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>
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/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node);
76 EXPORT_PER_CPU_SYMBOL(numa_node);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_);
88 int _node_numa_mem_[MAX_NUMNODES];
92 * Array of node states.
94 nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
95 [N_POSSIBLE] = NODE_MASK_ALL,
96 [N_ONLINE] = { { [0] = 1UL } },
98 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
100 [N_HIGH_MEMORY] = { { [0] = 1UL } },
102 #ifdef CONFIG_MOVABLE_NODE
103 [N_MEMORY] = { { [0] = 1UL } },
105 [N_CPU] = { { [0] = 1UL } },
108 EXPORT_SYMBOL(node_states);
110 /* Protect totalram_pages and zone->managed_pages */
111 static DEFINE_SPINLOCK(managed_page_count_lock);
113 unsigned long totalram_pages __read_mostly;
114 unsigned long totalreserve_pages __read_mostly;
115 unsigned long totalcma_pages __read_mostly;
117 * When calculating the number of globally allowed dirty pages, there
118 * is a certain number of per-zone reserves that should not be
119 * considered dirtyable memory. This is the sum of those reserves
120 * over all existing zones that contribute dirtyable memory.
122 unsigned long dirty_balance_reserve __read_mostly;
124 int percpu_pagelist_fraction;
125 gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
127 #ifdef CONFIG_PM_SLEEP
129 * The following functions are used by the suspend/hibernate code to temporarily
130 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
131 * while devices are suspended. To avoid races with the suspend/hibernate code,
132 * they should always be called with pm_mutex held (gfp_allowed_mask also should
133 * only be modified with pm_mutex held, unless the suspend/hibernate code is
134 * guaranteed not to run in parallel with that modification).
137 static gfp_t saved_gfp_mask;
139 void pm_restore_gfp_mask(void)
141 WARN_ON(!mutex_is_locked(&pm_mutex));
142 if (saved_gfp_mask) {
143 gfp_allowed_mask = saved_gfp_mask;
148 void pm_restrict_gfp_mask(void)
150 WARN_ON(!mutex_is_locked(&pm_mutex));
151 WARN_ON(saved_gfp_mask);
152 saved_gfp_mask = gfp_allowed_mask;
153 gfp_allowed_mask &= ~GFP_IOFS;
156 bool pm_suspended_storage(void)
158 if ((gfp_allowed_mask & GFP_IOFS) == GFP_IOFS)
162 #endif /* CONFIG_PM_SLEEP */
164 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
165 int pageblock_order __read_mostly;
168 static void __free_pages_ok(struct page *page, unsigned int order);
171 * results with 256, 32 in the lowmem_reserve sysctl:
172 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
173 * 1G machine -> (16M dma, 784M normal, 224M high)
174 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
175 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
176 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
178 * TBD: should special case ZONE_DMA32 machines here - in those we normally
179 * don't need any ZONE_NORMAL reservation
181 int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
182 #ifdef CONFIG_ZONE_DMA
185 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 EXPORT_SYMBOL(totalram_pages);
196 static char * const zone_names[MAX_NR_ZONES] = {
197 #ifdef CONFIG_ZONE_DMA
200 #ifdef CONFIG_ZONE_DMA32
204 #ifdef CONFIG_HIGHMEM
210 int min_free_kbytes = 1024;
211 int user_min_free_kbytes = -1;
213 static unsigned long __meminitdata nr_kernel_pages;
214 static unsigned long __meminitdata nr_all_pages;
215 static unsigned long __meminitdata dma_reserve;
217 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
218 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
219 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
220 static unsigned long __initdata required_kernelcore;
221 static unsigned long __initdata required_movablecore;
222 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
224 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
226 EXPORT_SYMBOL(movable_zone);
227 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
230 int nr_node_ids __read_mostly = MAX_NUMNODES;
231 int nr_online_nodes __read_mostly = 1;
232 EXPORT_SYMBOL(nr_node_ids);
233 EXPORT_SYMBOL(nr_online_nodes);
236 int page_group_by_mobility_disabled __read_mostly;
238 void set_pageblock_migratetype(struct page *page, int migratetype)
240 if (unlikely(page_group_by_mobility_disabled &&
241 migratetype < MIGRATE_PCPTYPES))
242 migratetype = MIGRATE_UNMOVABLE;
244 set_pageblock_flags_group(page, (unsigned long)migratetype,
245 PB_migrate, PB_migrate_end);
248 #ifdef CONFIG_DEBUG_VM
249 static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
253 unsigned long pfn = page_to_pfn(page);
254 unsigned long sp, start_pfn;
257 seq = zone_span_seqbegin(zone);
258 start_pfn = zone->zone_start_pfn;
259 sp = zone->spanned_pages;
260 if (!zone_spans_pfn(zone, pfn))
262 } while (zone_span_seqretry(zone, seq));
265 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
266 pfn, zone_to_nid(zone), zone->name,
267 start_pfn, start_pfn + sp);
272 static int page_is_consistent(struct zone *zone, struct page *page)
274 if (!pfn_valid_within(page_to_pfn(page)))
276 if (zone != page_zone(page))
282 * Temporary debugging check for pages not lying within a given zone.
284 static int bad_range(struct zone *zone, struct page *page)
286 if (page_outside_zone_boundaries(zone, page))
288 if (!page_is_consistent(zone, page))
294 static inline int bad_range(struct zone *zone, struct page *page)
300 static void bad_page(struct page *page, const char *reason,
301 unsigned long bad_flags)
303 static unsigned long resume;
304 static unsigned long nr_shown;
305 static unsigned long nr_unshown;
307 /* Don't complain about poisoned pages */
308 if (PageHWPoison(page)) {
309 page_mapcount_reset(page); /* remove PageBuddy */
314 * Allow a burst of 60 reports, then keep quiet for that minute;
315 * or allow a steady drip of one report per second.
317 if (nr_shown == 60) {
318 if (time_before(jiffies, resume)) {
324 "BUG: Bad page state: %lu messages suppressed\n",
331 resume = jiffies + 60 * HZ;
333 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
334 current->comm, page_to_pfn(page));
335 dump_page_badflags(page, reason, bad_flags);
340 /* Leave bad fields for debug, except PageBuddy could make trouble */
341 page_mapcount_reset(page); /* remove PageBuddy */
342 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
346 * Higher-order pages are called "compound pages". They are structured thusly:
348 * The first PAGE_SIZE page is called the "head page".
350 * The remaining PAGE_SIZE pages are called "tail pages".
352 * All pages have PG_compound set. All tail pages have their ->first_page
353 * pointing at the head page.
355 * The first tail page's ->lru.next holds the address of the compound page's
356 * put_page() function. Its ->lru.prev holds the order of allocation.
357 * This usage means that zero-order pages may not be compound.
360 static void free_compound_page(struct page *page)
362 __free_pages_ok(page, compound_order(page));
365 void prep_compound_page(struct page *page, unsigned long order)
368 int nr_pages = 1 << order;
370 set_compound_page_dtor(page, free_compound_page);
371 set_compound_order(page, order);
373 for (i = 1; i < nr_pages; i++) {
374 struct page *p = page + i;
375 set_page_count(p, 0);
376 p->first_page = page;
377 /* Make sure p->first_page is always valid for PageTail() */
383 static inline void prep_zero_page(struct page *page, unsigned int order,
389 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
390 * and __GFP_HIGHMEM from hard or soft interrupt context.
392 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
393 for (i = 0; i < (1 << order); i++)
394 clear_highpage(page + i);
397 #ifdef CONFIG_DEBUG_PAGEALLOC
398 unsigned int _debug_guardpage_minorder;
399 bool _debug_pagealloc_enabled __read_mostly;
400 bool _debug_guardpage_enabled __read_mostly;
402 static int __init early_debug_pagealloc(char *buf)
407 if (strcmp(buf, "on") == 0)
408 _debug_pagealloc_enabled = true;
412 early_param("debug_pagealloc", early_debug_pagealloc);
414 static bool need_debug_guardpage(void)
416 /* If we don't use debug_pagealloc, we don't need guard page */
417 if (!debug_pagealloc_enabled())
423 static void init_debug_guardpage(void)
425 if (!debug_pagealloc_enabled())
428 _debug_guardpage_enabled = true;
431 struct page_ext_operations debug_guardpage_ops = {
432 .need = need_debug_guardpage,
433 .init = init_debug_guardpage,
436 static int __init debug_guardpage_minorder_setup(char *buf)
440 if (kstrtoul(buf, 10, &res) < 0 || res > MAX_ORDER / 2) {
441 printk(KERN_ERR "Bad debug_guardpage_minorder value\n");
444 _debug_guardpage_minorder = res;
445 printk(KERN_INFO "Setting debug_guardpage_minorder to %lu\n", res);
448 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup);
450 static inline void set_page_guard(struct zone *zone, struct page *page,
451 unsigned int order, int migratetype)
453 struct page_ext *page_ext;
455 if (!debug_guardpage_enabled())
458 page_ext = lookup_page_ext(page);
459 __set_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
461 INIT_LIST_HEAD(&page->lru);
462 set_page_private(page, order);
463 /* Guard pages are not available for any usage */
464 __mod_zone_freepage_state(zone, -(1 << order), migratetype);
467 static inline void clear_page_guard(struct zone *zone, struct page *page,
468 unsigned int order, int migratetype)
470 struct page_ext *page_ext;
472 if (!debug_guardpage_enabled())
475 page_ext = lookup_page_ext(page);
476 __clear_bit(PAGE_EXT_DEBUG_GUARD, &page_ext->flags);
478 set_page_private(page, 0);
479 if (!is_migrate_isolate(migratetype))
480 __mod_zone_freepage_state(zone, (1 << order), migratetype);
483 struct page_ext_operations debug_guardpage_ops = { NULL, };
484 static inline void set_page_guard(struct zone *zone, struct page *page,
485 unsigned int order, int migratetype) {}
486 static inline void clear_page_guard(struct zone *zone, struct page *page,
487 unsigned int order, int migratetype) {}
490 static inline void set_page_order(struct page *page, unsigned int order)
492 set_page_private(page, order);
493 __SetPageBuddy(page);
496 static inline void rmv_page_order(struct page *page)
498 __ClearPageBuddy(page);
499 set_page_private(page, 0);
503 * This function checks whether a page is free && is the buddy
504 * we can do coalesce a page and its buddy if
505 * (a) the buddy is not in a hole &&
506 * (b) the buddy is in the buddy system &&
507 * (c) a page and its buddy have the same order &&
508 * (d) a page and its buddy are in the same zone.
510 * For recording whether a page is in the buddy system, we set ->_mapcount
511 * PAGE_BUDDY_MAPCOUNT_VALUE.
512 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
513 * serialized by zone->lock.
515 * For recording page's order, we use page_private(page).
517 static inline int page_is_buddy(struct page *page, struct page *buddy,
520 if (!pfn_valid_within(page_to_pfn(buddy)))
523 if (page_is_guard(buddy) && page_order(buddy) == order) {
524 if (page_zone_id(page) != page_zone_id(buddy))
527 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
532 if (PageBuddy(buddy) && page_order(buddy) == order) {
534 * zone check is done late to avoid uselessly
535 * calculating zone/node ids for pages that could
538 if (page_zone_id(page) != page_zone_id(buddy))
541 VM_BUG_ON_PAGE(page_count(buddy) != 0, buddy);
549 * Freeing function for a buddy system allocator.
551 * The concept of a buddy system is to maintain direct-mapped table
552 * (containing bit values) for memory blocks of various "orders".
553 * The bottom level table contains the map for the smallest allocatable
554 * units of memory (here, pages), and each level above it describes
555 * pairs of units from the levels below, hence, "buddies".
556 * At a high level, all that happens here is marking the table entry
557 * at the bottom level available, and propagating the changes upward
558 * as necessary, plus some accounting needed to play nicely with other
559 * parts of the VM system.
560 * At each level, we keep a list of pages, which are heads of continuous
561 * free pages of length of (1 << order) and marked with _mapcount
562 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
564 * So when we are allocating or freeing one, we can derive the state of the
565 * other. That is, if we allocate a small block, and both were
566 * free, the remainder of the region must be split into blocks.
567 * If a block is freed, and its buddy is also free, then this
568 * triggers coalescing into a block of larger size.
