Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs
[linux/fpc-iii.git] / mm / page_alloc.c
blobe3758a09a009747bd17cb75442d0fcae69a74cc4
1 /*
2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
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)
17 #include <linux/stddef.h>
18 #include <linux/mm.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
68 #include "internal.h"
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
73 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74 DEFINE_PER_CPU(int, numa_node);
75 EXPORT_PER_CPU_SYMBOL(numa_node);
76 #endif
78 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
85 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
86 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
87 #endif
90 * Array of node states.
92 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
93 [N_POSSIBLE] = NODE_MASK_ALL,
94 [N_ONLINE] = { { [0] = 1UL } },
95 #ifndef CONFIG_NUMA
96 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
97 #ifdef CONFIG_HIGHMEM
98 [N_HIGH_MEMORY] = { { [0] = 1UL } },
99 #endif
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY] = { { [0] = 1UL } },
102 #endif
103 [N_CPU] = { { [0] = 1UL } },
104 #endif /* NUMA */
106 EXPORT_SYMBOL(node_states);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock);
111 unsigned long totalram_pages __read_mostly;
112 unsigned long totalreserve_pages __read_mostly;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly;
121 int percpu_pagelist_fraction;
122 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex));
139 if (saved_gfp_mask) {
140 gfp_allowed_mask = saved_gfp_mask;
141 saved_gfp_mask = 0;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex));
148 WARN_ON(saved_gfp_mask);
149 saved_gfp_mask = gfp_allowed_mask;
150 gfp_allowed_mask &= ~GFP_IOFS;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
156 return false;
157 return true;
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly;
163 #endif
165 static void __free_pages_ok(struct page *page, unsigned int order);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
179 #ifdef CONFIG_ZONE_DMA
180 256,
181 #endif
182 #ifdef CONFIG_ZONE_DMA32
183 256,
184 #endif
185 #ifdef CONFIG_HIGHMEM
187 #endif
191 EXPORT_SYMBOL(totalram_pages);
193 static char * const zone_names[MAX_NR_ZONES] = {
194 #ifdef CONFIG_ZONE_DMA
195 "DMA",
196 #endif
197 #ifdef CONFIG_ZONE_DMA32
198 "DMA32",
199 #endif
200 "Normal",
201 #ifdef CONFIG_HIGHMEM
202 "HighMem",
203 #endif
204 "Movable",
207 int min_free_kbytes = 1024;
208 int user_min_free_kbytes = -1;
210 static unsigned long __meminitdata nr_kernel_pages;
211 static unsigned long __meminitdata nr_all_pages;
212 static unsigned long __meminitdata dma_reserve;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
217 static unsigned long __initdata required_kernelcore;
218 static unsigned long __initdata required_movablecore;
219 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
222 int movable_zone;
223 EXPORT_SYMBOL(movable_zone);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
226 #if MAX_NUMNODES > 1
227 int nr_node_ids __read_mostly = MAX_NUMNODES;
228 int nr_online_nodes __read_mostly = 1;
229 EXPORT_SYMBOL(nr_node_ids);
230 EXPORT_SYMBOL(nr_online_nodes);
231 #endif
233 int page_group_by_mobility_disabled __read_mostly;
235 void set_pageblock_migratetype(struct page *page, int migratetype)
237 if (unlikely(page_group_by_mobility_disabled &&
238 migratetype < MIGRATE_PCPTYPES))
239 migratetype = MIGRATE_UNMOVABLE;
241 set_pageblock_flags_group(page, (unsigned long)migratetype,
242 PB_migrate, PB_migrate_end);
245 bool oom_killer_disabled __read_mostly;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
250 int ret = 0;
251 unsigned seq;
252 unsigned long pfn = page_to_pfn(page);
253 unsigned long sp, start_pfn;
255 do {
256 seq = zone_span_seqbegin(zone);
257 start_pfn = zone->zone_start_pfn;
258 sp = zone->spanned_pages;
259 if (!zone_spans_pfn(zone, pfn))
260 ret = 1;
261 } while (zone_span_seqretry(zone, seq));
263 if (ret)
264 pr_err("page %lu outside zone [ %lu - %lu ]\n",
265 pfn, start_pfn, start_pfn + sp);
267 return ret;
270 static int page_is_consistent(struct zone *zone, struct page *page)
272 if (!pfn_valid_within(page_to_pfn(page)))
273 return 0;
274 if (zone != page_zone(page))
275 return 0;
277 return 1;
280 * Temporary debugging check for pages not lying within a given zone.
282 static int bad_range(struct zone *zone, struct page *page)
284 if (page_outside_zone_boundaries(zone, page))
285 return 1;
286 if (!page_is_consistent(zone, page))
287 return 1;
289 return 0;
291 #else
292 static inline int bad_range(struct zone *zone, struct page *page)
294 return 0;
296 #endif
298 static void bad_page(struct page *page, char *reason, unsigned long bad_flags)
300 static unsigned long resume;
301 static unsigned long nr_shown;
302 static unsigned long nr_unshown;
304 /* Don't complain about poisoned pages */
305 if (PageHWPoison(page)) {
306 page_mapcount_reset(page); /* remove PageBuddy */
307 return;
311 * Allow a burst of 60 reports, then keep quiet for that minute;
312 * or allow a steady drip of one report per second.
314 if (nr_shown == 60) {
315 if (time_before(jiffies, resume)) {
316 nr_unshown++;
317 goto out;
319 if (nr_unshown) {
320 printk(KERN_ALERT
321 "BUG: Bad page state: %lu messages suppressed\n",
322 nr_unshown);
323 nr_unshown = 0;
325 nr_shown = 0;
327 if (nr_shown++ == 0)
328 resume = jiffies + 60 * HZ;
330 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
331 current->comm, page_to_pfn(page));
332 dump_page_badflags(page, reason, bad_flags);
334 print_modules();
335 dump_stack();
336 out:
337 /* Leave bad fields for debug, except PageBuddy could make trouble */
338 page_mapcount_reset(page); /* remove PageBuddy */
339 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
343 * Higher-order pages are called "compound pages". They are structured thusly:
345 * The first PAGE_SIZE page is called the "head page".
347 * The remaining PAGE_SIZE pages are called "tail pages".
349 * All pages have PG_compound set. All tail pages have their ->first_page
350 * pointing at the head page.
352 * The first tail page's ->lru.next holds the address of the compound page's
353 * put_page() function. Its ->lru.prev holds the order of allocation.
354 * This usage means that zero-order pages may not be compound.
357 static void free_compound_page(struct page *page)
359 __free_pages_ok(page, compound_order(page));
362 void prep_compound_page(struct page *page, unsigned long order)
364 int i;
365 int nr_pages = 1 << order;
367 set_compound_page_dtor(page, free_compound_page);
368 set_compound_order(page, order);
369 __SetPageHead(page);
370 for (i = 1; i < nr_pages; i++) {
371 struct page *p = page + i;
372 __SetPageTail(p);
373 set_page_count(p, 0);
374 p->first_page = page;
378 /* update __split_huge_page_refcount if you change this function */
379 static int destroy_compound_page(struct page *page, unsigned long order)
381 int i;
382 int nr_pages = 1 << order;
383 int bad = 0;
385 if (unlikely(compound_order(page) != order)) {
386 bad_page(page, "wrong compound order", 0);
387 bad++;
390 __ClearPageHead(page);
392 for (i = 1; i < nr_pages; i++) {
393 struct page *p = page + i;
395 if (unlikely(!PageTail(p))) {
396 bad_page(page, "PageTail not set", 0);
397 bad++;
398 } else if (unlikely(p->first_page != page)) {
399 bad_page(page, "first_page not consistent", 0);
400 bad++;
402 __ClearPageTail(p);
405 return bad;
408 static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
410 int i;
413 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
414 * and __GFP_HIGHMEM from hard or soft interrupt context.
416 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
417 for (i = 0; i < (1 << order); i++)
418 clear_highpage(page + i);
421 #ifdef CONFIG_DEBUG_PAGEALLOC
422 unsigned int _debug_guardpage_minorder;
424 static int __init debug_guardpage_minorder_setup(char *buf)
426 unsigned long res;
428 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
429 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
430 return 0;
432 _debug_guardpage_minorder = res;
433 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
434 return 0;
436 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
438 static inline void set_page_guard_flag(struct page *page)
440 __set_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
443 static inline void clear_page_guard_flag(struct page *page)
445 __clear_bit(PAGE_DEBUG_FLAG_GUARD, &page->debug_flags);
447 #else
448 static inline void set_page_guard_flag(struct page *page) { }
449 static inline void clear_page_guard_flag(struct page *page) { }
450 #endif
452 static inline void set_page_order(struct page *page, int order)
454 set_page_private(page, order);
455 __SetPageBuddy(page);
458 static inline void rmv_page_order(struct page *page)
460 __ClearPageBuddy(page);
461 set_page_private(page, 0);
465 * Locate the struct page for both the matching buddy in our
466 * pair (buddy1) and the combined O(n+1) page they form (page).
468 * 1) Any buddy B1 will have an order O twin B2 which satisfies
469 * the following equation:
470 * B2 = B1 ^ (1 << O)
471 * For example, if the starting buddy (buddy2) is #8 its order
472 * 1 buddy is #10:
473 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
475 * 2) Any buddy B will have an order O+1 parent P which
476 * satisfies the following equation:
477 * P = B & ~(1 << O)
479 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
481 static inline unsigned long
482 __find_buddy_index(unsigned long page_idx, unsigned int order)
484 return page_idx ^ (1 << order);
488 * This function checks whether a page is free && is the buddy
489 * we can do coalesce a page and its buddy if
490 * (a) the buddy is not in a hole &&
491 * (b) the buddy is in the buddy system &&
492 * (c) a page and its buddy have the same order &&
493 * (d) a page and its buddy are in the same zone.
495 * For recording whether a page is in the buddy system, we set ->_mapcount
496 * PAGE_BUDDY_MAPCOUNT_VALUE.
497 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
498 * serialized by zone->lock.
500 * For recording page's order, we use page_private(page).
502 static inline int page_is_buddy(struct page *page, struct page *buddy,
503 int order)
505 if (!pfn_valid_within(page_to_pfn(buddy)))
506 return 0;
508 if (page_zone_id(page) != page_zone_id(buddy))
509 return 0;
511 if (page_is_guard(buddy) && page_order(buddy) == order) {
512 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
513 return 1;
516 if (PageBuddy(buddy) && page_order(buddy) == order) {
517 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
518 return 1;
520 return 0;
524 * Freeing function for a buddy system allocator.
526 * The concept of a buddy system is to maintain direct-mapped table
527 * (containing bit values) for memory blocks of various "orders".
528 * The bottom level table contains the map for the smallest allocatable
529 * units of memory (here, pages), and each level above it describes
530 * pairs of units from the levels below, hence, "buddies".
531 * At a high level, all that happens here is marking the table entry
532 * at the bottom level available, and propagating the changes upward
533 * as necessary, plus some accounting needed to play nicely with other
534 * parts of the VM system.
535 * At each level, we keep a list of pages, which are heads of continuous
536 * free pages of length of (1 << order) and marked with _mapcount
537 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
538 * field.
539 * So when we are allocating or freeing one, we can derive the state of the
540 * other. That is, if we allocate a small block, and both were
541 * free, the remainder of the region must be split into blocks.
542 * If a block is freed, and its buddy is also free, then this
543 * triggers coalescing into a block of larger size.
545 * -- nyc
548 static inline void __free_one_page(struct page *page,
549 struct zone *zone, unsigned int order,
550 int migratetype)
552 unsigned long page_idx;
553 unsigned long combined_idx;
554 unsigned long uninitialized_var(buddy_idx);
555 struct page *buddy;
557 VM_BUG_ON(!zone_is_initialized(zone));
559 if (unlikely(PageCompound(page)))
560 if (unlikely(destroy_compound_page(page, order)))
561 return;
563 VM_BUG_ON(migratetype == -1);
565 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
567 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
568 VM_BUG_ON_PAGE(bad_range(zone, page), page);
570 while (order < MAX_ORDER-1) {
571 buddy_idx = __find_buddy_index(page_idx, order);
572 buddy = page + (buddy_idx - page_idx);
573 if (!page_is_buddy(page, buddy, order))
574 break;
576 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
577 * merge with it and move up one order.
579 if (page_is_guard(buddy)) {
580 clear_page_guard_flag(buddy);
581 set_page_private(page, 0);
582 __mod_zone_freepage_state(zone, 1 << order,
583 migratetype);
584 } else {
585 list_del(&buddy->lru);
586 zone->free_area[order].nr_free--;
587 rmv_page_order(buddy);
589 combined_idx = buddy_idx & page_idx;
590 page = page + (combined_idx - page_idx);
591 page_idx = combined_idx;
592 order++;
594 set_page_order(page, order);
597 * If this is not the largest possible page, check if the buddy
598 * of the next-highest order is free. If it is, it's possible
599 * that pages are being freed that will coalesce soon. In case,
600 * that is happening, add the free page to the tail of the list
601 * so it's less likely to be used soon and more likely to be merged
602 * as a higher order page
604 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
605 struct page *higher_page, *higher_buddy;
606 combined_idx = buddy_idx & page_idx;
607 higher_page = page + (combined_idx - page_idx);
608 buddy_idx = __find_buddy_index(combined_idx, order + 1);
609 higher_buddy = higher_page + (buddy_idx - combined_idx);
610 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
611 list_add_tail(&page->lru,
612 &zone->free_area[order].free_list[migratetype]);
613 goto out;
617 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
618 out:
619 zone->free_area[order].nr_free++;
622 static inline int free_pages_check(struct page *page)
624 char *bad_reason = NULL;
625 unsigned long bad_flags = 0;
627 if (unlikely(page_mapcount(page)))
628 bad_reason = "nonzero mapcount";
629 if (unlikely(page->mapping != NULL))
630 bad_reason = "non-NULL mapping";
631 if (unlikely(atomic_read(&page->_count) != 0))
632 bad_reason = "nonzero _count";
633 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
634 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
635 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
637 if (unlikely(mem_cgroup_bad_page_check(page)))
638 bad_reason = "cgroup check failed";
639 if (unlikely(bad_reason)) {
640 bad_page(page, bad_reason, bad_flags);
641 return 1;
643 page_cpupid_reset_last(page);
644 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
645 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
646 return 0;
650 * Frees a number of pages from the PCP lists
651 * Assumes all pages on list are in same zone, and of same order.
652 * count is the number of pages to free.
654 * If the zone was previously in an "all pages pinned" state then look to
655 * see if this freeing clears that state.
657 * And clear the zone's pages_scanned counter, to hold off the "all pages are
658 * pinned" detection logic.
660 static void free_pcppages_bulk(struct zone *zone, int count,
661 struct per_cpu_pages *pcp)
663 int migratetype = 0;
664 int batch_free = 0;
665 int to_free = count;
667 spin_lock(&zone->lock);
668 zone->pages_scanned = 0;
670 while (to_free) {
671 struct page *page;
672 struct list_head *list;
675 * Remove pages from lists in a round-robin fashion. A
676 * batch_free count is maintained that is incremented when an
677 * empty list is encountered. This is so more pages are freed
678 * off fuller lists instead of spinning excessively around empty
679 * lists
681 do {
682 batch_free++;
683 if (++migratetype == MIGRATE_PCPTYPES)
684 migratetype = 0;
685 list = &pcp->lists[migratetype];
686 } while (list_empty(list));
688 /* This is the only non-empty list. Free them all. */
689 if (batch_free == MIGRATE_PCPTYPES)
690 batch_free = to_free;
692 do {
693 int mt; /* migratetype of the to-be-freed page */
695 page = list_entry(list->prev, struct page, lru);
696 /* must delete as __free_one_page list manipulates */
697 list_del(&page->lru);
698 mt = get_freepage_migratetype(page);
699 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
700 __free_one_page(page, zone, 0, mt);
701 trace_mm_page_pcpu_drain(page, 0, mt);
702 if (likely(!is_migrate_isolate_page(page))) {
703 __mod_zone_page_state(zone, NR_FREE_PAGES, 1);
704 if (is_migrate_cma(mt))
705 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, 1);
707 } while (--to_free && --batch_free && !list_empty(list));
709 spin_unlock(&zone->lock);
712 static void free_one_page(struct zone *zone, struct page *page, int order,
713 int migratetype)
715 spin_lock(&zone->lock);
716 zone->pages_scanned = 0;
718 __free_one_page(page, zone, order, migratetype);
719 if (unlikely(!is_migrate_isolate(migratetype)))
720 __mod_zone_freepage_state(zone, 1 << order, migratetype);
721 spin_unlock(&zone->lock);
724 static bool free_pages_prepare(struct page *page, unsigned int order)
726 int i;
727 int bad = 0;
729 trace_mm_page_free(page, order);
730 kmemcheck_free_shadow(page, order);
732 if (PageAnon(page))
733 page->mapping = NULL;
734 for (i = 0; i < (1 << order); i++)
735 bad += free_pages_check(page + i);
736 if (bad)
737 return false;
739 if (!PageHighMem(page)) {
740 debug_check_no_locks_freed(page_address(page),
741 PAGE_SIZE << order);
742 debug_check_no_obj_freed(page_address(page),
743 PAGE_SIZE << order);
745 arch_free_page(page, order);
746 kernel_map_pages(page, 1 << order, 0);
748 return true;
751 static void __free_pages_ok(struct page *page, unsigned int order)
753 unsigned long flags;
754 int migratetype;
756 if (!free_pages_prepare(page, order))
757 return;
759 local_irq_save(flags);
760 __count_vm_events(PGFREE, 1 << order);
761 migratetype = get_pageblock_migratetype(page);
762 set_freepage_migratetype(page, migratetype);
763 free_one_page(page_zone(page), page, order, migratetype);
764 local_irq_restore(flags);
767 void __init __free_pages_bootmem(struct page *page, unsigned int order)
769 unsigned int nr_pages = 1 << order;
770 struct page *p = page;
771 unsigned int loop;
773 prefetchw(p);
774 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
775 prefetchw(p + 1);
776 __ClearPageReserved(p);
777 set_page_count(p, 0);
779 __ClearPageReserved(p);
780 set_page_count(p, 0);
782 page_zone(page)->managed_pages += nr_pages;
783 set_page_refcounted(page);
784 __free_pages(page, order);
787 #ifdef CONFIG_CMA
788 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
789 void __init init_cma_reserved_pageblock(struct page *page)
791 unsigned i = pageblock_nr_pages;
792 struct page *p = page;
794 do {
795 __ClearPageReserved(p);
796 set_page_count(p, 0);
797 } while (++p, --i);
799 set_page_refcounted(page);
800 set_pageblock_migratetype(page, MIGRATE_CMA);
801 __free_pages(page, pageblock_order);
802 adjust_managed_page_count(page, pageblock_nr_pages);
804 #endif
807 * The order of subdivision here is critical for the IO subsystem.
808 * Please do not alter this order without good reasons and regression
809 * testing. Specifically, as large blocks of memory are subdivided,
810 * the order in which smaller blocks are delivered depends on the order
811 * they're subdivided in this function. This is the primary factor
812 * influencing the order in which pages are delivered to the IO
813 * subsystem according to empirical testing, and this is also justified
814 * by considering the behavior of a buddy system containing a single
815 * large block of memory acted on by a series of small allocations.
816 * This behavior is a critical factor in sglist merging's success.
818 * -- nyc
820 static inline void expand(struct zone *zone, struct page *page,
821 int low, int high, struct free_area *area,
822 int migratetype)
824 unsigned long size = 1 << high;
826 while (high > low) {
827 area--;
828 high--;
829 size >>= 1;
830 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
832 #ifdef CONFIG_DEBUG_PAGEALLOC
833 if (high < debug_guardpage_minorder()) {
835 * Mark as guard pages (or page), that will allow to
836 * merge back to allocator when buddy will be freed.
837 * Corresponding page table entries will not be touched,
838 * pages will stay not present in virtual address space
840 INIT_LIST_HEAD(&page[size].lru);
841 set_page_guard_flag(&page[size]);
842 set_page_private(&page[size], high);
843 /* Guard pages are not available for any usage */
844 __mod_zone_freepage_state(zone, -(1 << high),
845 migratetype);
846 continue;
848 #endif
849 list_add(&page[size].lru, &area->free_list[migratetype]);
850 area->nr_free++;
851 set_page_order(&page[size], high);
856 * This page is about to be returned from the page allocator
858 static inline int check_new_page(struct page *page)
860 char *bad_reason = NULL;
861 unsigned long bad_flags = 0;
863 if (unlikely(page_mapcount(page)))
864 bad_reason = "nonzero mapcount";
865 if (unlikely(page->mapping != NULL))
866 bad_reason = "non-NULL mapping";
867 if (unlikely(atomic_read(&page->_count) != 0))
868 bad_reason = "nonzero _count";
869 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
870 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
871 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
873 if (unlikely(mem_cgroup_bad_page_check(page)))
874 bad_reason = "cgroup check failed";
875 if (unlikely(bad_reason)) {
876 bad_page(page, bad_reason, bad_flags);
877 return 1;
879 return 0;
882 static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
884 int i;
886 for (i = 0; i < (1 << order); i++) {
887 struct page *p = page + i;
888 if (unlikely(check_new_page(p)))
889 return 1;
892 set_page_private(page, 0);
893 set_page_refcounted(page);
895 arch_alloc_page(page, order);
896 kernel_map_pages(page, 1 << order, 1);
898 if (gfp_flags & __GFP_ZERO)
899 prep_zero_page(page, order, gfp_flags);
901 if (order && (gfp_flags & __GFP_COMP))
902 prep_compound_page(page, order);
904 return 0;
908 * Go through the free lists for the given migratetype and remove
909 * the smallest available page from the freelists
911 static inline
912 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
913 int migratetype)
915 unsigned int current_order;
916 struct free_area *area;
917 struct page *page;
919 /* Find a page of the appropriate size in the preferred list */
920 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
921 area = &(zone->free_area[current_order]);
922 if (list_empty(&area->free_list[migratetype]))
923 continue;
925 page = list_entry(area->free_list[migratetype].next,
926 struct page, lru);
927 list_del(&page->lru);
928 rmv_page_order(page);
929 area->nr_free--;
930 expand(zone, page, order, current_order, area, migratetype);
931 return page;
934 return NULL;
939 * This array describes the order lists are fallen back to when
940 * the free lists for the desirable migrate type are depleted
942 static int fallbacks[MIGRATE_TYPES][4] = {
943 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
944 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
945 #ifdef CONFIG_CMA
946 [MIGRATE_MOVABLE] = { MIGRATE_CMA, MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
947 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
948 #else
949 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
950 #endif
951 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
952 #ifdef CONFIG_MEMORY_ISOLATION
953 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
954 #endif
958 * Move the free pages in a range to the free lists of the requested type.
