node_page_state(pgdat, NR_UNEVICTABLE)),
4923 K(node_page_state(pgdat, NR_ISOLATED_ANON)),
4924 K(node_page_state(pgdat, NR_ISOLATED_FILE)),
4925 K(node_page_state(pgdat, NR_FILE_MAPPED)),
4926 K(node_page_state(pgdat, NR_FILE_DIRTY)),
4927 K(node_page_state(pgdat, NR_WRITEBACK)),
4928 K(node_page_state(pgdat, NR_SHMEM)),
4929#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4930 K(node_page_state(pgdat, NR_SHMEM_THPS) * HPAGE_PMD_NR),
4931 K(node_page_state(pgdat, NR_SHMEM_PMDMAPPED)
4932 * HPAGE_PMD_NR),
4933 K(node_page_state(pgdat, NR_ANON_THPS) * HPAGE_PMD_NR),
4934#endif
4935 K(node_page_state(pgdat, NR_WRITEBACK_TEMP)),
4936 K(node_page_state(pgdat, NR_UNSTABLE_NFS)),
4937 pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES ?
4938 "yes" : "no");
4939 }
4940
4941 for_each_populated_zone(zone) {
4942 int i;
4943
4944 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
4945 continue;
4946
4947 free_pcp = 0;
4948 for_each_online_cpu(cpu)
4949 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
4950
4951 show_node(zone);
4952 printk(KERN_CONT
4953 "%s"
4954 " free:%lukB"
4955 " min:%lukB"
4956 " low:%lukB"
4957 " high:%lukB"
4958 " active_anon:%lukB"
4959 " inactive_anon:%lukB"
4960 " active_file:%lukB"
4961 " inactive_file:%lukB"
4962 " unevictable:%lukB"
4963 " writepending:%lukB"
4964 " present:%lukB"
4965 " managed:%lukB"
4966 " mlocked:%lukB"
4967 " kernel_stack:%lukB"
4968 " pagetables:%lukB"
4969 " bounce:%lukB"
4970 " free_pcp:%lukB"
4971 " local_pcp:%ukB"
4972 " free_cma:%lukB"
4973 "\n",
4974 zone->name,
4975 K(zone_page_state(zone, NR_FREE_PAGES)),
4976 K(min_wmark_pages(zone)),
4977 K(low_wmark_pages(zone)),
4978 K(high_wmark_pages(zone)),
4979 K(zone_page_state(zone, NR_ZONE_ACTIVE_ANON)),
4980 K(zone_page_state(zone, NR_ZONE_INACTIVE_ANON)),
4981 K(zone_page_state(zone, NR_ZONE_ACTIVE_FILE)),
4982 K(zone_page_state(zone, NR_ZONE_INACTIVE_FILE)),
4983 K(zone_page_state(zone, NR_ZONE_UNEVICTABLE)),
4984 K(zone_page_state(zone, NR_ZONE_WRITE_PENDING)),
4985 K(zone->present_pages),
4986 K(zone->managed_pages),
4987 K(zone_page_state(zone, NR_MLOCK)),
4988 zone_page_state(zone, NR_KERNEL_STACK_KB),
4989 K(zone_page_state(zone, NR_PAGETABLE)),
4990 K(zone_page_state(zone, NR_BOUNCE)),
4991 K(free_pcp),
4992 K(this_cpu_read(zone->pageset->pcp.count)),
4993 K(zone_page_state(zone, NR_FREE_CMA_PAGES)));
4994 printk("lowmem_reserve[]:");
4995 for (i = 0; i < MAX_NR_ZONES; i++)
4996 printk(KERN_CONT " %ld", zone->lowmem_reserve[i]);
4997 printk(KERN_CONT "\n");
4998 }
4999
5000 for_each_populated_zone(zone) {
5001 unsigned int order;
5002 unsigned long nr[MAX_ORDER], flags, total = 0;
5003 unsigned char types[MAX_ORDER];
5004
5005 if (show_mem_node_skip(filter, zone_to_nid(zone), nodemask))
5006 continue;
5007 show_node(zone);
5008 printk(KERN_CONT "%s: ", zone->name);
5009
5010 spin_lock_irqsave(&zone->lock, flags);
5011 for (order = 0; order < MAX_ORDER; order++) {
5012 struct free_area *area = &zone->free_area[order];
5013 int type;
5014
5015 nr[order] = area->nr_free;
5016 total += nr[order] << order;
5017
5018 types[order] = 0;
5019 for (type = 0; type < MIGRATE_TYPES; type++) {
5020 if (!list_empty(&area->free_list[type]))
5021 types[order] |= 1 << type;
5022 }
5023 }
5024 spin_unlock_irqrestore(&zone->lock, flags);
5025 for (order = 0; order < MAX_ORDER; order++) {
5026 printk(KERN_CONT "%lu*%lukB ",
5027 nr[order], K(1UL) << order);
5028 if (nr[order])
5029 show_migration_types(types[order]);
5030 }
5031 printk(KERN_CONT "= %lukB\n", K(total));
5032 }
5033
5034 hugetlb_show_meminfo();
5035
5036 printk("%ld total pagecache pages\n", global_node_page_state(NR_FILE_PAGES));
5037
5038 show_swap_cache_info();
5039}
5040
5041static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5042{
5043 zoneref->zone = zone;
5044 zoneref->zone_idx = zone_idx(zone);
5045}
5046
5047/*
5048 * Builds allocation fallback zone lists.
5049 *
5050 * Add all populated zones of a node to the zonelist.
5051 */
5052static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5053{
5054 struct zone *zone;
5055 enum zone_type zone_type = MAX_NR_ZONES;
5056 int nr_zones = 0;
5057
5058 do {
5059 zone_type--;
5060 zone = pgdat->node_zones + zone_type;
5061 if (managed_zone(zone)) {
5062 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5063 check_highest_zone(zone_type);
5064 }
5065 } while (zone_type);
5066
5067 return nr_zones;
5068}
5069
5070#ifdef CONFIG_NUMA
5071
5072static int __parse_numa_zonelist_order(char *s)
5073{
5074 /*
5075 * We used to support different zonlists modes but they turned
5076 * out to be just not useful. Let's keep the warning in place
5077 * if somebody still use the cmd line parameter so that we do
5078 * not fail it silently
5079 */
5080 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5081 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5082 return -EINVAL;
5083 }
5084 return 0;
5085}
5086
5087static __init int setup_numa_zonelist_order(char *s)
5088{
5089 if (!s)
5090 return 0;
5091
5092 return __parse_numa_zonelist_order(s);
5093}
5094early_param("numa_zonelist_order", setup_numa_zonelist_order);
5095
5096char numa_zonelist_order[] = "Node";
5097
5098/*
5099 * sysctl handler for numa_zonelist_order
5100 */
5101int numa_zonelist_order_handler(struct ctl_table *table, int write,
5102 void __user *buffer, size_t *length,
5103 loff_t *ppos)
5104{
5105 char *str;
5106 int ret;
5107
5108 if (!write)
5109 return proc_dostring(table, write, buffer, length, ppos);
5110 str = memdup_user_nul(buffer, 16);
5111 if (IS_ERR(str))
5112 return PTR_ERR(str);
5113
5114 ret = __parse_numa_zonelist_order(str);
5115 kfree(str);
5116 return ret;
5117}
5118
5119
5120#define MAX_NODE_LOAD (nr_online_nodes)
5121static int node_load[MAX_NUMNODES];
5122
5123/**
5124 * find_next_best_node - find the next node that should appear in a given node's fallback list
5125 * @node: node whose fallback list we're appending
5126 * @used_node_mask: nodemask_t of already used nodes
5127 *
5128 * We use a number of factors to determine which is the next node that should
5129 * appear on a given node's fallback list. The node should not have appeared
5130 * already in @node's fallback list, and it should be the next closest node
5131 * according to the distance array (which contains arbitrary distance values
5132 * from each node to each node in the system), and should also prefer nodes
5133 * with no CPUs, since presumably they'll have very little allocation pressure
5134 * on them otherwise.
5135 * It returns -1 if no node is found.
5136 */
5137static int find_next_best_node(int node, nodemask_t *used_node_mask)
5138{
5139 int n, val;
5140 int min_val = INT_MAX;
5141 int best_node = NUMA_NO_NODE;
5142 const struct cpumask *tmp = cpumask_of_node(0);
5143
5144 /* Use the local node if we haven't already */
5145 if (!node_isset(node, *used_node_mask)) {
5146 node_set(node, *used_node_mask);
5147 return node;
5148 }
5149
5150 for_each_node_state(n, N_MEMORY) {
5151
5152 /* Don't want a node to appear more than once */
5153 if (node_isset(n, *used_node_mask))
5154 continue;
5155
5156 /* Use the distance array to find the distance */
5157 val = node_distance(node, n);
5158
5159 /* Penalize nodes under us ("prefer the next node") */
5160 val += (n < node);
5161
5162 /* Give preference to headless and unused nodes */
5163 tmp = cpumask_of_node(n);
5164 if (!cpumask_empty(tmp))
5165 val += PENALTY_FOR_NODE_WITH_CPUS;
5166
5167 /* Slight preference for less loaded node */
5168 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
5169 val += node_load[n];
5170
5171 if (val < min_val) {
5172 min_val = val;
5173 best_node = n;
5174 }
5175 }
5176
5177 if (best_node >= 0)
5178 node_set(best_node, *used_node_mask);
5179
5180 return best_node;
5181}
5182
5183
5184/*
5185 * Build zonelists ordered by node and zones within node.
5186 * This results in maximum locality--normal zone overflows into local
5187 * DMA zone, if any--but risks exhausting DMA zone.
5188 */
5189static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5190 unsigned nr_nodes)
5191{
5192 struct zoneref *zonerefs;
5193 int i;
5194
5195 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5196
5197 for (i = 0; i < nr_nodes; i++) {
5198 int nr_zones;
5199
5200 pg_data_t *node = NODE_DATA(node_order[i]);
5201
5202 nr_zones = build_zonerefs_node(node, zonerefs);
5203 zonerefs += nr_zones;
5204 }
5205 zonerefs->zone = NULL;
5206 zonerefs->zone_idx = 0;
5207}
5208
5209/*
5210 * Build gfp_thisnode zonelists
5211 */
5212static void build_thisnode_zonelists(pg_data_t *pgdat)
5213{
5214 struct zoneref *zonerefs;
5215 int nr_zones;
5216
5217 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5218 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5219 zonerefs += nr_zones;
5220 zonerefs->zone = NULL;
5221 zonerefs->zone_idx = 0;
5222}
5223
5224/*
5225 * Build zonelists ordered by zone and nodes within zones.
5226 * This results in conserving DMA zone[s] until all Normal memory is
5227 * exhausted, but results in overflowing to remote node while memory
5228 * may still exist in local DMA zone.
5229 */
5230
5231static void build_zonelists(pg_data_t *pgdat)
5232{
5233 static int node_order[MAX_NUMNODES];
5234 int node, load, nr_nodes = 0;
5235 nodemask_t used_mask;
5236 int local_node, prev_node;
5237
5238 /* NUMA-aware ordering of nodes */
5239 local_node = pgdat->node_id;
5240 load = nr_online_nodes;
5241 prev_node = local_node;
5242 nodes_clear(used_mask);
5243
5244 memset(node_order, 0, sizeof(node_order));
5245 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5246 /*
5247 * We don't want to pressure a particular node.
5248 * So adding penalty to the first node in same
5249 * distance group to make it round-robin.
5250 */
5251 if (node_distance(local_node, node) !=
5252 node_distance(local_node, prev_node))
5253 node_load[node] = load;
5254
5255 node_order[nr_nodes++] = node;
5256 prev_node = node;
5257 load--;
5258 }
5259
5260 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5261 build_thisnode_zonelists(pgdat);
5262}
5263
5264#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5265/*
5266 * Return node id of node used for "local" allocations.
5267 * I.e., first node id of first zone in arg node's generic zonelist.
5268 * Used for initializing percpu 'numa_mem', which is used primarily
5269 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5270 */
5271int local_memory_node(int node)
5272{
5273 struct zoneref *z;
5274
5275 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5276 gfp_zone(GFP_KERNEL),
5277 NULL);
5278 return z->zone->node;
5279}
5280#endif
5281
5282static void setup_min_unmapped_ratio(void);
5283static void setup_min_slab_ratio(void);
5284#else /* CONFIG_NUMA */
5285
5286static void build_zonelists(pg_data_t *pgdat)
5287{
5288 int node, local_node;
5289 struct zoneref *zonerefs;
5290 int nr_zones;
5291
5292 local_node = pgdat->node_id;
5293
5294 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5295 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5296 zonerefs += nr_zones;
5297
5298 /*
5299 * Now we build the zonelist so that it contains the zones
5300 * of all the other nodes.