573 static inline void __free_one_page(struct page *page,
575 struct zone *zone, unsigned int order,
578 unsigned long page_idx;
579 unsigned long combined_idx;
580 unsigned long uninitialized_var(buddy_idx);
582 int max_order = MAX_ORDER;
584 VM_BUG_ON(!zone_is_initialized(zone));
585 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
587 VM_BUG_ON(migratetype == -1);
588 if (is_migrate_isolate(migratetype)) {
590 * We restrict max order of merging to prevent merge
591 * between freepages on isolate pageblock and normal
592 * pageblock. Without this, pageblock isolation
593 * could cause incorrect freepage accounting.
595 max_order = min(MAX_ORDER, pageblock_order + 1);
597 __mod_zone_freepage_state(zone, 1 << order, migratetype);
600 page_idx = pfn & ((1 << max_order) - 1);
602 VM_BUG_ON_PAGE(page_idx & ((1 << order) - 1), page);
603 VM_BUG_ON_PAGE(bad_range(zone, page), page);
605 while (order < max_order - 1) {
606 buddy_idx = __find_buddy_index(page_idx, order);
607 buddy = page + (buddy_idx - page_idx);
608 if (!page_is_buddy(page, buddy, order))
611 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
612 * merge with it and move up one order.
614 if (page_is_guard(buddy)) {
615 clear_page_guard(zone, buddy, order, migratetype);
617 list_del(&buddy->lru);
618 zone->free_area[order].nr_free--;
619 rmv_page_order(buddy);
621 combined_idx = buddy_idx & page_idx;
622 page = page + (combined_idx - page_idx);
623 page_idx = combined_idx;
626 set_page_order(page, order);
629 * If this is not the largest possible page, check if the buddy
630 * of the next-highest order is free. If it is, it's possible
631 * that pages are being freed that will coalesce soon. In case,
632 * that is happening, add the free page to the tail of the list
633 * so it's less likely to be used soon and more likely to be merged
634 * as a higher order page
636 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
637 struct page *higher_page, *higher_buddy;
638 combined_idx = buddy_idx & page_idx;
639 higher_page = page + (combined_idx - page_idx);
640 buddy_idx = __find_buddy_index(combined_idx, order + 1);
641 higher_buddy = higher_page + (buddy_idx - combined_idx);
642 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
643 list_add_tail(&page->lru,
644 &zone->free_area[order].free_list[migratetype]);
649 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
651 zone->free_area[order].nr_free++;
654 static inline int free_pages_check(struct page *page)
656 const char *bad_reason = NULL;
657 unsigned long bad_flags = 0;
659 if (unlikely(page_mapcount(page)))
660 bad_reason = "nonzero mapcount";
661 if (unlikely(page->mapping != NULL))
662 bad_reason = "non-NULL mapping";
663 if (unlikely(atomic_read(&page->_count) != 0))
664 bad_reason = "nonzero _count";
665 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_FREE)) {
666 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
667 bad_flags = PAGE_FLAGS_CHECK_AT_FREE;
670 if (unlikely(page->mem_cgroup))
671 bad_reason = "page still charged to cgroup";
673 if (unlikely(bad_reason)) {
674 bad_page(page, bad_reason, bad_flags);
677 page_cpupid_reset_last(page);
678 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
679 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
684 * Frees a number of pages from the PCP lists
685 * Assumes all pages on list are in same zone, and of same order.
686 * count is the number of pages to free.
688 * If the zone was previously in an "all pages pinned" state then look to
689 * see if this freeing clears that state.
691 * And clear the zone's pages_scanned counter, to hold off the "all pages are
692 * pinned" detection logic.
694 static void free_pcppages_bulk(struct zone *zone, int count,
695 struct per_cpu_pages *pcp)
700 unsigned long nr_scanned;
702 spin_lock(&zone->lock);
703 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
705 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
709 struct list_head *list;
712 * Remove pages from lists in a round-robin fashion. A
713 * batch_free count is maintained that is incremented when an
714 * empty list is encountered. This is so more pages are freed
715 * off fuller lists instead of spinning excessively around empty
720 if (++migratetype == MIGRATE_PCPTYPES)
722 list = &pcp->lists[migratetype];
723 } while (list_empty(list));
725 /* This is the only non-empty list. Free them all. */
726 if (batch_free == MIGRATE_PCPTYPES)
727 batch_free = to_free;
730 int mt; /* migratetype of the to-be-freed page */
732 page = list_entry(list->prev, struct page, lru);
733 /* must delete as __free_one_page list manipulates */
734 list_del(&page->lru);
735 mt = get_freepage_migratetype(page);
736 if (unlikely(has_isolate_pageblock(zone)))
737 mt = get_pageblock_migratetype(page);
739 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
740 __free_one_page(page, page_to_pfn(page), zone, 0, mt);
741 trace_mm_page_pcpu_drain(page, 0, mt);
742 } while (--to_free && --batch_free && !list_empty(list));
744 spin_unlock(&zone->lock);
747 static void free_one_page(struct zone *zone,
748 struct page *page, unsigned long pfn,
752 unsigned long nr_scanned;
753 spin_lock(&zone->lock);
754 nr_scanned = zone_page_state(zone, NR_PAGES_SCANNED);
756 __mod_zone_page_state(zone, NR_PAGES_SCANNED, -nr_scanned);
758 if (unlikely(has_isolate_pageblock(zone) ||
759 is_migrate_isolate(migratetype))) {
760 migratetype = get_pfnblock_migratetype(page, pfn);
762 __free_one_page(page, pfn, zone, order, migratetype);
763 spin_unlock(&zone->lock);
766 static int free_tail_pages_check(struct page *head_page, struct page *page)
768 if (!IS_ENABLED(CONFIG_DEBUG_VM))
770 if (unlikely(!PageTail(page))) {
771 bad_page(page, "PageTail not set", 0);
774 if (unlikely(page->first_page != head_page)) {
775 bad_page(page, "first_page not consistent", 0);
781 static bool free_pages_prepare(struct page *page, unsigned int order)
783 bool compound = PageCompound(page);
786 VM_BUG_ON_PAGE(PageTail(page), page);
787 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
789 trace_mm_page_free(page, order);
790 kmemcheck_free_shadow(page, order);
791 kasan_free_pages(page, order);
794 page->mapping = NULL;
795 bad += free_pages_check(page);
796 for (i = 1; i < (1 << order); i++) {
798 bad += free_tail_pages_check(page, page + i);
799 bad += free_pages_check(page + i);
804 reset_page_owner(page, order);
806 if (!PageHighMem(page)) {
807 debug_check_no_locks_freed(page_address(page),
809 debug_check_no_obj_freed(page_address(page),
812 arch_free_page(page, order);
813 kernel_map_pages(page, 1 << order, 0);
818 static void __free_pages_ok(struct page *page, unsigned int order)
822 unsigned long pfn = page_to_pfn(page);
824 if (!free_pages_prepare(page, order))
827 migratetype = get_pfnblock_migratetype(page, pfn);
828 local_irq_save(flags);
829 __count_vm_events(PGFREE, 1 << order);
830 set_freepage_migratetype(page, migratetype);
831 free_one_page(page_zone(page), page, pfn, order, migratetype);
832 local_irq_restore(flags);
835 void __init __free_pages_bootmem(struct page *page, unsigned int order)
837 unsigned int nr_pages = 1 << order;
838 struct page *p = page;
842 for (loop = 0; loop < (nr_pages - 1); loop++, p++) {
844 __ClearPageReserved(p);
845 set_page_count(p, 0);
847 __ClearPageReserved(p);
848 set_page_count(p, 0);
850 page_zone(page)->managed_pages += nr_pages;
851 set_page_refcounted(page);
852 __free_pages(page, order);
856 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
857 void __init init_cma_reserved_pageblock(struct page *page)
859 unsigned i = pageblock_nr_pages;
860 struct page *p = page;
863 __ClearPageReserved(p);
864 set_page_count(p, 0);
867 set_pageblock_migratetype(page, MIGRATE_CMA);
869 if (pageblock_order >= MAX_ORDER) {
870 i = pageblock_nr_pages;
873 set_page_refcounted(p);
874 __free_pages(p, MAX_ORDER - 1);
875 p += MAX_ORDER_NR_PAGES;
876 } while (i -= MAX_ORDER_NR_PAGES);
878 set_page_refcounted(page);
879 __free_pages(page, pageblock_order);
882 adjust_managed_page_count(page, pageblock_nr_pages);
887 * The order of subdivision here is critical for the IO subsystem.
888 * Please do not alter this order without good reasons and regression
889 * testing. Specifically, as large blocks of memory are subdivided,
890 * the order in which smaller blocks are delivered depends on the order
891 * they're subdivided in this function. This is the primary factor
892 * influencing the order in which pages are delivered to the IO
893 * subsystem according to empirical testing, and this is also justified
894 * by considering the behavior of a buddy system containing a single
895 * large block of memory acted on by a series of small allocations.
896 * This behavior is a critical factor in sglist merging's success.
900 static inline void expand(struct zone *zone, struct page *page,
901 int low, int high, struct free_area *area,
904 unsigned long size = 1 << high;
910 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
912 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC) &&
913 debug_guardpage_enabled() &&
914 high < debug_guardpage_minorder()) {
916 * Mark as guard pages (or page), that will allow to
917 * merge back to allocator when buddy will be freed.
918 * Corresponding page table entries will not be touched,
919 * pages will stay not present in virtual address space
921 set_page_guard(zone, &page[size], high, migratetype);
924 list_add(&page[size].lru, &area->free_list[migratetype]);
926 set_page_order(&page[size], high);
931 * This page is about to be returned from the page allocator
933 static inline int check_new_page(struct page *page)
935 const char *bad_reason = NULL;
936 unsigned long bad_flags = 0;
938 if (unlikely(page_mapcount(page)))
939 bad_reason = "nonzero mapcount";
940 if (unlikely(page->mapping != NULL))
941 bad_reason = "non-NULL mapping";
942 if (unlikely(atomic_read(&page->_count) != 0))
943 bad_reason = "nonzero _count";
944 if (unlikely(page->flags & PAGE_FLAGS_CHECK_AT_PREP)) {
945 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag set";
946 bad_flags = PAGE_FLAGS_CHECK_AT_PREP;
949 if (unlikely(page->mem_cgroup))
950 bad_reason = "page still charged to cgroup";
952 if (unlikely(bad_reason)) {
953 bad_page(page, bad_reason, bad_flags);
959 static int prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
964 for (i = 0; i < (1 << order); i++) {
965 struct page *p = page + i;
966 if (unlikely(check_new_page(p)))
970 set_page_private(page, 0);
971 set_page_refcounted(page);
973 arch_alloc_page(page, order);
974 kernel_map_pages(page, 1 << order, 1);
975 kasan_alloc_pages(page, order);
977 if (gfp_flags & __GFP_ZERO)
978 prep_zero_page(page, order, gfp_flags);
980 if (order && (gfp_flags & __GFP_COMP))
981 prep_compound_page(page, order);
983 set_page_owner(page, order, gfp_flags);
986 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
987 * allocate the page. The expectation is that the caller is taking
988 * steps that will free more memory. The caller should avoid the page
989 * being used for !PFMEMALLOC purposes.
991 page->pfmemalloc = !!(alloc_flags & ALLOC_NO_WATERMARKS);
997 * Go through the free lists for the given migratetype and remove
998 * the smallest available page from the freelists
1001 struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1004 unsigned int current_order;
1005 struct free_area *area;
1008 /* Find a page of the appropriate size in the preferred list */
1009 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
1010 area = &(zone->free_area[current_order]);
1011 if (list_empty(&area->free_list[migratetype]))
1014 page = list_entry(area->free_list[migratetype].next,
1016 list_del(&page->lru);
1017 rmv_page_order(page);
1019 expand(zone, page, order, current_order, area, migratetype);
1020 set_freepage_migratetype(page, migratetype);
1029 * This array describes the order lists are fallen back to when
1030 * the free lists for the desirable migrate type are depleted
1032 static int fallbacks[MIGRATE_TYPES][4] = {
1033 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1034 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
1035 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
1037 [MIGRATE_CMA] = { MIGRATE_RESERVE }, /* Never used */
1039 [MIGRATE_RESERVE] = { MIGRATE_RESERVE }, /* Never used */
1040 #ifdef CONFIG_MEMORY_ISOLATION
1041 [MIGRATE_ISOLATE] = { MIGRATE_RESERVE }, /* Never used */
1046 static struct page *__rmqueue_cma_fallback(struct zone *zone,
1049 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1052 static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1053 unsigned int order) { return NULL; }
1057 * Move the free pages in a range to the free lists of the requested type.