959 * Note that start_page and end_pages are not aligned on a pageblock
960 * boundary. If alignment is required, use move_freepages_block()
962 int move_freepages(struct zone *zone,
963 struct page *start_page, struct page *end_page,
964 int migratetype)
966 struct page *page;
967 unsigned long order;
968 int pages_moved = 0;
970 #ifndef CONFIG_HOLES_IN_ZONE
972 * page_zone is not safe to call in this context when
973 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
974 * anyway as we check zone boundaries in move_freepages_block().
975 * Remove at a later date when no bug reports exist related to
976 * grouping pages by mobility
978 BUG_ON(page_zone(start_page) != page_zone(end_page));
979 #endif
981 for (page = start_page; page <= end_page;) {
982 /* Make sure we are not inadvertently changing nodes */
983 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
985 if (!pfn_valid_within(page_to_pfn(page))) {
986 page++;
987 continue;
990 if (!PageBuddy(page)) {
991 page++;
992 continue;
995 order = page_order(page);
996 list_move(&page->lru,
997 &zone->free_area[order].free_list[migratetype]);
998 set_freepage_migratetype(page, migratetype);
999 page += 1 << order;
1000 pages_moved += 1 << order;
1003 return pages_moved;
1006 int move_freepages_block(struct zone *zone, struct page *page,
1007 int migratetype)
1009 unsigned long start_pfn, end_pfn;
1010 struct page *start_page, *end_page;
1012 start_pfn = page_to_pfn(page);
1013 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1014 start_page = pfn_to_page(start_pfn);
1015 end_page = start_page + pageblock_nr_pages - 1;
1016 end_pfn = start_pfn + pageblock_nr_pages - 1;
1018 /* Do not cross zone boundaries */
1019 if (!zone_spans_pfn(zone, start_pfn))
1020 start_page = page;
1021 if (!zone_spans_pfn(zone, end_pfn))
1022 return 0;
1024 return move_freepages(zone, start_page, end_page, migratetype);
1027 static void change_pageblock_range(struct page *pageblock_page,
1028 int start_order, int migratetype)
1030 int nr_pageblocks = 1 << (start_order - pageblock_order);
1032 while (nr_pageblocks--) {
1033 set_pageblock_migratetype(pageblock_page, migratetype);
1034 pageblock_page += pageblock_nr_pages;
1039 * If breaking a large block of pages, move all free pages to the preferred
1040 * allocation list. If falling back for a reclaimable kernel allocation, be
1041 * more aggressive about taking ownership of free pages.
1043 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1044 * nor move CMA pages to different free lists. We don't want unmovable pages
1045 * to be allocated from MIGRATE_CMA areas.
1047 * Returns the new migratetype of the pageblock (or the same old migratetype
1048 * if it was unchanged).
1050 static int try_to_steal_freepages(struct zone *zone, struct page *page,
1051 int start_type, int fallback_type)
1053 int current_order = page_order(page);
1056 * When borrowing from MIGRATE_CMA, we need to release the excess
1057 * buddy pages to CMA itself.
1059 if (is_migrate_cma(fallback_type))
1060 return fallback_type;
1062 /* Take ownership for orders >= pageblock_order */
1063 if (current_order >= pageblock_order) {
1064 change_pageblock_range(page, current_order, start_type);
1065 return start_type;
1068 if (current_order >= pageblock_order / 2 ||
1069 start_type == MIGRATE_RECLAIMABLE ||
1070 page_group_by_mobility_disabled) {
1071 int pages;
1073 pages = move_freepages_block(zone, page, start_type);
1075 /* Claim the whole block if over half of it is free */
1076 if (pages >= (1 << (pageblock_order-1)) ||
1077 page_group_by_mobility_disabled) {
1079 set_pageblock_migratetype(page, start_type);
1080 return start_type;
1085 return fallback_type;
1088 /* Remove an element from the buddy allocator from the fallback list */
1089 static inline struct page *
1090 __rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
1092 struct free_area *area;
1093 int current_order;
1094 struct page *page;
1095 int migratetype, new_type, i;
1097 /* Find the largest possible block of pages in the other list */
1098 for (current_order = MAX_ORDER-1; current_order >= order;
1099 --current_order) {
1100 for (i = 0;; i++) {
1101 migratetype = fallbacks[start_migratetype][i];
1103 /* MIGRATE_RESERVE handled later if necessary */
1104 if (migratetype == MIGRATE_RESERVE)
1105 break;
1107 area = &(zone->free_area[current_order]);
1108 if (list_empty(&area->free_list[migratetype]))
1109 continue;
1111 page = list_entry(area->free_list[migratetype].next,
1112 struct page, lru);
1113 area->nr_free--;
1115 new_type = try_to_steal_freepages(zone, page,
1116 start_migratetype,
1117 migratetype);
1119 /* Remove the page from the freelists */
1120 list_del(&page->lru);
1121 rmv_page_order(page);
1123 expand(zone, page, order, current_order, area,
1124 new_type);
1126 trace_mm_page_alloc_extfrag(page, order, current_order,
1127 start_migratetype, migratetype, new_type);
1129 return page;
1133 return NULL;
1137 * Do the hard work of removing an element from the buddy allocator.
1138 * Call me with the zone->lock already held.
1140 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1141 int migratetype)
1143 struct page *page;
1145 retry_reserve:
1146 page = __rmqueue_smallest(zone, order, migratetype);
1148 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1149 page = __rmqueue_fallback(zone, order, migratetype);
1152 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1153 * is used because __rmqueue_smallest is an inline function
1154 * and we want just one call site
1156 if (!page) {
1157 migratetype = MIGRATE_RESERVE;
1158 goto retry_reserve;
1162 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1163 return page;
1167 * Obtain a specified number of elements from the buddy allocator, all under
1168 * a single hold of the lock, for efficiency. Add them to the supplied list.
1169 * Returns the number of new pages which were placed at *list.
1171 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1172 unsigned long count, struct list_head *list,
1173 int migratetype, int cold)
1175 int mt = migratetype, i;
1177 spin_lock(&zone->lock);
1178 for (i = 0; i < count; ++i) {
1179 struct page *page = __rmqueue(zone, order, migratetype);
1180 if (unlikely(page == NULL))
1181 break;
1184 * Split buddy pages returned by expand() are received here
1185 * in physical page order. The page is added to the callers and
1186 * list and the list head then moves forward. From the callers
1187 * perspective, the linked list is ordered by page number in
1188 * some conditions. This is useful for IO devices that can
1189 * merge IO requests if the physical pages are ordered
1190 * properly.
1192 if (likely(cold == 0))
1193 list_add(&page->lru, list);
1194 else
1195 list_add_tail(&page->lru, list);
1196 if (IS_ENABLED(CONFIG_CMA)) {
1197 mt = get_pageblock_migratetype(page);
1198 if (!is_migrate_cma(mt) && !is_migrate_isolate(mt))
1199 mt = migratetype;
1201 set_freepage_migratetype(page, mt);
1202 list = &page->lru;
1203 if (is_migrate_cma(mt))
1204 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1205 -(1 << order));
1207 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1208 spin_unlock(&zone->lock);
1209 return i;
1212 #ifdef CONFIG_NUMA
1214 * Called from the vmstat counter updater to drain pagesets of this
1215 * currently executing processor on remote nodes after they have
1216 * expired.
1218 * Note that this function must be called with the thread pinned to
1219 * a single processor.
1221 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1223 unsigned long flags;
1224 int to_drain;
1225 unsigned long batch;
1227 local_irq_save(flags);
1228 batch = ACCESS_ONCE(pcp->batch);
1229 if (pcp->count >= batch)
1230 to_drain = batch;
1231 else
1232 to_drain = pcp->count;
1233 if (to_drain > 0) {
1234 free_pcppages_bulk(zone, to_drain, pcp);
1235 pcp->count -= to_drain;
1237 local_irq_restore(flags);
1239 #endif
1242 * Drain pages of the indicated processor.
1244 * The processor must either be the current processor and the
1245 * thread pinned to the current processor or a processor that
1246 * is not online.
1248 static void drain_pages(unsigned int cpu)
1250 unsigned long flags;
1251 struct zone *zone;
1253 for_each_populated_zone(zone) {
1254 struct per_cpu_pageset *pset;
1255 struct per_cpu_pages *pcp;
1257 local_irq_save(flags);
1258 pset = per_cpu_ptr(zone->pageset, cpu);
1260 pcp = &pset->pcp;
1261 if (pcp->count) {
1262 free_pcppages_bulk(zone, pcp->count, pcp);
1263 pcp->count = 0;
1265 local_irq_restore(flags);
1270 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1272 void drain_local_pages(void *arg)
1274 drain_pages(smp_processor_id());
1278 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1280 * Note that this code is protected against sending an IPI to an offline
1281 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1282 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1283 * nothing keeps CPUs from showing up after we populated the cpumask and
1284 * before the call to on_each_cpu_mask().
1286 void drain_all_pages(void)
1288 int cpu;
1289 struct per_cpu_pageset *pcp;
1290 struct zone *zone;
1293 * Allocate in the BSS so we wont require allocation in
1294 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1296 static cpumask_t cpus_with_pcps;
1299 * We don't care about racing with CPU hotplug event
1300 * as offline notification will cause the notified
1301 * cpu to drain that CPU pcps and on_each_cpu_mask
1302 * disables preemption as part of its processing
1304 for_each_online_cpu(cpu) {
1305 bool has_pcps = false;
1306 for_each_populated_zone(zone) {
1307 pcp = per_cpu_ptr(zone->pageset, cpu);
1308 if (pcp->pcp.count) {
1309 has_pcps = true;
1310 break;
1313 if (has_pcps)
1314 cpumask_set_cpu(cpu, &cpus_with_pcps);
1315 else
1316 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1318 on_each_cpu_mask(&cpus_with_pcps, drain_local_pages, NULL, 1);
1321 #ifdef CONFIG_HIBERNATION
1323 void mark_free_pages(struct zone *zone)
1325 unsigned long pfn, max_zone_pfn;
1326 unsigned long flags;
1327 int order, t;
1328 struct list_head *curr;
1330 if (zone_is_empty(zone))
1331 return;
1333 spin_lock_irqsave(&zone->lock, flags);
1335 max_zone_pfn = zone_end_pfn(zone);
1336 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1337 if (pfn_valid(pfn)) {
1338 struct page *page = pfn_to_page(pfn);
1340 if (!swsusp_page_is_forbidden(page))
1341 swsusp_unset_page_free(page);
1344 for_each_migratetype_order(order, t) {
1345 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1346 unsigned long i;
1348 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1349 for (i = 0; i < (1UL << order); i++)
1350 swsusp_set_page_free(pfn_to_page(pfn + i));
1353 spin_unlock_irqrestore(&zone->lock, flags);
1355 #endif /* CONFIG_PM */
1358 * Free a 0-order page
1359 * cold == 1 ? free a cold page : free a hot page
1361 void free_hot_cold_page(struct page *page, int cold)
1363 struct zone *zone = page_zone(page);
1364 struct per_cpu_pages *pcp;
1365 unsigned long flags;
1366 int migratetype;
1368 if (!free_pages_prepare(page, 0))
1369 return;
1371 migratetype = get_pageblock_migratetype(page);
1372 set_freepage_migratetype(page, migratetype);
1373 local_irq_save(flags);
1374 __count_vm_event(PGFREE);
1377 * We only track unmovable, reclaimable and movable on pcp lists.
1378 * Free ISOLATE pages back to the allocator because they are being
1379 * offlined but treat RESERVE as movable pages so we can get those
1380 * areas back if necessary. Otherwise, we may have to free
1381 * excessively into the page allocator
1383 if (migratetype >= MIGRATE_PCPTYPES) {
1384 if (unlikely(is_migrate_isolate(migratetype))) {
1385 free_one_page(zone, page, 0, migratetype);
1386 goto out;
1388 migratetype = MIGRATE_MOVABLE;
1391 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1392 if (cold)
1393 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1394 else
1395 list_add(&page->lru, &pcp->lists[migratetype]);
1396 pcp->count++;
1397 if (pcp->count >= pcp->high) {
1398 unsigned long batch = ACCESS_ONCE(pcp->batch);
1399 free_pcppages_bulk(zone, batch, pcp);
1400 pcp->count -= batch;
1403 out:
1404 local_irq_restore(flags);
1408 * Free a list of 0-order pages
1410 void free_hot_cold_page_list(struct list_head *list, int cold)
1412 struct page *page, *next;
1414 list_for_each_entry_safe(page, next, list, lru) {
1415 trace_mm_page_free_batched(page, cold);
1416 free_hot_cold_page(page, cold);
1421 * split_page takes a non-compound higher-order page, and splits it into
1422 * n (1<<order) sub-pages: page[0..n]
1423 * Each sub-page must be freed individually.
1425 * Note: this is probably too low level an operation for use in drivers.
1426 * Please consult with lkml before using this in your driver.
1428 void split_page(struct page *page, unsigned int order)
1430 int i;
1432 VM_BUG_ON_PAGE(PageCompound(page), page);
1433 VM_BUG_ON_PAGE(!page_count(page), page);
1435 #ifdef CONFIG_KMEMCHECK
1437 * Split shadow pages too, because free(page[0]) would
1438 * otherwise free the whole shadow.
1440 if (kmemcheck_page_is_tracked(page))
1441 split_page(virt_to_page(page[0].shadow), order);
1442 #endif
1444 for (i = 1; i < (1 << order); i++)
1445 set_page_refcounted(page + i);
1447 EXPORT_SYMBOL_GPL(split_page);
1449 static int __isolate_free_page(struct page *page, unsigned int order)
1451 unsigned long watermark;
1452 struct zone *zone;
1453 int mt;
1455 BUG_ON(!PageBuddy(page));
1457 zone = page_zone(page);
1458 mt = get_pageblock_migratetype(page);
1460 if (!is_migrate_isolate(mt)) {
1461 /* Obey watermarks as if the page was being allocated */
1462 watermark = low_wmark_pages(zone) + (1 << order);
1463 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1464 return 0;
1466 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1469 /* Remove page from free list */
1470 list_del(&page->lru);
1471 zone->free_area[order].nr_free--;
1472 rmv_page_order(page);
1474 /* Set the pageblock if the isolated page is at least a pageblock */
1475 if (order >= pageblock_order - 1) {
1476 struct page *endpage = page + (1 << order) - 1;
1477 for (; page < endpage; page += pageblock_nr_pages) {
1478 int mt = get_pageblock_migratetype(page);
1479 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1480 set_pageblock_migratetype(page,
1481 MIGRATE_MOVABLE);
1485 return 1UL << order;
1489 * Similar to split_page except the page is already free. As this is only
1490 * being used for migration, the migratetype of the block also changes.
1491 * As this is called with interrupts disabled, the caller is responsible
1492 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1493 * are enabled.
1495 * Note: this is probably too low level an operation for use in drivers.
1496 * Please consult with lkml before using this in your driver.
1498 int split_free_page(struct page *page)
1500 unsigned int order;
1501 int nr_pages;
1503 order = page_order(page);
1505 nr_pages = __isolate_free_page(page, order);
1506 if (!nr_pages)
1507 return 0;
1509 /* Split into individual pages */
1510 set_page_refcounted(page);
1511 split_page(page, order);
1512 return nr_pages;
1516 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1517 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1518 * or two.
1520 static inline
1521 struct page *buffered_rmqueue(struct zone *preferred_zone,
1522 struct zone *zone, int order, gfp_t gfp_flags,
1523 int migratetype)
1525 unsigned long flags;
1526 struct page *page;
1527 int cold = !!(gfp_flags & __GFP_COLD);
1529 again:
1530 if (likely(order == 0)) {
1531 struct per_cpu_pages *pcp;
1532 struct list_head *list;
1534 local_irq_save(flags);
1535 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1536 list = &pcp->lists[migratetype];
1537 if (list_empty(list)) {
1538 pcp->count += rmqueue_bulk(zone, 0,
1539 pcp->batch, list,
1540 migratetype, cold);
1541 if (unlikely(list_empty(list)))
1542 goto failed;
1545 if (cold)
1546 page = list_entry(list->prev, struct page, lru);
1547 else
1548 page = list_entry(list->next, struct page, lru);
1550 list_del(&page->lru);
1551 pcp->count--;
1552 } else {
1553 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1555 * __GFP_NOFAIL is not to be used in new code.
1557 * All __GFP_NOFAIL callers should be fixed so that they
1558 * properly detect and handle allocation failures.
1560 * We most definitely don't want callers attempting to
1561 * allocate greater than order-1 page units with
1562 * __GFP_NOFAIL.
1564 WARN_ON_ONCE(order > 1);
1566 spin_lock_irqsave(&zone->lock, flags);
1567 page = __rmqueue(zone, order, migratetype);
1568 spin_unlock(&zone->lock);
1569 if (!page)
1570 goto failed;
1571 __mod_zone_freepage_state(zone, -(1 << order),
1572 get_pageblock_migratetype(page));
1575 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1576 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1577 zone_statistics(preferred_zone, zone, gfp_flags);
1578 local_irq_restore(flags);
1580 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1581 if (prep_new_page(page, order, gfp_flags))
1582 goto again;
1583 return page;
1585 failed:
1586 local_irq_restore(flags);
1587 return NULL;
1590 #ifdef CONFIG_FAIL_PAGE_ALLOC
1592 static struct {
1593 struct fault_attr attr;
1595 u32 ignore_gfp_highmem;
1596 u32 ignore_gfp_wait;
1597 u32 min_order;
1598 } fail_page_alloc = {
1599 .attr = FAULT_ATTR_INITIALIZER,
1600 .ignore_gfp_wait = 1,
1601 .ignore_gfp_highmem = 1,
1602 .min_order = 1,
1605 static int __init setup_fail_page_alloc(char *str)
1607 return setup_fault_attr(&fail_page_alloc.attr, str);
1609 __setup("fail_page_alloc=", setup_fail_page_alloc);
1611 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1613 if (order < fail_page_alloc.min_order)
1614 return false;
1615 if (gfp_mask & __GFP_NOFAIL)
1616 return false;
1617 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1618 return false;
1619 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1620 return false;
1622 return should_fail(&fail_page_alloc.attr, 1 << order);
1625 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1627 static int __init fail_page_alloc_debugfs(void)
1629 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1630 struct dentry *dir;
1632 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1633 &fail_page_alloc.attr);
1634 if (IS_ERR(dir))
1635 return PTR_ERR(dir);
1637 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1638 &fail_page_alloc.ignore_gfp_wait))
1639 goto fail;
1640 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1641 &fail_page_alloc.ignore_gfp_highmem))
1642 goto fail;
1643 if (!debugfs_create_u32("min-order", mode, dir,
1644 &fail_page_alloc.min_order))
1645 goto fail;
1647 return 0;
1648 fail:
1649 debugfs_remove_recursive(dir);
1651 return -ENOMEM;
1654 late_initcall(fail_page_alloc_debugfs);
1656 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1658 #else /* CONFIG_FAIL_PAGE_ALLOC */
1660 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1662 return false;
1665 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1668 * Return true if free pages are above 'mark'. This takes into account the order
1669 * of the allocation.
1671 static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1672 int classzone_idx, int alloc_flags, long free_pages)
1674 /* free_pages my go negative - that's OK */
1675 long min = mark;
1676 long lowmem_reserve = z->lowmem_reserve[classzone_idx];
1677 int o;
1678 long free_cma = 0;
1680 free_pages -= (1 << order) - 1;
1681 if (alloc_flags & ALLOC_HIGH)
1682 min -= min / 2;
1683 if (alloc_flags & ALLOC_HARDER)
1684 min -= min / 4;
1685 #ifdef CONFIG_CMA
1686 /* If allocation can't use CMA areas don't use free CMA pages */
1687 if (!(alloc_flags & ALLOC_CMA))
1688 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1689 #endif
1691 if (free_pages - free_cma <= min + lowmem_reserve)
1692 return false;
1693 for (o = 0; o < order; o++) {
1694 /* At the next order, this order's pages become unavailable */
1695 free_pages -= z->free_area[o].nr_free << o;
1697 /* Require fewer higher order pages to be free */
1698 min >>= 1;
1700 if (free_pages <= min)
1701 return false;
1703 return true;
1706 bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1707 int classzone_idx, int alloc_flags)
1709 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1710 zone_page_state(z, NR_FREE_PAGES));
1713 bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1714 int classzone_idx, int alloc_flags)
1716 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1718 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1719 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1721 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1722 free_pages);
1725 #ifdef CONFIG_NUMA
1727 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1728 * skip over zones that are not allowed by the cpuset, or that have
1729 * been recently (in last second) found to be nearly full. See further
1730 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1731 * that have to skip over a lot of full or unallowed zones.
1733 * If the zonelist cache is present in the passed zonelist, then
1734 * returns a pointer to the allowed node mask (either the current
1735 * tasks mems_allowed, or node_states[N_MEMORY].)
1737 * If the zonelist cache is not available for this zonelist, does
1738 * nothing and returns NULL.
1740 * If the fullzones BITMAP in the zonelist cache is stale (more than
1741 * a second since last zap'd) then we zap it out (clear its bits.)
1743 * We hold off even calling zlc_setup, until after we've checked the
1744 * first zone in the zonelist, on the theory that most allocations will
1745 * be satisfied from that first zone, so best to examine that zone as
1746 * quickly as we can.
1748 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1750 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1751 nodemask_t *allowednodes; /* zonelist_cache approximation */
1753 zlc = zonelist->zlcache_ptr;
1754 if (!zlc)
1755 return NULL;
1757 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1758 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1759 zlc->last_full_zap = jiffies;
1762 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1763 &cpuset_current_mems_allowed :
1764 &node_states[N_MEMORY];
1765 return allowednodes;
1769 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1770 * if it is worth looking at further for free memory:
1771 * 1) Check that the zone isn't thought to be full (doesn't have its
1772 * bit set in the zonelist_cache fullzones BITMAP).
1773 * 2) Check that the zones node (obtained from the zonelist_cache
1774 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1775 * Return true (non-zero) if zone is worth looking at further, or
1776 * else return false (zero) if it is not.