5301 * We don't want to pressure a particular node, so when
5302 * building the zones for node N, we make sure that the
5303 * zones coming right after the local ones are those from
5304 * node N+1 (modulo N)
5305 */
5306 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
5307 if (!node_online(node))
5308 continue;
5309 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5310 zonerefs += nr_zones;
5311 }
5312 for (node = 0; node < local_node; node++) {
5313 if (!node_online(node))
5314 continue;
5315 nr_zones = build_zonerefs_node(NODE_DATA(node), zonerefs);
5316 zonerefs += nr_zones;
5317 }
5318
5319 zonerefs->zone = NULL;
5320 zonerefs->zone_idx = 0;
5321}
5322
5323#endif /* CONFIG_NUMA */
5324
5325/*
5326 * Boot pageset table. One per cpu which is going to be used for all
5327 * zones and all nodes. The parameters will be set in such a way
5328 * that an item put on a list will immediately be handed over to
5329 * the buddy list. This is safe since pageset manipulation is done
5330 * with interrupts disabled.
5331 *
5332 * The boot_pagesets must be kept even after bootup is complete for
5333 * unused processors and/or zones. They do play a role for bootstrapping
5334 * hotplugged processors.
5335 *
5336 * zoneinfo_show() and maybe other functions do
5337 * not check if the processor is online before following the pageset pointer.
5338 * Other parts of the kernel may not check if the zone is available.
5339 */
5340static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
5341static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
5342static DEFINE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
5343
5344static void __build_all_zonelists(void *data)
5345{
5346 int nid;
5347 int __maybe_unused cpu;
5348 pg_data_t *self = data;
5349 static DEFINE_SPINLOCK(lock);
5350
5351 spin_lock(&lock);
5352
5353#ifdef CONFIG_NUMA
5354 memset(node_load, 0, sizeof(node_load));
5355#endif
5356
5357 /*
5358 * This node is hotadded and no memory is yet present. So just
5359 * building zonelists is fine - no need to touch other nodes.
5360 */
5361 if (self && !node_online(self->node_id)) {
5362 build_zonelists(self);
5363 } else {
5364 for_each_online_node(nid) {
5365 pg_data_t *pgdat = NODE_DATA(nid);
5366
5367 build_zonelists(pgdat);
5368 }
5369
5370#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5371 /*
5372 * We now know the "local memory node" for each node--
5373 * i.e., the node of the first zone in the generic zonelist.
5374 * Set up numa_mem percpu variable for on-line cpus. During
5375 * boot, only the boot cpu should be on-line; we'll init the
5376 * secondary cpus' numa_mem as they come on-line. During
5377 * node/memory hotplug, we'll fixup all on-line cpus.
5378 */
5379 for_each_online_cpu(cpu)
5380 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5381#endif
5382 }
5383
5384 spin_unlock(&lock);
5385}
5386
5387static noinline void __init
5388build_all_zonelists_init(void)
5389{
5390 int cpu;
5391
5392 __build_all_zonelists(NULL);
5393
5394 /*
5395 * Initialize the boot_pagesets that are going to be used
5396 * for bootstrapping processors. The real pagesets for
5397 * each zone will be allocated later when the per cpu
5398 * allocator is available.
5399 *
5400 * boot_pagesets are used also for bootstrapping offline
5401 * cpus if the system is already booted because the pagesets
5402 * are needed to initialize allocators on a specific cpu too.
5403 * F.e. the percpu allocator needs the page allocator which
5404 * needs the percpu allocator in order to allocate its pagesets
5405 * (a chicken-egg dilemma).
5406 */
5407 for_each_possible_cpu(cpu)
5408 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
5409
5410 mminit_verify_zonelist();
5411 cpuset_init_current_mems_allowed();
5412}
5413
5414/*
5415 * unless system_state == SYSTEM_BOOTING.
5416 *
5417 * __ref due to call of __init annotated helper build_all_zonelists_init
5418 * [protected by SYSTEM_BOOTING].
5419 */
5420void __ref build_all_zonelists(pg_data_t *pgdat)
5421{
5422 if (system_state == SYSTEM_BOOTING) {
5423 build_all_zonelists_init();
5424 } else {
5425 __build_all_zonelists(pgdat);
5426 /* cpuset refresh routine should be here */
5427 }
5428 vm_total_pages = nr_free_pagecache_pages();
5429 /*
5430 * Disable grouping by mobility if the number of pages in the
5431 * system is too low to allow the mechanism to work. It would be
5432 * more accurate, but expensive to check per-zone. This check is
5433 * made on memory-hotadd so a system can start with mobility
5434 * disabled and enable it later
5435 */
5436 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5437 page_group_by_mobility_disabled = 1;
5438 else
5439 page_group_by_mobility_disabled = 0;
5440
5441 pr_info("Built %i zonelists, mobility grouping %s. Total pages: %ld\n",
5442 nr_online_nodes,
5443 page_group_by_mobility_disabled ? "off" : "on",
5444 vm_total_pages);
5445#ifdef CONFIG_NUMA
5446 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5447#endif
5448}
5449
5450/*
5451 * Initially all pages are reserved - free ones are freed
5452 * up by free_all_bootmem() once the early boot process is
5453 * done. Non-atomic initialization, single-pass.
5454 */
5455void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
5456 unsigned long start_pfn, enum memmap_context context,
5457 struct vmem_altmap *altmap)
5458{
5459 unsigned long end_pfn = start_pfn + size;
5460 pg_data_t *pgdat = NODE_DATA(nid);
5461 unsigned long pfn;
5462 unsigned long nr_initialised = 0;
5463 struct page *page;
5464#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5465 struct memblock_region *r = NULL, *tmp;
5466#endif
5467
5468 if (highest_memmap_pfn < end_pfn - 1)
5469 highest_memmap_pfn = end_pfn - 1;
5470
5471 /*
5472 * Honor reservation requested by the driver for this ZONE_DEVICE
5473 * memory
5474 */
5475 if (altmap && start_pfn == altmap->base_pfn)
5476 start_pfn += altmap->reserve;
5477
5478 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
5479 /*
5480 * There can be holes in boot-time mem_map[]s handed to this
5481 * function. They do not exist on hotplugged memory.
5482 */
5483 if (context != MEMMAP_EARLY)
5484 goto not_early;
5485
5486 if (!early_pfn_valid(pfn))
5487 continue;
5488 if (!early_pfn_in_nid(pfn, nid))
5489 continue;
5490 if (!update_defer_init(pgdat, pfn, end_pfn, &nr_initialised))
5491 break;
5492
5493#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5494 /*
5495 * Check given memblock attribute by firmware which can affect
5496 * kernel memory layout. If zone==ZONE_MOVABLE but memory is
5497 * mirrored, it's an overlapped memmap init. skip it.
5498 */
5499 if (mirrored_kernelcore && zone == ZONE_MOVABLE) {
5500 if (!r || pfn >= memblock_region_memory_end_pfn(r)) {
5501 for_each_memblock(memory, tmp)
5502 if (pfn < memblock_region_memory_end_pfn(tmp))
5503 break;
5504 r = tmp;
5505 }
5506 if (pfn >= memblock_region_memory_base_pfn(r) &&
5507 memblock_is_mirror(r)) {
5508 /* already initialized as NORMAL */
5509 pfn = memblock_region_memory_end_pfn(r);
5510 continue;
5511 }
5512 }
5513#endif
5514
5515not_early:
5516 page = pfn_to_page(pfn);
5517 __init_single_page(page, pfn, zone, nid);
5518 if (context == MEMMAP_HOTPLUG)
5519 SetPageReserved(page);
5520
5521 /*
5522 * Mark the block movable so that blocks are reserved for
5523 * movable at startup. This will force kernel allocations
5524 * to reserve their blocks rather than leaking throughout
5525 * the address space during boot when many long-lived
5526 * kernel allocations are made.
5527 *
5528 * bitmap is created for zone's valid pfn range. but memmap
5529 * can be created for invalid pages (for alignment)
5530 * check here not to call set_pageblock_migratetype() against
5531 * pfn out of zone.
5532 *
5533 * Please note that MEMMAP_HOTPLUG path doesn't clear memmap
5534 * because this is done early in sparse_add_one_section
5535 */
5536 if (!(pfn & (pageblock_nr_pages - 1))) {
5537 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5538 cond_resched();
5539 }
5540 }
5541}
5542
5543static void __meminit zone_init_free_lists(struct zone *zone)
5544{
5545 unsigned int order, t;
5546 for_each_migratetype_order(order, t) {
5547 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
5548 zone->free_area[order].nr_free = 0;
5549 }
5550}
5551
5552#ifndef __HAVE_ARCH_MEMMAP_INIT
5553#define memmap_init(size, nid, zone, start_pfn) \
5554 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY, NULL)
5555#endif
5556
5557static int zone_batchsize(struct zone *zone)
5558{
5559#ifdef CONFIG_MMU
5560 int batch;
5561
5562 /*
5563 * The per-cpu-pages pools are set to around 1000th of the
5564 * size of the zone. But no more than 1/2 of a meg.
5565 *
5566 * OK, so we don't know how big the cache is. So guess.
5567 */
5568 batch = zone->managed_pages / 1024;
5569 if (batch * PAGE_SIZE > 512 * 1024)
5570 batch = (512 * 1024) / PAGE_SIZE;
5571 batch /= 4; /* We effectively *= 4 below */
5572 if (batch < 1)
5573 batch = 1;
5574
5575 /*
5576 * Clamp the batch to a 2^n - 1 value. Having a power
5577 * of 2 value was found to be more likely to have
5578 * suboptimal cache aliasing properties in some cases.
5579 *
5580 * For example if 2 tasks are alternately allocating
5581 * batches of pages, one task can end up with a lot
5582 * of pages of one half of the possible page colors
5583 * and the other with pages of the other colors.
5584 */
5585 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5586
5587 return batch;
5588
5589#else
5590 /* The deferral and batching of frees should be suppressed under NOMMU
5591 * conditions.
5592 *
5593 * The problem is that NOMMU needs to be able to allocate large chunks
5594 * of contiguous memory as there's no hardware page translation to
5595 * assemble apparent contiguous memory from discontiguous pages.
5596 *
5597 * Queueing large contiguous runs of pages for batching, however,
5598 * causes the pages to actually be freed in smaller chunks. As there
5599 * can be a significant delay between the individual batches being
5600 * recycled, this leads to the once large chunks of space being
5601 * fragmented and becoming unavailable for high-order allocations.
5602 */
5603 return 0;
5604#endif
5605}
5606
5607/*
5608 * pcp->high and pcp->batch values are related and dependent on one another:
5609 * ->batch must never be higher then ->high.
5610 * The following function updates them in a safe manner without read side
5611 * locking.
5612 *
5613 * Any new users of pcp->batch and pcp->high should ensure they can cope with
5614 * those fields changing asynchronously (acording the the above rule).
5615 *
5616 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5617 * outside of boot time (or some other assurance that no concurrent updaters
5618 * exist).
5619 */
5620static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
5621 unsigned long batch)
5622{
5623 /* start with a fail safe value for batch */
5624 pcp->batch = 1;
5625 smp_wmb();
5626
5627 /* Update high, then batch, in order */
5628 pcp->high = high;
5629 smp_wmb();
5630
5631 pcp->batch = batch;
5632}
5633
5634/* a companion to pageset_set_high() */
5635static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
5636{
5637 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
5638}
5639
5640static void pageset_init(struct per_cpu_pageset *p)
5641{
5642 struct per_cpu_pages *pcp;
5643 int migratetype;
5644
5645 memset(p, 0, sizeof(*p));
5646
5647 pcp = &p->pcp;
5648 pcp->count = 0;
5649 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
5650 INIT_LIST_HEAD(&pcp->lists[migratetype]);
5651}
5652
5653static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
5654{
5655 pageset_init(p);
5656 pageset_set_batch(p, batch);
5657}
5658
5659/*
5660 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
5661 * to the value high for the pageset p.
5662 */
5663static void pageset_set_high(struct per_cpu_pageset *p,
5664 unsigned long high)
5665{
5666 unsigned long batch = max(1UL, high / 4);
5667 if ((high / 4) > (PAGE_SHIFT * 8))
5668 batch = PAGE_SHIFT * 8;
5669
5670 pageset_update(&p->pcp, high, batch);
5671}
5672
5673static void pageset_set_high_and_batch(struct zone *zone,
5674 struct per_cpu_pageset *pcp)
5675{
5676 if (percpu_pagelist_fraction)
5677 pageset_set_high(pcp,
5678 (zone->managed_pages /
5679 percpu_pagelist_fraction));
5680 else
5681 pageset_set_batch(pcp, zone_batchsize(zone));
5682}
5683
5684static void __meminit zone_pageset_init(struct zone *zone, int cpu)
5685{
5686 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
5687
5688 pageset_init(pcp);
5689 pageset_set_high_and_batch(zone, pcp);
5690}
5691
5692void __meminit setup_zone_pageset(struct zone *zone)
5693{
5694 int cpu;
5695 zone->pageset = alloc_percpu(struct per_cpu_pageset);
5696 for_each_possible_cpu(cpu)
5697 zone_pageset_init(zone, cpu);
5698}
5699
5700/*
5701 * Allocate per cpu pagesets and initialize them.
5702 * Before this call only boot pagesets were available.