1058 * Note that start_page and end_pages are not aligned on a pageblock
1059 * boundary. If alignment is required, use move_freepages_block()
1061 int move_freepages(struct zone *zone,
1062 struct page *start_page, struct page *end_page,
1066 unsigned long order;
1067 int pages_moved = 0;
1069 #ifndef CONFIG_HOLES_IN_ZONE
1071 * page_zone is not safe to call in this context when
1072 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1073 * anyway as we check zone boundaries in move_freepages_block().
1074 * Remove at a later date when no bug reports exist related to
1075 * grouping pages by mobility
1077 VM_BUG_ON(page_zone(start_page) != page_zone(end_page));
1080 for (page = start_page; page <= end_page;) {
1081 /* Make sure we are not inadvertently changing nodes */
1082 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1084 if (!pfn_valid_within(page_to_pfn(page))) {
1089 if (!PageBuddy(page)) {
1094 order = page_order(page);
1095 list_move(&page->lru,
1096 &zone->free_area[order].free_list[migratetype]);
1097 set_freepage_migratetype(page, migratetype);
1099 pages_moved += 1 << order;
1105 int move_freepages_block(struct zone *zone, struct page *page,
1108 unsigned long start_pfn, end_pfn;
1109 struct page *start_page, *end_page;
1111 start_pfn = page_to_pfn(page);
1112 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
1113 start_page = pfn_to_page(start_pfn);
1114 end_page = start_page + pageblock_nr_pages - 1;
1115 end_pfn = start_pfn + pageblock_nr_pages - 1;
1117 /* Do not cross zone boundaries */
1118 if (!zone_spans_pfn(zone, start_pfn))
1120 if (!zone_spans_pfn(zone, end_pfn))
1123 return move_freepages(zone, start_page, end_page, migratetype);
1126 static void change_pageblock_range(struct page *pageblock_page,
1127 int start_order, int migratetype)
1129 int nr_pageblocks = 1 << (start_order - pageblock_order);
1131 while (nr_pageblocks--) {
1132 set_pageblock_migratetype(pageblock_page, migratetype);
1133 pageblock_page += pageblock_nr_pages;
1138 * When we are falling back to another migratetype during allocation, try to
1139 * steal extra free pages from the same pageblocks to satisfy further
1140 * allocations, instead of polluting multiple pageblocks.
1142 * If we are stealing a relatively large buddy page, it is likely there will
1143 * be more free pages in the pageblock, so try to steal them all. For
1144 * reclaimable and unmovable allocations, we steal regardless of page size,
1145 * as fragmentation caused by those allocations polluting movable pageblocks
1146 * is worse than movable allocations stealing from unmovable and reclaimable
1149 static bool can_steal_fallback(unsigned int order, int start_mt)
1152 * Leaving this order check is intended, although there is
1153 * relaxed order check in next check. The reason is that
1154 * we can actually steal whole pageblock if this condition met,
1155 * but, below check doesn't guarantee it and that is just heuristic
1156 * so could be changed anytime.
1158 if (order >= pageblock_order)
1161 if (order >= pageblock_order / 2 ||
1162 start_mt == MIGRATE_RECLAIMABLE ||
1163 start_mt == MIGRATE_UNMOVABLE ||
1164 page_group_by_mobility_disabled)
1171 * This function implements actual steal behaviour. If order is large enough,
1172 * we can steal whole pageblock. If not, we first move freepages in this
1173 * pageblock and check whether half of pages are moved or not. If half of
1174 * pages are moved, we can change migratetype of pageblock and permanently
1175 * use it's pages as requested migratetype in the future.
1177 static void steal_suitable_fallback(struct zone *zone, struct page *page,
1180 int current_order = page_order(page);
1183 /* Take ownership for orders >= pageblock_order */
1184 if (current_order >= pageblock_order) {
1185 change_pageblock_range(page, current_order, start_type);
1189 pages = move_freepages_block(zone, page, start_type);
1191 /* Claim the whole block if over half of it is free */
1192 if (pages >= (1 << (pageblock_order-1)) ||
1193 page_group_by_mobility_disabled)
1194 set_pageblock_migratetype(page, start_type);
1198 * Check whether there is a suitable fallback freepage with requested order.
1199 * If only_stealable is true, this function returns fallback_mt only if
1200 * we can steal other freepages all together. This would help to reduce
1201 * fragmentation due to mixed migratetype pages in one pageblock.
1203 int find_suitable_fallback(struct free_area *area, unsigned int order,
1204 int migratetype, bool only_stealable, bool *can_steal)
1209 if (area->nr_free == 0)
1214 fallback_mt = fallbacks[migratetype][i];
1215 if (fallback_mt == MIGRATE_RESERVE)
1218 if (list_empty(&area->free_list[fallback_mt]))
1221 if (can_steal_fallback(order, migratetype))
1224 if (!only_stealable)
1234 /* Remove an element from the buddy allocator from the fallback list */
1235 static inline struct page *
1236 __rmqueue_fallback(struct zone *zone, unsigned int order, int start_migratetype)
1238 struct free_area *area;
1239 unsigned int current_order;
1244 /* Find the largest possible block of pages in the other list */
1245 for (current_order = MAX_ORDER-1;
1246 current_order >= order && current_order <= MAX_ORDER-1;
1248 area = &(zone->free_area[current_order]);
1249 fallback_mt = find_suitable_fallback(area, current_order,
1250 start_migratetype, false, &can_steal);
1251 if (fallback_mt == -1)
1254 page = list_entry(area->free_list[fallback_mt].next,
1257 steal_suitable_fallback(zone, page, start_migratetype);
1259 /* Remove the page from the freelists */
1261 list_del(&page->lru);
1262 rmv_page_order(page);
1264 expand(zone, page, order, current_order, area,
1267 * The freepage_migratetype may differ from pageblock's
1268 * migratetype depending on the decisions in
1269 * try_to_steal_freepages(). This is OK as long as it
1270 * does not differ for MIGRATE_CMA pageblocks. For CMA
1271 * we need to make sure unallocated pages flushed from
1272 * pcp lists are returned to the correct freelist.
1274 set_freepage_migratetype(page, start_migratetype);
1276 trace_mm_page_alloc_extfrag(page, order, current_order,
1277 start_migratetype, fallback_mt);
1286 * Do the hard work of removing an element from the buddy allocator.
1287 * Call me with the zone->lock already held.
1289 static struct page *__rmqueue(struct zone *zone, unsigned int order,
1295 page = __rmqueue_smallest(zone, order, migratetype);
1297 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1298 if (migratetype == MIGRATE_MOVABLE)
1299 page = __rmqueue_cma_fallback(zone, order);
1302 page = __rmqueue_fallback(zone, order, migratetype);
1305 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1306 * is used because __rmqueue_smallest is an inline function
1307 * and we want just one call site
1310 migratetype = MIGRATE_RESERVE;
1315 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1320 * Obtain a specified number of elements from the buddy allocator, all under
1321 * a single hold of the lock, for efficiency. Add them to the supplied list.
1322 * Returns the number of new pages which were placed at *list.
1324 static int rmqueue_bulk(struct zone *zone, unsigned int order,
1325 unsigned long count, struct list_head *list,
1326 int migratetype, bool cold)
1330 spin_lock(&zone->lock);
1331 for (i = 0; i < count; ++i) {
1332 struct page *page = __rmqueue(zone, order, migratetype);
1333 if (unlikely(page == NULL))
1337 * Split buddy pages returned by expand() are received here
1338 * in physical page order. The page is added to the callers and
1339 * list and the list head then moves forward. From the callers
1340 * perspective, the linked list is ordered by page number in
1341 * some conditions. This is useful for IO devices that can
1342 * merge IO requests if the physical pages are ordered
1346 list_add(&page->lru, list);
1348 list_add_tail(&page->lru, list);
1350 if (is_migrate_cma(get_freepage_migratetype(page)))
1351 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES,
1354 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1355 spin_unlock(&zone->lock);
1361 * Called from the vmstat counter updater to drain pagesets of this
1362 * currently executing processor on remote nodes after they have
1365 * Note that this function must be called with the thread pinned to
1366 * a single processor.
1368 void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1370 unsigned long flags;
1371 int to_drain, batch;
1373 local_irq_save(flags);
1374 batch = READ_ONCE(pcp->batch);
1375 to_drain = min(pcp->count, batch);
1377 free_pcppages_bulk(zone, to_drain, pcp);
1378 pcp->count -= to_drain;
1380 local_irq_restore(flags);
1385 * Drain pcplists of the indicated processor and zone.
1387 * The processor must either be the current processor and the
1388 * thread pinned to the current processor or a processor that
1391 static void drain_pages_zone(unsigned int cpu, struct zone *zone)
1393 unsigned long flags;
1394 struct per_cpu_pageset *pset;
1395 struct per_cpu_pages *pcp;
1397 local_irq_save(flags);
1398 pset = per_cpu_ptr(zone->pageset, cpu);
1402 free_pcppages_bulk(zone, pcp->count, pcp);
1405 local_irq_restore(flags);
1409 * Drain pcplists of all zones on the indicated processor.
1411 * The processor must either be the current processor and the
1412 * thread pinned to the current processor or a processor that
1415 static void drain_pages(unsigned int cpu)
1419 for_each_populated_zone(zone) {
1420 drain_pages_zone(cpu, zone);
1425 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1427 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1428 * the single zone's pages.
1430 void drain_local_pages(struct zone *zone)
1432 int cpu = smp_processor_id();
1435 drain_pages_zone(cpu, zone);
1441 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1443 * When zone parameter is non-NULL, spill just the single zone's pages.
1445 * Note that this code is protected against sending an IPI to an offline
1446 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1447 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1448 * nothing keeps CPUs from showing up after we populated the cpumask and
1449 * before the call to on_each_cpu_mask().
1451 void drain_all_pages(struct zone *zone)
1456 * Allocate in the BSS so we wont require allocation in
1457 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1459 static cpumask_t cpus_with_pcps;
1462 * We don't care about racing with CPU hotplug event
1463 * as offline notification will cause the notified
1464 * cpu to drain that CPU pcps and on_each_cpu_mask
1465 * disables preemption as part of its processing
1467 for_each_online_cpu(cpu) {
1468 struct per_cpu_pageset *pcp;
1470 bool has_pcps = false;
1473 pcp = per_cpu_ptr(zone->pageset, cpu);
1477 for_each_populated_zone(z) {
1478 pcp = per_cpu_ptr(z->pageset, cpu);
1479 if (pcp->pcp.count) {
1487 cpumask_set_cpu(cpu, &cpus_with_pcps);
1489 cpumask_clear_cpu(cpu, &cpus_with_pcps);
1491 on_each_cpu_mask(&cpus_with_pcps, (smp_call_func_t) drain_local_pages,
1495 #ifdef CONFIG_HIBERNATION
1497 void mark_free_pages(struct zone *zone)
1499 unsigned long pfn, max_zone_pfn;
1500 unsigned long flags;
1501 unsigned int order, t;
1502 struct list_head *curr;
1504 if (zone_is_empty(zone))
1507 spin_lock_irqsave(&zone->lock, flags);
1509 max_zone_pfn = zone_end_pfn(zone);
1510 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1511 if (pfn_valid(pfn)) {
1512 struct page *page = pfn_to_page(pfn);
1514 if (!swsusp_page_is_forbidden(page))
1515 swsusp_unset_page_free(page);
1518 for_each_migratetype_order(order, t) {
1519 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1522 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1523 for (i = 0; i < (1UL << order); i++)
1524 swsusp_set_page_free(pfn_to_page(pfn + i));
1527 spin_unlock_irqrestore(&zone->lock, flags);
1529 #endif /* CONFIG_PM */
1532 * Free a 0-order page
1533 * cold == true ? free a cold page : free a hot page
1535 void free_hot_cold_page(struct page *page, bool cold)
1537 struct zone *zone = page_zone(page);
1538 struct per_cpu_pages *pcp;
1539 unsigned long flags;
1540 unsigned long pfn = page_to_pfn(page);
1543 if (!free_pages_prepare(page, 0))
1546 migratetype = get_pfnblock_migratetype(page, pfn);
1547 set_freepage_migratetype(page, migratetype);
1548 local_irq_save(flags);
1549 __count_vm_event(PGFREE);
1552 * We only track unmovable, reclaimable and movable on pcp lists.