1778 * This check -ignores- the distinction between various watermarks,
1779 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1780 * found to be full for any variation of these watermarks, it will
1781 * be considered full for up to one second by all requests, unless
1782 * we are so low on memory on all allowed nodes that we are forced
1783 * into the second scan of the zonelist.
1785 * In the second scan we ignore this zonelist cache and exactly
1786 * apply the watermarks to all zones, even it is slower to do so.
1787 * We are low on memory in the second scan, and should leave no stone
1788 * unturned looking for a free page.
1790 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1791 nodemask_t *allowednodes)
1793 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1794 int i; /* index of *z in zonelist zones */
1795 int n; /* node that zone *z is on */
1797 zlc = zonelist->zlcache_ptr;
1798 if (!zlc)
1799 return 1;
1801 i = z - zonelist->_zonerefs;
1802 n = zlc->z_to_n[i];
1804 /* This zone is worth trying if it is allowed but not full */
1805 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1809 * Given 'z' scanning a zonelist, set the corresponding bit in
1810 * zlc->fullzones, so that subsequent attempts to allocate a page
1811 * from that zone don't waste time re-examining it.
1813 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1815 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1816 int i; /* index of *z in zonelist zones */
1818 zlc = zonelist->zlcache_ptr;
1819 if (!zlc)
1820 return;
1822 i = z - zonelist->_zonerefs;
1824 set_bit(i, zlc->fullzones);
1828 * clear all zones full, called after direct reclaim makes progress so that
1829 * a zone that was recently full is not skipped over for up to a second
1831 static void zlc_clear_zones_full(struct zonelist *zonelist)
1833 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1835 zlc = zonelist->zlcache_ptr;
1836 if (!zlc)
1837 return;
1839 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1842 static bool zone_local(struct zone *local_zone, struct zone *zone)
1844 return local_zone->node == zone->node;
1847 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1849 return node_isset(local_zone->node, zone->zone_pgdat->reclaim_nodes);
1852 static void __paginginit init_zone_allows_reclaim(int nid)
1854 int i;
1856 for_each_online_node(i)
1857 if (node_distance(nid, i) <= RECLAIM_DISTANCE)
1858 node_set(i, NODE_DATA(nid)->reclaim_nodes);
1859 else
1860 zone_reclaim_mode = 1;
1863 #else /* CONFIG_NUMA */
1865 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1867 return NULL;
1870 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1871 nodemask_t *allowednodes)
1873 return 1;
1876 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1880 static void zlc_clear_zones_full(struct zonelist *zonelist)
1884 static bool zone_local(struct zone *local_zone, struct zone *zone)
1886 return true;
1889 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
1891 return true;
1894 static inline void init_zone_allows_reclaim(int nid)
1897 #endif /* CONFIG_NUMA */
1900 * get_page_from_freelist goes through the zonelist trying to allocate
1901 * a page.
1903 static struct page *
1904 get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1905 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1906 struct zone *preferred_zone, int migratetype)
1908 struct zoneref *z;
1909 struct page *page = NULL;
1910 int classzone_idx;
1911 struct zone *zone;
1912 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1913 int zlc_active = 0; /* set if using zonelist_cache */
1914 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1916 classzone_idx = zone_idx(preferred_zone);
1917 zonelist_scan:
1919 * Scan zonelist, looking for a zone with enough free.
1920 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1922 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1923 high_zoneidx, nodemask) {
1924 unsigned long mark;
1926 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
1927 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1928 continue;
1929 if ((alloc_flags & ALLOC_CPUSET) &&
1930 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1931 continue;
1932 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1933 if (unlikely(alloc_flags & ALLOC_NO_WATERMARKS))
1934 goto try_this_zone;
1936 * Distribute pages in proportion to the individual
1937 * zone size to ensure fair page aging. The zone a
1938 * page was allocated in should have no effect on the
1939 * time the page has in memory before being reclaimed.
1941 * Try to stay in local zones in the fastpath. If
1942 * that fails, the slowpath is entered, which will do
1943 * another pass starting with the local zones, but
1944 * ultimately fall back to remote zones that do not
1945 * partake in the fairness round-robin cycle of this
1946 * zonelist.
1948 if (alloc_flags & ALLOC_WMARK_LOW) {
1949 if (zone_page_state(zone, NR_ALLOC_BATCH) <= 0)
1950 continue;
1951 if (!zone_local(preferred_zone, zone))
1952 continue;
1955 * When allocating a page cache page for writing, we
1956 * want to get it from a zone that is within its dirty
1957 * limit, such that no single zone holds more than its
1958 * proportional share of globally allowed dirty pages.
1959 * The dirty limits take into account the zone's
1960 * lowmem reserves and high watermark so that kswapd
1961 * should be able to balance it without having to
1962 * write pages from its LRU list.
1964 * This may look like it could increase pressure on
1965 * lower zones by failing allocations in higher zones
1966 * before they are full. But the pages that do spill
1967 * over are limited as the lower zones are protected
1968 * by this very same mechanism. It should not become
1969 * a practical burden to them.
1971 * XXX: For now, allow allocations to potentially
1972 * exceed the per-zone dirty limit in the slowpath
1973 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1974 * which is important when on a NUMA setup the allowed
1975 * zones are together not big enough to reach the
1976 * global limit. The proper fix for these situations
1977 * will require awareness of zones in the
1978 * dirty-throttling and the flusher threads.
1980 if ((alloc_flags & ALLOC_WMARK_LOW) &&
1981 (gfp_mask & __GFP_WRITE) && !zone_dirty_ok(zone))
1982 goto this_zone_full;
1984 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1985 if (!zone_watermark_ok(zone, order, mark,
1986 classzone_idx, alloc_flags)) {
1987 int ret;
1989 if (IS_ENABLED(CONFIG_NUMA) &&
1990 !did_zlc_setup && nr_online_nodes > 1) {
1992 * we do zlc_setup if there are multiple nodes
1993 * and before considering the first zone allowed
1994 * by the cpuset.
1996 allowednodes = zlc_setup(zonelist, alloc_flags);
1997 zlc_active = 1;
1998 did_zlc_setup = 1;
2001 if (zone_reclaim_mode == 0 ||
2002 !zone_allows_reclaim(preferred_zone, zone))
2003 goto this_zone_full;
2006 * As we may have just activated ZLC, check if the first
2007 * eligible zone has failed zone_reclaim recently.
2009 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2010 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2011 continue;
2013 ret = zone_reclaim(zone, gfp_mask, order);
2014 switch (ret) {
2015 case ZONE_RECLAIM_NOSCAN:
2016 /* did not scan */
2017 continue;
2018 case ZONE_RECLAIM_FULL:
2019 /* scanned but unreclaimable */
2020 continue;
2021 default:
2022 /* did we reclaim enough */
2023 if (zone_watermark_ok(zone, order, mark,
2024 classzone_idx, alloc_flags))
2025 goto try_this_zone;
2028 * Failed to reclaim enough to meet watermark.
2029 * Only mark the zone full if checking the min
2030 * watermark or if we failed to reclaim just
2031 * 1<<order pages or else the page allocator
2032 * fastpath will prematurely mark zones full
2033 * when the watermark is between the low and
2034 * min watermarks.
2036 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2037 ret == ZONE_RECLAIM_SOME)
2038 goto this_zone_full;
2040 continue;
2044 try_this_zone:
2045 page = buffered_rmqueue(preferred_zone, zone, order,
2046 gfp_mask, migratetype);
2047 if (page)
2048 break;
2049 this_zone_full:
2050 if (IS_ENABLED(CONFIG_NUMA))
2051 zlc_mark_zone_full(zonelist, z);
2054 if (unlikely(IS_ENABLED(CONFIG_NUMA) && page == NULL && zlc_active)) {
2055 /* Disable zlc cache for second zonelist scan */
2056 zlc_active = 0;
2057 goto zonelist_scan;
2060 if (page)
2062 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2063 * necessary to allocate the page. The expectation is
2064 * that the caller is taking steps that will free more
2065 * memory. The caller should avoid the page being used
2066 * for !PFMEMALLOC purposes.
2068 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
2070 return page;
2074 * Large machines with many possible nodes should not always dump per-node
2075 * meminfo in irq context.
2077 static inline bool should_suppress_show_mem(void)
2079 bool ret = false;
2081 #if NODES_SHIFT > 8
2082 ret = in_interrupt();
2083 #endif
2084 return ret;
2087 static DEFINE_RATELIMIT_STATE(nopage_rs,
2088 DEFAULT_RATELIMIT_INTERVAL,
2089 DEFAULT_RATELIMIT_BURST);
2091 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2093 unsigned int filter = SHOW_MEM_FILTER_NODES;
2095 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2096 debug_guardpage_minorder() > 0)
2097 return;
2100 * This documents exceptions given to allocations in certain
2101 * contexts that are allowed to allocate outside current's set
2102 * of allowed nodes.
2104 if (!(gfp_mask & __GFP_NOMEMALLOC))
2105 if (test_thread_flag(TIF_MEMDIE) ||
2106 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2107 filter &= ~SHOW_MEM_FILTER_NODES;
2108 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2109 filter &= ~SHOW_MEM_FILTER_NODES;
2111 if (fmt) {
2112 struct va_format vaf;
2113 va_list args;
2115 va_start(args, fmt);
2117 vaf.fmt = fmt;
2118 vaf.va = &args;
2120 pr_warn("%pV", &vaf);
2122 va_end(args);
2125 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2126 current->comm, order, gfp_mask);
2128 dump_stack();
2129 if (!should_suppress_show_mem())
2130 show_mem(filter);
2133 static inline int
2134 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2135 unsigned long did_some_progress,
2136 unsigned long pages_reclaimed)
2138 /* Do not loop if specifically requested */
2139 if (gfp_mask & __GFP_NORETRY)
2140 return 0;
2142 /* Always retry if specifically requested */
2143 if (gfp_mask & __GFP_NOFAIL)
2144 return 1;
2147 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2148 * making forward progress without invoking OOM. Suspend also disables
2149 * storage devices so kswapd will not help. Bail if we are suspending.
2151 if (!did_some_progress && pm_suspended_storage())
2152 return 0;
2155 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2156 * means __GFP_NOFAIL, but that may not be true in other
2157 * implementations.
2159 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2160 return 1;
2163 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2164 * specified, then we retry until we no longer reclaim any pages
2165 * (above), or we've reclaimed an order of pages at least as
2166 * large as the allocation's order. In both cases, if the
2167 * allocation still fails, we stop retrying.
2169 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2170 return 1;
2172 return 0;
2175 static inline struct page *
2176 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2177 struct zonelist *zonelist, enum zone_type high_zoneidx,
2178 nodemask_t *nodemask, struct zone *preferred_zone,
2179 int migratetype)
2181 struct page *page;
2183 /* Acquire the OOM killer lock for the zones in zonelist */
2184 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
2185 schedule_timeout_uninterruptible(1);
2186 return NULL;
2190 * Go through the zonelist yet one more time, keep very high watermark
2191 * here, this is only to catch a parallel oom killing, we must fail if
2192 * we're still under heavy pressure.
2194 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
2195 order, zonelist, high_zoneidx,
2196 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
2197 preferred_zone, migratetype);
2198 if (page)
2199 goto out;
2201 if (!(gfp_mask & __GFP_NOFAIL)) {
2202 /* The OOM killer will not help higher order allocs */
2203 if (order > PAGE_ALLOC_COSTLY_ORDER)
2204 goto out;
2205 /* The OOM killer does not needlessly kill tasks for lowmem */
2206 if (high_zoneidx < ZONE_NORMAL)
2207 goto out;
2209 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2210 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2211 * The caller should handle page allocation failure by itself if
2212 * it specifies __GFP_THISNODE.
2213 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2215 if (gfp_mask & __GFP_THISNODE)
2216 goto out;
2218 /* Exhausted what can be done so it's blamo time */
2219 out_of_memory(zonelist, gfp_mask, order, nodemask, false);
2221 out:
2222 clear_zonelist_oom(zonelist, gfp_mask);
2223 return page;
2226 #ifdef CONFIG_COMPACTION
2227 /* Try memory compaction for high-order allocations before reclaim */
2228 static struct page *
2229 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2230 struct zonelist *zonelist, enum zone_type high_zoneidx,
2231 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2232 int migratetype, bool sync_migration,
2233 bool *contended_compaction, bool *deferred_compaction,
2234 unsigned long *did_some_progress)
2236 if (!order)
2237 return NULL;
2239 if (compaction_deferred(preferred_zone, order)) {
2240 *deferred_compaction = true;
2241 return NULL;
2244 current->flags |= PF_MEMALLOC;
2245 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
2246 nodemask, sync_migration,
2247 contended_compaction);
2248 current->flags &= ~PF_MEMALLOC;
2250 if (*did_some_progress != COMPACT_SKIPPED) {
2251 struct page *page;
2253 /* Page migration frees to the PCP lists but we want merging */
2254 drain_pages(get_cpu());
2255 put_cpu();
2257 page = get_page_from_freelist(gfp_mask, nodemask,
2258 order, zonelist, high_zoneidx,
2259 alloc_flags & ~ALLOC_NO_WATERMARKS,
2260 preferred_zone, migratetype);
2261 if (page) {
2262 preferred_zone->compact_blockskip_flush = false;
2263 compaction_defer_reset(preferred_zone, order, true);
2264 count_vm_event(COMPACTSUCCESS);
2265 return page;
2269 * It's bad if compaction run occurs and fails.
2270 * The most likely reason is that pages exist,
2271 * but not enough to satisfy watermarks.
2273 count_vm_event(COMPACTFAIL);
2276 * As async compaction considers a subset of pageblocks, only
2277 * defer if the failure was a sync compaction failure.
2279 if (sync_migration)
2280 defer_compaction(preferred_zone, order);
2282 cond_resched();
2285 return NULL;
2287 #else
2288 static inline struct page *
2289 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2290 struct zonelist *zonelist, enum zone_type high_zoneidx,
2291 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2292 int migratetype, bool sync_migration,
2293 bool *contended_compaction, bool *deferred_compaction,
2294 unsigned long *did_some_progress)
2296 return NULL;
2298 #endif /* CONFIG_COMPACTION */
2300 /* Perform direct synchronous page reclaim */
2301 static int
2302 __perform_reclaim(gfp_t gfp_mask, unsigned int order, struct zonelist *zonelist,
2303 nodemask_t *nodemask)
2305 struct reclaim_state reclaim_state;
2306 int progress;
2308 cond_resched();
2310 /* We now go into synchronous reclaim */
2311 cpuset_memory_pressure_bump();
2312 current->flags |= PF_MEMALLOC;
2313 lockdep_set_current_reclaim_state(gfp_mask);
2314 reclaim_state.reclaimed_slab = 0;
2315 current->reclaim_state = &reclaim_state;
2317 progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
2319 current->reclaim_state = NULL;
2320 lockdep_clear_current_reclaim_state();
2321 current->flags &= ~PF_MEMALLOC;
2323 cond_resched();
2325 return progress;
2328 /* The really slow allocator path where we enter direct reclaim */
2329 static inline struct page *
2330 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2331 struct zonelist *zonelist, enum zone_type high_zoneidx,
2332 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
2333 int migratetype, unsigned long *did_some_progress)
2335 struct page *page = NULL;
2336 bool drained = false;
2338 *did_some_progress = __perform_reclaim(gfp_mask, order, zonelist,
2339 nodemask);
2340 if (unlikely(!(*did_some_progress)))
2341 return NULL;
2343 /* After successful reclaim, reconsider all zones for allocation */
2344 if (IS_ENABLED(CONFIG_NUMA))
2345 zlc_clear_zones_full(zonelist);
2347 retry:
2348 page = get_page_from_freelist(gfp_mask, nodemask, order,
2349 zonelist, high_zoneidx,
2350 alloc_flags & ~ALLOC_NO_WATERMARKS,
2351 preferred_zone, migratetype);
2354 * If an allocation failed after direct reclaim, it could be because
2355 * pages are pinned on the per-cpu lists. Drain them and try again
2357 if (!page && !drained) {
2358 drain_all_pages();
2359 drained = true;
2360 goto retry;
2363 return page;
2367 * This is called in the allocator slow-path if the allocation request is of
2368 * sufficient urgency to ignore watermarks and take other desperate measures
2370 static inline struct page *
2371 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2372 struct zonelist *zonelist, enum zone_type high_zoneidx,
2373 nodemask_t *nodemask, struct zone *preferred_zone,
2374 int migratetype)
2376 struct page *page;
2378 do {
2379 page = get_page_from_freelist(gfp_mask, nodemask, order,
2380 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2381 preferred_zone, migratetype);
2383 if (!page && gfp_mask & __GFP_NOFAIL)
2384 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2385 } while (!page && (gfp_mask & __GFP_NOFAIL));
2387 return page;
2390 static void prepare_slowpath(gfp_t gfp_mask, unsigned int order,
2391 struct zonelist *zonelist,
2392 enum zone_type high_zoneidx,
2393 struct zone *preferred_zone)
2395 struct zoneref *z;
2396 struct zone *zone;
2398 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
2399 if (!(gfp_mask & __GFP_NO_KSWAPD))
2400 wakeup_kswapd(zone, order, zone_idx(preferred_zone));
2402 * Only reset the batches of zones that were actually
2403 * considered in the fast path, we don't want to
2404 * thrash fairness information for zones that are not
2405 * actually part of this zonelist's round-robin cycle.
2407 if (!zone_local(preferred_zone, zone))
2408 continue;
2409 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2410 high_wmark_pages(zone) -
2411 low_wmark_pages(zone) -
2412 zone_page_state(zone, NR_ALLOC_BATCH));
2416 static inline int
2417 gfp_to_alloc_flags(gfp_t gfp_mask)
2419 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2420 const gfp_t wait = gfp_mask & __GFP_WAIT;
2422 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2423 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2426 * The caller may dip into page reserves a bit more if the caller
2427 * cannot run direct reclaim, or if the caller has realtime scheduling
2428 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2429 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2431 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2433 if (!wait) {
2435 * Not worth trying to allocate harder for
2436 * __GFP_NOMEMALLOC even if it can't schedule.
2438 if (!(gfp_mask & __GFP_NOMEMALLOC))
2439 alloc_flags |= ALLOC_HARDER;
2441 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2442 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2444 alloc_flags &= ~ALLOC_CPUSET;
2445 } else if (unlikely(rt_task(current)) && !in_interrupt())
2446 alloc_flags |= ALLOC_HARDER;
2448 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2449 if (gfp_mask & __GFP_MEMALLOC)
2450 alloc_flags |= ALLOC_NO_WATERMARKS;
2451 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2452 alloc_flags |= ALLOC_NO_WATERMARKS;
2453 else if (!in_interrupt() &&
2454 ((current->flags & PF_MEMALLOC) ||
2455 unlikely(test_thread_flag(TIF_MEMDIE))))
2456 alloc_flags |= ALLOC_NO_WATERMARKS;
2458 #ifdef CONFIG_CMA
2459 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2460 alloc_flags |= ALLOC_CMA;
2461 #endif
2462 return alloc_flags;
2465 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2467 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2470 static inline struct page *
2471 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2472 struct zonelist *zonelist, enum zone_type high_zoneidx,
2473 nodemask_t *nodemask, struct zone *preferred_zone,
2474 int migratetype)
2476 const gfp_t wait = gfp_mask & __GFP_WAIT;
2477 struct page *page = NULL;
2478 int alloc_flags;
2479 unsigned long pages_reclaimed = 0;
2480 unsigned long did_some_progress;
2481 bool sync_migration = false;
2482 bool deferred_compaction = false;
2483 bool contended_compaction = false;
2486 * In the slowpath, we sanity check order to avoid ever trying to
2487 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2488 * be using allocators in order of preference for an area that is
2489 * too large.
2491 if (order >= MAX_ORDER) {
2492 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2493 return NULL;
2497 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2498 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2499 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2500 * using a larger set of nodes after it has established that the
2501 * allowed per node queues are empty and that nodes are
2502 * over allocated.
2504 if (IS_ENABLED(CONFIG_NUMA) &&
2505 (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2506 goto nopage;
2508 restart:
2509 prepare_slowpath(gfp_mask, order, zonelist,
2510 high_zoneidx, preferred_zone);
2513 * OK, we're below the kswapd watermark and have kicked background
2514 * reclaim. Now things get more complex, so set up alloc_flags according
2515 * to how we want to proceed.
2517 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2520 * Find the true preferred zone if the allocation is unconstrained by
2521 * cpusets.
2523 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2524 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2525 &preferred_zone);
2527 rebalance:
2528 /* This is the last chance, in general, before the goto nopage. */
2529 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2530 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2531 preferred_zone, migratetype);
2532 if (page)
2533 goto got_pg;
2535 /* Allocate without watermarks if the context allows */
2536 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2538 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2539 * the allocation is high priority and these type of
2540 * allocations are system rather than user orientated
2542 zonelist = node_zonelist(numa_node_id(), gfp_mask);
2544 page = __alloc_pages_high_priority(gfp_mask, order,
2545 zonelist, high_zoneidx, nodemask,
2546 preferred_zone, migratetype);
2547 if (page) {
2548 goto got_pg;
2552 /* Atomic allocations - we can't balance anything */
2553 if (!wait) {
2555 * All existing users of the deprecated __GFP_NOFAIL are
2556 * blockable, so warn of any new users that actually allow this
2557 * type of allocation to fail.
2559 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2560 goto nopage;
2563 /* Avoid recursion of direct reclaim */
2564 if (current->flags & PF_MEMALLOC)
2565 goto nopage;
2567 /* Avoid allocations with no watermarks from looping endlessly */
2568 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2569 goto nopage;
2572 * Try direct compaction. The first pass is asynchronous. Subsequent
2573 * attempts after direct reclaim are synchronous
2575 page = __alloc_pages_direct_compact(gfp_mask, order,
2576 zonelist, high_zoneidx,
2577 nodemask,
2578 alloc_flags, preferred_zone,
2579 migratetype, sync_migration,
2580 &contended_compaction,
2581 &deferred_compaction,
2582 &did_some_progress);
2583 if (page)
2584 goto got_pg;
2585 sync_migration = true;
2588 * If compaction is deferred for high-order allocations, it is because
2589 * sync compaction recently failed. In this is the case and the caller
2590 * requested a movable allocation that does not heavily disrupt the
2591 * system then fail the allocation instead of entering direct reclaim.