5703 */
5704void __init setup_per_cpu_pageset(void)
5705{
5706 struct pglist_data *pgdat;
5707 struct zone *zone;
5708
5709 for_each_populated_zone(zone)
5710 setup_zone_pageset(zone);
5711
5712 for_each_online_pgdat(pgdat)
5713 pgdat->per_cpu_nodestats =
5714 alloc_percpu(struct per_cpu_nodestat);
5715}
5716
5717static __meminit void zone_pcp_init(struct zone *zone)
5718{
5719 /*
5720 * per cpu subsystem is not up at this point. The following code
5721 * relies on the ability of the linker to provide the
5722 * offset of a (static) per cpu variable into the per cpu area.
5723 */
5724 zone->pageset = &boot_pageset;
5725
5726 if (populated_zone(zone))
5727 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
5728 zone->name, zone->present_pages,
5729 zone_batchsize(zone));
5730}
5731
5732void __meminit init_currently_empty_zone(struct zone *zone,
5733 unsigned long zone_start_pfn,
5734 unsigned long size)
5735{
5736 struct pglist_data *pgdat = zone->zone_pgdat;
5737
5738 pgdat->nr_zones = zone_idx(zone) + 1;
5739
5740 zone->zone_start_pfn = zone_start_pfn;
5741
5742 mminit_dprintk(MMINIT_TRACE, "memmap_init",
5743 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
5744 pgdat->node_id,
5745 (unsigned long)zone_idx(zone),
5746 zone_start_pfn, (zone_start_pfn + size));
5747
5748 zone_init_free_lists(zone);
5749 zone->initialized = 1;
5750}
5751
5752#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5753#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
5754
5755/*
5756 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
5757 */
5758int __meminit __early_pfn_to_nid(unsigned long pfn,
5759 struct mminit_pfnnid_cache *state)
5760{
5761 unsigned long start_pfn, end_pfn;
5762 int nid;
5763
5764 if (state->last_start <= pfn && pfn < state->last_end)
5765 return state->last_nid;
5766
5767 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
5768 if (nid != -1) {
5769 state->last_start = start_pfn;
5770 state->last_end = end_pfn;
5771 state->last_nid = nid;
5772 }
5773
5774 return nid;
5775}
5776#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
5777
5778/**
5779 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
5780 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
5781 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
5782 *
5783 * If an architecture guarantees that all ranges registered contain no holes
5784 * and may be freed, this this function may be used instead of calling
5785 * memblock_free_early_nid() manually.
5786 */
5787void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
5788{
5789 unsigned long start_pfn, end_pfn;
5790 int i, this_nid;
5791
5792 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
5793 start_pfn = min(start_pfn, max_low_pfn);
5794 end_pfn = min(end_pfn, max_low_pfn);
5795
5796 if (start_pfn < end_pfn)
5797 memblock_free_early_nid(PFN_PHYS(start_pfn),
5798 (end_pfn - start_pfn) << PAGE_SHIFT,
5799 this_nid);
5800 }
5801}
5802
5803/**
5804 * sparse_memory_present_with_active_regions - Call memory_present for each active range
5805 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
5806 *
5807 * If an architecture guarantees that all ranges registered contain no holes and may
5808 * be freed, this function may be used instead of calling memory_present() manually.
5809 */
5810void __init sparse_memory_present_with_active_regions(int nid)
5811{
5812 unsigned long start_pfn, end_pfn;
5813 int i, this_nid;
5814
5815 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
5816 memory_present(this_nid, start_pfn, end_pfn);
5817}
5818
5819/**
5820 * get_pfn_range_for_nid - Return the start and end page frames for a node
5821 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
5822 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
5823 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
5824 *
5825 * It returns the start and end page frame of a node based on information
5826 * provided by memblock_set_node(). If called for a node
5827 * with no available memory, a warning is printed and the start and end
5828 * PFNs will be 0.
5829 */
5830void __meminit get_pfn_range_for_nid(unsigned int nid,
5831 unsigned long *start_pfn, unsigned long *end_pfn)
5832{
5833 unsigned long this_start_pfn, this_end_pfn;
5834 int i;
5835
5836 *start_pfn = -1UL;
5837 *end_pfn = 0;
5838
5839 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
5840 *start_pfn = min(*start_pfn, this_start_pfn);
5841 *end_pfn = max(*end_pfn, this_end_pfn);
5842 }
5843
5844 if (*start_pfn == -1UL)
5845 *start_pfn = 0;
5846}
5847
5848/*
5849 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5850 * assumption is made that zones within a node are ordered in monotonic
5851 * increasing memory addresses so that the "highest" populated zone is used
5852 */
5853static void __init find_usable_zone_for_movable(void)
5854{
5855 int zone_index;
5856 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
5857 if (zone_index == ZONE_MOVABLE)
5858 continue;
5859
5860 if (arch_zone_highest_possible_pfn[zone_index] >
5861 arch_zone_lowest_possible_pfn[zone_index])
5862 break;
5863 }
5864
5865 VM_BUG_ON(zone_index == -1);
5866 movable_zone = zone_index;
5867}
5868
5869/*
5870 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5871 * because it is sized independent of architecture. Unlike the other zones,
5872 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5873 * in each node depending on the size of each node and how evenly kernelcore
5874 * is distributed. This helper function adjusts the zone ranges
5875 * provided by the architecture for a given node by using the end of the
5876 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5877 * zones within a node are in order of monotonic increases memory addresses
5878 */
5879static void __meminit adjust_zone_range_for_zone_movable(int nid,
5880 unsigned long zone_type,
5881 unsigned long node_start_pfn,
5882 unsigned long node_end_pfn,
5883 unsigned long *zone_start_pfn,
5884 unsigned long *zone_end_pfn)
5885{
5886 /* Only adjust if ZONE_MOVABLE is on this node */
5887 if (zone_movable_pfn[nid]) {
5888 /* Size ZONE_MOVABLE */
5889 if (zone_type == ZONE_MOVABLE) {
5890 *zone_start_pfn = zone_movable_pfn[nid];
5891 *zone_end_pfn = min(node_end_pfn,
5892 arch_zone_highest_possible_pfn[movable_zone]);
5893
5894 /* Adjust for ZONE_MOVABLE starting within this range */
5895 } else if (!mirrored_kernelcore &&
5896 *zone_start_pfn < zone_movable_pfn[nid] &&
5897 *zone_end_pfn > zone_movable_pfn[nid]) {
5898 *zone_end_pfn = zone_movable_pfn[nid];
5899
5900 /* Check if this whole range is within ZONE_MOVABLE */
5901 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
5902 *zone_start_pfn = *zone_end_pfn;
5903 }
5904}
5905
5906/*
5907 * Return the number of pages a zone spans in a node, including holes
5908 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5909 */
5910static unsigned long __meminit zone_spanned_pages_in_node(int nid,
5911 unsigned long zone_type,
5912 unsigned long node_start_pfn,
5913 unsigned long node_end_pfn,
5914 unsigned long *zone_start_pfn,
5915 unsigned long *zone_end_pfn,
5916 unsigned long *ignored)
5917{
5918 /* When hotadd a new node from cpu_up(), the node should be empty */
5919 if (!node_start_pfn && !node_end_pfn)
5920 return 0;
5921
5922 /* Get the start and end of the zone */
5923 *zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
5924 *zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
5925 adjust_zone_range_for_zone_movable(nid, zone_type,
5926 node_start_pfn, node_end_pfn,
5927 zone_start_pfn, zone_end_pfn);
5928
5929 /* Check that this node has pages within the zone's required range */
5930 if (*zone_end_pfn < node_start_pfn || *zone_start_pfn > node_end_pfn)
5931 return 0;
5932
5933 /* Move the zone boundaries inside the node if necessary */
5934 *zone_end_pfn = min(*zone_end_pfn, node_end_pfn);
5935 *zone_start_pfn = max(*zone_start_pfn, node_start_pfn);
5936
5937 /* Return the spanned pages */
5938 return *zone_end_pfn - *zone_start_pfn;
5939}
5940
5941/*
5942 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5943 * then all holes in the requested range will be accounted for.
5944 */
5945unsigned long __meminit __absent_pages_in_range(int nid,
5946 unsigned long range_start_pfn,
5947 unsigned long range_end_pfn)
5948{
5949 unsigned long nr_absent = range_end_pfn - range_start_pfn;
5950 unsigned long start_pfn, end_pfn;
5951 int i;
5952
5953 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5954 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
5955 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
5956 nr_absent -= end_pfn - start_pfn;
5957 }
5958 return nr_absent;
5959}
5960
5961/**
5962 * absent_pages_in_range - Return number of page frames in holes within a range
5963 * @start_pfn: The start PFN to start searching for holes
5964 * @end_pfn: The end PFN to stop searching for holes
5965 *
5966 * It returns the number of pages frames in memory holes within a range.
5967 */
5968unsigned long __init absent_pages_in_range(unsigned long start_pfn,
5969 unsigned long end_pfn)
5970{
5971 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
5972}
5973
5974/* Return the number of page frames in holes in a zone on a node */
5975static unsigned long __meminit zone_absent_pages_in_node(int nid,
5976 unsigned long zone_type,
5977 unsigned long node_start_pfn,
5978 unsigned long node_end_pfn,
5979 unsigned long *ignored)
5980{
5981 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
5982 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
5983 unsigned long zone_start_pfn, zone_end_pfn;
5984 unsigned long nr_absent;
5985
5986 /* When hotadd a new node from cpu_up(), the node should be empty */
5987 if (!node_start_pfn && !node_end_pfn)
5988 return 0;
5989
5990 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
5991 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
5992
5993 adjust_zone_range_for_zone_movable(nid, zone_type,
5994 node_start_pfn, node_end_pfn,
5995 &zone_start_pfn, &zone_end_pfn);
5996 nr_absent = __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
5997
5998 /*
5999 * ZONE_MOVABLE handling.
6000 * Treat pages to be ZONE_MOVABLE in ZONE_NORMAL as absent pages
6001 * and vice versa.
6002 */
6003 if (mirrored_kernelcore && zone_movable_pfn[nid]) {
6004 unsigned long start_pfn, end_pfn;
6005 struct memblock_region *r;
6006
6007 for_each_memblock(memory, r) {
6008 start_pfn = clamp(memblock_region_memory_base_pfn(r),
6009 zone_start_pfn, zone_end_pfn);
6010 end_pfn = clamp(memblock_region_memory_end_pfn(r),
6011 zone_start_pfn, zone_end_pfn);
6012
6013 if (zone_type == ZONE_MOVABLE &&
6014 memblock_is_mirror(r))
6015 nr_absent += end_pfn - start_pfn;
6016
6017 if (zone_type == ZONE_NORMAL &&
6018 !memblock_is_mirror(r))
6019 nr_absent += end_pfn - start_pfn;
6020 }
6021 }
6022
6023 return nr_absent;
6024}
6025
6026#else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6027static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
6028 unsigned long zone_type,
6029 unsigned long node_start_pfn,
6030 unsigned long node_end_pfn,
6031 unsigned long *zone_start_pfn,
6032 unsigned long *zone_end_pfn,
6033 unsigned long *zones_size)
6034{
6035 unsigned int zone;
6036
6037 *zone_start_pfn = node_start_pfn;
6038 for (zone = 0; zone < zone_type; zone++)
6039 *zone_start_pfn += zones_size[zone];
6040
6041 *zone_end_pfn = *zone_start_pfn + zones_size[zone_type];
6042
6043 return zones_size[zone_type];
6044}
6045
6046static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
6047 unsigned long zone_type,
6048 unsigned long node_start_pfn,
6049 unsigned long node_end_pfn,
6050 unsigned long *zholes_size)
6051{
6052 if (!zholes_size)
6053 return 0;
6054
6055 return zholes_size[zone_type];
6056}
6057
6058#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6059
6060static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
6061 unsigned long node_start_pfn,
6062 unsigned long node_end_pfn,
6063 unsigned long *zones_size,
6064 unsigned long *zholes_size)
6065{
6066 unsigned long realtotalpages = 0, totalpages = 0;
6067 enum zone_type i;
6068
6069 for (i = 0; i < MAX_NR_ZONES; i++) {
6070 struct zone *zone = pgdat->node_zones + i;
6071 unsigned long zone_start_pfn, zone_end_pfn;
6072 unsigned long size, real_size;
6073
6074 size = zone_spanned_pages_in_node(pgdat->node_id, i,
6075 node_start_pfn,
6076 node_end_pfn,
6077 &zone_start_pfn,
6078 &zone_end_pfn,
6079 zones_size);
6080 real_size = size - zone_absent_pages_in_node(pgdat->node_id, i,
6081 node_start_pfn, node_end_pfn,
6082 zholes_size);
6083 if (size)
6084 zone->zone_start_pfn = zone_start_pfn;
6085 else
6086 zone->zone_start_pfn = 0;
6087 zone->spanned_pages = size;
6088 zone->present_pages = real_size;
6089
6090 totalpages += size;
6091 realtotalpages += real_size;
6092 }
6093
6094 pgdat->node_spanned_pages = totalpages;
6095 pgdat->node_present_pages = realtotalpages;
6096 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
6097 realtotalpages);
6098}
6099
6100#ifndef CONFIG_SPARSEMEM
6101/*
6102 * Calculate the size of the zone->blockflags rounded to an unsigned long
6103 * Start by making sure zonesize is a multiple of pageblock_order by rounding
6104 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
6105 * round what is now in bits to nearest long in bits, then return it in
6106 * bytes.