1553 * Free ISOLATE pages back to the allocator because they are being
1554 * offlined but treat RESERVE as movable pages so we can get those
1555 * areas back if necessary. Otherwise, we may have to free
1556 * excessively into the page allocator
1558 if (migratetype >= MIGRATE_PCPTYPES) {
1559 if (unlikely(is_migrate_isolate(migratetype))) {
1560 free_one_page(zone, page, pfn, 0, migratetype);
1563 migratetype = MIGRATE_MOVABLE;
1566 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1568 list_add(&page->lru, &pcp->lists[migratetype]);
1570 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1572 if (pcp->count >= pcp->high) {
1573 unsigned long batch = READ_ONCE(pcp->batch);
1574 free_pcppages_bulk(zone, batch, pcp);
1575 pcp->count -= batch;
1579 local_irq_restore(flags);
1583 * Free a list of 0-order pages
1585 void free_hot_cold_page_list(struct list_head *list, bool cold)
1587 struct page *page, *next;
1589 list_for_each_entry_safe(page, next, list, lru) {
1590 trace_mm_page_free_batched(page, cold);
1591 free_hot_cold_page(page, cold);
1596 * split_page takes a non-compound higher-order page, and splits it into
1597 * n (1<<order) sub-pages: page[0..n]
1598 * Each sub-page must be freed individually.
1600 * Note: this is probably too low level an operation for use in drivers.
1601 * Please consult with lkml before using this in your driver.
1603 void split_page(struct page *page, unsigned int order)
1607 VM_BUG_ON_PAGE(PageCompound(page), page);
1608 VM_BUG_ON_PAGE(!page_count(page), page);
1610 #ifdef CONFIG_KMEMCHECK
1612 * Split shadow pages too, because free(page[0]) would
1613 * otherwise free the whole shadow.
1615 if (kmemcheck_page_is_tracked(page))
1616 split_page(virt_to_page(page[0].shadow), order);
1619 set_page_owner(page, 0, 0);
1620 for (i = 1; i < (1 << order); i++) {
1621 set_page_refcounted(page + i);
1622 set_page_owner(page + i, 0, 0);
1625 EXPORT_SYMBOL_GPL(split_page);
1627 int __isolate_free_page(struct page *page, unsigned int order)
1629 unsigned long watermark;
1633 BUG_ON(!PageBuddy(page));
1635 zone = page_zone(page);
1636 mt = get_pageblock_migratetype(page);
1638 if (!is_migrate_isolate(mt)) {
1639 /* Obey watermarks as if the page was being allocated */
1640 watermark = low_wmark_pages(zone) + (1 << order);
1641 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1644 __mod_zone_freepage_state(zone, -(1UL << order), mt);
1647 /* Remove page from free list */
1648 list_del(&page->lru);
1649 zone->free_area[order].nr_free--;
1650 rmv_page_order(page);
1652 /* Set the pageblock if the isolated page is at least a pageblock */
1653 if (order >= pageblock_order - 1) {
1654 struct page *endpage = page + (1 << order) - 1;
1655 for (; page < endpage; page += pageblock_nr_pages) {
1656 int mt = get_pageblock_migratetype(page);
1657 if (!is_migrate_isolate(mt) && !is_migrate_cma(mt))
1658 set_pageblock_migratetype(page,
1663 set_page_owner(page, order, 0);
1664 return 1UL << order;
1668 * Similar to split_page except the page is already free. As this is only
1669 * being used for migration, the migratetype of the block also changes.
1670 * As this is called with interrupts disabled, the caller is responsible
1671 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1674 * Note: this is probably too low level an operation for use in drivers.
1675 * Please consult with lkml before using this in your driver.
1677 int split_free_page(struct page *page)
1682 order = page_order(page);
1684 nr_pages = __isolate_free_page(page, order);
1688 /* Split into individual pages */
1689 set_page_refcounted(page);
1690 split_page(page, order);
1695 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1698 struct page *buffered_rmqueue(struct zone *preferred_zone,
1699 struct zone *zone, unsigned int order,
1700 gfp_t gfp_flags, int migratetype)
1702 unsigned long flags;
1704 bool cold = ((gfp_flags & __GFP_COLD) != 0);
1706 if (likely(order == 0)) {
1707 struct per_cpu_pages *pcp;
1708 struct list_head *list;
1710 local_irq_save(flags);
1711 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1712 list = &pcp->lists[migratetype];
1713 if (list_empty(list)) {
1714 pcp->count += rmqueue_bulk(zone, 0,
1717 if (unlikely(list_empty(list)))
1722 page = list_entry(list->prev, struct page, lru);
1724 page = list_entry(list->next, struct page, lru);
1726 list_del(&page->lru);
1729 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1731 * __GFP_NOFAIL is not to be used in new code.
1733 * All __GFP_NOFAIL callers should be fixed so that they
1734 * properly detect and handle allocation failures.
1736 * We most definitely don't want callers attempting to
1737 * allocate greater than order-1 page units with
1740 WARN_ON_ONCE(order > 1);
1742 spin_lock_irqsave(&zone->lock, flags);
1743 page = __rmqueue(zone, order, migratetype);
1744 spin_unlock(&zone->lock);
1747 __mod_zone_freepage_state(zone, -(1 << order),
1748 get_freepage_migratetype(page));
1751 __mod_zone_page_state(zone, NR_ALLOC_BATCH, -(1 << order));
1752 if (atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]) <= 0 &&
1753 !test_bit(ZONE_FAIR_DEPLETED, &zone->flags))
1754 set_bit(ZONE_FAIR_DEPLETED, &zone->flags);
1756 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1757 zone_statistics(preferred_zone, zone, gfp_flags);
1758 local_irq_restore(flags);
1760 VM_BUG_ON_PAGE(bad_range(zone, page), page);
1764 local_irq_restore(flags);
1768 #ifdef CONFIG_FAIL_PAGE_ALLOC
1771 struct fault_attr attr;
1773 u32 ignore_gfp_highmem;
1774 u32 ignore_gfp_wait;
1776 } fail_page_alloc = {
1777 .attr = FAULT_ATTR_INITIALIZER,
1778 .ignore_gfp_wait = 1,
1779 .ignore_gfp_highmem = 1,
1783 static int __init setup_fail_page_alloc(char *str)
1785 return setup_fault_attr(&fail_page_alloc.attr, str);
1787 __setup("fail_page_alloc=", setup_fail_page_alloc);
1789 static bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1791 if (order < fail_page_alloc.min_order)
1793 if (gfp_mask & __GFP_NOFAIL)
1795 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1797 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1800 return should_fail(&fail_page_alloc.attr, 1 << order);
1803 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1805 static int __init fail_page_alloc_debugfs(void)
1807 umode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1810 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1811 &fail_page_alloc.attr);
1813 return PTR_ERR(dir);
1815 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1816 &fail_page_alloc.ignore_gfp_wait))
1818 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1819 &fail_page_alloc.ignore_gfp_highmem))
1821 if (!debugfs_create_u32("min-order", mode, dir,
1822 &fail_page_alloc.min_order))
1827 debugfs_remove_recursive(dir);
1832 late_initcall(fail_page_alloc_debugfs);
1834 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1836 #else /* CONFIG_FAIL_PAGE_ALLOC */
1838 static inline bool should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1843 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1846 * Return true if free pages are above 'mark'. This takes into account the order
1847 * of the allocation.
1849 static bool __zone_watermark_ok(struct zone *z, unsigned int order,
1850 unsigned long mark, int classzone_idx, int alloc_flags,
1853 /* free_pages may go negative - that's OK */
1858 free_pages -= (1 << order) - 1;
1859 if (alloc_flags & ALLOC_HIGH)
1861 if (alloc_flags & ALLOC_HARDER)
1864 /* If allocation can't use CMA areas don't use free CMA pages */
1865 if (!(alloc_flags & ALLOC_CMA))
1866 free_cma = zone_page_state(z, NR_FREE_CMA_PAGES);
1869 if (free_pages - free_cma <= min + z->lowmem_reserve[classzone_idx])
1871 for (o = 0; o < order; o++) {
1872 /* At the next order, this order's pages become unavailable */
1873 free_pages -= z->free_area[o].nr_free << o;
1875 /* Require fewer higher order pages to be free */
1878 if (free_pages <= min)
1884 bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
1885 int classzone_idx, int alloc_flags)
1887 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1888 zone_page_state(z, NR_FREE_PAGES));
1891 bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
1892 unsigned long mark, int classzone_idx, int alloc_flags)
1894 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1896 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1897 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1899 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1905 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1906 * skip over zones that are not allowed by the cpuset, or that have
1907 * been recently (in last second) found to be nearly full. See further
1908 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1909 * that have to skip over a lot of full or unallowed zones.
1911 * If the zonelist cache is present in the passed zonelist, then
1912 * returns a pointer to the allowed node mask (either the current
1913 * tasks mems_allowed, or node_states[N_MEMORY].)
1915 * If the zonelist cache is not available for this zonelist, does
1916 * nothing and returns NULL.
1918 * If the fullzones BITMAP in the zonelist cache is stale (more than
1919 * a second since last zap'd) then we zap it out (clear its bits.)
1921 * We hold off even calling zlc_setup, until after we've checked the
1922 * first zone in the zonelist, on the theory that most allocations will
1923 * be satisfied from that first zone, so best to examine that zone as
1924 * quickly as we can.
1926 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1928 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1929 nodemask_t *allowednodes; /* zonelist_cache approximation */
1931 zlc = zonelist->zlcache_ptr;
1935 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1936 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1937 zlc->last_full_zap = jiffies;
1940 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1941 &cpuset_current_mems_allowed :
1942 &node_states[N_MEMORY];
1943 return allowednodes;
1947 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1948 * if it is worth looking at further for free memory:
1949 * 1) Check that the zone isn't thought to be full (doesn't have its
1950 * bit set in the zonelist_cache fullzones BITMAP).
1951 * 2) Check that the zones node (obtained from the zonelist_cache
1952 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1953 * Return true (non-zero) if zone is worth looking at further, or
1954 * else return false (zero) if it is not.
1956 * This check -ignores- the distinction between various watermarks,
1957 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1958 * found to be full for any variation of these watermarks, it will
1959 * be considered full for up to one second by all requests, unless
1960 * we are so low on memory on all allowed nodes that we are forced
1961 * into the second scan of the zonelist.
1963 * In the second scan we ignore this zonelist cache and exactly
1964 * apply the watermarks to all zones, even it is slower to do so.
1965 * We are low on memory in the second scan, and should leave no stone
1966 * unturned looking for a free page.
1968 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1969 nodemask_t *allowednodes)
1971 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1972 int i; /* index of *z in zonelist zones */
1973 int n; /* node that zone *z is on */
1975 zlc = zonelist->zlcache_ptr;
1979 i = z - zonelist->_zonerefs;
1982 /* This zone is worth trying if it is allowed but not full */
1983 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1987 * Given 'z' scanning a zonelist, set the corresponding bit in
1988 * zlc->fullzones, so that subsequent attempts to allocate a page
1989 * from that zone don't waste time re-examining it.
1991 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1993 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1994 int i; /* index of *z in zonelist zones */
1996 zlc = zonelist->zlcache_ptr;
2000 i = z - zonelist->_zonerefs;
2002 set_bit(i, zlc->fullzones);
2006 * clear all zones full, called after direct reclaim makes progress so that
2007 * a zone that was recently full is not skipped over for up to a second
2009 static void zlc_clear_zones_full(struct zonelist *zonelist)
2011 struct zonelist_cache *zlc; /* cached zonelist speedup info */
2013 zlc = zonelist->zlcache_ptr;
2017 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
2020 static bool zone_local(struct zone *local_zone, struct zone *zone)
2022 return local_zone->node == zone->node;
2025 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2027 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <
2031 #else /* CONFIG_NUMA */
2033 static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
2038 static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
2039 nodemask_t *allowednodes)
2044 static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
2048 static void zlc_clear_zones_full(struct zonelist *zonelist)
2052 static bool zone_local(struct zone *local_zone, struct zone *zone)
2057 static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
2062 #endif /* CONFIG_NUMA */
2064 static void reset_alloc_batches(struct zone *preferred_zone)
2066 struct zone *zone = preferred_zone->zone_pgdat->node_zones;
2069 mod_zone_page_state(zone, NR_ALLOC_BATCH,
2070 high_wmark_pages(zone) - low_wmark_pages(zone) -
2071 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
2072 clear_bit(ZONE_FAIR_DEPLETED, &zone->flags);
2073 } while (zone++ != preferred_zone);
2077 * get_page_from_freelist goes through the zonelist trying to allocate
2080 static struct page *
2081 get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
2082 const struct alloc_context *ac)
2084 struct zonelist *zonelist = ac->zonelist;
2086 struct page *page = NULL;
2088 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
2089 int zlc_active = 0; /* set if using zonelist_cache */
2090 int did_zlc_setup = 0; /* just call zlc_setup() one time */
2091 bool consider_zone_dirty = (alloc_flags & ALLOC_WMARK_LOW) &&
2092 (gfp_mask & __GFP_WRITE);
2093 int nr_fair_skipped = 0;
2094 bool zonelist_rescan;
2097 zonelist_rescan = false;
2100 * Scan zonelist, looking for a zone with enough free.