2593 if ((deferred_compaction || contended_compaction) &&
2594 (gfp_mask & __GFP_NO_KSWAPD))
2595 goto nopage;
2597 /* Try direct reclaim and then allocating */
2598 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2599 zonelist, high_zoneidx,
2600 nodemask,
2601 alloc_flags, preferred_zone,
2602 migratetype, &did_some_progress);
2603 if (page)
2604 goto got_pg;
2607 * If we failed to make any progress reclaiming, then we are
2608 * running out of options and have to consider going OOM
2610 if (!did_some_progress) {
2611 if (oom_gfp_allowed(gfp_mask)) {
2612 if (oom_killer_disabled)
2613 goto nopage;
2614 /* Coredumps can quickly deplete all memory reserves */
2615 if ((current->flags & PF_DUMPCORE) &&
2616 !(gfp_mask & __GFP_NOFAIL))
2617 goto nopage;
2618 page = __alloc_pages_may_oom(gfp_mask, order,
2619 zonelist, high_zoneidx,
2620 nodemask, preferred_zone,
2621 migratetype);
2622 if (page)
2623 goto got_pg;
2625 if (!(gfp_mask & __GFP_NOFAIL)) {
2627 * The oom killer is not called for high-order
2628 * allocations that may fail, so if no progress
2629 * is being made, there are no other options and
2630 * retrying is unlikely to help.
2632 if (order > PAGE_ALLOC_COSTLY_ORDER)
2633 goto nopage;
2635 * The oom killer is not called for lowmem
2636 * allocations to prevent needlessly killing
2637 * innocent tasks.
2639 if (high_zoneidx < ZONE_NORMAL)
2640 goto nopage;
2643 goto restart;
2647 /* Check if we should retry the allocation */
2648 pages_reclaimed += did_some_progress;
2649 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2650 pages_reclaimed)) {
2651 /* Wait for some write requests to complete then retry */
2652 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2653 goto rebalance;
2654 } else {
2656 * High-order allocations do not necessarily loop after
2657 * direct reclaim and reclaim/compaction depends on compaction
2658 * being called after reclaim so call directly if necessary
2660 page = __alloc_pages_direct_compact(gfp_mask, order,
2661 zonelist, high_zoneidx,
2662 nodemask,
2663 alloc_flags, preferred_zone,
2664 migratetype, sync_migration,
2665 &contended_compaction,
2666 &deferred_compaction,
2667 &did_some_progress);
2668 if (page)
2669 goto got_pg;
2672 nopage:
2673 warn_alloc_failed(gfp_mask, order, NULL);
2674 return page;
2675 got_pg:
2676 if (kmemcheck_enabled)
2677 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2679 return page;
2683 * This is the 'heart' of the zoned buddy allocator.
2685 struct page *
2686 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2687 struct zonelist *zonelist, nodemask_t *nodemask)
2689 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2690 struct zone *preferred_zone;
2691 struct page *page = NULL;
2692 int migratetype = allocflags_to_migratetype(gfp_mask);
2693 unsigned int cpuset_mems_cookie;
2694 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET;
2695 struct mem_cgroup *memcg = NULL;
2697 gfp_mask &= gfp_allowed_mask;
2699 lockdep_trace_alloc(gfp_mask);
2701 might_sleep_if(gfp_mask & __GFP_WAIT);
2703 if (should_fail_alloc_page(gfp_mask, order))
2704 return NULL;
2707 * Check the zones suitable for the gfp_mask contain at least one
2708 * valid zone. It's possible to have an empty zonelist as a result
2709 * of GFP_THISNODE and a memoryless node
2711 if (unlikely(!zonelist->_zonerefs->zone))
2712 return NULL;
2715 * Will only have any effect when __GFP_KMEMCG is set. This is
2716 * verified in the (always inline) callee
2718 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2719 return NULL;
2721 retry_cpuset:
2722 cpuset_mems_cookie = get_mems_allowed();
2724 /* The preferred zone is used for statistics later */
2725 first_zones_zonelist(zonelist, high_zoneidx,
2726 nodemask ? : &cpuset_current_mems_allowed,
2727 &preferred_zone);
2728 if (!preferred_zone)
2729 goto out;
2731 #ifdef CONFIG_CMA
2732 if (allocflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2733 alloc_flags |= ALLOC_CMA;
2734 #endif
2735 /* First allocation attempt */
2736 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2737 zonelist, high_zoneidx, alloc_flags,
2738 preferred_zone, migratetype);
2739 if (unlikely(!page)) {
2741 * Runtime PM, block IO and its error handling path
2742 * can deadlock because I/O on the device might not
2743 * complete.
2745 gfp_mask = memalloc_noio_flags(gfp_mask);
2746 page = __alloc_pages_slowpath(gfp_mask, order,
2747 zonelist, high_zoneidx, nodemask,
2748 preferred_zone, migratetype);
2751 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2753 out:
2755 * When updating a task's mems_allowed, it is possible to race with
2756 * parallel threads in such a way that an allocation can fail while
2757 * the mask is being updated. If a page allocation is about to fail,
2758 * check if the cpuset changed during allocation and if so, retry.
2760 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
2761 goto retry_cpuset;
2763 memcg_kmem_commit_charge(page, memcg, order);
2765 return page;
2767 EXPORT_SYMBOL(__alloc_pages_nodemask);
2770 * Common helper functions.
2772 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2774 struct page *page;
2777 * __get_free_pages() returns a 32-bit address, which cannot represent
2778 * a highmem page
2780 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2782 page = alloc_pages(gfp_mask, order);
2783 if (!page)
2784 return 0;
2785 return (unsigned long) page_address(page);
2787 EXPORT_SYMBOL(__get_free_pages);
2789 unsigned long get_zeroed_page(gfp_t gfp_mask)
2791 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2793 EXPORT_SYMBOL(get_zeroed_page);
2795 void __free_pages(struct page *page, unsigned int order)
2797 if (put_page_testzero(page)) {
2798 if (order == 0)
2799 free_hot_cold_page(page, 0);
2800 else
2801 __free_pages_ok(page, order);
2805 EXPORT_SYMBOL(__free_pages);
2807 void free_pages(unsigned long addr, unsigned int order)
2809 if (addr != 0) {
2810 VM_BUG_ON(!virt_addr_valid((void *)addr));
2811 __free_pages(virt_to_page((void *)addr), order);
2815 EXPORT_SYMBOL(free_pages);
2818 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2819 * pages allocated with __GFP_KMEMCG.
2821 * Those pages are accounted to a particular memcg, embedded in the
2822 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2823 * for that information only to find out that it is NULL for users who have no
2824 * interest in that whatsoever, we provide these functions.
2826 * The caller knows better which flags it relies on.
2828 void __free_memcg_kmem_pages(struct page *page, unsigned int order)
2830 memcg_kmem_uncharge_pages(page, order);
2831 __free_pages(page, order);
2834 void free_memcg_kmem_pages(unsigned long addr, unsigned int order)
2836 if (addr != 0) {
2837 VM_BUG_ON(!virt_addr_valid((void *)addr));
2838 __free_memcg_kmem_pages(virt_to_page((void *)addr), order);
2842 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2844 if (addr) {
2845 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2846 unsigned long used = addr + PAGE_ALIGN(size);
2848 split_page(virt_to_page((void *)addr), order);
2849 while (used < alloc_end) {
2850 free_page(used);
2851 used += PAGE_SIZE;
2854 return (void *)addr;
2858 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2859 * @size: the number of bytes to allocate
2860 * @gfp_mask: GFP flags for the allocation
2862 * This function is similar to alloc_pages(), except that it allocates the
2863 * minimum number of pages to satisfy the request. alloc_pages() can only
2864 * allocate memory in power-of-two pages.
2866 * This function is also limited by MAX_ORDER.
2868 * Memory allocated by this function must be released by free_pages_exact().
2870 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2872 unsigned int order = get_order(size);
2873 unsigned long addr;
2875 addr = __get_free_pages(gfp_mask, order);
2876 return make_alloc_exact(addr, order, size);
2878 EXPORT_SYMBOL(alloc_pages_exact);
2881 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2882 * pages on a node.
2883 * @nid: the preferred node ID where memory should be allocated
2884 * @size: the number of bytes to allocate
2885 * @gfp_mask: GFP flags for the allocation
2887 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2888 * back.
2889 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2890 * but is not exact.
2892 void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2894 unsigned order = get_order(size);
2895 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2896 if (!p)
2897 return NULL;
2898 return make_alloc_exact((unsigned long)page_address(p), order, size);
2900 EXPORT_SYMBOL(alloc_pages_exact_nid);
2903 * free_pages_exact - release memory allocated via alloc_pages_exact()
2904 * @virt: the value returned by alloc_pages_exact.
2905 * @size: size of allocation, same value as passed to alloc_pages_exact().
2907 * Release the memory allocated by a previous call to alloc_pages_exact.
2909 void free_pages_exact(void *virt, size_t size)
2911 unsigned long addr = (unsigned long)virt;
2912 unsigned long end = addr + PAGE_ALIGN(size);
2914 while (addr < end) {
2915 free_page(addr);
2916 addr += PAGE_SIZE;
2919 EXPORT_SYMBOL(free_pages_exact);
2922 * nr_free_zone_pages - count number of pages beyond high watermark
2923 * @offset: The zone index of the highest zone
2925 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2926 * high watermark within all zones at or below a given zone index. For each
2927 * zone, the number of pages is calculated as:
2928 * managed_pages - high_pages
2930 static unsigned long nr_free_zone_pages(int offset)
2932 struct zoneref *z;
2933 struct zone *zone;
2935 /* Just pick one node, since fallback list is circular */
2936 unsigned long sum = 0;
2938 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2940 for_each_zone_zonelist(zone, z, zonelist, offset) {
2941 unsigned long size = zone->managed_pages;
2942 unsigned long high = high_wmark_pages(zone);
2943 if (size > high)
2944 sum += size - high;
2947 return sum;
2951 * nr_free_buffer_pages - count number of pages beyond high watermark
2953 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2954 * watermark within ZONE_DMA and ZONE_NORMAL.
2956 unsigned long nr_free_buffer_pages(void)
2958 return nr_free_zone_pages(gfp_zone(GFP_USER));
2960 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2963 * nr_free_pagecache_pages - count number of pages beyond high watermark
2965 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2966 * high watermark within all zones.
2968 unsigned long nr_free_pagecache_pages(void)
2970 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2973 static inline void show_node(struct zone *zone)
2975 if (IS_ENABLED(CONFIG_NUMA))
2976 printk("Node %d ", zone_to_nid(zone));
2979 void si_meminfo(struct sysinfo *val)
2981 val->totalram = totalram_pages;
2982 val->sharedram = 0;
2983 val->freeram = global_page_state(NR_FREE_PAGES);
2984 val->bufferram = nr_blockdev_pages();
2985 val->totalhigh = totalhigh_pages;
2986 val->freehigh = nr_free_highpages();
2987 val->mem_unit = PAGE_SIZE;
2990 EXPORT_SYMBOL(si_meminfo);
2992 #ifdef CONFIG_NUMA
2993 void si_meminfo_node(struct sysinfo *val, int nid)
2995 int zone_type; /* needs to be signed */
2996 unsigned long managed_pages = 0;
2997 pg_data_t *pgdat = NODE_DATA(nid);
2999 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3000 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3001 val->totalram = managed_pages;
3002 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3003 #ifdef CONFIG_HIGHMEM
3004 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3005 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3006 NR_FREE_PAGES);
3007 #else
3008 val->totalhigh = 0;
3009 val->freehigh = 0;
3010 #endif
3011 val->mem_unit = PAGE_SIZE;
3013 #endif
3016 * Determine whether the node should be displayed or not, depending on whether
3017 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3019 bool skip_free_areas_node(unsigned int flags, int nid)
3021 bool ret = false;
3022 unsigned int cpuset_mems_cookie;
3024 if (!(flags & SHOW_MEM_FILTER_NODES))
3025 goto out;
3027 do {
3028 cpuset_mems_cookie = get_mems_allowed();
3029 ret = !node_isset(nid, cpuset_current_mems_allowed);
3030 } while (!put_mems_allowed(cpuset_mems_cookie));
3031 out:
3032 return ret;
3035 #define K(x) ((x) << (PAGE_SHIFT-10))
3037 static void show_migration_types(unsigned char type)
3039 static const char types[MIGRATE_TYPES] = {
3040 [MIGRATE_UNMOVABLE] = 'U',
3041 [MIGRATE_RECLAIMABLE] = 'E',
3042 [MIGRATE_MOVABLE] = 'M',
3043 [MIGRATE_RESERVE] = 'R',
3044 #ifdef CONFIG_CMA
3045 [MIGRATE_CMA] = 'C',
3046 #endif
3047 #ifdef CONFIG_MEMORY_ISOLATION
3048 [MIGRATE_ISOLATE] = 'I',
3049 #endif
3051 char tmp[MIGRATE_TYPES + 1];
3052 char *p = tmp;
3053 int i;
3055 for (i = 0; i < MIGRATE_TYPES; i++) {
3056 if (type & (1 << i))
3057 *p++ = types[i];
3060 *p = '\0';
3061 printk("(%s) ", tmp);
3065 * Show free area list (used inside shift_scroll-lock stuff)
3066 * We also calculate the percentage fragmentation. We do this by counting the
3067 * memory on each free list with the exception of the first item on the list.
3068 * Suppresses nodes that are not allowed by current's cpuset if
3069 * SHOW_MEM_FILTER_NODES is passed.
3071 void show_free_areas(unsigned int filter)
3073 int cpu;
3074 struct zone *zone;
3076 for_each_populated_zone(zone) {
3077 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3078 continue;
3079 show_node(zone);
3080 printk("%s per-cpu:\n", zone->name);
3082 for_each_online_cpu(cpu) {
3083 struct per_cpu_pageset *pageset;
3085 pageset = per_cpu_ptr(zone->pageset, cpu);
3087 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3088 cpu, pageset->pcp.high,
3089 pageset->pcp.batch, pageset->pcp.count);
3093 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3094 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3095 " unevictable:%lu"
3096 " dirty:%lu writeback:%lu unstable:%lu\n"
3097 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3098 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3099 " free_cma:%lu\n",
3100 global_page_state(NR_ACTIVE_ANON),
3101 global_page_state(NR_INACTIVE_ANON),
3102 global_page_state(NR_ISOLATED_ANON),
3103 global_page_state(NR_ACTIVE_FILE),
3104 global_page_state(NR_INACTIVE_FILE),
3105 global_page_state(NR_ISOLATED_FILE),
3106 global_page_state(NR_UNEVICTABLE),
3107 global_page_state(NR_FILE_DIRTY),
3108 global_page_state(NR_WRITEBACK),
3109 global_page_state(NR_UNSTABLE_NFS),
3110 global_page_state(NR_FREE_PAGES),
3111 global_page_state(NR_SLAB_RECLAIMABLE),
3112 global_page_state(NR_SLAB_UNRECLAIMABLE),
3113 global_page_state(NR_FILE_MAPPED),
3114 global_page_state(NR_SHMEM),
3115 global_page_state(NR_PAGETABLE),
3116 global_page_state(NR_BOUNCE),
3117 global_page_state(NR_FREE_CMA_PAGES));
3119 for_each_populated_zone(zone) {
3120 int i;
3122 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3123 continue;
3124 show_node(zone);
3125 printk("%s"
3126 " free:%lukB"
3127 " min:%lukB"
3128 " low:%lukB"
3129 " high:%lukB"
3130 " active_anon:%lukB"
3131 " inactive_anon:%lukB"
3132 " active_file:%lukB"
3133 " inactive_file:%lukB"
3134 " unevictable:%lukB"
3135 " isolated(anon):%lukB"
3136 " isolated(file):%lukB"
3137 " present:%lukB"
3138 " managed:%lukB"
3139 " mlocked:%lukB"
3140 " dirty:%lukB"
3141 " writeback:%lukB"
3142 " mapped:%lukB"
3143 " shmem:%lukB"
3144 " slab_reclaimable:%lukB"
3145 " slab_unreclaimable:%lukB"
3146 " kernel_stack:%lukB"
3147 " pagetables:%lukB"
3148 " unstable:%lukB"
3149 " bounce:%lukB"
3150 " free_cma:%lukB"
3151 " writeback_tmp:%lukB"
3152 " pages_scanned:%lu"
3153 " all_unreclaimable? %s"
3154 "\n",
3155 zone->name,
3156 K(zone_page_state(zone, NR_FREE_PAGES)),
3157 K(min_wmark_pages(zone)),
3158 K(low_wmark_pages(zone)),
3159 K(high_wmark_pages(zone)),
3160 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3161 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3162 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3163 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3164 K(zone_page_state(zone, NR_UNEVICTABLE)),
3165 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3166 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3167 K(zone->present_pages),
3168 K(zone->managed_pages),
3169 K(zone_page_state(zone, NR_MLOCK)),
3170 K(zone_page_state(zone, NR_FILE_DIRTY)),
3171 K(zone_page_state(zone, NR_WRITEBACK)),
3172 K(zone_page_state(zone, NR_FILE_MAPPED)),
3173 K(zone_page_state(zone, NR_SHMEM)),
3174 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3175 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3176 zone_page_state(zone, NR_KERNEL_STACK) *
3177 THREAD_SIZE / 1024,
3178 K(zone_page_state(zone, NR_PAGETABLE)),
3179 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3180 K(zone_page_state(zone, NR_BOUNCE)),
3181 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3182 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3183 zone->pages_scanned,
3184 (!zone_reclaimable(zone) ? "yes" : "no")
3186 printk("lowmem_reserve[]:");
3187 for (i = 0; i < MAX_NR_ZONES; i++)
3188 printk(" %lu", zone->lowmem_reserve[i]);
3189 printk("\n");
3192 for_each_populated_zone(zone) {
3193 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3194 unsigned char types[MAX_ORDER];
3196 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3197 continue;
3198 show_node(zone);
3199 printk("%s: ", zone->name);
3201 spin_lock_irqsave(&zone->lock, flags);
3202 for (order = 0; order < MAX_ORDER; order++) {
3203 struct free_area *area = &zone->free_area[order];
3204 int type;
3206 nr[order] = area->nr_free;
3207 total += nr[order] << order;
3209 types[order] = 0;
3210 for (type = 0; type < MIGRATE_TYPES; type++) {
3211 if (!list_empty(&area->free_list[type]))
3212 types[order] |= 1 << type;
3215 spin_unlock_irqrestore(&zone->lock, flags);
3216 for (order = 0; order < MAX_ORDER; order++) {
3217 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3218 if (nr[order])
3219 show_migration_types(types[order]);
3221 printk("= %lukB\n", K(total));
3224 hugetlb_show_meminfo();
3226 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3228 show_swap_cache_info();
3231 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3233 zoneref->zone = zone;
3234 zoneref->zone_idx = zone_idx(zone);
3238 * Builds allocation fallback zone lists.
3240 * Add all populated zones of a node to the zonelist.
3242 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3243 int nr_zones)
3245 struct zone *zone;
3246 enum zone_type zone_type = MAX_NR_ZONES;
3248 do {
3249 zone_type--;
3250 zone = pgdat->node_zones + zone_type;
3251 if (populated_zone(zone)) {
3252 zoneref_set_zone(zone,
3253 &zonelist->_zonerefs[nr_zones++]);
3254 check_highest_zone(zone_type);
3256 } while (zone_type);
3258 return nr_zones;
3263 * zonelist_order:
3264 * 0 = automatic detection of better ordering.
3265 * 1 = order by ([node] distance, -zonetype)
3266 * 2 = order by (-zonetype, [node] distance)
3268 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3269 * the same zonelist. So only NUMA can configure this param.
3271 #define ZONELIST_ORDER_DEFAULT 0
3272 #define ZONELIST_ORDER_NODE 1
3273 #define ZONELIST_ORDER_ZONE 2
3275 /* zonelist order in the kernel.
3276 * set_zonelist_order() will set this to NODE or ZONE.
3278 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3279 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3282 #ifdef CONFIG_NUMA
3283 /* The value user specified ....changed by config */
3284 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3285 /* string for sysctl */
3286 #define NUMA_ZONELIST_ORDER_LEN 16
3287 char numa_zonelist_order[16] = "default";
3290 * interface for configure zonelist ordering.
3291 * command line option "numa_zonelist_order"
3292 * = "[dD]efault - default, automatic configuration.
3293 * = "[nN]ode - order by node locality, then by zone within node
3294 * = "[zZ]one - order by zone, then by locality within zone
3297 static int __parse_numa_zonelist_order(char *s)
3299 if (*s == 'd' || *s == 'D') {
3300 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3301 } else if (*s == 'n' || *s == 'N') {
3302 user_zonelist_order = ZONELIST_ORDER_NODE;
3303 } else if (*s == 'z' || *s == 'Z') {
3304 user_zonelist_order = ZONELIST_ORDER_ZONE;
3305 } else {
3306 printk(KERN_WARNING
3307 "Ignoring invalid numa_zonelist_order value: "
3308 "%s\n", s);
3309 return -EINVAL;
3311 return 0;
3314 static __init int setup_numa_zonelist_order(char *s)
3316 int ret;
3318 if (!s)
3319 return 0;
3321 ret = __parse_numa_zonelist_order(s);
3322 if (ret == 0)
3323 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3325 return ret;
3327 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3330 * sysctl handler for numa_zonelist_order
3332 int numa_zonelist_order_handler(ctl_table *table, int write,
3333 void __user *buffer, size_t *length,
3334 loff_t *ppos)
3336 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3337 int ret;
3338 static DEFINE_MUTEX(zl_order_mutex);
3340 mutex_lock(&zl_order_mutex);
3341 if (write) {
3342 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3343 ret = -EINVAL;
3344 goto out;
3346 strcpy(saved_string, (char *)table->data);
3348 ret = proc_dostring(table, write, buffer, length, ppos);
3349 if (ret)
3350 goto out;
3351 if (write) {
3352 int oldval = user_zonelist_order;
3354 ret = __parse_numa_zonelist_order((char *)table->data);
3355 if (ret) {
3357 * bogus value. restore saved string
3359 strncpy((char *)table->data, saved_string,
3360 NUMA_ZONELIST_ORDER_LEN);
3361 user_zonelist_order = oldval;
3362 } else if (oldval != user_zonelist_order) {
3363 mutex_lock(&zonelists_mutex);
3364 build_all_zonelists(NULL, NULL);
3365 mutex_unlock(&zonelists_mutex);
3368 out:
3369 mutex_unlock(&zl_order_mutex);
3370 return ret;
3374 #define MAX_NODE_LOAD (nr_online_nodes)
3375 static int node_load[MAX_NUMNODES];
3378 * find_next_best_node - find the next node that should appear in a given node's fallback list
3379 * @node: node whose fallback list we're appending
3380 * @used_node_mask: nodemask_t of already used nodes
3382 * We use a number of factors to determine which is the next node that should
3383 * appear on a given node's fallback list. The node should not have appeared
3384 * already in @node's fallback list, and it should be the next closest node
3385 * according to the distance array (which contains arbitrary distance values
3386 * from each node to each node in the system), and should also prefer nodes
3387 * with no CPUs, since presumably they'll have very little allocation pressure
3388 * on them otherwise.