6107 */
6108static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
6109{
6110 unsigned long usemapsize;
6111
6112 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
6113 usemapsize = roundup(zonesize, pageblock_nr_pages);
6114 usemapsize = usemapsize >> pageblock_order;
6115 usemapsize *= NR_PAGEBLOCK_BITS;
6116 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
6117
6118 return usemapsize / 8;
6119}
6120
6121static void __init setup_usemap(struct pglist_data *pgdat,
6122 struct zone *zone,
6123 unsigned long zone_start_pfn,
6124 unsigned long zonesize)
6125{
6126 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
6127 zone->pageblock_flags = NULL;
6128 if (usemapsize)
6129 zone->pageblock_flags =
6130 memblock_virt_alloc_node_nopanic(usemapsize,
6131 pgdat->node_id);
6132}
6133#else
6134static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
6135 unsigned long zone_start_pfn, unsigned long zonesize) {}
6136#endif /* CONFIG_SPARSEMEM */
6137
6138#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
6139
6140/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
6141void __paginginit set_pageblock_order(void)
6142{
6143 unsigned int order;
6144
6145 /* Check that pageblock_nr_pages has not already been setup */
6146 if (pageblock_order)
6147 return;
6148
6149 if (HPAGE_SHIFT > PAGE_SHIFT)
6150 order = HUGETLB_PAGE_ORDER;
6151 else
6152 order = MAX_ORDER - 1;
6153
6154 /*
6155 * Assume the largest contiguous order of interest is a huge page.
6156 * This value may be variable depending on boot parameters on IA64 and
6157 * powerpc.
6158 */
6159 pageblock_order = order;
6160}
6161#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6162
6163/*
6164 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
6165 * is unused as pageblock_order is set at compile-time. See
6166 * include/linux/pageblock-flags.h for the values of pageblock_order based on
6167 * the kernel config
6168 */
6169void __paginginit set_pageblock_order(void)
6170{
6171}
6172
6173#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
6174
6175static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
6176 unsigned long present_pages)
6177{
6178 unsigned long pages = spanned_pages;
6179
6180 /*
6181 * Provide a more accurate estimation if there are holes within
6182 * the zone and SPARSEMEM is in use. If there are holes within the
6183 * zone, each populated memory region may cost us one or two extra
6184 * memmap pages due to alignment because memmap pages for each
6185 * populated regions may not be naturally aligned on page boundary.
6186 * So the (present_pages >> 4) heuristic is a tradeoff for that.
6187 */
6188 if (spanned_pages > present_pages + (present_pages >> 4) &&
6189 IS_ENABLED(CONFIG_SPARSEMEM))
6190 pages = present_pages;
6191
6192 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
6193}
6194
6195/*
6196 * Set up the zone data structures:
6197 * - mark all pages reserved
6198 * - mark all memory queues empty
6199 * - clear the memory bitmaps
6200 *
6201 * NOTE: pgdat should get zeroed by caller.
6202 */
6203static void __paginginit free_area_init_core(struct pglist_data *pgdat)
6204{
6205 enum zone_type j;
6206 int nid = pgdat->node_id;
6207
6208 pgdat_resize_init(pgdat);
6209#ifdef CONFIG_NUMA_BALANCING
6210 spin_lock_init(&pgdat->numabalancing_migrate_lock);
6211 pgdat->numabalancing_migrate_nr_pages = 0;
6212 pgdat->numabalancing_migrate_next_window = jiffies;
6213#endif
6214#ifdef CONFIG_TRANSPARENT_HUGEPAGE
6215 spin_lock_init(&pgdat->split_queue_lock);
6216 INIT_LIST_HEAD(&pgdat->split_queue);
6217 pgdat->split_queue_len = 0;
6218#endif
6219 init_waitqueue_head(&pgdat->kswapd_wait);
6220 init_waitqueue_head(&pgdat->pfmemalloc_wait);
6221#ifdef CONFIG_COMPACTION
6222 init_waitqueue_head(&pgdat->kcompactd_wait);
6223#endif
6224 pgdat_page_ext_init(pgdat);
6225 spin_lock_init(&pgdat->lru_lock);
6226 lruvec_init(node_lruvec(pgdat));
6227
6228 pgdat->per_cpu_nodestats = &boot_nodestats;
6229
6230 for (j = 0; j < MAX_NR_ZONES; j++) {
6231 struct zone *zone = pgdat->node_zones + j;
6232 unsigned long size, realsize, freesize, memmap_pages;
6233 unsigned long zone_start_pfn = zone->zone_start_pfn;
6234
6235 size = zone->spanned_pages;
6236 realsize = freesize = zone->present_pages;
6237
6238 /*
6239 * Adjust freesize so that it accounts for how much memory
6240 * is used by this zone for memmap. This affects the watermark
6241 * and per-cpu initialisations
6242 */
6243 memmap_pages = calc_memmap_size(size, realsize);
6244 if (!is_highmem_idx(j)) {
6245 if (freesize >= memmap_pages) {
6246 freesize -= memmap_pages;
6247 if (memmap_pages)
6248 printk(KERN_DEBUG
6249 " %s zone: %lu pages used for memmap\n",
6250 zone_names[j], memmap_pages);
6251 } else
6252 pr_warn(" %s zone: %lu pages exceeds freesize %lu\n",
6253 zone_names[j], memmap_pages, freesize);
6254 }
6255
6256 /* Account for reserved pages */
6257 if (j == 0 && freesize > dma_reserve) {
6258 freesize -= dma_reserve;
6259 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
6260 zone_names[0], dma_reserve);
6261 }
6262
6263 if (!is_highmem_idx(j))
6264 nr_kernel_pages += freesize;
6265 /* Charge for highmem memmap if there are enough kernel pages */
6266 else if (nr_kernel_pages > memmap_pages * 2)
6267 nr_kernel_pages -= memmap_pages;
6268 nr_all_pages += freesize;
6269
6270 /*
6271 * Set an approximate value for lowmem here, it will be adjusted
6272 * when the bootmem allocator frees pages into the buddy system.
6273 * And all highmem pages will be managed by the buddy system.
6274 */
6275 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
6276#ifdef CONFIG_NUMA
6277 zone->node = nid;
6278#endif
6279 zone->name = zone_names[j];
6280 zone->zone_pgdat = pgdat;
6281 spin_lock_init(&zone->lock);
6282 zone_seqlock_init(zone);
6283 zone_pcp_init(zone);
6284
6285 if (!size)
6286 continue;
6287
6288 set_pageblock_order();
6289 setup_usemap(pgdat, zone, zone_start_pfn, size);
6290 init_currently_empty_zone(zone, zone_start_pfn, size);
6291 memmap_init(size, nid, j, zone_start_pfn);
6292 }
6293}
6294
6295#ifdef CONFIG_FLAT_NODE_MEM_MAP
6296static void __ref alloc_node_mem_map(struct pglist_data *pgdat)
6297{
6298 unsigned long __maybe_unused start = 0;
6299 unsigned long __maybe_unused offset = 0;
6300
6301 /* Skip empty nodes */
6302 if (!pgdat->node_spanned_pages)
6303 return;
6304
6305 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
6306 offset = pgdat->node_start_pfn - start;
6307 /* ia64 gets its own node_mem_map, before this, without bootmem */
6308 if (!pgdat->node_mem_map) {
6309 unsigned long size, end;
6310 struct page *map;
6311
6312 /*
6313 * The zone's endpoints aren't required to be MAX_ORDER
6314 * aligned but the node_mem_map endpoints must be in order
6315 * for the buddy allocator to function correctly.
6316 */
6317 end = pgdat_end_pfn(pgdat);
6318 end = ALIGN(end, MAX_ORDER_NR_PAGES);
6319 size = (end - start) * sizeof(struct page);
6320 map = memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
6321 pgdat->node_mem_map = map + offset;
6322 }
6323 pr_debug("%s: node %d, pgdat %08lx, node_mem_map %08lx\n",
6324 __func__, pgdat->node_id, (unsigned long)pgdat,
6325 (unsigned long)pgdat->node_mem_map);
6326#ifndef CONFIG_NEED_MULTIPLE_NODES
6327 /*
6328 * With no DISCONTIG, the global mem_map is just set as node 0's
6329 */
6330 if (pgdat == NODE_DATA(0)) {
6331 mem_map = NODE_DATA(0)->node_mem_map;
6332#if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
6333 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
6334 mem_map -= offset;
6335#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6336 }
6337#endif
6338}
6339#else
6340static void __ref alloc_node_mem_map(struct pglist_data *pgdat) { }
6341#endif /* CONFIG_FLAT_NODE_MEM_MAP */
6342
6343void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
6344 unsigned long node_start_pfn, unsigned long *zholes_size)
6345{
6346 pg_data_t *pgdat = NODE_DATA(nid);
6347 unsigned long start_pfn = 0;
6348 unsigned long end_pfn = 0;
6349
6350 /* pg_data_t should be reset to zero when it's allocated */
6351 WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
6352
6353 pgdat->node_id = nid;
6354 pgdat->node_start_pfn = node_start_pfn;
6355 pgdat->per_cpu_nodestats = NULL;
6356#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6357 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
6358 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
6359 (u64)start_pfn << PAGE_SHIFT,
6360 end_pfn ? ((u64)end_pfn << PAGE_SHIFT) - 1 : 0);
6361#else
6362 start_pfn = node_start_pfn;
6363#endif
6364 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
6365 zones_size, zholes_size);
6366
6367 alloc_node_mem_map(pgdat);
6368
6369#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
6370 /*
6371 * We start only with one section of pages, more pages are added as
6372 * needed until the rest of deferred pages are initialized.
6373 */
6374 pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
6375 pgdat->node_spanned_pages);
6376 pgdat->first_deferred_pfn = ULONG_MAX;
6377#endif
6378 free_area_init_core(pgdat);
6379}
6380
6381#ifdef CONFIG_HAVE_MEMBLOCK
6382/*
6383 * Only struct pages that are backed by physical memory are zeroed and
6384 * initialized by going through __init_single_page(). But, there are some
6385 * struct pages which are reserved in memblock allocator and their fields
6386 * may be accessed (for example page_to_pfn() on some configuration accesses
6387 * flags). We must explicitly zero those struct pages.
6388 */
6389void __paginginit zero_resv_unavail(void)
6390{
6391 phys_addr_t start, end;
6392 unsigned long pfn;
6393 u64 i, pgcnt;
6394
6395 /*
6396 * Loop through ranges that are reserved, but do not have reported
6397 * physical memory backing.
6398 */
6399 pgcnt = 0;
6400 for_each_resv_unavail_range(i, &start, &end) {
6401 for (pfn = PFN_DOWN(start); pfn < PFN_UP(end); pfn++) {
6402 if (!pfn_valid(ALIGN_DOWN(pfn, pageblock_nr_pages)))
6403 continue;
6404 mm_zero_struct_page(pfn_to_page(pfn));
6405 pgcnt++;
6406 }
6407 }
6408
6409 /*
6410 * Struct pages that do not have backing memory. This could be because
6411 * firmware is using some of this memory, or for some other reasons.
6412 * Once memblock is changed so such behaviour is not allowed: i.e.
6413 * list of "reserved" memory must be a subset of list of "memory", then
6414 * this code can be removed.
6415 */
6416 if (pgcnt)
6417 pr_info("Reserved but unavailable: %lld pages", pgcnt);
6418}
6419#endif /* CONFIG_HAVE_MEMBLOCK */
6420
6421#ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
6422
6423#if MAX_NUMNODES > 1
6424/*
6425 * Figure out the number of possible node ids.
6426 */
6427void __init setup_nr_node_ids(void)
6428{
6429 unsigned int highest;
6430
6431 highest = find_last_bit(node_possible_map.bits, MAX_NUMNODES);
6432 nr_node_ids = highest + 1;
6433}
6434#endif
6435
6436/**
6437 * node_map_pfn_alignment - determine the maximum internode alignment
6438 *
6439 * This function should be called after node map is populated and sorted.
6440 * It calculates the maximum power of two alignment which can distinguish
6441 * all the nodes.
6442 *
6443 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
6444 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
6445 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
6446 * shifted, 1GiB is enough and this function will indicate so.
6447 *
6448 * This is used to test whether pfn -> nid mapping of the chosen memory
6449 * model has fine enough granularity to avoid incorrect mapping for the
6450 * populated node map.
6451 *
6452 * Returns the determined alignment in pfn's. 0 if there is no alignment
6453 * requirement (single node).
6454 */
6455unsigned long __init node_map_pfn_alignment(void)
6456{
6457 unsigned long accl_mask = 0, last_end = 0;
6458 unsigned long start, end, mask;
6459 int last_nid = -1;
6460 int i, nid;
6461
6462 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
6463 if (!start || last_nid < 0 || last_nid == nid) {
6464 last_nid = nid;
6465 last_end = end;
6466 continue;
6467 }
6468
6469 /*
6470 * Start with a mask granular enough to pin-point to the
6471 * start pfn and tick off bits one-by-one until it becomes
6472 * too coarse to separate the current node from the last.