2101 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2103 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->high_zoneidx,
2107 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2108 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2110 if (cpusets_enabled() &&
2111 (alloc_flags & ALLOC_CPUSET) &&
2112 !cpuset_zone_allowed(zone, gfp_mask))
2115 * Distribute pages in proportion to the individual
2116 * zone size to ensure fair page aging. The zone a
2117 * page was allocated in should have no effect on the
2118 * time the page has in memory before being reclaimed.
2120 if (alloc_flags & ALLOC_FAIR) {
2121 if (!zone_local(ac->preferred_zone, zone))
2123 if (test_bit(ZONE_FAIR_DEPLETED, &zone->flags)) {
2129 * When allocating a page cache page for writing, we
2130 * want to get it from a zone that is within its dirty
2131 * limit, such that no single zone holds more than its
2132 * proportional share of globally allowed dirty pages.
2133 * The dirty limits take into account the zone's
2134 * lowmem reserves and high watermark so that kswapd
2135 * should be able to balance it without having to
2136 * write pages from its LRU list.
2138 * This may look like it could increase pressure on
2139 * lower zones by failing allocations in higher zones
2140 * before they are full. But the pages that do spill
2141 * over are limited as the lower zones are protected
2142 * by this very same mechanism. It should not become
2143 * a practical burden to them.
2145 * XXX: For now, allow allocations to potentially
2146 * exceed the per-zone dirty limit in the slowpath
2147 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2148 * which is important when on a NUMA setup the allowed
2149 * zones are together not big enough to reach the
2150 * global limit. The proper fix for these situations
2151 * will require awareness of zones in the
2152 * dirty-throttling and the flusher threads.
2154 if (consider_zone_dirty && !zone_dirty_ok(zone))
2157 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
2158 if (!zone_watermark_ok(zone, order, mark,
2159 ac->classzone_idx, alloc_flags)) {
2162 /* Checked here to keep the fast path fast */
2163 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
2164 if (alloc_flags & ALLOC_NO_WATERMARKS)
2167 if (IS_ENABLED(CONFIG_NUMA) &&
2168 !did_zlc_setup && nr_online_nodes > 1) {
2170 * we do zlc_setup if there are multiple nodes
2171 * and before considering the first zone allowed
2174 allowednodes = zlc_setup(zonelist, alloc_flags);
2179 if (zone_reclaim_mode == 0 ||
2180 !zone_allows_reclaim(ac->preferred_zone, zone))
2181 goto this_zone_full;
2184 * As we may have just activated ZLC, check if the first
2185 * eligible zone has failed zone_reclaim recently.
2187 if (IS_ENABLED(CONFIG_NUMA) && zlc_active &&
2188 !zlc_zone_worth_trying(zonelist, z, allowednodes))
2191 ret = zone_reclaim(zone, gfp_mask, order);
2193 case ZONE_RECLAIM_NOSCAN:
2196 case ZONE_RECLAIM_FULL:
2197 /* scanned but unreclaimable */
2200 /* did we reclaim enough */
2201 if (zone_watermark_ok(zone, order, mark,
2202 ac->classzone_idx, alloc_flags))
2206 * Failed to reclaim enough to meet watermark.
2207 * Only mark the zone full if checking the min
2208 * watermark or if we failed to reclaim just
2209 * 1<<order pages or else the page allocator
2210 * fastpath will prematurely mark zones full
2211 * when the watermark is between the low and
2214 if (((alloc_flags & ALLOC_WMARK_MASK) == ALLOC_WMARK_MIN) ||
2215 ret == ZONE_RECLAIM_SOME)
2216 goto this_zone_full;
2223 page = buffered_rmqueue(ac->preferred_zone, zone, order,
2224 gfp_mask, ac->migratetype);
2226 if (prep_new_page(page, order, gfp_mask, alloc_flags))
2231 if (IS_ENABLED(CONFIG_NUMA) && zlc_active)
2232 zlc_mark_zone_full(zonelist, z);
2236 * The first pass makes sure allocations are spread fairly within the
2237 * local node. However, the local node might have free pages left
2238 * after the fairness batches are exhausted, and remote zones haven't
2239 * even been considered yet. Try once more without fairness, and
2240 * include remote zones now, before entering the slowpath and waking
2241 * kswapd: prefer spilling to a remote zone over swapping locally.
2243 if (alloc_flags & ALLOC_FAIR) {
2244 alloc_flags &= ~ALLOC_FAIR;
2245 if (nr_fair_skipped) {
2246 zonelist_rescan = true;
2247 reset_alloc_batches(ac->preferred_zone);
2249 if (nr_online_nodes > 1)
2250 zonelist_rescan = true;
2253 if (unlikely(IS_ENABLED(CONFIG_NUMA) && zlc_active)) {
2254 /* Disable zlc cache for second zonelist scan */
2256 zonelist_rescan = true;
2259 if (zonelist_rescan)
2266 * Large machines with many possible nodes should not always dump per-node
2267 * meminfo in irq context.
2269 static inline bool should_suppress_show_mem(void)
2274 ret = in_interrupt();
2279 static DEFINE_RATELIMIT_STATE(nopage_rs,
2280 DEFAULT_RATELIMIT_INTERVAL,
2281 DEFAULT_RATELIMIT_BURST);
2283 void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
2285 unsigned int filter = SHOW_MEM_FILTER_NODES;
2287 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs) ||
2288 debug_guardpage_minorder() > 0)
2292 * This documents exceptions given to allocations in certain
2293 * contexts that are allowed to allocate outside current's set
2296 if (!(gfp_mask & __GFP_NOMEMALLOC))
2297 if (test_thread_flag(TIF_MEMDIE) ||
2298 (current->flags & (PF_MEMALLOC | PF_EXITING)))
2299 filter &= ~SHOW_MEM_FILTER_NODES;
2300 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
2301 filter &= ~SHOW_MEM_FILTER_NODES;
2304 struct va_format vaf;
2307 va_start(args, fmt);
2312 pr_warn("%pV", &vaf);
2317 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2318 current->comm, order, gfp_mask);
2321 if (!should_suppress_show_mem())
2326 should_alloc_retry(gfp_t gfp_mask, unsigned int order,
2327 unsigned long did_some_progress,
2328 unsigned long pages_reclaimed)
2330 /* Do not loop if specifically requested */
2331 if (gfp_mask & __GFP_NORETRY)
2334 /* Always retry if specifically requested */
2335 if (gfp_mask & __GFP_NOFAIL)
2339 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2340 * making forward progress without invoking OOM. Suspend also disables
2341 * storage devices so kswapd will not help. Bail if we are suspending.
2343 if (!did_some_progress && pm_suspended_storage())
2347 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2348 * means __GFP_NOFAIL, but that may not be true in other
2351 if (order <= PAGE_ALLOC_COSTLY_ORDER)
2355 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2356 * specified, then we retry until we no longer reclaim any pages
2357 * (above), or we've reclaimed an order of pages at least as
2358 * large as the allocation's order. In both cases, if the
2359 * allocation still fails, we stop retrying.
2361 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
2367 static inline struct page *
2368 __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
2369 const struct alloc_context *ac, unsigned long *did_some_progress)
2373 *did_some_progress = 0;
2376 * Acquire the per-zone oom lock for each zone. If that
2377 * fails, somebody else is making progress for us.
2379 if (!oom_zonelist_trylock(ac->zonelist, gfp_mask)) {
2380 *did_some_progress = 1;
2381 schedule_timeout_uninterruptible(1);
2386 * Go through the zonelist yet one more time, keep very high watermark
2387 * here, this is only to catch a parallel oom killing, we must fail if
2388 * we're still under heavy pressure.
2390 page = get_page_from_freelist(gfp_mask | __GFP_HARDWALL, order,
2391 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
2395 if (!(gfp_mask & __GFP_NOFAIL)) {
2396 /* Coredumps can quickly deplete all memory reserves */
2397 if (current->flags & PF_DUMPCORE)
2399 /* The OOM killer will not help higher order allocs */
2400 if (order > PAGE_ALLOC_COSTLY_ORDER)
2402 /* The OOM killer does not needlessly kill tasks for lowmem */
2403 if (ac->high_zoneidx < ZONE_NORMAL)
2405 /* The OOM killer does not compensate for light reclaim */
2406 if (!(gfp_mask & __GFP_FS)) {
2408 * XXX: Page reclaim didn't yield anything,
2409 * and the OOM killer can't be invoked, but
2410 * keep looping as per should_alloc_retry().
2412 *did_some_progress = 1;
2415 /* The OOM killer may not free memory on a specific node */
2416 if (gfp_mask & __GFP_THISNODE)
2419 /* Exhausted what can be done so it's blamo time */
2420 if (out_of_memory(ac->zonelist, gfp_mask, order, ac->nodemask, false)
2421 || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL))
2422 *did_some_progress = 1;
2424 oom_zonelist_unlock(ac->zonelist, gfp_mask);
2428 #ifdef CONFIG_COMPACTION
2429 /* Try memory compaction for high-order allocations before reclaim */
2430 static struct page *
2431 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2432 int alloc_flags, const struct alloc_context *ac,
2433 enum migrate_mode mode, int *contended_compaction,
2434 bool *deferred_compaction)
2436 unsigned long compact_result;
2442 current->flags |= PF_MEMALLOC;
2443 compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
2444 mode, contended_compaction);
2445 current->flags &= ~PF_MEMALLOC;
2447 switch (compact_result) {
2448 case COMPACT_DEFERRED:
2449 *deferred_compaction = true;
2451 case COMPACT_SKIPPED:
2458 * At least in one zone compaction wasn't deferred or skipped, so let's
2459 * count a compaction stall
2461 count_vm_event(COMPACTSTALL);
2463 page = get_page_from_freelist(gfp_mask, order,
2464 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2467 struct zone *zone = page_zone(page);
2469 zone->compact_blockskip_flush = false;
2470 compaction_defer_reset(zone, order, true);
2471 count_vm_event(COMPACTSUCCESS);
2476 * It's bad if compaction run occurs and fails. The most likely reason
2477 * is that pages exist, but not enough to satisfy watermarks.