3389 * It returns -1 if no node is found.
3391 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3393 int n, val;
3394 int min_val = INT_MAX;
3395 int best_node = NUMA_NO_NODE;
3396 const struct cpumask *tmp = cpumask_of_node(0);
3398 /* Use the local node if we haven't already */
3399 if (!node_isset(node, *used_node_mask)) {
3400 node_set(node, *used_node_mask);
3401 return node;
3404 for_each_node_state(n, N_MEMORY) {
3406 /* Don't want a node to appear more than once */
3407 if (node_isset(n, *used_node_mask))
3408 continue;
3410 /* Use the distance array to find the distance */
3411 val = node_distance(node, n);
3413 /* Penalize nodes under us ("prefer the next node") */
3414 val += (n < node);
3416 /* Give preference to headless and unused nodes */
3417 tmp = cpumask_of_node(n);
3418 if (!cpumask_empty(tmp))
3419 val += PENALTY_FOR_NODE_WITH_CPUS;
3421 /* Slight preference for less loaded node */
3422 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3423 val += node_load[n];
3425 if (val < min_val) {
3426 min_val = val;
3427 best_node = n;
3431 if (best_node >= 0)
3432 node_set(best_node, *used_node_mask);
3434 return best_node;
3439 * Build zonelists ordered by node and zones within node.
3440 * This results in maximum locality--normal zone overflows into local
3441 * DMA zone, if any--but risks exhausting DMA zone.
3443 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3445 int j;
3446 struct zonelist *zonelist;
3448 zonelist = &pgdat->node_zonelists[0];
3449 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3451 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3452 zonelist->_zonerefs[j].zone = NULL;
3453 zonelist->_zonerefs[j].zone_idx = 0;
3457 * Build gfp_thisnode zonelists
3459 static void build_thisnode_zonelists(pg_data_t *pgdat)
3461 int j;
3462 struct zonelist *zonelist;
3464 zonelist = &pgdat->node_zonelists[1];
3465 j = build_zonelists_node(pgdat, zonelist, 0);
3466 zonelist->_zonerefs[j].zone = NULL;
3467 zonelist->_zonerefs[j].zone_idx = 0;
3471 * Build zonelists ordered by zone and nodes within zones.
3472 * This results in conserving DMA zone[s] until all Normal memory is
3473 * exhausted, but results in overflowing to remote node while memory
3474 * may still exist in local DMA zone.
3476 static int node_order[MAX_NUMNODES];
3478 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3480 int pos, j, node;
3481 int zone_type; /* needs to be signed */
3482 struct zone *z;
3483 struct zonelist *zonelist;
3485 zonelist = &pgdat->node_zonelists[0];
3486 pos = 0;
3487 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3488 for (j = 0; j < nr_nodes; j++) {
3489 node = node_order[j];
3490 z = &NODE_DATA(node)->node_zones[zone_type];
3491 if (populated_zone(z)) {
3492 zoneref_set_zone(z,
3493 &zonelist->_zonerefs[pos++]);
3494 check_highest_zone(zone_type);
3498 zonelist->_zonerefs[pos].zone = NULL;
3499 zonelist->_zonerefs[pos].zone_idx = 0;
3502 static int default_zonelist_order(void)
3504 int nid, zone_type;
3505 unsigned long low_kmem_size, total_size;
3506 struct zone *z;
3507 int average_size;
3509 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3510 * If they are really small and used heavily, the system can fall
3511 * into OOM very easily.
3512 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3514 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3515 low_kmem_size = 0;
3516 total_size = 0;
3517 for_each_online_node(nid) {
3518 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3519 z = &NODE_DATA(nid)->node_zones[zone_type];
3520 if (populated_zone(z)) {
3521 if (zone_type < ZONE_NORMAL)
3522 low_kmem_size += z->managed_pages;
3523 total_size += z->managed_pages;
3524 } else if (zone_type == ZONE_NORMAL) {
3526 * If any node has only lowmem, then node order
3527 * is preferred to allow kernel allocations
3528 * locally; otherwise, they can easily infringe
3529 * on other nodes when there is an abundance of
3530 * lowmem available to allocate from.
3532 return ZONELIST_ORDER_NODE;
3536 if (!low_kmem_size || /* there are no DMA area. */
3537 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
3538 return ZONELIST_ORDER_NODE;
3540 * look into each node's config.
3541 * If there is a node whose DMA/DMA32 memory is very big area on
3542 * local memory, NODE_ORDER may be suitable.
3544 average_size = total_size /
3545 (nodes_weight(node_states[N_MEMORY]) + 1);
3546 for_each_online_node(nid) {
3547 low_kmem_size = 0;
3548 total_size = 0;
3549 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
3550 z = &NODE_DATA(nid)->node_zones[zone_type];
3551 if (populated_zone(z)) {
3552 if (zone_type < ZONE_NORMAL)
3553 low_kmem_size += z->present_pages;
3554 total_size += z->present_pages;
3557 if (low_kmem_size &&
3558 total_size > average_size && /* ignore small node */
3559 low_kmem_size > total_size * 70/100)
3560 return ZONELIST_ORDER_NODE;
3562 return ZONELIST_ORDER_ZONE;
3565 static void set_zonelist_order(void)
3567 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3568 current_zonelist_order = default_zonelist_order();
3569 else
3570 current_zonelist_order = user_zonelist_order;
3573 static void build_zonelists(pg_data_t *pgdat)
3575 int j, node, load;
3576 enum zone_type i;
3577 nodemask_t used_mask;
3578 int local_node, prev_node;
3579 struct zonelist *zonelist;
3580 int order = current_zonelist_order;
3582 /* initialize zonelists */
3583 for (i = 0; i < MAX_ZONELISTS; i++) {
3584 zonelist = pgdat->node_zonelists + i;
3585 zonelist->_zonerefs[0].zone = NULL;
3586 zonelist->_zonerefs[0].zone_idx = 0;
3589 /* NUMA-aware ordering of nodes */
3590 local_node = pgdat->node_id;
3591 load = nr_online_nodes;
3592 prev_node = local_node;
3593 nodes_clear(used_mask);
3595 memset(node_order, 0, sizeof(node_order));
3596 j = 0;
3598 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3600 * We don't want to pressure a particular node.
3601 * So adding penalty to the first node in same
3602 * distance group to make it round-robin.
3604 if (node_distance(local_node, node) !=
3605 node_distance(local_node, prev_node))
3606 node_load[node] = load;
3608 prev_node = node;
3609 load--;
3610 if (order == ZONELIST_ORDER_NODE)
3611 build_zonelists_in_node_order(pgdat, node);
3612 else
3613 node_order[j++] = node; /* remember order */
3616 if (order == ZONELIST_ORDER_ZONE) {
3617 /* calculate node order -- i.e., DMA last! */
3618 build_zonelists_in_zone_order(pgdat, j);
3621 build_thisnode_zonelists(pgdat);
3624 /* Construct the zonelist performance cache - see further mmzone.h */
3625 static void build_zonelist_cache(pg_data_t *pgdat)
3627 struct zonelist *zonelist;
3628 struct zonelist_cache *zlc;
3629 struct zoneref *z;
3631 zonelist = &pgdat->node_zonelists[0];
3632 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3633 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3634 for (z = zonelist->_zonerefs; z->zone; z++)
3635 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3638 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3640 * Return node id of node used for "local" allocations.
3641 * I.e., first node id of first zone in arg node's generic zonelist.
3642 * Used for initializing percpu 'numa_mem', which is used primarily
3643 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3645 int local_memory_node(int node)
3647 struct zone *zone;
3649 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3650 gfp_zone(GFP_KERNEL),
3651 NULL,
3652 &zone);
3653 return zone->node;
3655 #endif
3657 #else /* CONFIG_NUMA */
3659 static void set_zonelist_order(void)
3661 current_zonelist_order = ZONELIST_ORDER_ZONE;
3664 static void build_zonelists(pg_data_t *pgdat)
3666 int node, local_node;
3667 enum zone_type j;
3668 struct zonelist *zonelist;
3670 local_node = pgdat->node_id;
3672 zonelist = &pgdat->node_zonelists[0];
3673 j = build_zonelists_node(pgdat, zonelist, 0);
3676 * Now we build the zonelist so that it contains the zones
3677 * of all the other nodes.
3678 * We don't want to pressure a particular node, so when
3679 * building the zones for node N, we make sure that the
3680 * zones coming right after the local ones are those from
3681 * node N+1 (modulo N)
3683 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3684 if (!node_online(node))
3685 continue;
3686 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3688 for (node = 0; node < local_node; node++) {
3689 if (!node_online(node))
3690 continue;
3691 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3694 zonelist->_zonerefs[j].zone = NULL;
3695 zonelist->_zonerefs[j].zone_idx = 0;
3698 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3699 static void build_zonelist_cache(pg_data_t *pgdat)
3701 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3704 #endif /* CONFIG_NUMA */
3707 * Boot pageset table. One per cpu which is going to be used for all
3708 * zones and all nodes. The parameters will be set in such a way
3709 * that an item put on a list will immediately be handed over to
3710 * the buddy list. This is safe since pageset manipulation is done
3711 * with interrupts disabled.
3713 * The boot_pagesets must be kept even after bootup is complete for
3714 * unused processors and/or zones. They do play a role for bootstrapping
3715 * hotplugged processors.
3717 * zoneinfo_show() and maybe other functions do
3718 * not check if the processor is online before following the pageset pointer.
3719 * Other parts of the kernel may not check if the zone is available.
3721 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3722 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3723 static void setup_zone_pageset(struct zone *zone);
3726 * Global mutex to protect against size modification of zonelists
3727 * as well as to serialize pageset setup for the new populated zone.
3729 DEFINE_MUTEX(zonelists_mutex);
3731 /* return values int ....just for stop_machine() */
3732 static int __build_all_zonelists(void *data)
3734 int nid;
3735 int cpu;
3736 pg_data_t *self = data;
3738 #ifdef CONFIG_NUMA
3739 memset(node_load, 0, sizeof(node_load));
3740 #endif
3742 if (self && !node_online(self->node_id)) {
3743 build_zonelists(self);
3744 build_zonelist_cache(self);
3747 for_each_online_node(nid) {
3748 pg_data_t *pgdat = NODE_DATA(nid);
3750 build_zonelists(pgdat);
3751 build_zonelist_cache(pgdat);
3755 * Initialize the boot_pagesets that are going to be used
3756 * for bootstrapping processors. The real pagesets for
3757 * each zone will be allocated later when the per cpu
3758 * allocator is available.
3760 * boot_pagesets are used also for bootstrapping offline
3761 * cpus if the system is already booted because the pagesets
3762 * are needed to initialize allocators on a specific cpu too.
3763 * F.e. the percpu allocator needs the page allocator which
3764 * needs the percpu allocator in order to allocate its pagesets
3765 * (a chicken-egg dilemma).
3767 for_each_possible_cpu(cpu) {
3768 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3770 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3772 * We now know the "local memory node" for each node--
3773 * i.e., the node of the first zone in the generic zonelist.
3774 * Set up numa_mem percpu variable for on-line cpus. During
3775 * boot, only the boot cpu should be on-line; we'll init the
3776 * secondary cpus' numa_mem as they come on-line. During
3777 * node/memory hotplug, we'll fixup all on-line cpus.
3779 if (cpu_online(cpu))
3780 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3781 #endif
3784 return 0;
3788 * Called with zonelists_mutex held always
3789 * unless system_state == SYSTEM_BOOTING.
3791 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3793 set_zonelist_order();
3795 if (system_state == SYSTEM_BOOTING) {
3796 __build_all_zonelists(NULL);
3797 mminit_verify_zonelist();
3798 cpuset_init_current_mems_allowed();
3799 } else {
3800 #ifdef CONFIG_MEMORY_HOTPLUG
3801 if (zone)
3802 setup_zone_pageset(zone);
3803 #endif
3804 /* we have to stop all cpus to guarantee there is no user
3805 of zonelist */
3806 stop_machine(__build_all_zonelists, pgdat, NULL);
3807 /* cpuset refresh routine should be here */
3809 vm_total_pages = nr_free_pagecache_pages();
3811 * Disable grouping by mobility if the number of pages in the
3812 * system is too low to allow the mechanism to work. It would be
3813 * more accurate, but expensive to check per-zone. This check is
3814 * made on memory-hotadd so a system can start with mobility
3815 * disabled and enable it later
3817 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3818 page_group_by_mobility_disabled = 1;
3819 else
3820 page_group_by_mobility_disabled = 0;
3822 printk("Built %i zonelists in %s order, mobility grouping %s. "
3823 "Total pages: %ld\n",
3824 nr_online_nodes,
3825 zonelist_order_name[current_zonelist_order],
3826 page_group_by_mobility_disabled ? "off" : "on",
3827 vm_total_pages);
3828 #ifdef CONFIG_NUMA
3829 printk("Policy zone: %s\n", zone_names[policy_zone]);
3830 #endif
3834 * Helper functions to size the waitqueue hash table.
3835 * Essentially these want to choose hash table sizes sufficiently
3836 * large so that collisions trying to wait on pages are rare.
3837 * But in fact, the number of active page waitqueues on typical
3838 * systems is ridiculously low, less than 200. So this is even
3839 * conservative, even though it seems large.
3841 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3842 * waitqueues, i.e. the size of the waitq table given the number of pages.
3844 #define PAGES_PER_WAITQUEUE 256
3846 #ifndef CONFIG_MEMORY_HOTPLUG
3847 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3849 unsigned long size = 1;
3851 pages /= PAGES_PER_WAITQUEUE;
3853 while (size < pages)
3854 size <<= 1;
3857 * Once we have dozens or even hundreds of threads sleeping
3858 * on IO we've got bigger problems than wait queue collision.
3859 * Limit the size of the wait table to a reasonable size.
3861 size = min(size, 4096UL);
3863 return max(size, 4UL);
3865 #else
3867 * A zone's size might be changed by hot-add, so it is not possible to determine
3868 * a suitable size for its wait_table. So we use the maximum size now.
3870 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3872 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3873 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3874 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3876 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3877 * or more by the traditional way. (See above). It equals:
3879 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3880 * ia64(16K page size) : = ( 8G + 4M)byte.
3881 * powerpc (64K page size) : = (32G +16M)byte.
3883 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3885 return 4096UL;
3887 #endif
3890 * This is an integer logarithm so that shifts can be used later
3891 * to extract the more random high bits from the multiplicative
3892 * hash function before the remainder is taken.
3894 static inline unsigned long wait_table_bits(unsigned long size)
3896 return ffz(~size);
3900 * Check if a pageblock contains reserved pages
3902 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3904 unsigned long pfn;
3906 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3907 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3908 return 1;
3910 return 0;
3914 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3915 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3916 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3917 * higher will lead to a bigger reserve which will get freed as contiguous
3918 * blocks as reclaim kicks in
3920 static void setup_zone_migrate_reserve(struct zone *zone)
3922 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3923 struct page *page;
3924 unsigned long block_migratetype;
3925 int reserve;
3926 int old_reserve;
3929 * Get the start pfn, end pfn and the number of blocks to reserve
3930 * We have to be careful to be aligned to pageblock_nr_pages to
3931 * make sure that we always check pfn_valid for the first page in
3932 * the block.
3934 start_pfn = zone->zone_start_pfn;
3935 end_pfn = zone_end_pfn(zone);
3936 start_pfn = roundup(start_pfn, pageblock_nr_pages);
3937 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3938 pageblock_order;
3941 * Reserve blocks are generally in place to help high-order atomic
3942 * allocations that are short-lived. A min_free_kbytes value that
3943 * would result in more than 2 reserve blocks for atomic allocations
3944 * is assumed to be in place to help anti-fragmentation for the
3945 * future allocation of hugepages at runtime.
3947 reserve = min(2, reserve);
3948 old_reserve = zone->nr_migrate_reserve_block;
3950 /* When memory hot-add, we almost always need to do nothing */
3951 if (reserve == old_reserve)
3952 return;
3953 zone->nr_migrate_reserve_block = reserve;
3955 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3956 if (!pfn_valid(pfn))
3957 continue;
3958 page = pfn_to_page(pfn);
3960 /* Watch out for overlapping nodes */
3961 if (page_to_nid(page) != zone_to_nid(zone))
3962 continue;
3964 block_migratetype = get_pageblock_migratetype(page);
3966 /* Only test what is necessary when the reserves are not met */
3967 if (reserve > 0) {
3969 * Blocks with reserved pages will never free, skip
3970 * them.
3972 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3973 if (pageblock_is_reserved(pfn, block_end_pfn))
3974 continue;
3976 /* If this block is reserved, account for it */
3977 if (block_migratetype == MIGRATE_RESERVE) {
3978 reserve--;
3979 continue;
3982 /* Suitable for reserving if this block is movable */
3983 if (block_migratetype == MIGRATE_MOVABLE) {
3984 set_pageblock_migratetype(page,
3985 MIGRATE_RESERVE);
3986 move_freepages_block(zone, page,
3987 MIGRATE_RESERVE);
3988 reserve--;
3989 continue;
3991 } else if (!old_reserve) {
3993 * At boot time we don't need to scan the whole zone
3994 * for turning off MIGRATE_RESERVE.
3996 break;
4000 * If the reserve is met and this is a previous reserved block,
4001 * take it back
4003 if (block_migratetype == MIGRATE_RESERVE) {
4004 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4005 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4011 * Initially all pages are reserved - free ones are freed
4012 * up by free_all_bootmem() once the early boot process is
4013 * done. Non-atomic initialization, single-pass.
4015 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4016 unsigned long start_pfn, enum memmap_context context)
4018 struct page *page;
4019 unsigned long end_pfn = start_pfn + size;
4020 unsigned long pfn;
4021 struct zone *z;
4023 if (highest_memmap_pfn < end_pfn - 1)
4024 highest_memmap_pfn = end_pfn - 1;
4026 z = &NODE_DATA(nid)->node_zones[zone];
4027 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4029 * There can be holes in boot-time mem_map[]s
4030 * handed to this function. They do not
4031 * exist on hotplugged memory.
4033 if (context == MEMMAP_EARLY) {
4034 if (!early_pfn_valid(pfn))
4035 continue;
4036 if (!early_pfn_in_nid(pfn, nid))
4037 continue;
4039 page = pfn_to_page(pfn);
4040 set_page_links(page, zone, nid, pfn);
4041 mminit_verify_page_links(page, zone, nid, pfn);
4042 init_page_count(page);
4043 page_mapcount_reset(page);
4044 page_cpupid_reset_last(page);
4045 SetPageReserved(page);
4047 * Mark the block movable so that blocks are reserved for
4048 * movable at startup. This will force kernel allocations
4049 * to reserve their blocks rather than leaking throughout
4050 * the address space during boot when many long-lived
4051 * kernel allocations are made. Later some blocks near
4052 * the start are marked MIGRATE_RESERVE by
4053 * setup_zone_migrate_reserve()
4055 * bitmap is created for zone's valid pfn range. but memmap
4056 * can be created for invalid pages (for alignment)
4057 * check here not to call set_pageblock_migratetype() against
4058 * pfn out of zone.
4060 if ((z->zone_start_pfn <= pfn)
4061 && (pfn < zone_end_pfn(z))
4062 && !(pfn & (pageblock_nr_pages - 1)))
4063 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4065 INIT_LIST_HEAD(&page->lru);
4066 #ifdef WANT_PAGE_VIRTUAL
4067 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4068 if (!is_highmem_idx(zone))
4069 set_page_address(page, __va(pfn << PAGE_SHIFT));
4070 #endif
4074 static void __meminit zone_init_free_lists(struct zone *zone)
4076 int order, t;
4077 for_each_migratetype_order(order, t) {
4078 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4079 zone->free_area[order].nr_free = 0;
4083 #ifndef __HAVE_ARCH_MEMMAP_INIT
4084 #define memmap_init(size, nid, zone, start_pfn) \
4085 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4086 #endif
4088 static int __meminit zone_batchsize(struct zone *zone)
4090 #ifdef CONFIG_MMU
4091 int batch;
4094 * The per-cpu-pages pools are set to around 1000th of the
4095 * size of the zone. But no more than 1/2 of a meg.
4097 * OK, so we don't know how big the cache is. So guess.
4099 batch = zone->managed_pages / 1024;
4100 if (batch * PAGE_SIZE > 512 * 1024)
4101 batch = (512 * 1024) / PAGE_SIZE;
4102 batch /= 4; /* We effectively *= 4 below */
4103 if (batch < 1)
4104 batch = 1;
4107 * Clamp the batch to a 2^n - 1 value. Having a power
4108 * of 2 value was found to be more likely to have
4109 * suboptimal cache aliasing properties in some cases.
4111 * For example if 2 tasks are alternately allocating
4112 * batches of pages, one task can end up with a lot
4113 * of pages of one half of the possible page colors
4114 * and the other with pages of the other colors.
4116 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4118 return batch;
4120 #else
4121 /* The deferral and batching of frees should be suppressed under NOMMU
4122 * conditions.
4124 * The problem is that NOMMU needs to be able to allocate large chunks
4125 * of contiguous memory as there's no hardware page translation to
4126 * assemble apparent contiguous memory from discontiguous pages.
4128 * Queueing large contiguous runs of pages for batching, however,
4129 * causes the pages to actually be freed in smaller chunks. As there
4130 * can be a significant delay between the individual batches being
4131 * recycled, this leads to the once large chunks of space being
4132 * fragmented and becoming unavailable for high-order allocations.
4134 return 0;
4135 #endif
4139 * pcp->high and pcp->batch values are related and dependent on one another:
4140 * ->batch must never be higher then ->high.
4141 * The following function updates them in a safe manner without read side
4142 * locking.
4144 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4145 * those fields changing asynchronously (acording the the above rule).
4147 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4148 * outside of boot time (or some other assurance that no concurrent updaters
4149 * exist).