6473 */
6474 mask = ~((1 << __ffs(start)) - 1);
6475 while (mask && last_end <= (start & (mask << 1)))
6476 mask <<= 1;
6477
6478 /* accumulate all internode masks */
6479 accl_mask |= mask;
6480 }
6481
6482 /* convert mask to number of pages */
6483 return ~accl_mask + 1;
6484}
6485
6486/* Find the lowest pfn for a node */
6487static unsigned long __init find_min_pfn_for_node(int nid)
6488{
6489 unsigned long min_pfn = ULONG_MAX;
6490 unsigned long start_pfn;
6491 int i;
6492
6493 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
6494 min_pfn = min(min_pfn, start_pfn);
6495
6496 if (min_pfn == ULONG_MAX) {
6497 pr_warn("Could not find start_pfn for node %d\n", nid);
6498 return 0;
6499 }
6500
6501 return min_pfn;
6502}
6503
6504/**
6505 * find_min_pfn_with_active_regions - Find the minimum PFN registered
6506 *
6507 * It returns the minimum PFN based on information provided via
6508 * memblock_set_node().
6509 */
6510unsigned long __init find_min_pfn_with_active_regions(void)
6511{
6512 return find_min_pfn_for_node(MAX_NUMNODES);
6513}
6514
6515/*
6516 * early_calculate_totalpages()
6517 * Sum pages in active regions for movable zone.
6518 * Populate N_MEMORY for calculating usable_nodes.
6519 */
6520static unsigned long __init early_calculate_totalpages(void)
6521{
6522 unsigned long totalpages = 0;
6523 unsigned long start_pfn, end_pfn;
6524 int i, nid;
6525
6526 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
6527 unsigned long pages = end_pfn - start_pfn;
6528
6529 totalpages += pages;
6530 if (pages)
6531 node_set_state(nid, N_MEMORY);
6532 }
6533 return totalpages;
6534}
6535
6536/*
6537 * Find the PFN the Movable zone begins in each node. Kernel memory
6538 * is spread evenly between nodes as long as the nodes have enough
6539 * memory. When they don't, some nodes will have more kernelcore than
6540 * others
6541 */
6542static void __init find_zone_movable_pfns_for_nodes(void)
6543{
6544 int i, nid;
6545 unsigned long usable_startpfn;
6546 unsigned long kernelcore_node, kernelcore_remaining;
6547 /* save the state before borrow the nodemask */
6548 nodemask_t saved_node_state = node_states[N_MEMORY];
6549 unsigned long totalpages = early_calculate_totalpages();
6550 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
6551 struct memblock_region *r;
6552
6553 /* Need to find movable_zone earlier when movable_node is specified. */
6554 find_usable_zone_for_movable();
6555
6556 /*
6557 * If movable_node is specified, ignore kernelcore and movablecore
6558 * options.
6559 */
6560 if (movable_node_is_enabled()) {
6561 for_each_memblock(memory, r) {
6562 if (!memblock_is_hotpluggable(r))
6563 continue;
6564
6565 nid = r->nid;
6566
6567 usable_startpfn = PFN_DOWN(r->base);
6568 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6569 min(usable_startpfn, zone_movable_pfn[nid]) :
6570 usable_startpfn;
6571 }
6572
6573 goto out2;
6574 }
6575
6576 /*
6577 * If kernelcore=mirror is specified, ignore movablecore option
6578 */
6579 if (mirrored_kernelcore) {
6580 bool mem_below_4gb_not_mirrored = false;
6581
6582 for_each_memblock(memory, r) {
6583 if (memblock_is_mirror(r))
6584 continue;
6585
6586 nid = r->nid;
6587
6588 usable_startpfn = memblock_region_memory_base_pfn(r);
6589
6590 if (usable_startpfn < 0x100000) {
6591 mem_below_4gb_not_mirrored = true;
6592 continue;
6593 }
6594
6595 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
6596 min(usable_startpfn, zone_movable_pfn[nid]) :
6597 usable_startpfn;
6598 }
6599
6600 if (mem_below_4gb_not_mirrored)
6601 pr_warn("This configuration results in unmirrored kernel memory.");
6602
6603 goto out2;
6604 }
6605
6606 /*
6607 * If kernelcore=nn% or movablecore=nn% was specified, calculate the
6608 * amount of necessary memory.
6609 */
6610 if (required_kernelcore_percent)
6611 required_kernelcore = (totalpages * 100 * required_kernelcore_percent) /
6612 10000UL;
6613 if (required_movablecore_percent)
6614 required_movablecore = (totalpages * 100 * required_movablecore_percent) /
6615 10000UL;
6616
6617 /*
6618 * If movablecore= was specified, calculate what size of
6619 * kernelcore that corresponds so that memory usable for
6620 * any allocation type is evenly spread. If both kernelcore
6621 * and movablecore are specified, then the value of kernelcore
6622 * will be used for required_kernelcore if it's greater than
6623 * what movablecore would have allowed.
6624 */
6625 if (required_movablecore) {
6626 unsigned long corepages;
6627
6628 /*
6629 * Round-up so that ZONE_MOVABLE is at least as large as what
6630 * was requested by the user
6631 */
6632 required_movablecore =
6633 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
6634 required_movablecore = min(totalpages, required_movablecore);
6635 corepages = totalpages - required_movablecore;
6636
6637 required_kernelcore = max(required_kernelcore, corepages);
6638 }
6639
6640 /*
6641 * If kernelcore was not specified or kernelcore size is larger
6642 * than totalpages, there is no ZONE_MOVABLE.
6643 */
6644 if (!required_kernelcore || required_kernelcore >= totalpages)
6645 goto out;
6646
6647 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
6648 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
6649
6650restart:
6651 /* Spread kernelcore memory as evenly as possible throughout nodes */
6652 kernelcore_node = required_kernelcore / usable_nodes;
6653 for_each_node_state(nid, N_MEMORY) {
6654 unsigned long start_pfn, end_pfn;
6655
6656 /*
6657 * Recalculate kernelcore_node if the division per node
6658 * now exceeds what is necessary to satisfy the requested
6659 * amount of memory for the kernel
6660 */
6661 if (required_kernelcore < kernelcore_node)
6662 kernelcore_node = required_kernelcore / usable_nodes;
6663
6664 /*
6665 * As the map is walked, we track how much memory is usable
6666 * by the kernel using kernelcore_remaining. When it is
6667 * 0, the rest of the node is usable by ZONE_MOVABLE
6668 */
6669 kernelcore_remaining = kernelcore_node;
6670
6671 /* Go through each range of PFNs within this node */
6672 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
6673 unsigned long size_pages;
6674
6675 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
6676 if (start_pfn >= end_pfn)
6677 continue;
6678
6679 /* Account for what is only usable for kernelcore */
6680 if (start_pfn < usable_startpfn) {
6681 unsigned long kernel_pages;
6682 kernel_pages = min(end_pfn, usable_startpfn)
6683 - start_pfn;
6684
6685 kernelcore_remaining -= min(kernel_pages,
6686 kernelcore_remaining);
6687 required_kernelcore -= min(kernel_pages,
6688 required_kernelcore);
6689
6690 /* Continue if range is now fully accounted */
6691 if (end_pfn <= usable_startpfn) {
6692
6693 /*
6694 * Push zone_movable_pfn to the end so
6695 * that if we have to rebalance
6696 * kernelcore across nodes, we will
6697 * not double account here
6698 */
6699 zone_movable_pfn[nid] = end_pfn;
6700 continue;
6701 }
6702 start_pfn = usable_startpfn;
6703 }
6704
6705 /*
6706 * The usable PFN range for ZONE_MOVABLE is from
6707 * start_pfn->end_pfn. Calculate size_pages as the
6708 * number of pages used as kernelcore
6709 */
6710 size_pages = end_pfn - start_pfn;
6711 if (size_pages > kernelcore_remaining)
6712 size_pages = kernelcore_remaining;
6713 zone_movable_pfn[nid] = start_pfn + size_pages;
6714
6715 /*
6716 * Some kernelcore has been met, update counts and
6717 * break if the kernelcore for this node has been
6718 * satisfied
6719 */
6720 required_kernelcore -= min(required_kernelcore,
6721 size_pages);
6722 kernelcore_remaining -= size_pages;
6723 if (!kernelcore_remaining)
6724 break;
6725 }
6726 }
6727
6728 /*
6729 * If there is still required_kernelcore, we do another pass with one
6730 * less node in the count. This will push zone_movable_pfn[nid] further
6731 * along on the nodes that still have memory until kernelcore is
6732 * satisfied
6733 */
6734 usable_nodes--;
6735 if (usable_nodes && required_kernelcore > usable_nodes)
6736 goto restart;
6737
6738out2:
6739 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
6740 for (nid = 0; nid < MAX_NUMNODES; nid++)
6741 zone_movable_pfn[nid] =
6742 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
6743
6744out:
6745 /* restore the node_state */
6746 node_states[N_MEMORY] = saved_node_state;
6747}
6748
6749/* Any regular or high memory on that node ? */
6750static void check_for_memory(pg_data_t *pgdat, int nid)
6751{
6752 enum zone_type zone_type;
6753
6754 if (N_MEMORY == N_NORMAL_MEMORY)
6755 return;
6756
6757 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
6758 struct zone *zone = &pgdat->node_zones[zone_type];
6759 if (populated_zone(zone)) {
6760 node_set_state(nid, N_HIGH_MEMORY);
6761 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
6762 zone_type <= ZONE_NORMAL)
6763 node_set_state(nid, N_NORMAL_MEMORY);
6764 break;
6765 }
6766 }
6767}
6768
6769/**
6770 * free_area_init_nodes - Initialise all pg_data_t and zone data
6771 * @max_zone_pfn: an array of max PFNs for each zone
6772 *
6773 * This will call free_area_init_node() for each active node in the system.
6774 * Using the page ranges provided by memblock_set_node(), the size of each
6775 * zone in each node and their holes is calculated. If the maximum PFN
6776 * between two adjacent zones match, it is assumed that the zone is empty.
6777 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
6778 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
6779 * starts where the previous one ended. For example, ZONE_DMA32 starts
6780 * at arch_max_dma_pfn.
6781 */
6782void __init free_area_init_nodes(unsigned long *max_zone_pfn)
6783{
6784 unsigned long start_pfn, end_pfn;
6785 int i, nid;
6786
6787 /* Record where the zone boundaries are */
6788 memset(arch_zone_lowest_possible_pfn, 0,
6789 sizeof(arch_zone_lowest_possible_pfn));
6790 memset(arch_zone_highest_possible_pfn, 0,
6791 sizeof(arch_zone_highest_possible_pfn));
6792
6793 start_pfn = find_min_pfn_with_active_regions();
6794
6795 for (i = 0; i < MAX_NR_ZONES; i++) {
6796 if (i == ZONE_MOVABLE)
6797 continue;
6798
6799 end_pfn = max(max_zone_pfn[i], start_pfn);
6800 arch_zone_lowest_possible_pfn[i] = start_pfn;
6801 arch_zone_highest_possible_pfn[i] = end_pfn;
6802
6803 start_pfn = end_pfn;
6804 }
6805
6806 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
6807 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
6808 find_zone_movable_pfns_for_nodes();
6809
6810 /* Print out the zone ranges */
6811 pr_info("Zone ranges:\n");
6812 for (i = 0; i < MAX_NR_ZONES; i++) {
6813 if (i == ZONE_MOVABLE)
6814 continue;
6815 pr_info(" %-8s ", zone_names[i]);
6816 if (arch_zone_lowest_possible_pfn[i] ==
6817 arch_zone_highest_possible_pfn[i])
6818 pr_cont("empty\n");
6819 else
6820 pr_cont("[mem %#018Lx-%#018Lx]\n",
6821 (u64)arch_zone_lowest_possible_pfn[i]
6822 << PAGE_SHIFT,
6823 ((u64)arch_zone_highest_possible_pfn[i]
6824 << PAGE_SHIFT) - 1);
6825 }
6826
6827 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
6828 pr_info("Movable zone start for each node\n");
6829 for (i = 0; i < MAX_NUMNODES; i++) {
6830 if (zone_movable_pfn[i])
6831 pr_info(" Node %d: %#018Lx\n", i,
6832 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
6833 }
6834
6835 /* Print out the early node map */
6836 pr_info("Early memory node ranges\n");
6837 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
6838 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
6839 (u64)start_pfn << PAGE_SHIFT,
6840 ((u64)end_pfn << PAGE_SHIFT) - 1);
6841
6842 /* Initialise every node */
6843 mminit_verify_pageflags_layout();
6844 setup_nr_node_ids();
6845 for_each_online_node(nid) {
6846 pg_data_t *pgdat = NODE_DATA(nid);
6847 free_area_init_node(nid, NULL,
6848 find_min_pfn_for_node(nid), NULL);
6849
6850 /* Any memory on that node */
6851 if (pgdat->node_present_pages)
6852 node_set_state(nid, N_MEMORY);
6853 check_for_memory(pgdat, nid);
6854 }
6855 zero_resv_unavail();
6856}
6857
6858static int __init cmdline_parse_core(char *p, unsigned long *core,
6859 unsigned long *percent)
6860{
6861 unsigned long long coremem;
6862 char *endptr;
6863
6864 if (!p)
6865 return -EINVAL;
6866
6867 /* Value may be a percentage of total memory, otherwise bytes */
6868 coremem = simple_strtoull(p, &endptr, 0);
6869 if (*endptr == '%') {
6870 /* Paranoid check for percent values greater than 100 */
6871 WARN_ON(coremem > 100);
6872
6873 *percent = coremem;
6874 } else {
6875 coremem = memparse(p, &p);
6876 /* Paranoid check that UL is enough for the coremem value */
6877 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
6878
6879 *core = coremem >> PAGE_SHIFT;
6880 *percent = 0UL;
6881 }
6882 return 0;
6883}
6884
6885/*
6886 * kernelcore=size sets the amount of memory for use for allocations that
6887 * cannot be reclaimed or migrated.