2479 count_vm_event(COMPACTFAIL);
2486 static inline struct page *
2487 __alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
2488 int alloc_flags, const struct alloc_context *ac,
2489 enum migrate_mode mode, int *contended_compaction,
2490 bool *deferred_compaction)
2494 #endif /* CONFIG_COMPACTION */
2496 /* Perform direct synchronous page reclaim */
2498 __perform_reclaim(gfp_t gfp_mask, unsigned int order,
2499 const struct alloc_context *ac)
2501 struct reclaim_state reclaim_state;
2506 /* We now go into synchronous reclaim */
2507 cpuset_memory_pressure_bump();
2508 current->flags |= PF_MEMALLOC;
2509 lockdep_set_current_reclaim_state(gfp_mask);
2510 reclaim_state.reclaimed_slab = 0;
2511 current->reclaim_state = &reclaim_state;
2513 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
2516 current->reclaim_state = NULL;
2517 lockdep_clear_current_reclaim_state();
2518 current->flags &= ~PF_MEMALLOC;
2525 /* The really slow allocator path where we enter direct reclaim */
2526 static inline struct page *
2527 __alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
2528 int alloc_flags, const struct alloc_context *ac,
2529 unsigned long *did_some_progress)
2531 struct page *page = NULL;
2532 bool drained = false;
2534 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
2535 if (unlikely(!(*did_some_progress)))
2538 /* After successful reclaim, reconsider all zones for allocation */
2539 if (IS_ENABLED(CONFIG_NUMA))
2540 zlc_clear_zones_full(ac->zonelist);
2543 page = get_page_from_freelist(gfp_mask, order,
2544 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2547 * If an allocation failed after direct reclaim, it could be because
2548 * pages are pinned on the per-cpu lists. Drain them and try again
2550 if (!page && !drained) {
2551 drain_all_pages(NULL);
2560 * This is called in the allocator slow-path if the allocation request is of
2561 * sufficient urgency to ignore watermarks and take other desperate measures
2563 static inline struct page *
2564 __alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
2565 const struct alloc_context *ac)
2570 page = get_page_from_freelist(gfp_mask, order,
2571 ALLOC_NO_WATERMARKS, ac);
2573 if (!page && gfp_mask & __GFP_NOFAIL)
2574 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC,
2576 } while (!page && (gfp_mask & __GFP_NOFAIL));
2581 static void wake_all_kswapds(unsigned int order, const struct alloc_context *ac)
2586 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
2587 ac->high_zoneidx, ac->nodemask)
2588 wakeup_kswapd(zone, order, zone_idx(ac->preferred_zone));
2592 gfp_to_alloc_flags(gfp_t gfp_mask)
2594 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2595 const bool atomic = !(gfp_mask & (__GFP_WAIT | __GFP_NO_KSWAPD));
2597 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2598 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2601 * The caller may dip into page reserves a bit more if the caller
2602 * cannot run direct reclaim, or if the caller has realtime scheduling
2603 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2604 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2606 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2610 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2611 * if it can't schedule.
2613 if (!(gfp_mask & __GFP_NOMEMALLOC))
2614 alloc_flags |= ALLOC_HARDER;
2616 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2617 * comment for __cpuset_node_allowed().
2619 alloc_flags &= ~ALLOC_CPUSET;
2620 } else if (unlikely(rt_task(current)) && !in_interrupt())
2621 alloc_flags |= ALLOC_HARDER;
2623 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2624 if (gfp_mask & __GFP_MEMALLOC)
2625 alloc_flags |= ALLOC_NO_WATERMARKS;
2626 else if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
2627 alloc_flags |= ALLOC_NO_WATERMARKS;
2628 else if (!in_interrupt() &&
2629 ((current->flags & PF_MEMALLOC) ||
2630 unlikely(test_thread_flag(TIF_MEMDIE))))
2631 alloc_flags |= ALLOC_NO_WATERMARKS;
2634 if (gfpflags_to_migratetype(gfp_mask) == MIGRATE_MOVABLE)
2635 alloc_flags |= ALLOC_CMA;
2640 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
2642 return !!(gfp_to_alloc_flags(gfp_mask) & ALLOC_NO_WATERMARKS);
2645 static inline struct page *
2646 __alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2647 struct alloc_context *ac)
2649 const gfp_t wait = gfp_mask & __GFP_WAIT;
2650 struct page *page = NULL;
2652 unsigned long pages_reclaimed = 0;
2653 unsigned long did_some_progress;
2654 enum migrate_mode migration_mode = MIGRATE_ASYNC;
2655 bool deferred_compaction = false;
2656 int contended_compaction = COMPACT_CONTENDED_NONE;
2659 * In the slowpath, we sanity check order to avoid ever trying to
2660 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2661 * be using allocators in order of preference for an area that is
2664 if (order >= MAX_ORDER) {
2665 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2670 * If this allocation cannot block and it is for a specific node, then
2671 * fail early. There's no need to wakeup kswapd or retry for a
2672 * speculative node-specific allocation.
2674 if (IS_ENABLED(CONFIG_NUMA) && (gfp_mask & __GFP_THISNODE) && !wait)
2678 if (!(gfp_mask & __GFP_NO_KSWAPD))
2679 wake_all_kswapds(order, ac);
2682 * OK, we're below the kswapd watermark and have kicked background
2683 * reclaim. Now things get more complex, so set up alloc_flags according
2684 * to how we want to proceed.
2686 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2689 * Find the true preferred zone if the allocation is unconstrained by
2692 if (!(alloc_flags & ALLOC_CPUSET) && !ac->nodemask) {
2693 struct zoneref *preferred_zoneref;
2694 preferred_zoneref = first_zones_zonelist(ac->zonelist,
2695 ac->high_zoneidx, NULL, &ac->preferred_zone);
2696 ac->classzone_idx = zonelist_zone_idx(preferred_zoneref);
2699 /* This is the last chance, in general, before the goto nopage. */
2700 page = get_page_from_freelist(gfp_mask, order,
2701 alloc_flags & ~ALLOC_NO_WATERMARKS, ac);
2705 /* Allocate without watermarks if the context allows */
2706 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2708 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2709 * the allocation is high priority and these type of
2710 * allocations are system rather than user orientated
2712 ac->zonelist = node_zonelist(numa_node_id(), gfp_mask);
2714 page = __alloc_pages_high_priority(gfp_mask, order, ac);
2721 /* Atomic allocations - we can't balance anything */
2724 * All existing users of the deprecated __GFP_NOFAIL are
2725 * blockable, so warn of any new users that actually allow this
2726 * type of allocation to fail.
2728 WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL);
2732 /* Avoid recursion of direct reclaim */
2733 if (current->flags & PF_MEMALLOC)
2736 /* Avoid allocations with no watermarks from looping endlessly */
2737 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2741 * Try direct compaction. The first pass is asynchronous. Subsequent
2742 * attempts after direct reclaim are synchronous
2744 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
2746 &contended_compaction,
2747 &deferred_compaction);
2751 /* Checks for THP-specific high-order allocations */
2752 if ((gfp_mask & GFP_TRANSHUGE) == GFP_TRANSHUGE) {
2754 * If compaction is deferred for high-order allocations, it is
2755 * because sync compaction recently failed. If this is the case
2756 * and the caller requested a THP allocation, we do not want
2757 * to heavily disrupt the system, so we fail the allocation
2758 * instead of entering direct reclaim.
2760 if (deferred_compaction)
2764 * In all zones where compaction was attempted (and not
2765 * deferred or skipped), lock contention has been detected.
2766 * For THP allocation we do not want to disrupt the others
2767 * so we fallback to base pages instead.
2769 if (contended_compaction == COMPACT_CONTENDED_LOCK)
2773 * If compaction was aborted due to need_resched(), we do not
2774 * want to further increase allocation latency, unless it is
2775 * khugepaged trying to collapse.
2777 if (contended_compaction == COMPACT_CONTENDED_SCHED
2778 && !(current->flags & PF_KTHREAD))
2783 * It can become very expensive to allocate transparent hugepages at
2784 * fault, so use asynchronous memory compaction for THP unless it is
2785 * khugepaged trying to collapse.
2787 if ((gfp_mask & GFP_TRANSHUGE) != GFP_TRANSHUGE ||
2788 (current->flags & PF_KTHREAD))
2789 migration_mode = MIGRATE_SYNC_LIGHT;
2791 /* Try direct reclaim and then allocating */
2792 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
2793 &did_some_progress);
2797 /* Check if we should retry the allocation */
2798 pages_reclaimed += did_some_progress;
2799 if (should_alloc_retry(gfp_mask, order, did_some_progress,
2802 * If we fail to make progress by freeing individual
2803 * pages, but the allocation wants us to keep going,
2804 * start OOM killing tasks.
2806 if (!did_some_progress) {
2807 page = __alloc_pages_may_oom(gfp_mask, order, ac,
2808 &did_some_progress);
2811 if (!did_some_progress)
2814 /* Wait for some write requests to complete then retry */
2815 wait_iff_congested(ac->preferred_zone, BLK_RW_ASYNC, HZ/50);
2819 * High-order allocations do not necessarily loop after
2820 * direct reclaim and reclaim/compaction depends on compaction
2821 * being called after reclaim so call directly if necessary
2823 page = __alloc_pages_direct_compact(gfp_mask, order,
2824 alloc_flags, ac, migration_mode,
2825 &contended_compaction,
2826 &deferred_compaction);
2832 warn_alloc_failed(gfp_mask, order, NULL);
2838 * This is the 'heart' of the zoned buddy allocator.
2841 __alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2842 struct zonelist *zonelist, nodemask_t *nodemask)
2844 struct zoneref *preferred_zoneref;
2845 struct page *page = NULL;
2846 unsigned int cpuset_mems_cookie;
2847 int alloc_flags = ALLOC_WMARK_LOW|ALLOC_CPUSET|ALLOC_FAIR;
2848 gfp_t alloc_mask; /* The gfp_t that was actually used for allocation */
2849 struct alloc_context ac = {
2850 .high_zoneidx = gfp_zone(gfp_mask),
2851 .nodemask = nodemask,
2852 .migratetype = gfpflags_to_migratetype(gfp_mask),
2855 gfp_mask &= gfp_allowed_mask;
2857 lockdep_trace_alloc(gfp_mask);
2859 might_sleep_if(gfp_mask & __GFP_WAIT);
2861 if (should_fail_alloc_page(gfp_mask, order))
2865 * Check the zones suitable for the gfp_mask contain at least one
2866 * valid zone. It's possible to have an empty zonelist as a result
2867 * of __GFP_THISNODE and a memoryless node
2869 if (unlikely(!zonelist->_zonerefs->zone))
2872 if (IS_ENABLED(CONFIG_CMA) && ac.migratetype == MIGRATE_MOVABLE)
2873 alloc_flags |= ALLOC_CMA;
2876 cpuset_mems_cookie = read_mems_allowed_begin();
2878 /* We set it here, as __alloc_pages_slowpath might have changed it */
2879 ac.zonelist = zonelist;
2880 /* The preferred zone is used for statistics later */
2881 preferred_zoneref = first_zones_zonelist(ac.zonelist, ac.high_zoneidx,
2882 ac.nodemask ? : &cpuset_current_mems_allowed,
2883 &ac.preferred_zone);
2884 if (!ac.preferred_zone)
2886 ac.classzone_idx = zonelist_zone_idx(preferred_zoneref);
2888 /* First allocation attempt */
2889 alloc_mask = gfp_mask|__GFP_HARDWALL;
2890 page = get_page_from_freelist(alloc_mask, order, alloc_flags, &ac);
2891 if (unlikely(!page)) {
2893 * Runtime PM, block IO and its error handling path
2894 * can deadlock because I/O on the device might not
2897 alloc_mask = memalloc_noio_flags(gfp_mask);
2899 page = __alloc_pages_slowpath(alloc_mask, order, &ac);
2902 if (kmemcheck_enabled && page)
2903 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2905 trace_mm_page_alloc(page, order, alloc_mask, ac.migratetype);
2909 * When updating a task's mems_allowed, it is possible to race with
2910 * parallel threads in such a way that an allocation can fail while
2911 * the mask is being updated. If a page allocation is about to fail,
2912 * check if the cpuset changed during allocation and if so, retry.
2914 if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2919 EXPORT_SYMBOL(__alloc_pages_nodemask);
2922 * Common helper functions.
2924 unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2929 * __get_free_pages() returns a 32-bit address, which cannot represent
2932 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2934 page = alloc_pages(gfp_mask, order);
2937 return (unsigned long) page_address(page);
2939 EXPORT_SYMBOL(__get_free_pages);
2941 unsigned long get_zeroed_page(gfp_t gfp_mask)
2943 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2945 EXPORT_SYMBOL(get_zeroed_page);
2947 void __free_pages(struct page *page, unsigned int order)
2949 if (put_page_testzero(page)) {
2951 free_hot_cold_page(page, false);
2953 __free_pages_ok(page, order);
2957 EXPORT_SYMBOL(__free_pages);
2959 void free_pages(unsigned long addr, unsigned int order)
2962 VM_BUG_ON(!virt_addr_valid((void *)addr));
2963 __free_pages(virt_to_page((void *)addr), order);
2967 EXPORT_SYMBOL(free_pages);
2970 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2971 * of the current memory cgroup.
2973 * It should be used when the caller would like to use kmalloc, but since the
2974 * allocation is large, it has to fall back to the page allocator.