4151 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4152 unsigned long batch)
4154 /* start with a fail safe value for batch */
4155 pcp->batch = 1;
4156 smp_wmb();
4158 /* Update high, then batch, in order */
4159 pcp->high = high;
4160 smp_wmb();
4162 pcp->batch = batch;
4165 /* a companion to pageset_set_high() */
4166 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4168 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4171 static void pageset_init(struct per_cpu_pageset *p)
4173 struct per_cpu_pages *pcp;
4174 int migratetype;
4176 memset(p, 0, sizeof(*p));
4178 pcp = &p->pcp;
4179 pcp->count = 0;
4180 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4181 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4184 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4186 pageset_init(p);
4187 pageset_set_batch(p, batch);
4191 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4192 * to the value high for the pageset p.
4194 static void pageset_set_high(struct per_cpu_pageset *p,
4195 unsigned long high)
4197 unsigned long batch = max(1UL, high / 4);
4198 if ((high / 4) > (PAGE_SHIFT * 8))
4199 batch = PAGE_SHIFT * 8;
4201 pageset_update(&p->pcp, high, batch);
4204 static void __meminit pageset_set_high_and_batch(struct zone *zone,
4205 struct per_cpu_pageset *pcp)
4207 if (percpu_pagelist_fraction)
4208 pageset_set_high(pcp,
4209 (zone->managed_pages /
4210 percpu_pagelist_fraction));
4211 else
4212 pageset_set_batch(pcp, zone_batchsize(zone));
4215 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4217 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4219 pageset_init(pcp);
4220 pageset_set_high_and_batch(zone, pcp);
4223 static void __meminit setup_zone_pageset(struct zone *zone)
4225 int cpu;
4226 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4227 for_each_possible_cpu(cpu)
4228 zone_pageset_init(zone, cpu);
4232 * Allocate per cpu pagesets and initialize them.
4233 * Before this call only boot pagesets were available.
4235 void __init setup_per_cpu_pageset(void)
4237 struct zone *zone;
4239 for_each_populated_zone(zone)
4240 setup_zone_pageset(zone);
4243 static noinline __init_refok
4244 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4246 int i;
4247 size_t alloc_size;
4250 * The per-page waitqueue mechanism uses hashed waitqueues
4251 * per zone.
4253 zone->wait_table_hash_nr_entries =
4254 wait_table_hash_nr_entries(zone_size_pages);
4255 zone->wait_table_bits =
4256 wait_table_bits(zone->wait_table_hash_nr_entries);
4257 alloc_size = zone->wait_table_hash_nr_entries
4258 * sizeof(wait_queue_head_t);
4260 if (!slab_is_available()) {
4261 zone->wait_table = (wait_queue_head_t *)
4262 memblock_virt_alloc_node_nopanic(
4263 alloc_size, zone->zone_pgdat->node_id);
4264 } else {
4266 * This case means that a zone whose size was 0 gets new memory
4267 * via memory hot-add.
4268 * But it may be the case that a new node was hot-added. In
4269 * this case vmalloc() will not be able to use this new node's
4270 * memory - this wait_table must be initialized to use this new
4271 * node itself as well.
4272 * To use this new node's memory, further consideration will be
4273 * necessary.
4275 zone->wait_table = vmalloc(alloc_size);
4277 if (!zone->wait_table)
4278 return -ENOMEM;
4280 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4281 init_waitqueue_head(zone->wait_table + i);
4283 return 0;
4286 static __meminit void zone_pcp_init(struct zone *zone)
4289 * per cpu subsystem is not up at this point. The following code
4290 * relies on the ability of the linker to provide the
4291 * offset of a (static) per cpu variable into the per cpu area.
4293 zone->pageset = &boot_pageset;
4295 if (populated_zone(zone))
4296 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4297 zone->name, zone->present_pages,
4298 zone_batchsize(zone));
4301 int __meminit init_currently_empty_zone(struct zone *zone,
4302 unsigned long zone_start_pfn,
4303 unsigned long size,
4304 enum memmap_context context)
4306 struct pglist_data *pgdat = zone->zone_pgdat;
4307 int ret;
4308 ret = zone_wait_table_init(zone, size);
4309 if (ret)
4310 return ret;
4311 pgdat->nr_zones = zone_idx(zone) + 1;
4313 zone->zone_start_pfn = zone_start_pfn;
4315 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4316 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4317 pgdat->node_id,
4318 (unsigned long)zone_idx(zone),
4319 zone_start_pfn, (zone_start_pfn + size));
4321 zone_init_free_lists(zone);
4323 return 0;
4326 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4327 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4329 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4330 * Architectures may implement their own version but if add_active_range()
4331 * was used and there are no special requirements, this is a convenient
4332 * alternative
4334 int __meminit __early_pfn_to_nid(unsigned long pfn)
4336 unsigned long start_pfn, end_pfn;
4337 int nid;
4339 * NOTE: The following SMP-unsafe globals are only used early in boot
4340 * when the kernel is running single-threaded.
4342 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4343 static int __meminitdata last_nid;
4345 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4346 return last_nid;
4348 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4349 if (nid != -1) {
4350 last_start_pfn = start_pfn;
4351 last_end_pfn = end_pfn;
4352 last_nid = nid;
4355 return nid;
4357 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4359 int __meminit early_pfn_to_nid(unsigned long pfn)
4361 int nid;
4363 nid = __early_pfn_to_nid(pfn);
4364 if (nid >= 0)
4365 return nid;
4366 /* just returns 0 */
4367 return 0;
4370 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4371 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4373 int nid;
4375 nid = __early_pfn_to_nid(pfn);
4376 if (nid >= 0 && nid != node)
4377 return false;
4378 return true;
4380 #endif
4383 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4384 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4385 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4387 * If an architecture guarantees that all ranges registered with
4388 * add_active_ranges() contain no holes and may be freed, this
4389 * this function may be used instead of calling memblock_free_early_nid()
4390 * manually.
4392 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4394 unsigned long start_pfn, end_pfn;
4395 int i, this_nid;
4397 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4398 start_pfn = min(start_pfn, max_low_pfn);
4399 end_pfn = min(end_pfn, max_low_pfn);
4401 if (start_pfn < end_pfn)
4402 memblock_free_early_nid(PFN_PHYS(start_pfn),
4403 (end_pfn - start_pfn) << PAGE_SHIFT,
4404 this_nid);
4409 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4410 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4412 * If an architecture guarantees that all ranges registered with
4413 * add_active_ranges() contain no holes and may be freed, this
4414 * function may be used instead of calling memory_present() manually.
4416 void __init sparse_memory_present_with_active_regions(int nid)
4418 unsigned long start_pfn, end_pfn;
4419 int i, this_nid;
4421 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4422 memory_present(this_nid, start_pfn, end_pfn);
4426 * get_pfn_range_for_nid - Return the start and end page frames for a node
4427 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4428 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4429 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4431 * It returns the start and end page frame of a node based on information
4432 * provided by an arch calling add_active_range(). If called for a node
4433 * with no available memory, a warning is printed and the start and end
4434 * PFNs will be 0.
4436 void __meminit get_pfn_range_for_nid(unsigned int nid,
4437 unsigned long *start_pfn, unsigned long *end_pfn)
4439 unsigned long this_start_pfn, this_end_pfn;
4440 int i;
4442 *start_pfn = -1UL;
4443 *end_pfn = 0;
4445 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4446 *start_pfn = min(*start_pfn, this_start_pfn);
4447 *end_pfn = max(*end_pfn, this_end_pfn);
4450 if (*start_pfn == -1UL)
4451 *start_pfn = 0;
4455 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4456 * assumption is made that zones within a node are ordered in monotonic
4457 * increasing memory addresses so that the "highest" populated zone is used
4459 static void __init find_usable_zone_for_movable(void)
4461 int zone_index;
4462 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4463 if (zone_index == ZONE_MOVABLE)
4464 continue;
4466 if (arch_zone_highest_possible_pfn[zone_index] >
4467 arch_zone_lowest_possible_pfn[zone_index])
4468 break;
4471 VM_BUG_ON(zone_index == -1);
4472 movable_zone = zone_index;
4476 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4477 * because it is sized independent of architecture. Unlike the other zones,
4478 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4479 * in each node depending on the size of each node and how evenly kernelcore
4480 * is distributed. This helper function adjusts the zone ranges
4481 * provided by the architecture for a given node by using the end of the
4482 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4483 * zones within a node are in order of monotonic increases memory addresses
4485 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4486 unsigned long zone_type,
4487 unsigned long node_start_pfn,
4488 unsigned long node_end_pfn,
4489 unsigned long *zone_start_pfn,
4490 unsigned long *zone_end_pfn)
4492 /* Only adjust if ZONE_MOVABLE is on this node */
4493 if (zone_movable_pfn[nid]) {
4494 /* Size ZONE_MOVABLE */
4495 if (zone_type == ZONE_MOVABLE) {
4496 *zone_start_pfn = zone_movable_pfn[nid];
4497 *zone_end_pfn = min(node_end_pfn,
4498 arch_zone_highest_possible_pfn[movable_zone]);
4500 /* Adjust for ZONE_MOVABLE starting within this range */
4501 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4502 *zone_end_pfn > zone_movable_pfn[nid]) {
4503 *zone_end_pfn = zone_movable_pfn[nid];
4505 /* Check if this whole range is within ZONE_MOVABLE */
4506 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4507 *zone_start_pfn = *zone_end_pfn;
4512 * Return the number of pages a zone spans in a node, including holes
4513 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4515 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4516 unsigned long zone_type,
4517 unsigned long node_start_pfn,
4518 unsigned long node_end_pfn,
4519 unsigned long *ignored)
4521 unsigned long zone_start_pfn, zone_end_pfn;
4523 /* Get the start and end of the zone */
4524 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4525 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4526 adjust_zone_range_for_zone_movable(nid, zone_type,
4527 node_start_pfn, node_end_pfn,
4528 &zone_start_pfn, &zone_end_pfn);
4530 /* Check that this node has pages within the zone's required range */
4531 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4532 return 0;
4534 /* Move the zone boundaries inside the node if necessary */
4535 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4536 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4538 /* Return the spanned pages */
4539 return zone_end_pfn - zone_start_pfn;
4543 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4544 * then all holes in the requested range will be accounted for.
4546 unsigned long __meminit __absent_pages_in_range(int nid,
4547 unsigned long range_start_pfn,
4548 unsigned long range_end_pfn)
4550 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4551 unsigned long start_pfn, end_pfn;
4552 int i;
4554 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4555 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4556 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4557 nr_absent -= end_pfn - start_pfn;
4559 return nr_absent;
4563 * absent_pages_in_range - Return number of page frames in holes within a range
4564 * @start_pfn: The start PFN to start searching for holes
4565 * @end_pfn: The end PFN to stop searching for holes
4567 * It returns the number of pages frames in memory holes within a range.
4569 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4570 unsigned long end_pfn)
4572 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4575 /* Return the number of page frames in holes in a zone on a node */
4576 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4577 unsigned long zone_type,
4578 unsigned long node_start_pfn,
4579 unsigned long node_end_pfn,
4580 unsigned long *ignored)
4582 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4583 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4584 unsigned long zone_start_pfn, zone_end_pfn;
4586 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4587 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4589 adjust_zone_range_for_zone_movable(nid, zone_type,
4590 node_start_pfn, node_end_pfn,
4591 &zone_start_pfn, &zone_end_pfn);
4592 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4595 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4596 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4597 unsigned long zone_type,
4598 unsigned long node_start_pfn,
4599 unsigned long node_end_pfn,
4600 unsigned long *zones_size)
4602 return zones_size[zone_type];
4605 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4606 unsigned long zone_type,
4607 unsigned long node_start_pfn,
4608 unsigned long node_end_pfn,
4609 unsigned long *zholes_size)
4611 if (!zholes_size)
4612 return 0;
4614 return zholes_size[zone_type];
4617 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4619 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4620 unsigned long node_start_pfn,
4621 unsigned long node_end_pfn,
4622 unsigned long *zones_size,
4623 unsigned long *zholes_size)
4625 unsigned long realtotalpages, totalpages = 0;
4626 enum zone_type i;
4628 for (i = 0; i < MAX_NR_ZONES; i++)
4629 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4630 node_start_pfn,
4631 node_end_pfn,
4632 zones_size);
4633 pgdat->node_spanned_pages = totalpages;
4635 realtotalpages = totalpages;
4636 for (i = 0; i < MAX_NR_ZONES; i++)
4637 realtotalpages -=
4638 zone_absent_pages_in_node(pgdat->node_id, i,
4639 node_start_pfn, node_end_pfn,
4640 zholes_size);
4641 pgdat->node_present_pages = realtotalpages;
4642 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4643 realtotalpages);
4646 #ifndef CONFIG_SPARSEMEM
4648 * Calculate the size of the zone->blockflags rounded to an unsigned long
4649 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4650 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4651 * round what is now in bits to nearest long in bits, then return it in
4652 * bytes.
4654 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4656 unsigned long usemapsize;
4658 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4659 usemapsize = roundup(zonesize, pageblock_nr_pages);
4660 usemapsize = usemapsize >> pageblock_order;
4661 usemapsize *= NR_PAGEBLOCK_BITS;
4662 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4664 return usemapsize / 8;
4667 static void __init setup_usemap(struct pglist_data *pgdat,
4668 struct zone *zone,
4669 unsigned long zone_start_pfn,
4670 unsigned long zonesize)
4672 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4673 zone->pageblock_flags = NULL;
4674 if (usemapsize)
4675 zone->pageblock_flags =
4676 memblock_virt_alloc_node_nopanic(usemapsize,
4677 pgdat->node_id);
4679 #else
4680 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4681 unsigned long zone_start_pfn, unsigned long zonesize) {}
4682 #endif /* CONFIG_SPARSEMEM */
4684 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4686 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4687 void __paginginit set_pageblock_order(void)
4689 unsigned int order;
4691 /* Check that pageblock_nr_pages has not already been setup */
4692 if (pageblock_order)
4693 return;
4695 if (HPAGE_SHIFT > PAGE_SHIFT)
4696 order = HUGETLB_PAGE_ORDER;
4697 else
4698 order = MAX_ORDER - 1;
4701 * Assume the largest contiguous order of interest is a huge page.
4702 * This value may be variable depending on boot parameters on IA64 and
4703 * powerpc.
4705 pageblock_order = order;
4707 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4710 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4711 * is unused as pageblock_order is set at compile-time. See
4712 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4713 * the kernel config
4715 void __paginginit set_pageblock_order(void)
4719 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4721 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4722 unsigned long present_pages)
4724 unsigned long pages = spanned_pages;
4727 * Provide a more accurate estimation if there are holes within
4728 * the zone and SPARSEMEM is in use. If there are holes within the
4729 * zone, each populated memory region may cost us one or two extra
4730 * memmap pages due to alignment because memmap pages for each
4731 * populated regions may not naturally algined on page boundary.
4732 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4734 if (spanned_pages > present_pages + (present_pages >> 4) &&
4735 IS_ENABLED(CONFIG_SPARSEMEM))
4736 pages = present_pages;
4738 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4742 * Set up the zone data structures:
4743 * - mark all pages reserved
4744 * - mark all memory queues empty
4745 * - clear the memory bitmaps
4747 * NOTE: pgdat should get zeroed by caller.
4749 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4750 unsigned long node_start_pfn, unsigned long node_end_pfn,
4751 unsigned long *zones_size, unsigned long *zholes_size)
4753 enum zone_type j;
4754 int nid = pgdat->node_id;
4755 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4756 int ret;
4758 pgdat_resize_init(pgdat);
4759 #ifdef CONFIG_NUMA_BALANCING
4760 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4761 pgdat->numabalancing_migrate_nr_pages = 0;
4762 pgdat->numabalancing_migrate_next_window = jiffies;
4763 #endif
4764 init_waitqueue_head(&pgdat->kswapd_wait);
4765 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4766 pgdat_page_cgroup_init(pgdat);
4768 for (j = 0; j < MAX_NR_ZONES; j++) {
4769 struct zone *zone = pgdat->node_zones + j;
4770 unsigned long size, realsize, freesize, memmap_pages;
4772 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4773 node_end_pfn, zones_size);
4774 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4775 node_start_pfn,
4776 node_end_pfn,
4777 zholes_size);
4780 * Adjust freesize so that it accounts for how much memory
4781 * is used by this zone for memmap. This affects the watermark
4782 * and per-cpu initialisations
4784 memmap_pages = calc_memmap_size(size, realsize);
4785 if (freesize >= memmap_pages) {
4786 freesize -= memmap_pages;
4787 if (memmap_pages)
4788 printk(KERN_DEBUG
4789 " %s zone: %lu pages used for memmap\n",
4790 zone_names[j], memmap_pages);
4791 } else
4792 printk(KERN_WARNING
4793 " %s zone: %lu pages exceeds freesize %lu\n",
4794 zone_names[j], memmap_pages, freesize);
4796 /* Account for reserved pages */
4797 if (j == 0 && freesize > dma_reserve) {
4798 freesize -= dma_reserve;
4799 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4800 zone_names[0], dma_reserve);
4803 if (!is_highmem_idx(j))
4804 nr_kernel_pages += freesize;
4805 /* Charge for highmem memmap if there are enough kernel pages */
4806 else if (nr_kernel_pages > memmap_pages * 2)
4807 nr_kernel_pages -= memmap_pages;
4808 nr_all_pages += freesize;
4810 zone->spanned_pages = size;
4811 zone->present_pages = realsize;
4813 * Set an approximate value for lowmem here, it will be adjusted
4814 * when the bootmem allocator frees pages into the buddy system.
4815 * And all highmem pages will be managed by the buddy system.
4817 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4818 #ifdef CONFIG_NUMA
4819 zone->node = nid;
4820 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4821 / 100;
4822 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4823 #endif
4824 zone->name = zone_names[j];
4825 spin_lock_init(&zone->lock);
4826 spin_lock_init(&zone->lru_lock);
4827 zone_seqlock_init(zone);
4828 zone->zone_pgdat = pgdat;
4829 zone_pcp_init(zone);
4831 /* For bootup, initialized properly in watermark setup */
4832 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4834 lruvec_init(&zone->lruvec);
4835 if (!size)
4836 continue;
4838 set_pageblock_order();
4839 setup_usemap(pgdat, zone, zone_start_pfn, size);
4840 ret = init_currently_empty_zone(zone, zone_start_pfn,
4841 size, MEMMAP_EARLY);
4842 BUG_ON(ret);
4843 memmap_init(size, nid, j, zone_start_pfn);
4844 zone_start_pfn += size;
4848 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4850 /* Skip empty nodes */
4851 if (!pgdat->node_spanned_pages)
4852 return;
4854 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4855 /* ia64 gets its own node_mem_map, before this, without bootmem */
4856 if (!pgdat->node_mem_map) {
4857 unsigned long size, start, end;
4858 struct page *map;
4861 * The zone's endpoints aren't required to be MAX_ORDER
4862 * aligned but the node_mem_map endpoints must be in order
4863 * for the buddy allocator to function correctly.
4865 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4866 end = pgdat_end_pfn(pgdat);
4867 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4868 size = (end - start) * sizeof(struct page);
4869 map = alloc_remap(pgdat->node_id, size);
4870 if (!map)
4871 map = memblock_virt_alloc_node_nopanic(size,
4872 pgdat->node_id);
4873 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4875 #ifndef CONFIG_NEED_MULTIPLE_NODES
4877 * With no DISCONTIG, the global mem_map is just set as node 0's
4879 if (pgdat == NODE_DATA(0)) {
4880 mem_map = NODE_DATA(0)->node_mem_map;
4881 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4882 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4883 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4884 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4886 #endif
4887 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4890 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4891 unsigned long node_start_pfn, unsigned long *zholes_size)
4893 pg_data_t *pgdat = NODE_DATA(nid);
4894 unsigned long start_pfn = 0;
4895 unsigned long end_pfn = 0;
4897 /* pg_data_t should be reset to zero when it's allocated */
4898 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
4900 pgdat->node_id = nid;
4901 pgdat->node_start_pfn = node_start_pfn;
4902 init_zone_allows_reclaim(nid);
4903 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4904 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
4905 #endif
4906 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
4907 zones_size, zholes_size);
4909 alloc_node_mem_map(pgdat);
4910 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4911 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4912 nid, (unsigned long)pgdat,
4913 (unsigned long)pgdat->node_mem_map);
4914 #endif
4916 free_area_init_core(pgdat, start_pfn, end_pfn,
4917 zones_size, zholes_size);
4920 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4922 #if MAX_NUMNODES > 1
4924 * Figure out the number of possible node ids.
4926 void __init setup_nr_node_ids(void)
4928 unsigned int node;
4929 unsigned int highest = 0;
4931 for_each_node_mask(node, node_possible_map)
4932 highest = node;
4933 nr_node_ids = highest + 1;
4935 #endif
4938 * node_map_pfn_alignment - determine the maximum internode alignment
4940 * This function should be called after node map is populated and sorted.
4941 * It calculates the maximum power of two alignment which can distinguish
4942 * all the nodes.
4944 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4945 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4946 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4947 * shifted, 1GiB is enough and this function will indicate so.
4949 * This is used to test whether pfn -> nid mapping of the chosen memory
4950 * model has fine enough granularity to avoid incorrect mapping for the
4951 * populated node map.
4953 * Returns the determined alignment in pfn's. 0 if there is no alignment
4954 * requirement (single node).
4956 unsigned long __init node_map_pfn_alignment(void)
4958 unsigned long accl_mask = 0, last_end = 0;
4959 unsigned long start, end, mask;
4960 int last_nid = -1;
4961 int i, nid;
4963 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
4964 if (!start || last_nid < 0 || last_nid == nid) {
4965 last_nid = nid;
4966 last_end = end;
4967 continue;
4971 * Start with a mask granular enough to pin-point to the
4972 * start pfn and tick off bits one-by-one until it becomes
4973 * too coarse to separate the current node from the last.
4975 mask = ~((1 << __ffs(start)) - 1);
4976 while (mask && last_end <= (start & (mask << 1)))
4977 mask <<= 1;
4979 /* accumulate all internode masks */
4980 accl_mask |= mask;
4983 /* convert mask to number of pages */
4984 return ~accl_mask + 1;
4987 /* Find the lowest pfn for a node */
4988 static unsigned long __init find_min_pfn_for_node(int nid)
4990 unsigned long min_pfn = ULONG_MAX;
4991 unsigned long start_pfn;
4992 int i;
4994 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
4995 min_pfn = min(min_pfn, start_pfn);
4997 if (min_pfn == ULONG_MAX) {
4998 printk(KERN_WARNING
4999 "Could not find start_pfn for node %d\n", nid);
5000 return 0;
5003 return min_pfn;
5007 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5009 * It returns the minimum PFN based on information provided via
5010 * add_active_range().