6888 */
6889static int __init cmdline_parse_kernelcore(char *p)
6890{
6891 /* parse kernelcore=mirror */
6892 if (parse_option_str(p, "mirror")) {
6893 mirrored_kernelcore = true;
6894 return 0;
6895 }
6896
6897 return cmdline_parse_core(p, &required_kernelcore,
6898 &required_kernelcore_percent);
6899}
6900
6901/*
6902 * movablecore=size sets the amount of memory for use for allocations that
6903 * can be reclaimed or migrated.
6904 */
6905static int __init cmdline_parse_movablecore(char *p)
6906{
6907 return cmdline_parse_core(p, &required_movablecore,
6908 &required_movablecore_percent);
6909}
6910
6911early_param("kernelcore", cmdline_parse_kernelcore);
6912early_param("movablecore", cmdline_parse_movablecore);
6913
6914#endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
6915
6916void adjust_managed_page_count(struct page *page, long count)
6917{
6918 spin_lock(&managed_page_count_lock);
6919 page_zone(page)->managed_pages += count;
6920 totalram_pages += count;
6921#ifdef CONFIG_HIGHMEM
6922 if (PageHighMem(page))
6923 totalhigh_pages += count;
6924#endif
6925 spin_unlock(&managed_page_count_lock);
6926}
6927EXPORT_SYMBOL(adjust_managed_page_count);
6928
6929unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
6930{
6931 void *pos;
6932 unsigned long pages = 0;
6933
6934 start = (void *)PAGE_ALIGN((unsigned long)start);
6935 end = (void *)((unsigned long)end & PAGE_MASK);
6936 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
6937 if ((unsigned int)poison <= 0xFF)
6938 memset(pos, poison, PAGE_SIZE);
6939 free_reserved_page(virt_to_page(pos));
6940 }
6941
6942 if (pages && s)
6943 pr_info("Freeing %s memory: %ldK\n",
6944 s, pages << (PAGE_SHIFT - 10));
6945
6946 return pages;
6947}
6948EXPORT_SYMBOL(free_reserved_area);
6949
6950#ifdef CONFIG_HIGHMEM
6951void free_highmem_page(struct page *page)
6952{
6953 __free_reserved_page(page);
6954 totalram_pages++;
6955 page_zone(page)->managed_pages++;
6956 totalhigh_pages++;
6957}
6958#endif
6959
6960
6961void __init mem_init_print_info(const char *str)
6962{
6963 unsigned long physpages, codesize, datasize, rosize, bss_size;
6964 unsigned long init_code_size, init_data_size;
6965
6966 physpages = get_num_physpages();
6967 codesize = _etext - _stext;
6968 datasize = _edata - _sdata;
6969 rosize = __end_rodata - __start_rodata;
6970 bss_size = __bss_stop - __bss_start;
6971 init_data_size = __init_end - __init_begin;
6972 init_code_size = _einittext - _sinittext;
6973
6974 /*
6975 * Detect special cases and adjust section sizes accordingly:
6976 * 1) .init.* may be embedded into .data sections
6977 * 2) .init.text.* may be out of [__init_begin, __init_end],
6978 * please refer to arch/tile/kernel/vmlinux.lds.S.
6979 * 3) .rodata.* may be embedded into .text or .data sections.
6980 */
6981#define adj_init_size(start, end, size, pos, adj) \
6982 do { \
6983 if (start <= pos && pos < end && size > adj) \
6984 size -= adj; \
6985 } while (0)
6986
6987 adj_init_size(__init_begin, __init_end, init_data_size,
6988 _sinittext, init_code_size);
6989 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
6990 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
6991 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
6992 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
6993
6994#undef adj_init_size
6995
6996 pr_info("Memory: %luK/%luK available (%luK kernel code, %luK rwdata, %luK rodata, %luK init, %luK bss, %luK reserved, %luK cma-reserved"
6997#ifdef CONFIG_HIGHMEM
6998 ", %luK highmem"
6999#endif
7000 "%s%s)\n",
7001 nr_free_pages() << (PAGE_SHIFT - 10),
7002 physpages << (PAGE_SHIFT - 10),
7003 codesize >> 10, datasize >> 10, rosize >> 10,
7004 (init_data_size + init_code_size) >> 10, bss_size >> 10,
7005 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT - 10),
7006 totalcma_pages << (PAGE_SHIFT - 10),
7007#ifdef CONFIG_HIGHMEM
7008 totalhigh_pages << (PAGE_SHIFT - 10),
7009#endif
7010 str ? ", " : "", str ? str : "");
7011}
7012
7013/**
7014 * set_dma_reserve - set the specified number of pages reserved in the first zone
7015 * @new_dma_reserve: The number of pages to mark reserved
7016 *
7017 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
7018 * In the DMA zone, a significant percentage may be consumed by kernel image
7019 * and other unfreeable allocations which can skew the watermarks badly. This
7020 * function may optionally be used to account for unfreeable pages in the
7021 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
7022 * smaller per-cpu batchsize.
7023 */
7024void __init set_dma_reserve(unsigned long new_dma_reserve)
7025{
7026 dma_reserve = new_dma_reserve;
7027}
7028
7029void __init free_area_init(unsigned long *zones_size)
7030{
7031 free_area_init_node(0, zones_size,
7032 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
7033 zero_resv_unavail();
7034}
7035
7036static int page_alloc_cpu_dead(unsigned int cpu)
7037{
7038
7039 lru_add_drain_cpu(cpu);
7040 drain_pages(cpu);
7041
7042 /*
7043 * Spill the event counters of the dead processor
7044 * into the current processors event counters.
7045 * This artificially elevates the count of the current
7046 * processor.
7047 */
7048 vm_events_fold_cpu(cpu);
7049
7050 /*
7051 * Zero the differential counters of the dead processor
7052 * so that the vm statistics are consistent.
7053 *
7054 * This is only okay since the processor is dead and cannot
7055 * race with what we are doing.
7056 */
7057 cpu_vm_stats_fold(cpu);
7058 return 0;
7059}
7060
7061void __init page_alloc_init(void)
7062{
7063 int ret;
7064
7065 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC_DEAD,
7066 "mm/page_alloc:dead", NULL,
7067 page_alloc_cpu_dead);
7068 WARN_ON(ret < 0);
7069}
7070
7071/*
7072 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
7073 * or min_free_kbytes changes.
7074 */
7075static void calculate_totalreserve_pages(void)
7076{
7077 struct pglist_data *pgdat;
7078 unsigned long reserve_pages = 0;
7079 enum zone_type i, j;
7080
7081 for_each_online_pgdat(pgdat) {
7082
7083 pgdat->totalreserve_pages = 0;
7084
7085 for (i = 0; i < MAX_NR_ZONES; i++) {
7086 struct zone *zone = pgdat->node_zones + i;
7087 long max = 0;
7088
7089 /* Find valid and maximum lowmem_reserve in the zone */
7090 for (j = i; j < MAX_NR_ZONES; j++) {
7091 if (zone->lowmem_reserve[j] > max)
7092 max = zone->lowmem_reserve[j];
7093 }
7094
7095 /* we treat the high watermark as reserved pages. */
7096 max += high_wmark_pages(zone);
7097
7098 if (max > zone->managed_pages)
7099 max = zone->managed_pages;
7100
7101 pgdat->totalreserve_pages += max;
7102
7103 reserve_pages += max;
7104 }
7105 }
7106 totalreserve_pages = reserve_pages;
7107}
7108
7109/*
7110 * setup_per_zone_lowmem_reserve - called whenever
7111 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
7112 * has a correct pages reserved value, so an adequate number of
7113 * pages are left in the zone after a successful __alloc_pages().
7114 */
7115static void setup_per_zone_lowmem_reserve(void)
7116{
7117 struct pglist_data *pgdat;
7118 enum zone_type j, idx;
7119
7120 for_each_online_pgdat(pgdat) {
7121 for (j = 0; j < MAX_NR_ZONES; j++) {
7122 struct zone *zone = pgdat->node_zones + j;
7123 unsigned long managed_pages = zone->managed_pages;
7124
7125 zone->lowmem_reserve[j] = 0;
7126
7127 idx = j;
7128 while (idx) {
7129 struct zone *lower_zone;
7130
7131 idx--;
7132 lower_zone = pgdat->node_zones + idx;
7133
7134 if (sysctl_lowmem_reserve_ratio[idx] < 1) {
7135 sysctl_lowmem_reserve_ratio[idx] = 0;
7136 lower_zone->lowmem_reserve[j] = 0;
7137 } else {
7138 lower_zone->lowmem_reserve[j] =
7139 managed_pages / sysctl_lowmem_reserve_ratio[idx];
7140 }
7141 managed_pages += lower_zone->managed_pages;
7142 }
7143 }
7144 }
7145
7146 /* update totalreserve_pages */
7147 calculate_totalreserve_pages();
7148}
7149
7150static void __setup_per_zone_wmarks(void)
7151{
7152 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
7153 unsigned long lowmem_pages = 0;
7154 struct zone *zone;
7155 unsigned long flags;
7156
7157 /* Calculate total number of !ZONE_HIGHMEM pages */
7158 for_each_zone(zone) {
7159 if (!is_highmem(zone))
7160 lowmem_pages += zone->managed_pages;
7161 }
7162
7163 for_each_zone(zone) {
7164 u64 tmp;
7165
7166 spin_lock_irqsave(&zone->lock, flags);
7167 tmp = (u64)pages_min * zone->managed_pages;
7168 do_div(tmp, lowmem_pages);
7169 if (is_highmem(zone)) {
7170 /*
7171 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
7172 * need highmem pages, so cap pages_min to a small
7173 * value here.
7174 *
7175 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
7176 * deltas control asynch page reclaim, and so should
7177 * not be capped for highmem.
7178 */
7179 unsigned long min_pages;
7180
7181 min_pages = zone->managed_pages / 1024;
7182 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
7183 zone->watermark[WMARK_MIN] = min_pages;
7184 } else {
7185 /*
7186 * If it's a lowmem zone, reserve a number of pages
7187 * proportionate to the zone's size.
7188 */
7189 zone->watermark[WMARK_MIN] = tmp;
7190 }
7191
7192 /*
7193 * Set the kswapd watermarks distance according to the
7194 * scale factor in proportion to available memory, but
7195 * ensure a minimum size on small systems.
7196 */
7197 tmp = max_t(u64, tmp >> 2,
7198 mult_frac(zone->managed_pages,
7199 watermark_scale_factor, 10000));
7200
7201 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
7202 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + tmp * 2;
7203
7204 spin_unlock_irqrestore(&zone->lock, flags);
7205 }
7206
7207 /* update totalreserve_pages */
7208 calculate_totalreserve_pages();
7209}
7210
7211/**
7212 * setup_per_zone_wmarks - called when min_free_kbytes changes
7213 * or when memory is hot-{added|removed}
7214 *
7215 * Ensures that the watermark[min,low,high] values for each zone are set
7216 * correctly with respect to min_free_kbytes.
7217 */
7218void setup_per_zone_wmarks(void)
7219{
7220 static DEFINE_SPINLOCK(lock);
7221
7222 spin_lock(&lock);
7223 __setup_per_zone_wmarks();
7224 spin_unlock(&lock);
7225}
7226
7227/*
7228 * Initialise min_free_kbytes.