2976 struct page *alloc_kmem_pages(gfp_t gfp_mask, unsigned int order)
2979 struct mem_cgroup *memcg = NULL;
2981 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2983 page = alloc_pages(gfp_mask, order);
2984 memcg_kmem_commit_charge(page, memcg, order);
2988 struct page *alloc_kmem_pages_node(int nid, gfp_t gfp_mask, unsigned int order)
2991 struct mem_cgroup *memcg = NULL;
2993 if (!memcg_kmem_newpage_charge(gfp_mask, &memcg, order))
2995 page = alloc_pages_node(nid, gfp_mask, order);
2996 memcg_kmem_commit_charge(page, memcg, order);
3001 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3004 void __free_kmem_pages(struct page *page, unsigned int order)
3006 memcg_kmem_uncharge_pages(page, order);
3007 __free_pages(page, order);
3010 void free_kmem_pages(unsigned long addr, unsigned int order)
3013 VM_BUG_ON(!virt_addr_valid((void *)addr));
3014 __free_kmem_pages(virt_to_page((void *)addr), order);
3018 static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
3021 unsigned long alloc_end = addr + (PAGE_SIZE << order);
3022 unsigned long used = addr + PAGE_ALIGN(size);
3024 split_page(virt_to_page((void *)addr), order);
3025 while (used < alloc_end) {
3030 return (void *)addr;
3034 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3035 * @size: the number of bytes to allocate
3036 * @gfp_mask: GFP flags for the allocation
3038 * This function is similar to alloc_pages(), except that it allocates the
3039 * minimum number of pages to satisfy the request. alloc_pages() can only
3040 * allocate memory in power-of-two pages.
3042 * This function is also limited by MAX_ORDER.
3044 * Memory allocated by this function must be released by free_pages_exact().
3046 void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
3048 unsigned int order = get_order(size);
3051 addr = __get_free_pages(gfp_mask, order);
3052 return make_alloc_exact(addr, order, size);
3054 EXPORT_SYMBOL(alloc_pages_exact);
3057 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3059 * @nid: the preferred node ID where memory should be allocated
3060 * @size: the number of bytes to allocate
3061 * @gfp_mask: GFP flags for the allocation
3063 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3065 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3068 void * __meminit alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
3070 unsigned order = get_order(size);
3071 struct page *p = alloc_pages_node(nid, gfp_mask, order);
3074 return make_alloc_exact((unsigned long)page_address(p), order, size);
3078 * free_pages_exact - release memory allocated via alloc_pages_exact()
3079 * @virt: the value returned by alloc_pages_exact.
3080 * @size: size of allocation, same value as passed to alloc_pages_exact().
3082 * Release the memory allocated by a previous call to alloc_pages_exact.
3084 void free_pages_exact(void *virt, size_t size)
3086 unsigned long addr = (unsigned long)virt;
3087 unsigned long end = addr + PAGE_ALIGN(size);
3089 while (addr < end) {
3094 EXPORT_SYMBOL(free_pages_exact);
3097 * nr_free_zone_pages - count number of pages beyond high watermark
3098 * @offset: The zone index of the highest zone
3100 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3101 * high watermark within all zones at or below a given zone index. For each
3102 * zone, the number of pages is calculated as:
3103 * managed_pages - high_pages
3105 static unsigned long nr_free_zone_pages(int offset)
3110 /* Just pick one node, since fallback list is circular */
3111 unsigned long sum = 0;
3113 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
3115 for_each_zone_zonelist(zone, z, zonelist, offset) {
3116 unsigned long size = zone->managed_pages;
3117 unsigned long high = high_wmark_pages(zone);
3126 * nr_free_buffer_pages - count number of pages beyond high watermark
3128 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3129 * watermark within ZONE_DMA and ZONE_NORMAL.
3131 unsigned long nr_free_buffer_pages(void)
3133 return nr_free_zone_pages(gfp_zone(GFP_USER));
3135 EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
3138 * nr_free_pagecache_pages - count number of pages beyond high watermark
3140 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3141 * high watermark within all zones.
3143 unsigned long nr_free_pagecache_pages(void)
3145 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
3148 static inline void show_node(struct zone *zone)
3150 if (IS_ENABLED(CONFIG_NUMA))
3151 printk("Node %d ", zone_to_nid(zone));
3154 void si_meminfo(struct sysinfo *val)
3156 val->totalram = totalram_pages;
3157 val->sharedram = global_page_state(NR_SHMEM);
3158 val->freeram = global_page_state(NR_FREE_PAGES);
3159 val->bufferram = nr_blockdev_pages();
3160 val->totalhigh = totalhigh_pages;
3161 val->freehigh = nr_free_highpages();
3162 val->mem_unit = PAGE_SIZE;
3165 EXPORT_SYMBOL(si_meminfo);
3168 void si_meminfo_node(struct sysinfo *val, int nid)
3170 int zone_type; /* needs to be signed */
3171 unsigned long managed_pages = 0;
3172 pg_data_t *pgdat = NODE_DATA(nid);
3174 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++)
3175 managed_pages += pgdat->node_zones[zone_type].managed_pages;
3176 val->totalram = managed_pages;
3177 val->sharedram = node_page_state(nid, NR_SHMEM);
3178 val->freeram = node_page_state(nid, NR_FREE_PAGES);
3179 #ifdef CONFIG_HIGHMEM
3180 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].managed_pages;
3181 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
3187 val->mem_unit = PAGE_SIZE;
3192 * Determine whether the node should be displayed or not, depending on whether
3193 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3195 bool skip_free_areas_node(unsigned int flags, int nid)
3198 unsigned int cpuset_mems_cookie;
3200 if (!(flags & SHOW_MEM_FILTER_NODES))
3204 cpuset_mems_cookie = read_mems_allowed_begin();
3205 ret = !node_isset(nid, cpuset_current_mems_allowed);
3206 } while (read_mems_allowed_retry(cpuset_mems_cookie));
3211 #define K(x) ((x) << (PAGE_SHIFT-10))
3213 static void show_migration_types(unsigned char type)
3215 static const char types[MIGRATE_TYPES] = {
3216 [MIGRATE_UNMOVABLE] = 'U',
3217 [MIGRATE_RECLAIMABLE] = 'E',
3218 [MIGRATE_MOVABLE] = 'M',
3219 [MIGRATE_RESERVE] = 'R',
3221 [MIGRATE_CMA] = 'C',
3223 #ifdef CONFIG_MEMORY_ISOLATION
3224 [MIGRATE_ISOLATE] = 'I',
3227 char tmp[MIGRATE_TYPES + 1];
3231 for (i = 0; i < MIGRATE_TYPES; i++) {
3232 if (type & (1 << i))
3237 printk("(%s) ", tmp);
3241 * Show free area list (used inside shift_scroll-lock stuff)
3242 * We also calculate the percentage fragmentation. We do this by counting the
3243 * memory on each free list with the exception of the first item on the list.
3246 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3249 void show_free_areas(unsigned int filter)
3251 unsigned long free_pcp = 0;
3255 for_each_populated_zone(zone) {
3256 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3259 for_each_online_cpu(cpu)
3260 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3263 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3264 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3265 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3266 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3267 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3268 " free:%lu free_pcp:%lu free_cma:%lu\n",
3269 global_page_state(NR_ACTIVE_ANON),
3270 global_page_state(NR_INACTIVE_ANON),
3271 global_page_state(NR_ISOLATED_ANON),
3272 global_page_state(NR_ACTIVE_FILE),
3273 global_page_state(NR_INACTIVE_FILE),
3274 global_page_state(NR_ISOLATED_FILE),
3275 global_page_state(NR_UNEVICTABLE),
3276 global_page_state(NR_FILE_DIRTY),
3277 global_page_state(NR_WRITEBACK),
3278 global_page_state(NR_UNSTABLE_NFS),
3279 global_page_state(NR_SLAB_RECLAIMABLE),
3280 global_page_state(NR_SLAB_UNRECLAIMABLE),
3281 global_page_state(NR_FILE_MAPPED),
3282 global_page_state(NR_SHMEM),
3283 global_page_state(NR_PAGETABLE),
3284 global_page_state(NR_BOUNCE),
3285 global_page_state(NR_FREE_PAGES),
3287 global_page_state(NR_FREE_CMA_PAGES));
3289 for_each_populated_zone(zone) {
3292 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3296 for_each_online_cpu(cpu)
3297 free_pcp += per_cpu_ptr(zone->pageset, cpu)->pcp.count;
3305 " active_anon:%lukB"
3306 " inactive_anon:%lukB"
3307 " active_file:%lukB"
3308 " inactive_file:%lukB"
3309 " unevictable:%lukB"
3310 " isolated(anon):%lukB"
3311 " isolated(file):%lukB"
3319 " slab_reclaimable:%lukB"
3320 " slab_unreclaimable:%lukB"
3321 " kernel_stack:%lukB"
3328 " writeback_tmp:%lukB"
3329 " pages_scanned:%lu"
3330 " all_unreclaimable? %s"
3333 K(zone_page_state(zone, NR_FREE_PAGES)),
3334 K(min_wmark_pages(zone)),
3335 K(low_wmark_pages(zone)),
3336 K(high_wmark_pages(zone)),
3337 K(zone_page_state(zone, NR_ACTIVE_ANON)),
3338 K(zone_page_state(zone, NR_INACTIVE_ANON)),
3339 K(zone_page_state(zone, NR_ACTIVE_FILE)),
3340 K(zone_page_state(zone, NR_INACTIVE_FILE)),
3341 K(zone_page_state(zone, NR_UNEVICTABLE)),
3342 K(zone_page_state(zone, NR_ISOLATED_ANON)),
3343 K(zone_page_state(zone, NR_ISOLATED_FILE)),
3344 K(zone->present_pages),
3345 K(zone->managed_pages),
3346 K(zone_page_state(zone, NR_MLOCK)),
3347 K(zone_page_state(zone, NR_FILE_DIRTY)),
3348 K(zone_page_state(zone, NR_WRITEBACK)),
3349 K(zone_page_state(zone, NR_FILE_MAPPED)),
3350 K(zone_page_state(zone, NR_SHMEM)),
3351 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
3352 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
3353 zone_page_state(zone, NR_KERNEL_STACK) *
3355 K(zone_page_state(zone, NR_PAGETABLE)),
3356 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
3357 K(zone_page_state(zone, NR_BOUNCE)),
3359 K(this_cpu_read(zone->pageset->pcp.count)),
3360 K(zone_page_state(zone, NR_FREE_CMA_PAGES)),
3361 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
3362 K(zone_page_state(zone, NR_PAGES_SCANNED)),
3363 (!zone_reclaimable(zone) ? "yes" : "no")
3365 printk("lowmem_reserve[]:");
3366 for (i = 0; i < MAX_NR_ZONES; i++)
3367 printk(" %ld", zone->lowmem_reserve[i]);
3371 for_each_populated_zone(zone) {
3372 unsigned long nr[MAX_ORDER], flags, order, total = 0;
3373 unsigned char types[MAX_ORDER];
3375 if (skip_free_areas_node(filter, zone_to_nid(zone)))
3378 printk("%s: ", zone->name);
3380 spin_lock_irqsave(&zone->lock, flags);
3381 for (order = 0; order < MAX_ORDER; order++) {
3382 struct free_area *area = &zone->free_area[order];
3385 nr[order] = area->nr_free;
3386 total += nr[order] << order;
3389 for (type = 0; type < MIGRATE_TYPES; type++) {
3390 if (!list_empty(&area->free_list[type]))
3391 types[order] |= 1 << type;
3394 spin_unlock_irqrestore(&zone->lock, flags);
3395 for (order = 0; order < MAX_ORDER; order++) {
3396 printk("%lu*%lukB ", nr[order], K(1UL) << order);
3398 show_migration_types(types[order]);
3400 printk("= %lukB\n", K(total));
3403 hugetlb_show_meminfo();
3405 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
3407 show_swap_cache_info();
3410 static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
3412 zoneref->zone = zone;
3413 zoneref->zone_idx = zone_idx(zone);
3417 * Builds allocation fallback zone lists.
3419 * Add all populated zones of a node to the zonelist.
3421 static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
3425 enum zone_type zone_type = MAX_NR_ZONES;
3429 zone = pgdat->node_zones + zone_type;
3430 if (populated_zone(zone)) {
3431 zoneref_set_zone(zone,
3432 &zonelist->_zonerefs[nr_zones++]);
3433 check_highest_zone(zone_type);
3435 } while (zone_type);
3443 * 0 = automatic detection of better ordering.