5012 unsigned long __init find_min_pfn_with_active_regions(void)
5014 return find_min_pfn_for_node(MAX_NUMNODES);
5018 * early_calculate_totalpages()
5019 * Sum pages in active regions for movable zone.
5020 * Populate N_MEMORY for calculating usable_nodes.
5022 static unsigned long __init early_calculate_totalpages(void)
5024 unsigned long totalpages = 0;
5025 unsigned long start_pfn, end_pfn;
5026 int i, nid;
5028 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5029 unsigned long pages = end_pfn - start_pfn;
5031 totalpages += pages;
5032 if (pages)
5033 node_set_state(nid, N_MEMORY);
5035 return totalpages;
5039 * Find the PFN the Movable zone begins in each node. Kernel memory
5040 * is spread evenly between nodes as long as the nodes have enough
5041 * memory. When they don't, some nodes will have more kernelcore than
5042 * others
5044 static void __init find_zone_movable_pfns_for_nodes(void)
5046 int i, nid;
5047 unsigned long usable_startpfn;
5048 unsigned long kernelcore_node, kernelcore_remaining;
5049 /* save the state before borrow the nodemask */
5050 nodemask_t saved_node_state = node_states[N_MEMORY];
5051 unsigned long totalpages = early_calculate_totalpages();
5052 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5053 struct memblock_type *type = &memblock.memory;
5055 /* Need to find movable_zone earlier when movable_node is specified. */
5056 find_usable_zone_for_movable();
5059 * If movable_node is specified, ignore kernelcore and movablecore
5060 * options.
5062 if (movable_node_is_enabled()) {
5063 for (i = 0; i < type->cnt; i++) {
5064 if (!memblock_is_hotpluggable(&type->regions[i]))
5065 continue;
5067 nid = type->regions[i].nid;
5069 usable_startpfn = PFN_DOWN(type->regions[i].base);
5070 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5071 min(usable_startpfn, zone_movable_pfn[nid]) :
5072 usable_startpfn;
5075 goto out2;
5079 * If movablecore=nn[KMG] was specified, calculate what size of
5080 * kernelcore that corresponds so that memory usable for
5081 * any allocation type is evenly spread. If both kernelcore
5082 * and movablecore are specified, then the value of kernelcore
5083 * will be used for required_kernelcore if it's greater than
5084 * what movablecore would have allowed.
5086 if (required_movablecore) {
5087 unsigned long corepages;
5090 * Round-up so that ZONE_MOVABLE is at least as large as what
5091 * was requested by the user
5093 required_movablecore =
5094 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5095 corepages = totalpages - required_movablecore;
5097 required_kernelcore = max(required_kernelcore, corepages);
5100 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5101 if (!required_kernelcore)
5102 goto out;
5104 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5105 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5107 restart:
5108 /* Spread kernelcore memory as evenly as possible throughout nodes */
5109 kernelcore_node = required_kernelcore / usable_nodes;
5110 for_each_node_state(nid, N_MEMORY) {
5111 unsigned long start_pfn, end_pfn;
5114 * Recalculate kernelcore_node if the division per node
5115 * now exceeds what is necessary to satisfy the requested
5116 * amount of memory for the kernel
5118 if (required_kernelcore < kernelcore_node)
5119 kernelcore_node = required_kernelcore / usable_nodes;
5122 * As the map is walked, we track how much memory is usable
5123 * by the kernel using kernelcore_remaining. When it is
5124 * 0, the rest of the node is usable by ZONE_MOVABLE
5126 kernelcore_remaining = kernelcore_node;
5128 /* Go through each range of PFNs within this node */
5129 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5130 unsigned long size_pages;
5132 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5133 if (start_pfn >= end_pfn)
5134 continue;
5136 /* Account for what is only usable for kernelcore */
5137 if (start_pfn < usable_startpfn) {
5138 unsigned long kernel_pages;
5139 kernel_pages = min(end_pfn, usable_startpfn)
5140 - start_pfn;
5142 kernelcore_remaining -= min(kernel_pages,
5143 kernelcore_remaining);
5144 required_kernelcore -= min(kernel_pages,
5145 required_kernelcore);
5147 /* Continue if range is now fully accounted */
5148 if (end_pfn <= usable_startpfn) {
5151 * Push zone_movable_pfn to the end so
5152 * that if we have to rebalance
5153 * kernelcore across nodes, we will
5154 * not double account here
5156 zone_movable_pfn[nid] = end_pfn;
5157 continue;
5159 start_pfn = usable_startpfn;
5163 * The usable PFN range for ZONE_MOVABLE is from
5164 * start_pfn->end_pfn. Calculate size_pages as the
5165 * number of pages used as kernelcore
5167 size_pages = end_pfn - start_pfn;
5168 if (size_pages > kernelcore_remaining)
5169 size_pages = kernelcore_remaining;
5170 zone_movable_pfn[nid] = start_pfn + size_pages;
5173 * Some kernelcore has been met, update counts and
5174 * break if the kernelcore for this node has been
5175 * satisfied
5177 required_kernelcore -= min(required_kernelcore,
5178 size_pages);
5179 kernelcore_remaining -= size_pages;
5180 if (!kernelcore_remaining)
5181 break;
5186 * If there is still required_kernelcore, we do another pass with one
5187 * less node in the count. This will push zone_movable_pfn[nid] further
5188 * along on the nodes that still have memory until kernelcore is
5189 * satisfied
5191 usable_nodes--;
5192 if (usable_nodes && required_kernelcore > usable_nodes)
5193 goto restart;
5195 out2:
5196 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5197 for (nid = 0; nid < MAX_NUMNODES; nid++)
5198 zone_movable_pfn[nid] =
5199 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5201 out:
5202 /* restore the node_state */
5203 node_states[N_MEMORY] = saved_node_state;
5206 /* Any regular or high memory on that node ? */
5207 static void check_for_memory(pg_data_t *pgdat, int nid)
5209 enum zone_type zone_type;
5211 if (N_MEMORY == N_NORMAL_MEMORY)
5212 return;
5214 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5215 struct zone *zone = &pgdat->node_zones[zone_type];
5216 if (populated_zone(zone)) {
5217 node_set_state(nid, N_HIGH_MEMORY);
5218 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5219 zone_type <= ZONE_NORMAL)
5220 node_set_state(nid, N_NORMAL_MEMORY);
5221 break;
5227 * free_area_init_nodes - Initialise all pg_data_t and zone data
5228 * @max_zone_pfn: an array of max PFNs for each zone
5230 * This will call free_area_init_node() for each active node in the system.
5231 * Using the page ranges provided by add_active_range(), the size of each
5232 * zone in each node and their holes is calculated. If the maximum PFN
5233 * between two adjacent zones match, it is assumed that the zone is empty.
5234 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5235 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5236 * starts where the previous one ended. For example, ZONE_DMA32 starts
5237 * at arch_max_dma_pfn.
5239 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5241 unsigned long start_pfn, end_pfn;
5242 int i, nid;
5244 /* Record where the zone boundaries are */
5245 memset(arch_zone_lowest_possible_pfn, 0,
5246 sizeof(arch_zone_lowest_possible_pfn));
5247 memset(arch_zone_highest_possible_pfn, 0,
5248 sizeof(arch_zone_highest_possible_pfn));
5249 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5250 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5251 for (i = 1; i < MAX_NR_ZONES; i++) {
5252 if (i == ZONE_MOVABLE)
5253 continue;
5254 arch_zone_lowest_possible_pfn[i] =
5255 arch_zone_highest_possible_pfn[i-1];
5256 arch_zone_highest_possible_pfn[i] =
5257 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5259 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5260 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5262 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5263 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5264 find_zone_movable_pfns_for_nodes();
5266 /* Print out the zone ranges */
5267 printk("Zone ranges:\n");
5268 for (i = 0; i < MAX_NR_ZONES; i++) {
5269 if (i == ZONE_MOVABLE)
5270 continue;
5271 printk(KERN_CONT " %-8s ", zone_names[i]);
5272 if (arch_zone_lowest_possible_pfn[i] ==
5273 arch_zone_highest_possible_pfn[i])
5274 printk(KERN_CONT "empty\n");
5275 else
5276 printk(KERN_CONT "[mem %0#10lx-%0#10lx]\n",
5277 arch_zone_lowest_possible_pfn[i] << PAGE_SHIFT,
5278 (arch_zone_highest_possible_pfn[i]
5279 << PAGE_SHIFT) - 1);
5282 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5283 printk("Movable zone start for each node\n");
5284 for (i = 0; i < MAX_NUMNODES; i++) {
5285 if (zone_movable_pfn[i])
5286 printk(" Node %d: %#010lx\n", i,
5287 zone_movable_pfn[i] << PAGE_SHIFT);
5290 /* Print out the early node map */
5291 printk("Early memory node ranges\n");
5292 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5293 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid,
5294 start_pfn << PAGE_SHIFT, (end_pfn << PAGE_SHIFT) - 1);
5296 /* Initialise every node */
5297 mminit_verify_pageflags_layout();
5298 setup_nr_node_ids();
5299 for_each_online_node(nid) {
5300 pg_data_t *pgdat = NODE_DATA(nid);
5301 free_area_init_node(nid, NULL,
5302 find_min_pfn_for_node(nid), NULL);
5304 /* Any memory on that node */
5305 if (pgdat->node_present_pages)
5306 node_set_state(nid, N_MEMORY);
5307 check_for_memory(pgdat, nid);
5311 static int __init cmdline_parse_core(char *p, unsigned long *core)
5313 unsigned long long coremem;
5314 if (!p)
5315 return -EINVAL;
5317 coremem = memparse(p, &p);
5318 *core = coremem >> PAGE_SHIFT;
5320 /* Paranoid check that UL is enough for the coremem value */
5321 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5323 return 0;
5327 * kernelcore=size sets the amount of memory for use for allocations that
5328 * cannot be reclaimed or migrated.
5330 static int __init cmdline_parse_kernelcore(char *p)
5332 return cmdline_parse_core(p, &required_kernelcore);
5336 * movablecore=size sets the amount of memory for use for allocations that
5337 * can be reclaimed or migrated.
5339 static int __init cmdline_parse_movablecore(char *p)
5341 return cmdline_parse_core(p, &required_movablecore);
5344 early_param("kernelcore", cmdline_parse_kernelcore);
5345 early_param("movablecore", cmdline_parse_movablecore);
5347 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5349 void adjust_managed_page_count(struct page *page, long count)
5351 spin_lock(&managed_page_count_lock);
5352 page_zone(page)->managed_pages += count;
5353 totalram_pages += count;
5354 #ifdef CONFIG_HIGHMEM
5355 if (PageHighMem(page))
5356 totalhigh_pages += count;
5357 #endif
5358 spin_unlock(&managed_page_count_lock);
5360 EXPORT_SYMBOL(adjust_managed_page_count);
5362 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5364 void *pos;
5365 unsigned long pages = 0;
5367 start = (void *)PAGE_ALIGN((unsigned long)start);
5368 end = (void *)((unsigned long)end & PAGE_MASK);
5369 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5370 if ((unsigned int)poison <= 0xFF)
5371 memset(pos, poison, PAGE_SIZE);
5372 free_reserved_page(virt_to_page(pos));
5375 if (pages && s)
5376 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5377 s, pages << (PAGE_SHIFT - 10), start, end);
5379 return pages;
5381 EXPORT_SYMBOL(free_reserved_area);
5383 #ifdef CONFIG_HIGHMEM
5384 void free_highmem_page(struct page *page)
5386 __free_reserved_page(page);
5387 totalram_pages++;
5388 page_zone(page)->managed_pages++;
5389 totalhigh_pages++;
5391 #endif
5394 void __init mem_init_print_info(const char *str)
5396 unsigned long physpages, codesize, datasize, rosize, bss_size;
5397 unsigned long init_code_size, init_data_size;
5399 physpages = get_num_physpages();
5400 codesize = _etext - _stext;
5401 datasize = _edata - _sdata;
5402 rosize = __end_rodata - __start_rodata;
5403 bss_size = __bss_stop - __bss_start;
5404 init_data_size = __init_end - __init_begin;
5405 init_code_size = _einittext - _sinittext;
5408 * Detect special cases and adjust section sizes accordingly:
5409 * 1) .init.* may be embedded into .data sections
5410 * 2) .init.text.* may be out of [__init_begin, __init_end],
5411 * please refer to arch/tile/kernel/vmlinux.lds.S.
5412 * 3) .rodata.* may be embedded into .text or .data sections.
5414 #define adj_init_size(start, end, size, pos, adj) \
5415 do { \
5416 if (start <= pos && pos < end && size > adj) \
5417 size -= adj; \
5418 } while (0)
5420 adj_init_size(__init_begin, __init_end, init_data_size,
5421 _sinittext, init_code_size);
5422 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5423 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5424 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5425 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5427 #undef adj_init_size
5429 printk("Memory: %luK/%luK available "
5430 "(%luK kernel code, %luK rwdata, %luK rodata, "
5431 "%luK init, %luK bss, %luK reserved"
5432 #ifdef CONFIG_HIGHMEM
5433 ", %luK highmem"
5434 #endif
5435 "%s%s)\n",
5436 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5437 codesize >> 10, datasize >> 10, rosize >> 10,
5438 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5439 (physpages - totalram_pages) << (PAGE_SHIFT-10),
5440 #ifdef CONFIG_HIGHMEM
5441 totalhigh_pages << (PAGE_SHIFT-10),
5442 #endif
5443 str ? ", " : "", str ? str : "");
5447 * set_dma_reserve - set the specified number of pages reserved in the first zone
5448 * @new_dma_reserve: The number of pages to mark reserved
5450 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5451 * In the DMA zone, a significant percentage may be consumed by kernel image
5452 * and other unfreeable allocations which can skew the watermarks badly. This
5453 * function may optionally be used to account for unfreeable pages in the
5454 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5455 * smaller per-cpu batchsize.
5457 void __init set_dma_reserve(unsigned long new_dma_reserve)
5459 dma_reserve = new_dma_reserve;
5462 void __init free_area_init(unsigned long *zones_size)
5464 free_area_init_node(0, zones_size,
5465 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5468 static int page_alloc_cpu_notify(struct notifier_block *self,
5469 unsigned long action, void *hcpu)
5471 int cpu = (unsigned long)hcpu;
5473 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5474 lru_add_drain_cpu(cpu);
5475 drain_pages(cpu);
5478 * Spill the event counters of the dead processor
5479 * into the current processors event counters.
5480 * This artificially elevates the count of the current
5481 * processor.
5483 vm_events_fold_cpu(cpu);
5486 * Zero the differential counters of the dead processor
5487 * so that the vm statistics are consistent.
5489 * This is only okay since the processor is dead and cannot
5490 * race with what we are doing.
5492 cpu_vm_stats_fold(cpu);
5494 return NOTIFY_OK;
5497 void __init page_alloc_init(void)
5499 hotcpu_notifier(page_alloc_cpu_notify, 0);
5503 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5504 * or min_free_kbytes changes.
5506 static void calculate_totalreserve_pages(void)
5508 struct pglist_data *pgdat;
5509 unsigned long reserve_pages = 0;
5510 enum zone_type i, j;
5512 for_each_online_pgdat(pgdat) {
5513 for (i = 0; i < MAX_NR_ZONES; i++) {
5514 struct zone *zone = pgdat->node_zones + i;
5515 unsigned long max = 0;
5517 /* Find valid and maximum lowmem_reserve in the zone */
5518 for (j = i; j < MAX_NR_ZONES; j++) {
5519 if (zone->lowmem_reserve[j] > max)
5520 max = zone->lowmem_reserve[j];
5523 /* we treat the high watermark as reserved pages. */
5524 max += high_wmark_pages(zone);
5526 if (max > zone->managed_pages)
5527 max = zone->managed_pages;
5528 reserve_pages += max;
5530 * Lowmem reserves are not available to
5531 * GFP_HIGHUSER page cache allocations and
5532 * kswapd tries to balance zones to their high
5533 * watermark. As a result, neither should be
5534 * regarded as dirtyable memory, to prevent a
5535 * situation where reclaim has to clean pages
5536 * in order to balance the zones.
5538 zone->dirty_balance_reserve = max;
5541 dirty_balance_reserve = reserve_pages;
5542 totalreserve_pages = reserve_pages;
5546 * setup_per_zone_lowmem_reserve - called whenever
5547 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5548 * has a correct pages reserved value, so an adequate number of
5549 * pages are left in the zone after a successful __alloc_pages().
5551 static void setup_per_zone_lowmem_reserve(void)
5553 struct pglist_data *pgdat;
5554 enum zone_type j, idx;
5556 for_each_online_pgdat(pgdat) {
5557 for (j = 0; j < MAX_NR_ZONES; j++) {
5558 struct zone *zone = pgdat->node_zones + j;
5559 unsigned long managed_pages = zone->managed_pages;
5561 zone->lowmem_reserve[j] = 0;
5563 idx = j;
5564 while (idx) {
5565 struct zone *lower_zone;
5567 idx--;
5569 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5570 sysctl_lowmem_reserve_ratio[idx] = 1;
5572 lower_zone = pgdat->node_zones + idx;
5573 lower_zone->lowmem_reserve[j] = managed_pages /
5574 sysctl_lowmem_reserve_ratio[idx];
5575 managed_pages += lower_zone->managed_pages;
5580 /* update totalreserve_pages */
5581 calculate_totalreserve_pages();
5584 static void __setup_per_zone_wmarks(void)
5586 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5587 unsigned long lowmem_pages = 0;
5588 struct zone *zone;
5589 unsigned long flags;
5591 /* Calculate total number of !ZONE_HIGHMEM pages */
5592 for_each_zone(zone) {
5593 if (!is_highmem(zone))
5594 lowmem_pages += zone->managed_pages;
5597 for_each_zone(zone) {
5598 u64 tmp;
5600 spin_lock_irqsave(&zone->lock, flags);
5601 tmp = (u64)pages_min * zone->managed_pages;
5602 do_div(tmp, lowmem_pages);
5603 if (is_highmem(zone)) {
5605 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5606 * need highmem pages, so cap pages_min to a small
5607 * value here.
5609 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5610 * deltas controls asynch page reclaim, and so should
5611 * not be capped for highmem.
5613 unsigned long min_pages;
5615 min_pages = zone->managed_pages / 1024;
5616 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5617 zone->watermark[WMARK_MIN] = min_pages;
5618 } else {
5620 * If it's a lowmem zone, reserve a number of pages
5621 * proportionate to the zone's size.
5623 zone->watermark[WMARK_MIN] = tmp;
5626 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5627 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5629 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5630 high_wmark_pages(zone) -
5631 low_wmark_pages(zone) -
5632 zone_page_state(zone, NR_ALLOC_BATCH));
5634 setup_zone_migrate_reserve(zone);
5635 spin_unlock_irqrestore(&zone->lock, flags);
5638 /* update totalreserve_pages */
5639 calculate_totalreserve_pages();
5643 * setup_per_zone_wmarks - called when min_free_kbytes changes
5644 * or when memory is hot-{added|removed}
5646 * Ensures that the watermark[min,low,high] values for each zone are set
5647 * correctly with respect to min_free_kbytes.
5649 void setup_per_zone_wmarks(void)
5651 mutex_lock(&zonelists_mutex);
5652 __setup_per_zone_wmarks();
5653 mutex_unlock(&zonelists_mutex);
5657 * The inactive anon list should be small enough that the VM never has to
5658 * do too much work, but large enough that each inactive page has a chance
5659 * to be referenced again before it is swapped out.
5661 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5662 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5663 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5664 * the anonymous pages are kept on the inactive list.
5666 * total target max
5667 * memory ratio inactive anon
5668 * -------------------------------------
5669 * 10MB 1 5MB
5670 * 100MB 1 50MB
5671 * 1GB 3 250MB
5672 * 10GB 10 0.9GB
5673 * 100GB 31 3GB
5674 * 1TB 101 10GB
5675 * 10TB 320 32GB
5677 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5679 unsigned int gb, ratio;
5681 /* Zone size in gigabytes */
5682 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5683 if (gb)
5684 ratio = int_sqrt(10 * gb);
5685 else
5686 ratio = 1;
5688 zone->inactive_ratio = ratio;
5691 static void __meminit setup_per_zone_inactive_ratio(void)
5693 struct zone *zone;
5695 for_each_zone(zone)
5696 calculate_zone_inactive_ratio(zone);
5700 * Initialise min_free_kbytes.
5702 * For small machines we want it small (128k min). For large machines
5703 * we want it large (64MB max). But it is not linear, because network
5704 * bandwidth does not increase linearly with machine size. We use
5706 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5707 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5709 * which yields
5711 * 16MB: 512k
5712 * 32MB: 724k
5713 * 64MB: 1024k
5714 * 128MB: 1448k
5715 * 256MB: 2048k
5716 * 512MB: 2896k
5717 * 1024MB: 4096k
5718 * 2048MB: 5792k
5719 * 4096MB: 8192k
5720 * 8192MB: 11584k
5721 * 16384MB: 16384k
5723 int __meminit init_per_zone_wmark_min(void)
5725 unsigned long lowmem_kbytes;
5726 int new_min_free_kbytes;
5728 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5729 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5731 if (new_min_free_kbytes > user_min_free_kbytes) {
5732 min_free_kbytes = new_min_free_kbytes;
5733 if (min_free_kbytes < 128)
5734 min_free_kbytes = 128;
5735 if (min_free_kbytes > 65536)
5736 min_free_kbytes = 65536;
5737 } else {
5738 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5739 new_min_free_kbytes, user_min_free_kbytes);
5741 setup_per_zone_wmarks();
5742 refresh_zone_stat_thresholds();
5743 setup_per_zone_lowmem_reserve();
5744 setup_per_zone_inactive_ratio();
5745 return 0;
5747 module_init(init_per_zone_wmark_min)
5750 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5751 * that we can call two helper functions whenever min_free_kbytes
5752 * changes.