7229 *
7230 * For small machines we want it small (128k min). For large machines
7231 * we want it large (64MB max). But it is not linear, because network
7232 * bandwidth does not increase linearly with machine size. We use
7233 *
7234 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
7235 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
7236 *
7237 * which yields
7238 *
7239 * 16MB: 512k
7240 * 32MB: 724k
7241 * 64MB: 1024k
7242 * 128MB: 1448k
7243 * 256MB: 2048k
7244 * 512MB: 2896k
7245 * 1024MB: 4096k
7246 * 2048MB: 5792k
7247 * 4096MB: 8192k
7248 * 8192MB: 11584k
7249 * 16384MB: 16384k
7250 */
7251int __meminit init_per_zone_wmark_min(void)
7252{
7253 unsigned long lowmem_kbytes;
7254 int new_min_free_kbytes;
7255
7256 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
7257 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
7258
7259 if (new_min_free_kbytes > user_min_free_kbytes) {
7260 min_free_kbytes = new_min_free_kbytes;
7261 if (min_free_kbytes < 128)
7262 min_free_kbytes = 128;
7263 if (min_free_kbytes > 65536)
7264 min_free_kbytes = 65536;
7265 } else {
7266 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
7267 new_min_free_kbytes, user_min_free_kbytes);
7268 }
7269 setup_per_zone_wmarks();
7270 refresh_zone_stat_thresholds();
7271 setup_per_zone_lowmem_reserve();
7272
7273#ifdef CONFIG_NUMA
7274 setup_min_unmapped_ratio();
7275 setup_min_slab_ratio();
7276#endif
7277
7278 return 0;
7279}
7280core_initcall(init_per_zone_wmark_min)
7281
7282/*
7283 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
7284 * that we can call two helper functions whenever min_free_kbytes
7285 * changes.
7286 */
7287int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
7288 void __user *buffer, size_t *length, loff_t *ppos)
7289{
7290 int rc;
7291
7292 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7293 if (rc)
7294 return rc;
7295
7296 if (write) {
7297 user_min_free_kbytes = min_free_kbytes;
7298 setup_per_zone_wmarks();
7299 }
7300 return 0;
7301}
7302
7303int watermark_scale_factor_sysctl_handler(struct ctl_table *table, int write,
7304 void __user *buffer, size_t *length, loff_t *ppos)
7305{
7306 int rc;
7307
7308 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7309 if (rc)
7310 return rc;
7311
7312 if (write)
7313 setup_per_zone_wmarks();
7314
7315 return 0;
7316}
7317
7318#ifdef CONFIG_NUMA
7319static void setup_min_unmapped_ratio(void)
7320{
7321 pg_data_t *pgdat;
7322 struct zone *zone;
7323
7324 for_each_online_pgdat(pgdat)
7325 pgdat->min_unmapped_pages = 0;
7326
7327 for_each_zone(zone)
7328 zone->zone_pgdat->min_unmapped_pages += (zone->managed_pages *
7329 sysctl_min_unmapped_ratio) / 100;
7330}
7331
7332
7333int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
7334 void __user *buffer, size_t *length, loff_t *ppos)
7335{
7336 int rc;
7337
7338 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7339 if (rc)
7340 return rc;
7341
7342 setup_min_unmapped_ratio();
7343
7344 return 0;
7345}
7346
7347static void setup_min_slab_ratio(void)
7348{
7349 pg_data_t *pgdat;
7350 struct zone *zone;
7351
7352 for_each_online_pgdat(pgdat)
7353 pgdat->min_slab_pages = 0;
7354
7355 for_each_zone(zone)
7356 zone->zone_pgdat->min_slab_pages += (zone->managed_pages *
7357 sysctl_min_slab_ratio) / 100;
7358}
7359
7360int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
7361 void __user *buffer, size_t *length, loff_t *ppos)
7362{
7363 int rc;
7364
7365 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
7366 if (rc)
7367 return rc;
7368
7369 setup_min_slab_ratio();
7370
7371 return 0;
7372}
7373#endif
7374
7375/*
7376 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
7377 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
7378 * whenever sysctl_lowmem_reserve_ratio changes.
7379 *
7380 * The reserve ratio obviously has absolutely no relation with the
7381 * minimum watermarks. The lowmem reserve ratio can only make sense
7382 * if in function of the boot time zone sizes.
7383 */
7384int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
7385 void __user *buffer, size_t *length, loff_t *ppos)
7386{
7387 proc_dointvec_minmax(table, write, buffer, length, ppos);
7388 setup_per_zone_lowmem_reserve();
7389 return 0;
7390}
7391
7392/*
7393 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
7394 * cpu. It is the fraction of total pages in each zone that a hot per cpu
7395 * pagelist can have before it gets flushed back to buddy allocator.
7396 */
7397int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
7398 void __user *buffer, size_t *length, loff_t *ppos)
7399{
7400 struct zone *zone;
7401 int old_percpu_pagelist_fraction;
7402 int ret;
7403
7404 mutex_lock(&pcp_batch_high_lock);
7405 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
7406
7407 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
7408 if (!write || ret < 0)
7409 goto out;
7410
7411 /* Sanity checking to avoid pcp imbalance */
7412 if (percpu_pagelist_fraction &&
7413 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
7414 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
7415 ret = -EINVAL;
7416 goto out;
7417 }
7418
7419 /* No change? */
7420 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
7421 goto out;
7422
7423 for_each_populated_zone(zone) {
7424 unsigned int cpu;
7425
7426 for_each_possible_cpu(cpu)
7427 pageset_set_high_and_batch(zone,
7428 per_cpu_ptr(zone->pageset, cpu));
7429 }
7430out:
7431 mutex_unlock(&pcp_batch_high_lock);
7432 return ret;
7433}
7434
7435#ifdef CONFIG_NUMA
7436int hashdist = HASHDIST_DEFAULT;
7437
7438static int __init set_hashdist(char *str)
7439{
7440 if (!str)
7441 return 0;
7442 hashdist = simple_strtoul(str, &str, 0);
7443 return 1;
7444}
7445__setup("hashdist=", set_hashdist);
7446#endif
7447
7448#ifndef __HAVE_ARCH_RESERVED_KERNEL_PAGES
7449/*
7450 * Returns the number of pages that arch has reserved but
7451 * is not known to alloc_large_system_hash().
7452 */
7453static unsigned long __init arch_reserved_kernel_pages(void)
7454{
7455 return 0;
7456}
7457#endif
7458
7459/*
7460 * Adaptive scale is meant to reduce sizes of hash tables on large memory
7461 * machines. As memory size is increased the scale is also increased but at
7462 * slower pace. Starting from ADAPT_SCALE_BASE (64G), every time memory
7463 * quadruples the scale is increased by one, which means the size of hash table
7464 * only doubles, instead of quadrupling as well.
7465 * Because 32-bit systems cannot have large physical memory, where this scaling
7466 * makes sense, it is disabled on such platforms.
7467 */
7468#if __BITS_PER_LONG > 32
7469#define ADAPT_SCALE_BASE (64ul << 30)
7470#define ADAPT_SCALE_SHIFT 2
7471#define ADAPT_SCALE_NPAGES (ADAPT_SCALE_BASE >> PAGE_SHIFT)
7472#endif
7473
7474/*
7475 * allocate a large system hash table from bootmem
7476 * - it is assumed that the hash table must contain an exact power-of-2
7477 * quantity of entries
7478 * - limit is the number of hash buckets, not the total allocation size
7479 */
7480void *__init alloc_large_system_hash(const char *tablename,
7481 unsigned long bucketsize,
7482 unsigned long numentries,
7483 int scale,
7484 int flags,
7485 unsigned int *_hash_shift,
7486 unsigned int *_hash_mask,
7487 unsigned long low_limit,
7488 unsigned long high_limit)
7489{
7490 unsigned long long max = high_limit;
7491 unsigned long log2qty, size;
7492 void *table = NULL;
7493 gfp_t gfp_flags;
7494
7495 /* allow the kernel cmdline to have a say */
7496 if (!numentries) {
7497 /* round applicable memory size up to nearest megabyte */
7498 numentries = nr_kernel_pages;
7499 numentries -= arch_reserved_kernel_pages();
7500
7501 /* It isn't necessary when PAGE_SIZE >= 1MB */
7502 if (PAGE_SHIFT < 20)
7503 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
7504
7505#if __BITS_PER_LONG > 32
7506 if (!high_limit) {
7507 unsigned long adapt;
7508
7509 for (adapt = ADAPT_SCALE_NPAGES; adapt < numentries;
7510 adapt <<= ADAPT_SCALE_SHIFT)
7511 scale++;
7512 }
7513#endif
7514
7515 /* limit to 1 bucket per 2^scale bytes of low memory */
7516 if (scale > PAGE_SHIFT)
7517 numentries >>= (scale - PAGE_SHIFT);
7518 else
7519 numentries <<= (PAGE_SHIFT - scale);
7520
7521 /* Make sure we've got at least a 0-order allocation.. */
7522 if (unlikely(flags & HASH_SMALL)) {
7523 /* Makes no sense without HASH_EARLY */
7524 WARN_ON(!(flags & HASH_EARLY));
7525 if (!(numentries >> *_hash_shift)) {
7526 numentries = 1UL << *_hash_shift;
7527 BUG_ON(!numentries);
7528 }
7529 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
7530 numentries = PAGE_SIZE / bucketsize;
7531 }
7532 numentries = roundup_pow_of_two(numentries);
7533
7534 /* limit allocation size to 1/16 total memory by default */
7535 if (max == 0) {
7536 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
7537 do_div(max, bucketsize);
7538 }
7539 max = min(max, 0x80000000ULL);
7540
7541 if (numentries < low_limit)
7542 numentries = low_limit;
7543 if (numentries > max)
7544 numentries = max;
7545
7546 log2qty = ilog2(numentries);
7547
7548 gfp_flags = (flags & HASH_ZERO) ? GFP_ATOMIC | __GFP_ZERO : GFP_ATOMIC;
7549 do {
7550 size = bucketsize << log2qty;
7551 if (flags & HASH_EARLY) {
7552 if (flags & HASH_ZERO)
7553 table = memblock_virt_alloc_nopanic(size, 0);
7554 else
7555 table = memblock_virt_alloc_raw(size, 0);
7556 } else if (hashdist) {
7557 table = __vmalloc(size, gfp_flags, PAGE_KERNEL);
7558 } else {
7559 /*
7560 * If bucketsize is not a power-of-two, we may free
7561 * some pages at the end of hash table which
7562 * alloc_pages_exact() automatically does
7563 */
7564 if (get_order(size) < MAX_ORDER) {
7565 table = alloc_pages_exact(size, gfp_flags);
7566 kmemleak_alloc(table, size, 1, gfp_flags);
7567 }
7568 }
7569 } while (!table && size > PAGE_SIZE && --log2qty);
7570
7571 if (!table)
7572 panic("Failed to allocate %s hash table\n", tablename);
7573
7574 pr_info("%s hash table entries: %ld (order: %d, %lu bytes)\n",
7575 tablename, 1UL << log2qty, ilog2(size) - PAGE_SHIFT, size);
7576
7577 if (_hash_shift)
7578 *_hash_shift = log2qty;
7579 if (_hash_mask)
7580 *_hash_mask = (1 << log2qty) - 1;
7581
7582 return table;
7583}
7584
7585/*
7586 * This function checks whether pageblock includes unmovable pages or not.
7587 * If @count is not zero, it is okay to include less @count unmovable pages
7588 *
7589 * PageLRU check without isolation or lru_lock could race so that
7590 * MIGRATE_MOVABLE block might include unmovable pages. And __PageMovable
7591 * check without lock_page also may miss some movable non-lru pages at
7592 * race condition. So you can't expect this function should be exact.
7593 */
7594bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
7595 int migratetype,
7596 bool skip_hwpoisoned_pages)
7597{
7598 unsigned long pfn, iter, found;
7599
7600 /*
7601 * TODO we could make this much more efficient by not checking every
7602 * page in the range if we know all of them are in MOVABLE_ZONE and
7603 * that the movable zone guarantees that pages are migratable but
7604 * the later is not the case right now unfortunatelly. E.g. movablecore
7605 * can still lead to having bootmem allocations in zone_movable.
7606 */
7607
7608 /*
7609 * CMA allocations (alloc_contig_range) really need to mark isolate
7610 * CMA pageblocks even when they are not movable in fact so consider
7611 * them movable here.
7612 */
7613 if (is_migrate_cma(migratetype) &&
7614 is_migrate_cma(get_pageblock_migratetype(page)))
7615 return false;
7616
7617 pfn = page_to_pfn(page);
7618 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
7619 unsigned long check = pfn + iter;
7620
7621 if (!pfn_valid_within(check))
7622 continue;
7623
7624 page = pfn_to_page(check);
7625
7626 if (PageReserved(page))
7627 goto unmovable;
7628
7629 /*
7630 * Hugepages are not in LRU lists, but they're movable.
7631 * We need not scan over tail pages bacause we don't
7632 * handle each tail page individually in migration.
7633 */
7634 if (PageHuge(page)) {
7635 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
7636 continue;
7637 }
7638
7639 /*
7640 * We can't use page_count without pin a page
7641 * because another CPU can free compound page.
7642 * This check already skips compound tails of THP
7643 * because their page->_refcount is zero at all time.
7644 */
7645 if (!page_ref_count(page)) {
7646 if (PageBuddy(page))
7647 iter += (1 << page_order(page)) - 1;
7648 continue;
7649 }
7650
7651 /*
7652 * The HWPoisoned page may be not in buddy system, and
7653 * page_count() is not 0.
7654 */
7655 if (skip_hwpoisoned_pages && PageHWPoison(page))
7656 continue;
7657
7658 if (__PageMovable(page))
7659 continue;
7660
7661 if (!PageLRU(page))
7662 found++;
7663 /*
7664 * If there are RECLAIMABLE pages, we need to check
7665 * it. But now, memory offline itself doesn't call
7666 * shrink_node_slabs() and it still to be fixed.
7667 */
7668 /*
7669 * If the page is not RAM, page_count()should be 0.
7670 * we don't need more check. This is an _used_ not-movable page.
7671 *
7672 * The problematic thing here is PG_reserved pages. PG_reserved
7673 * is set to both of a memory hole page and a _used_ kernel
7674 * page at boot.
7675 */
7676 if (found > count)
7677 goto unmovable;
7678 }
7679 return false;
7680unmovable:
7681 WARN_ON_ONCE(zone_idx(zone) == ZONE_MOVABLE);
7682 return true;
7683}
7684
7685bool is_pageblock_removable_nolock(struct page *page)
7686{
7687 struct zone *zone;
7688 unsigned long pfn;
7689
7690 /*
7691 * We have to be careful here because we are iterating over memory
7692 * sections which are not zone aware so we might end up outside of
7693 * the zone but still within the section.
7694 * We have to take care about the node as well. If the node is offline
7695 * its NODE_DATA will be NULL - see page_zone.
7696 */
7697 if (!node_online(page_to_nid(page)))
7698 return false;
7699
7700 zone = page_zone(page);
7701 pfn = page_to_pfn(page);
7702 if (!zone_spans_pfn(zone, pfn))
7703 return false;
7704
7705 return !has_unmovable_pages(zone, page, 0, MIGRATE_MOVABLE, true);
7706}
7707
7708#if (defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || defined(CONFIG_CMA)
7709
7710static unsigned long pfn_max_align_down(unsigned long pfn)
7711{
7712 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
7713 pageblock_nr_pages) - 1);
7714}
7715
7716static unsigned long pfn_max_align_up(unsigned long pfn)
7717{
7718 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
7719 pageblock_nr_pages));
7720}
7721
7722/* [start, end) must belong to a single zone. */
7723static int __alloc_contig_migrate_range(struct compact_control *cc,
7724 unsigned long start, unsigned long end)
7725{
7726 /* This function is based on compact_zone() from compaction.c. */
7727 unsigned long nr_reclaimed;
7728 unsigned long pfn = start;
7729 unsigned int tries = 0;
7730 int ret = 0;
7731
7732 migrate_prep();
7733
7734 while (pfn < end || !list_empty(&cc->migratepages)) {
7735 if (fatal_signal_pending(current)) {
7736 ret = -EINTR;
7737 break;
7738 }
7739
7740 if (list_empty(&cc->migratepages)) {
7741 cc->nr_migratepages = 0;
7742 pfn = isolate_migratepages_range(cc, pfn, end);
7743 if (!pfn) {
7744 ret = -EINTR;
7745 break;
7746 }
7747 tries = 0;
7748 } else if (++tries == 5) {
7749 ret = ret < 0 ? ret : -EBUSY;
7750 break;
7751 }
7752
7753 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
7754 &cc->migratepages);
7755 cc->nr_migratepages -= nr_reclaimed;
7756
7757 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
7758 NULL, 0, cc->mode, MR_CONTIG_RANGE);
7759 }
7760 if (ret < 0) {
7761 putback_movable_pages(&cc->migratepages);
7762 return ret;
7763 }
7764 return 0;
7765}
7766
7767/**
7768 * alloc_contig_range() -- tries to allocate given range of pages
7769 * @start: start PFN to allocate
7770 * @end: one-past-the-last PFN to allocate
7771 * @migratetype: migratetype of the underlaying pageblocks (either
7772 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
7773 * in range must have the same migratetype and it must
7774 * be either of the two.
7775 * @gfp_mask: GFP mask to use during compaction
7776 *
7777 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
7778 * aligned. The PFN range must belong to a single zone.
7779 *
7780 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
7781 * pageblocks in the range. Once isolated, the pageblocks should not
7782 * be modified by others.
7783 *
7784 * Returns zero on success or negative error code. On success all
7785 * pages which PFN is in [start, end) are allocated for the caller and
7786 * need to be freed with free_contig_range().
7787 */
7788int alloc_contig_range(unsigned long start, unsigned long end,
7789 unsigned migratetype, gfp_t gfp_mask)
7790{
7791 unsigned long outer_start, outer_end;
7792 unsigned int order;
7793 int ret = 0;
7794
7795 struct compact_control cc = {
7796 .nr_migratepages = 0,
7797 .order = -1,
7798 .zone = page_zone(pfn_to_page(start)),
7799 .mode = MIGRATE_SYNC,
7800 .ignore_skip_hint = true,
7801 .no_set_skip_hint = true,
7802 .gfp_mask = current_gfp_context(gfp_mask),
7803 };
7804 INIT_LIST_HEAD(&cc.migratepages);
7805
7806 /*
7807 * What we do here is we mark all pageblocks in range as
7808 * MIGRATE_ISOLATE. Because pageblock and max order pages may
7809 * have different sizes, and due to the way page allocator
7810 * work, we align the range to biggest of the two pages so
7811 * that page allocator won't try to merge buddies from
7812 * different pageblocks and change MIGRATE_ISOLATE to some
7813 * other migration type.
7814 *
7815 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
7816 * migrate the pages from an unaligned range (ie. pages that
7817 * we are interested in). This will put all the pages in
7818 * range back to page allocator as MIGRATE_ISOLATE.
7819 *
7820 * When this is done, we take the pages in range from page
7821 * allocator removing them from the buddy system. This way
7822 * page allocator will never consider using them.
7823 *
7824 * This lets us mark the pageblocks back as
7825 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
7826 * aligned range but not in the unaligned, original range are
7827 * put back to page allocator so that buddy can use them.
7828 */
7829
7830 ret = start_isolate_page_range(pfn_max_align_down(start),
7831 pfn_max_align_up(end), migratetype,
7832 false);
7833 if (ret)
7834 return ret;
7835
7836 /*
7837 * In case of -EBUSY, we'd like to know which page causes problem.
7838 * So, just fall through. test_pages_isolated() has a tracepoint
7839 * which will report the busy page.
7840 *
7841 * It is possible that busy pages could become available before
7842 * the call to test_pages_isolated, and the range will actually be
7843 * allocated. So, if we fall through be sure to clear ret so that
7844 * -EBUSY is not accidentally used or returned to caller.
7845 */
7846 ret = __alloc_contig_migrate_range(&cc, start, end);
7847 if (ret && ret != -EBUSY)
7848 goto done;
7849 ret =0;
7850
7851 /*
7852 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
7853 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
7854 * more, all pages in [start, end) are free in page allocator.
7855 * What we are going to do is to allocate all pages from
7856 * [start, end) (that is remove them from page allocator).
7857 *
7858 * The only problem is that pages at the beginning and at the
7859 * end of interesting range may be not aligned with pages that
7860 * page allocator holds, ie. they can be part of higher order
7861 * pages. Because of this, we reserve the bigger range and
7862 * once this is done free the pages we are not interested in.
7863 *
7864 * We don't have to hold zone->lock here because the pages are
7865 * isolated thus they won't get removed from buddy.
7866 */
7867
7868 lru_add_drain_all();
7869 drain_all_pages(cc.zone);
7870
7871 order = 0;
7872 outer_start = start;
7873 while (!PageBuddy(pfn_to_page(outer_start))) {
7874 if (++order >= MAX_ORDER) {
7875 outer_start = start;
7876 break;
7877 }
7878 outer_start &= ~0UL << order;
7879 }
7880
7881 if (outer_start != start) {
7882 order = page_order(pfn_to_page(outer_start));
7883
7884 /*
7885 * outer_start page could be small order buddy page and
7886 * it doesn't include start page. Adjust outer_start
7887 * in this case to report failed page properly
7888 * on tracepoint in test_pages_isolated()
7889 */
7890 if (outer_start + (1UL << order) <= start)
7891 outer_start = start;
7892 }
7893
7894 /* Make sure the range is really isolated. */
7895 if (test_pages_isolated(outer_start, end, false)) {
7896 pr_info_ratelimited("%s: [%lx, %lx) PFNs busy\n",
7897 __func__, outer_start, end);
7898 ret = -EBUSY;
7899 goto done;
7900 }
7901
7902 /* Grab isolated pages from freelists. */
7903 outer_end = isolate_freepages_range(&cc, outer_start, end);
7904 if (!outer_end) {
7905 ret = -EBUSY;
7906 goto done;
7907 }
7908
7909 /* Free head and tail (if any) */
7910 if (start != outer_start)
7911 free_contig_range(outer_start, start - outer_start);
7912 if (end != outer_end)
7913 free_contig_range(end, outer_end - end);
7914
7915done:
7916 undo_isolate_page_range(pfn_max_align_down(start),
7917 pfn_max_align_up(end), migratetype);
7918 return ret;
7919}
7920
7921void free_contig_range(unsigned long pfn, unsigned nr_pages)
7922{
7923 unsigned int count = 0;
7924
7925 for (; nr_pages--; pfn++) {
7926 struct page *page = pfn_to_page(pfn);
7927
7928 count += page_count(page) != 1;
7929 __free_page(page);
7930 }
7931 WARN(count != 0, "%d pages are still in use!\n", count);
7932}
7933#endif
7934
7935#ifdef CONFIG_MEMORY_HOTPLUG
7936/*
7937 * The zone indicated has a new number of managed_pages; batch sizes and percpu
7938 * page high values need to be recalulated.
7939 */
7940void __meminit zone_pcp_update(struct zone *zone)
7941{
7942 unsigned cpu;
7943 mutex_lock(&pcp_batch_high_lock);
7944 for_each_possible_cpu(cpu)
7945 pageset_set_high_and_batch(zone,
7946 per_cpu_ptr(zone->pageset, cpu));
7947 mutex_unlock(&pcp_batch_high_lock);
7948}
7949#endif
7950
7951void zone_pcp_reset(struct zone *zone)
7952{
7953 unsigned long flags;
7954 int cpu;
7955 struct per_cpu_pageset *pset;
7956
7957 /* avoid races with drain_pages() */
7958 local_irq_save(flags);
7959 if (zone->pageset != &boot_pageset) {
7960 for_each_online_cpu(cpu) {
7961 pset = per_cpu_ptr(zone->pageset, cpu);
7962 drain_zonestat(zone, pset);
7963 }
7964 free_percpu(zone->pageset);
7965 zone->pageset = &boot_pageset;
7966 }
7967 local_irq_restore(flags);
7968}
7969
7970#ifdef CONFIG_MEMORY_HOTREMOVE
7971/*
7972 * All pages in the range must be in a single zone and isolated
7973 * before calling this.
7974 */
7975void
7976__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
7977{
7978 struct page *page;
7979 struct zone *zone;
7980 unsigned int order, i;
7981 unsigned long pfn;
7982 unsigned long flags;
7983 /* find the first valid pfn */
7984 for (pfn = start_pfn; pfn < end_pfn; pfn++)
7985 if (pfn_valid(pfn))
7986 break;
7987 if (pfn == end_pfn)
7988 return;
7989 offline_mem_sections(pfn, end_pfn);
7990 zone = page_zone(pfn_to_page(pfn));
7991 spin_lock_irqsave(&zone->lock, flags);
7992 pfn = start_pfn;
7993 while (pfn < end_pfn) {
7994 if (!pfn_valid(pfn)) {
7995 pfn++;
7996 continue;
7997 }
7998 page = pfn_to_page(pfn);
7999 /*
8000 * The HWPoisoned page may be not in buddy system, and
8001 * page_count() is not 0.
8002 */
8003 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
8004 pfn++;
8005 SetPageReserved(page);
8006 continue;
8007 }
8008
8009 BUG_ON(page_count(page));
8010 BUG_ON(!PageBuddy(page));
8011 order = page_order(page);
8012#ifdef CONFIG_DEBUG_VM
8013 pr_info("remove from free list %lx %d %lx\n",
8014 pfn, 1 << order, end_pfn);
8015#endif
8016 list_del(&page->lru);
8017 rmv_page_order(page);
8018 zone->free_area[order].nr_free--;
8019 for (i = 0; i < (1 << order); i++)
8020 SetPageReserved((page+i));
8021 pfn += (1 << order);
8022 }
8023 spin_unlock_irqrestore(&zone->lock, flags);
8024}
8025#endif
8026
8027bool is_free_buddy_page(struct page *page)
8028{
8029 struct zone *zone = page_zone(page);
8030 unsigned long pfn = page_to_pfn(page);
8031 unsigned long flags;
8032 unsigned int order;
8033
8034 spin_lock_irqsave(&zone->lock, flags);
8035 for (order = 0; order < MAX_ORDER; order++) {
8036 struct page *page_head = page - (pfn & ((1 << order) - 1));
8037
8038 if (PageBuddy(page_head) && page_order(page_head) >= order)
8039 break;
8040 }
8041 spin_unlock_irqrestore(&zone->lock, flags);
8042
8043 return order < MAX_ORDER;
8044}