5754 int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5755 void __user *buffer, size_t *length, loff_t *ppos)
5757 int rc;
5759 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5760 if (rc)
5761 return rc;
5763 if (write) {
5764 user_min_free_kbytes = min_free_kbytes;
5765 setup_per_zone_wmarks();
5767 return 0;
5770 #ifdef CONFIG_NUMA
5771 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5772 void __user *buffer, size_t *length, loff_t *ppos)
5774 struct zone *zone;
5775 int rc;
5777 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5778 if (rc)
5779 return rc;
5781 for_each_zone(zone)
5782 zone->min_unmapped_pages = (zone->managed_pages *
5783 sysctl_min_unmapped_ratio) / 100;
5784 return 0;
5787 int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5788 void __user *buffer, size_t *length, loff_t *ppos)
5790 struct zone *zone;
5791 int rc;
5793 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5794 if (rc)
5795 return rc;
5797 for_each_zone(zone)
5798 zone->min_slab_pages = (zone->managed_pages *
5799 sysctl_min_slab_ratio) / 100;
5800 return 0;
5802 #endif
5805 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5806 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5807 * whenever sysctl_lowmem_reserve_ratio changes.
5809 * The reserve ratio obviously has absolutely no relation with the
5810 * minimum watermarks. The lowmem reserve ratio can only make sense
5811 * if in function of the boot time zone sizes.
5813 int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5814 void __user *buffer, size_t *length, loff_t *ppos)
5816 proc_dointvec_minmax(table, write, buffer, length, ppos);
5817 setup_per_zone_lowmem_reserve();
5818 return 0;
5822 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5823 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5824 * pagelist can have before it gets flushed back to buddy allocator.
5826 int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5827 void __user *buffer, size_t *length, loff_t *ppos)
5829 struct zone *zone;
5830 unsigned int cpu;
5831 int ret;
5833 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5834 if (!write || (ret < 0))
5835 return ret;
5837 mutex_lock(&pcp_batch_high_lock);
5838 for_each_populated_zone(zone) {
5839 unsigned long high;
5840 high = zone->managed_pages / percpu_pagelist_fraction;
5841 for_each_possible_cpu(cpu)
5842 pageset_set_high(per_cpu_ptr(zone->pageset, cpu),
5843 high);
5845 mutex_unlock(&pcp_batch_high_lock);
5846 return 0;
5849 int hashdist = HASHDIST_DEFAULT;
5851 #ifdef CONFIG_NUMA
5852 static int __init set_hashdist(char *str)
5854 if (!str)
5855 return 0;
5856 hashdist = simple_strtoul(str, &str, 0);
5857 return 1;
5859 __setup("hashdist=", set_hashdist);
5860 #endif
5863 * allocate a large system hash table from bootmem
5864 * - it is assumed that the hash table must contain an exact power-of-2
5865 * quantity of entries
5866 * - limit is the number of hash buckets, not the total allocation size
5868 void *__init alloc_large_system_hash(const char *tablename,
5869 unsigned long bucketsize,
5870 unsigned long numentries,
5871 int scale,
5872 int flags,
5873 unsigned int *_hash_shift,
5874 unsigned int *_hash_mask,
5875 unsigned long low_limit,
5876 unsigned long high_limit)
5878 unsigned long long max = high_limit;
5879 unsigned long log2qty, size;
5880 void *table = NULL;
5882 /* allow the kernel cmdline to have a say */
5883 if (!numentries) {
5884 /* round applicable memory size up to nearest megabyte */
5885 numentries = nr_kernel_pages;
5887 /* It isn't necessary when PAGE_SIZE >= 1MB */
5888 if (PAGE_SHIFT < 20)
5889 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
5891 /* limit to 1 bucket per 2^scale bytes of low memory */
5892 if (scale > PAGE_SHIFT)
5893 numentries >>= (scale - PAGE_SHIFT);
5894 else
5895 numentries <<= (PAGE_SHIFT - scale);
5897 /* Make sure we've got at least a 0-order allocation.. */
5898 if (unlikely(flags & HASH_SMALL)) {
5899 /* Makes no sense without HASH_EARLY */
5900 WARN_ON(!(flags & HASH_EARLY));
5901 if (!(numentries >> *_hash_shift)) {
5902 numentries = 1UL << *_hash_shift;
5903 BUG_ON(!numentries);
5905 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5906 numentries = PAGE_SIZE / bucketsize;
5908 numentries = roundup_pow_of_two(numentries);
5910 /* limit allocation size to 1/16 total memory by default */
5911 if (max == 0) {
5912 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5913 do_div(max, bucketsize);
5915 max = min(max, 0x80000000ULL);
5917 if (numentries < low_limit)
5918 numentries = low_limit;
5919 if (numentries > max)
5920 numentries = max;
5922 log2qty = ilog2(numentries);
5924 do {
5925 size = bucketsize << log2qty;
5926 if (flags & HASH_EARLY)
5927 table = memblock_virt_alloc_nopanic(size, 0);
5928 else if (hashdist)
5929 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5930 else {
5932 * If bucketsize is not a power-of-two, we may free
5933 * some pages at the end of hash table which
5934 * alloc_pages_exact() automatically does
5936 if (get_order(size) < MAX_ORDER) {
5937 table = alloc_pages_exact(size, GFP_ATOMIC);
5938 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5941 } while (!table && size > PAGE_SIZE && --log2qty);
5943 if (!table)
5944 panic("Failed to allocate %s hash table\n", tablename);
5946 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5947 tablename,
5948 (1UL << log2qty),
5949 ilog2(size) - PAGE_SHIFT,
5950 size);
5952 if (_hash_shift)
5953 *_hash_shift = log2qty;
5954 if (_hash_mask)
5955 *_hash_mask = (1 << log2qty) - 1;
5957 return table;
5960 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5961 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5962 unsigned long pfn)
5964 #ifdef CONFIG_SPARSEMEM
5965 return __pfn_to_section(pfn)->pageblock_flags;
5966 #else
5967 return zone->pageblock_flags;
5968 #endif /* CONFIG_SPARSEMEM */
5971 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5973 #ifdef CONFIG_SPARSEMEM
5974 pfn &= (PAGES_PER_SECTION-1);
5975 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5976 #else
5977 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
5978 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5979 #endif /* CONFIG_SPARSEMEM */
5983 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5984 * @page: The page within the block of interest
5985 * @start_bitidx: The first bit of interest to retrieve
5986 * @end_bitidx: The last bit of interest
5987 * returns pageblock_bits flags
5989 unsigned long get_pageblock_flags_group(struct page *page,
5990 int start_bitidx, int end_bitidx)
5992 struct zone *zone;
5993 unsigned long *bitmap;
5994 unsigned long pfn, bitidx;
5995 unsigned long flags = 0;
5996 unsigned long value = 1;
5998 zone = page_zone(page);
5999 pfn = page_to_pfn(page);
6000 bitmap = get_pageblock_bitmap(zone, pfn);
6001 bitidx = pfn_to_bitidx(zone, pfn);
6003 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6004 if (test_bit(bitidx + start_bitidx, bitmap))
6005 flags |= value;
6007 return flags;
6011 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
6012 * @page: The page within the block of interest
6013 * @start_bitidx: The first bit of interest
6014 * @end_bitidx: The last bit of interest
6015 * @flags: The flags to set
6017 void set_pageblock_flags_group(struct page *page, unsigned long flags,
6018 int start_bitidx, int end_bitidx)
6020 struct zone *zone;
6021 unsigned long *bitmap;
6022 unsigned long pfn, bitidx;
6023 unsigned long value = 1;
6025 zone = page_zone(page);
6026 pfn = page_to_pfn(page);
6027 bitmap = get_pageblock_bitmap(zone, pfn);
6028 bitidx = pfn_to_bitidx(zone, pfn);
6029 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6031 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
6032 if (flags & value)
6033 __set_bit(bitidx + start_bitidx, bitmap);
6034 else
6035 __clear_bit(bitidx + start_bitidx, bitmap);
6039 * This function checks whether pageblock includes unmovable pages or not.
6040 * If @count is not zero, it is okay to include less @count unmovable pages
6042 * PageLRU check without isolation or lru_lock could race so that
6043 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6044 * expect this function should be exact.
6046 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6047 bool skip_hwpoisoned_pages)
6049 unsigned long pfn, iter, found;
6050 int mt;
6053 * For avoiding noise data, lru_add_drain_all() should be called
6054 * If ZONE_MOVABLE, the zone never contains unmovable pages
6056 if (zone_idx(zone) == ZONE_MOVABLE)
6057 return false;
6058 mt = get_pageblock_migratetype(page);
6059 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6060 return false;
6062 pfn = page_to_pfn(page);
6063 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6064 unsigned long check = pfn + iter;
6066 if (!pfn_valid_within(check))
6067 continue;
6069 page = pfn_to_page(check);
6072 * Hugepages are not in LRU lists, but they're movable.
6073 * We need not scan over tail pages bacause we don't
6074 * handle each tail page individually in migration.
6076 if (PageHuge(page)) {
6077 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6078 continue;
6082 * We can't use page_count without pin a page
6083 * because another CPU can free compound page.
6084 * This check already skips compound tails of THP
6085 * because their page->_count is zero at all time.
6087 if (!atomic_read(&page->_count)) {
6088 if (PageBuddy(page))
6089 iter += (1 << page_order(page)) - 1;
6090 continue;
6094 * The HWPoisoned page may be not in buddy system, and
6095 * page_count() is not 0.
6097 if (skip_hwpoisoned_pages && PageHWPoison(page))
6098 continue;
6100 if (!PageLRU(page))
6101 found++;
6103 * If there are RECLAIMABLE pages, we need to check it.
6104 * But now, memory offline itself doesn't call shrink_slab()
6105 * and it still to be fixed.
6108 * If the page is not RAM, page_count()should be 0.
6109 * we don't need more check. This is an _used_ not-movable page.
6111 * The problematic thing here is PG_reserved pages. PG_reserved
6112 * is set to both of a memory hole page and a _used_ kernel
6113 * page at boot.
6115 if (found > count)
6116 return true;
6118 return false;
6121 bool is_pageblock_removable_nolock(struct page *page)
6123 struct zone *zone;
6124 unsigned long pfn;
6127 * We have to be careful here because we are iterating over memory
6128 * sections which are not zone aware so we might end up outside of
6129 * the zone but still within the section.
6130 * We have to take care about the node as well. If the node is offline
6131 * its NODE_DATA will be NULL - see page_zone.
6133 if (!node_online(page_to_nid(page)))
6134 return false;
6136 zone = page_zone(page);
6137 pfn = page_to_pfn(page);
6138 if (!zone_spans_pfn(zone, pfn))
6139 return false;
6141 return !has_unmovable_pages(zone, page, 0, true);
6144 #ifdef CONFIG_CMA
6146 static unsigned long pfn_max_align_down(unsigned long pfn)
6148 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6149 pageblock_nr_pages) - 1);
6152 static unsigned long pfn_max_align_up(unsigned long pfn)
6154 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6155 pageblock_nr_pages));
6158 /* [start, end) must belong to a single zone. */
6159 static int __alloc_contig_migrate_range(struct compact_control *cc,
6160 unsigned long start, unsigned long end)
6162 /* This function is based on compact_zone() from compaction.c. */
6163 unsigned long nr_reclaimed;
6164 unsigned long pfn = start;
6165 unsigned int tries = 0;
6166 int ret = 0;
6168 migrate_prep();
6170 while (pfn < end || !list_empty(&cc->migratepages)) {
6171 if (fatal_signal_pending(current)) {
6172 ret = -EINTR;
6173 break;
6176 if (list_empty(&cc->migratepages)) {
6177 cc->nr_migratepages = 0;
6178 pfn = isolate_migratepages_range(cc->zone, cc,
6179 pfn, end, true);
6180 if (!pfn) {
6181 ret = -EINTR;
6182 break;
6184 tries = 0;
6185 } else if (++tries == 5) {
6186 ret = ret < 0 ? ret : -EBUSY;
6187 break;
6190 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6191 &cc->migratepages);
6192 cc->nr_migratepages -= nr_reclaimed;
6194 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6195 0, MIGRATE_SYNC, MR_CMA);
6197 if (ret < 0) {
6198 putback_movable_pages(&cc->migratepages);
6199 return ret;
6201 return 0;
6205 * alloc_contig_range() -- tries to allocate given range of pages
6206 * @start: start PFN to allocate
6207 * @end: one-past-the-last PFN to allocate
6208 * @migratetype: migratetype of the underlaying pageblocks (either
6209 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6210 * in range must have the same migratetype and it must
6211 * be either of the two.
6213 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6214 * aligned, however it's the caller's responsibility to guarantee that
6215 * we are the only thread that changes migrate type of pageblocks the
6216 * pages fall in.
6218 * The PFN range must belong to a single zone.
6220 * Returns zero on success or negative error code. On success all
6221 * pages which PFN is in [start, end) are allocated for the caller and
6222 * need to be freed with free_contig_range().
6224 int alloc_contig_range(unsigned long start, unsigned long end,
6225 unsigned migratetype)
6227 unsigned long outer_start, outer_end;
6228 int ret = 0, order;
6230 struct compact_control cc = {
6231 .nr_migratepages = 0,
6232 .order = -1,
6233 .zone = page_zone(pfn_to_page(start)),
6234 .sync = true,
6235 .ignore_skip_hint = true,
6237 INIT_LIST_HEAD(&cc.migratepages);
6240 * What we do here is we mark all pageblocks in range as
6241 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6242 * have different sizes, and due to the way page allocator
6243 * work, we align the range to biggest of the two pages so
6244 * that page allocator won't try to merge buddies from
6245 * different pageblocks and change MIGRATE_ISOLATE to some
6246 * other migration type.
6248 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6249 * migrate the pages from an unaligned range (ie. pages that
6250 * we are interested in). This will put all the pages in
6251 * range back to page allocator as MIGRATE_ISOLATE.
6253 * When this is done, we take the pages in range from page
6254 * allocator removing them from the buddy system. This way
6255 * page allocator will never consider using them.
6257 * This lets us mark the pageblocks back as
6258 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6259 * aligned range but not in the unaligned, original range are
6260 * put back to page allocator so that buddy can use them.
6263 ret = start_isolate_page_range(pfn_max_align_down(start),
6264 pfn_max_align_up(end), migratetype,
6265 false);
6266 if (ret)
6267 return ret;
6269 ret = __alloc_contig_migrate_range(&cc, start, end);
6270 if (ret)
6271 goto done;
6274 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6275 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6276 * more, all pages in [start, end) are free in page allocator.
6277 * What we are going to do is to allocate all pages from
6278 * [start, end) (that is remove them from page allocator).
6280 * The only problem is that pages at the beginning and at the
6281 * end of interesting range may be not aligned with pages that
6282 * page allocator holds, ie. they can be part of higher order
6283 * pages. Because of this, we reserve the bigger range and
6284 * once this is done free the pages we are not interested in.
6286 * We don't have to hold zone->lock here because the pages are
6287 * isolated thus they won't get removed from buddy.
6290 lru_add_drain_all();
6291 drain_all_pages();
6293 order = 0;
6294 outer_start = start;
6295 while (!PageBuddy(pfn_to_page(outer_start))) {
6296 if (++order >= MAX_ORDER) {
6297 ret = -EBUSY;
6298 goto done;
6300 outer_start &= ~0UL << order;
6303 /* Make sure the range is really isolated. */
6304 if (test_pages_isolated(outer_start, end, false)) {
6305 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6306 outer_start, end);
6307 ret = -EBUSY;
6308 goto done;
6312 /* Grab isolated pages from freelists. */
6313 outer_end = isolate_freepages_range(&cc, outer_start, end);
6314 if (!outer_end) {
6315 ret = -EBUSY;
6316 goto done;
6319 /* Free head and tail (if any) */
6320 if (start != outer_start)
6321 free_contig_range(outer_start, start - outer_start);
6322 if (end != outer_end)
6323 free_contig_range(end, outer_end - end);
6325 done:
6326 undo_isolate_page_range(pfn_max_align_down(start),
6327 pfn_max_align_up(end), migratetype);
6328 return ret;
6331 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6333 unsigned int count = 0;
6335 for (; nr_pages--; pfn++) {
6336 struct page *page = pfn_to_page(pfn);
6338 count += page_count(page) != 1;
6339 __free_page(page);
6341 WARN(count != 0, "%d pages are still in use!\n", count);
6343 #endif
6345 #ifdef CONFIG_MEMORY_HOTPLUG
6347 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6348 * page high values need to be recalulated.
6350 void __meminit zone_pcp_update(struct zone *zone)
6352 unsigned cpu;
6353 mutex_lock(&pcp_batch_high_lock);
6354 for_each_possible_cpu(cpu)
6355 pageset_set_high_and_batch(zone,
6356 per_cpu_ptr(zone->pageset, cpu));
6357 mutex_unlock(&pcp_batch_high_lock);
6359 #endif
6361 void zone_pcp_reset(struct zone *zone)
6363 unsigned long flags;
6364 int cpu;
6365 struct per_cpu_pageset *pset;
6367 /* avoid races with drain_pages() */
6368 local_irq_save(flags);
6369 if (zone->pageset != &boot_pageset) {
6370 for_each_online_cpu(cpu) {
6371 pset = per_cpu_ptr(zone->pageset, cpu);
6372 drain_zonestat(zone, pset);
6374 free_percpu(zone->pageset);
6375 zone->pageset = &boot_pageset;
6377 local_irq_restore(flags);
6380 #ifdef CONFIG_MEMORY_HOTREMOVE
6382 * All pages in the range must be isolated before calling this.
6384 void
6385 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6387 struct page *page;
6388 struct zone *zone;
6389 int order, i;
6390 unsigned long pfn;
6391 unsigned long flags;
6392 /* find the first valid pfn */
6393 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6394 if (pfn_valid(pfn))
6395 break;
6396 if (pfn == end_pfn)
6397 return;
6398 zone = page_zone(pfn_to_page(pfn));
6399 spin_lock_irqsave(&zone->lock, flags);
6400 pfn = start_pfn;
6401 while (pfn < end_pfn) {
6402 if (!pfn_valid(pfn)) {
6403 pfn++;
6404 continue;
6406 page = pfn_to_page(pfn);
6408 * The HWPoisoned page may be not in buddy system, and
6409 * page_count() is not 0.
6411 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6412 pfn++;
6413 SetPageReserved(page);
6414 continue;
6417 BUG_ON(page_count(page));
6418 BUG_ON(!PageBuddy(page));
6419 order = page_order(page);
6420 #ifdef CONFIG_DEBUG_VM
6421 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6422 pfn, 1 << order, end_pfn);
6423 #endif
6424 list_del(&page->lru);
6425 rmv_page_order(page);
6426 zone->free_area[order].nr_free--;
6427 for (i = 0; i < (1 << order); i++)
6428 SetPageReserved((page+i));
6429 pfn += (1 << order);
6431 spin_unlock_irqrestore(&zone->lock, flags);
6433 #endif
6435 #ifdef CONFIG_MEMORY_FAILURE
6436 bool is_free_buddy_page(struct page *page)
6438 struct zone *zone = page_zone(page);
6439 unsigned long pfn = page_to_pfn(page);
6440 unsigned long flags;
6441 int order;
6443 spin_lock_irqsave(&zone->lock, flags);
6444 for (order = 0; order < MAX_ORDER; order++) {
6445 struct page *page_head = page - (pfn & ((1 << order) - 1));
6447 if (PageBuddy(page_head) && page_order(page_head) >= order)
6448 break;
6450 spin_unlock_irqrestore(&zone->lock, flags);
6452 return order < MAX_ORDER;
6454 #endif
6456 static const struct trace_print_flags pageflag_names[] = {
6457 {1UL << PG_locked, "locked" },
6458 {1UL << PG_error, "error" },
6459 {1UL << PG_referenced, "referenced" },
6460 {1UL << PG_uptodate, "uptodate" },
6461 {1UL << PG_dirty, "dirty" },
6462 {1UL << PG_lru, "lru" },
6463 {1UL << PG_active, "active" },
6464 {1UL << PG_slab, "slab" },
6465 {1UL << PG_owner_priv_1, "owner_priv_1" },
6466 {1UL << PG_arch_1, "arch_1" },
6467 {1UL << PG_reserved, "reserved" },
6468 {1UL << PG_private, "private" },
6469 {1UL << PG_private_2, "private_2" },
6470 {1UL << PG_writeback, "writeback" },
6471 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6472 {1UL << PG_head, "head" },
6473 {1UL << PG_tail, "tail" },
6474 #else
6475 {1UL << PG_compound, "compound" },
6476 #endif
6477 {1UL << PG_swapcache, "swapcache" },
6478 {1UL << PG_mappedtodisk, "mappedtodisk" },
6479 {1UL << PG_reclaim, "reclaim" },
6480 {1UL << PG_swapbacked, "swapbacked" },
6481 {1UL << PG_unevictable, "unevictable" },
6482 #ifdef CONFIG_MMU
6483 {1UL << PG_mlocked, "mlocked" },
6484 #endif
6485 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6486 {1UL << PG_uncached, "uncached" },
6487 #endif
6488 #ifdef CONFIG_MEMORY_FAILURE
6489 {1UL << PG_hwpoison, "hwpoison" },
6490 #endif
6491 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6492 {1UL << PG_compound_lock, "compound_lock" },
6493 #endif
6496 static void dump_page_flags(unsigned long flags)
6498 const char *delim = "";
6499 unsigned long mask;
6500 int i;
6502 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names) != __NR_PAGEFLAGS);
6504 printk(KERN_ALERT "page flags: %#lx(", flags);
6506 /* remove zone id */
6507 flags &= (1UL << NR_PAGEFLAGS) - 1;
6509 for (i = 0; i < ARRAY_SIZE(pageflag_names) && flags; i++) {
6511 mask = pageflag_names[i].mask;
6512 if ((flags & mask) != mask)
6513 continue;
6515 flags &= ~mask;
6516 printk("%s%s", delim, pageflag_names[i].name);
6517 delim = "|";
6520 /* check for left over flags */
6521 if (flags)
6522 printk("%s%#lx", delim, flags);
6524 printk(")\n");
6527 void dump_page_badflags(struct page *page, char *reason, unsigned long badflags)
6529 printk(KERN_ALERT
6530 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6531 page, atomic_read(&page->_count), page_mapcount(page),
6532 page->mapping, page->index);
6533 dump_page_flags(page->flags);
6534 if (reason)
6535 pr_alert("page dumped because: %s\n", reason);
6536 if (page->flags & badflags) {
6537 pr_alert("bad because of flags:\n");
6538 dump_page_flags(page->flags & badflags);
6540 mem_cgroup_print_bad_page(page);
6543 void dump_page(struct page *page, char *reason)
6545 dump_page_badflags(page, reason, 0);
6547 EXPORT_SYMBOL_GPL(dump_page);