3444 * 1 = order by ([node] distance, -zonetype)
3445 * 2 = order by (-zonetype, [node] distance)
3447 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3448 * the same zonelist. So only NUMA can configure this param.
3450 #define ZONELIST_ORDER_DEFAULT 0
3451 #define ZONELIST_ORDER_NODE 1
3452 #define ZONELIST_ORDER_ZONE 2
3454 /* zonelist order in the kernel.
3455 * set_zonelist_order() will set this to NODE or ZONE.
3457 static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
3458 static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
3462 /* The value user specified ....changed by config */
3463 static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3464 /* string for sysctl */
3465 #define NUMA_ZONELIST_ORDER_LEN 16
3466 char numa_zonelist_order[16] = "default";
3469 * interface for configure zonelist ordering.
3470 * command line option "numa_zonelist_order"
3471 * = "[dD]efault - default, automatic configuration.
3472 * = "[nN]ode - order by node locality, then by zone within node
3473 * = "[zZ]one - order by zone, then by locality within zone
3476 static int __parse_numa_zonelist_order(char *s)
3478 if (*s == 'd' || *s == 'D') {
3479 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
3480 } else if (*s == 'n' || *s == 'N') {
3481 user_zonelist_order = ZONELIST_ORDER_NODE;
3482 } else if (*s == 'z' || *s == 'Z') {
3483 user_zonelist_order = ZONELIST_ORDER_ZONE;
3486 "Ignoring invalid numa_zonelist_order value: "
3493 static __init int setup_numa_zonelist_order(char *s)
3500 ret = __parse_numa_zonelist_order(s);
3502 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
3506 early_param("numa_zonelist_order", setup_numa_zonelist_order);
3509 * sysctl handler for numa_zonelist_order
3511 int numa_zonelist_order_handler(struct ctl_table *table, int write,
3512 void __user *buffer, size_t *length,
3515 char saved_string[NUMA_ZONELIST_ORDER_LEN];
3517 static DEFINE_MUTEX(zl_order_mutex);
3519 mutex_lock(&zl_order_mutex);
3521 if (strlen((char *)table->data) >= NUMA_ZONELIST_ORDER_LEN) {
3525 strcpy(saved_string, (char *)table->data);
3527 ret = proc_dostring(table, write, buffer, length, ppos);
3531 int oldval = user_zonelist_order;
3533 ret = __parse_numa_zonelist_order((char *)table->data);
3536 * bogus value. restore saved string
3538 strncpy((char *)table->data, saved_string,
3539 NUMA_ZONELIST_ORDER_LEN);
3540 user_zonelist_order = oldval;
3541 } else if (oldval != user_zonelist_order) {
3542 mutex_lock(&zonelists_mutex);
3543 build_all_zonelists(NULL, NULL);
3544 mutex_unlock(&zonelists_mutex);
3548 mutex_unlock(&zl_order_mutex);
3553 #define MAX_NODE_LOAD (nr_online_nodes)
3554 static int node_load[MAX_NUMNODES];
3557 * find_next_best_node - find the next node that should appear in a given node's fallback list
3558 * @node: node whose fallback list we're appending
3559 * @used_node_mask: nodemask_t of already used nodes
3561 * We use a number of factors to determine which is the next node that should
3562 * appear on a given node's fallback list. The node should not have appeared
3563 * already in @node's fallback list, and it should be the next closest node
3564 * according to the distance array (which contains arbitrary distance values
3565 * from each node to each node in the system), and should also prefer nodes
3566 * with no CPUs, since presumably they'll have very little allocation pressure
3567 * on them otherwise.
3568 * It returns -1 if no node is found.
3570 static int find_next_best_node(int node, nodemask_t *used_node_mask)
3573 int min_val = INT_MAX;
3574 int best_node = NUMA_NO_NODE;
3575 const struct cpumask *tmp = cpumask_of_node(0);
3577 /* Use the local node if we haven't already */
3578 if (!node_isset(node, *used_node_mask)) {
3579 node_set(node, *used_node_mask);
3583 for_each_node_state(n, N_MEMORY) {
3585 /* Don't want a node to appear more than once */
3586 if (node_isset(n, *used_node_mask))
3589 /* Use the distance array to find the distance */
3590 val = node_distance(node, n);
3592 /* Penalize nodes under us ("prefer the next node") */
3595 /* Give preference to headless and unused nodes */
3596 tmp = cpumask_of_node(n);
3597 if (!cpumask_empty(tmp))
3598 val += PENALTY_FOR_NODE_WITH_CPUS;
3600 /* Slight preference for less loaded node */
3601 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
3602 val += node_load[n];
3604 if (val < min_val) {
3611 node_set(best_node, *used_node_mask);
3618 * Build zonelists ordered by node and zones within node.
3619 * This results in maximum locality--normal zone overflows into local
3620 * DMA zone, if any--but risks exhausting DMA zone.
3622 static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
3625 struct zonelist *zonelist;
3627 zonelist = &pgdat->node_zonelists[0];
3628 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
3630 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3631 zonelist->_zonerefs[j].zone = NULL;
3632 zonelist->_zonerefs[j].zone_idx = 0;
3636 * Build gfp_thisnode zonelists
3638 static void build_thisnode_zonelists(pg_data_t *pgdat)
3641 struct zonelist *zonelist;
3643 zonelist = &pgdat->node_zonelists[1];
3644 j = build_zonelists_node(pgdat, zonelist, 0);
3645 zonelist->_zonerefs[j].zone = NULL;
3646 zonelist->_zonerefs[j].zone_idx = 0;
3650 * Build zonelists ordered by zone and nodes within zones.
3651 * This results in conserving DMA zone[s] until all Normal memory is
3652 * exhausted, but results in overflowing to remote node while memory
3653 * may still exist in local DMA zone.
3655 static int node_order[MAX_NUMNODES];
3657 static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
3660 int zone_type; /* needs to be signed */
3662 struct zonelist *zonelist;
3664 zonelist = &pgdat->node_zonelists[0];
3666 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
3667 for (j = 0; j < nr_nodes; j++) {
3668 node = node_order[j];
3669 z = &NODE_DATA(node)->node_zones[zone_type];
3670 if (populated_zone(z)) {
3672 &zonelist->_zonerefs[pos++]);
3673 check_highest_zone(zone_type);
3677 zonelist->_zonerefs[pos].zone = NULL;
3678 zonelist->_zonerefs[pos].zone_idx = 0;
3681 #if defined(CONFIG_64BIT)
3683 * Devices that require DMA32/DMA are relatively rare and do not justify a
3684 * penalty to every machine in case the specialised case applies. Default
3685 * to Node-ordering on 64-bit NUMA machines
3687 static int default_zonelist_order(void)
3689 return ZONELIST_ORDER_NODE;
3693 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3694 * by the kernel. If processes running on node 0 deplete the low memory zone
3695 * then reclaim will occur more frequency increasing stalls and potentially
3696 * be easier to OOM if a large percentage of the zone is under writeback or
3697 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3698 * Hence, default to zone ordering on 32-bit.
3700 static int default_zonelist_order(void)
3702 return ZONELIST_ORDER_ZONE;
3704 #endif /* CONFIG_64BIT */
3706 static void set_zonelist_order(void)
3708 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3709 current_zonelist_order = default_zonelist_order();
3711 current_zonelist_order = user_zonelist_order;
3714 static void build_zonelists(pg_data_t *pgdat)
3718 nodemask_t used_mask;
3719 int local_node, prev_node;
3720 struct zonelist *zonelist;
3721 int order = current_zonelist_order;
3723 /* initialize zonelists */
3724 for (i = 0; i < MAX_ZONELISTS; i++) {
3725 zonelist = pgdat->node_zonelists + i;
3726 zonelist->_zonerefs[0].zone = NULL;
3727 zonelist->_zonerefs[0].zone_idx = 0;
3730 /* NUMA-aware ordering of nodes */
3731 local_node = pgdat->node_id;
3732 load = nr_online_nodes;
3733 prev_node = local_node;
3734 nodes_clear(used_mask);
3736 memset(node_order, 0, sizeof(node_order));
3739 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3741 * We don't want to pressure a particular node.
3742 * So adding penalty to the first node in same
3743 * distance group to make it round-robin.
3745 if (node_distance(local_node, node) !=
3746 node_distance(local_node, prev_node))
3747 node_load[node] = load;
3751 if (order == ZONELIST_ORDER_NODE)
3752 build_zonelists_in_node_order(pgdat, node);
3754 node_order[j++] = node; /* remember order */
3757 if (order == ZONELIST_ORDER_ZONE) {
3758 /* calculate node order -- i.e., DMA last! */
3759 build_zonelists_in_zone_order(pgdat, j);
3762 build_thisnode_zonelists(pgdat);
3765 /* Construct the zonelist performance cache - see further mmzone.h */
3766 static void build_zonelist_cache(pg_data_t *pgdat)
3768 struct zonelist *zonelist;
3769 struct zonelist_cache *zlc;
3772 zonelist = &pgdat->node_zonelists[0];
3773 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3774 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3775 for (z = zonelist->_zonerefs; z->zone; z++)
3776 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3779 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3781 * Return node id of node used for "local" allocations.
3782 * I.e., first node id of first zone in arg node's generic zonelist.
3783 * Used for initializing percpu 'numa_mem', which is used primarily
3784 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3786 int local_memory_node(int node)
3790 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3791 gfp_zone(GFP_KERNEL),
3798 #else /* CONFIG_NUMA */
3800 static void set_zonelist_order(void)
3802 current_zonelist_order = ZONELIST_ORDER_ZONE;
3805 static void build_zonelists(pg_data_t *pgdat)
3807 int node, local_node;
3809 struct zonelist *zonelist;
3811 local_node = pgdat->node_id;
3813 zonelist = &pgdat->node_zonelists[0];
3814 j = build_zonelists_node(pgdat, zonelist, 0);
3817 * Now we build the zonelist so that it contains the zones
3818 * of all the other nodes.
3819 * We don't want to pressure a particular node, so when
3820 * building the zones for node N, we make sure that the
3821 * zones coming right after the local ones are those from
3822 * node N+1 (modulo N)
3824 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3825 if (!node_online(node))
3827 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3829 for (node = 0; node < local_node; node++) {
3830 if (!node_online(node))
3832 j = build_zonelists_node(NODE_DATA(node), zonelist, j);
3835 zonelist->_zonerefs[j].zone = NULL;
3836 zonelist->_zonerefs[j].zone_idx = 0;
3839 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3840 static void build_zonelist_cache(pg_data_t *pgdat)
3842 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3845 #endif /* CONFIG_NUMA */
3848 * Boot pageset table. One per cpu which is going to be used for all
3849 * zones and all nodes. The parameters will be set in such a way
3850 * that an item put on a list will immediately be handed over to
3851 * the buddy list. This is safe since pageset manipulation is done
3852 * with interrupts disabled.
3854 * The boot_pagesets must be kept even after bootup is complete for
3855 * unused processors and/or zones. They do play a role for bootstrapping
3856 * hotplugged processors.
3858 * zoneinfo_show() and maybe other functions do
3859 * not check if the processor is online before following the pageset pointer.
3860 * Other parts of the kernel may not check if the zone is available.
3862 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3863 static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3864 static void setup_zone_pageset(struct zone *zone);
3867 * Global mutex to protect against size modification of zonelists
3868 * as well as to serialize pageset setup for the new populated zone.
3870 DEFINE_MUTEX(zonelists_mutex);
3872 /* return values int ....just for stop_machine() */
3873 static int __build_all_zonelists(void *data)
3877 pg_data_t *self = data;
3880 memset(node_load, 0, sizeof(node_load));
3883 if (self && !node_online(self->node_id)) {
3884 build_zonelists(self);
3885 build_zonelist_cache(self);
3888 for_each_online_node(nid) {
3889 pg_data_t *pgdat = NODE_DATA(nid);
3891 build_zonelists(pgdat);
3892 build_zonelist_cache(pgdat);
3896 * Initialize the boot_pagesets that are going to be used
3897 * for bootstrapping processors. The real pagesets for
3898 * each zone will be allocated later when the per cpu
3899 * allocator is available.
3901 * boot_pagesets are used also for bootstrapping offline
3902 * cpus if the system is already booted because the pagesets
3903 * are needed to initialize allocators on a specific cpu too.
3904 * F.e. the percpu allocator needs the page allocator which
3905 * needs the percpu allocator in order to allocate its pagesets
3906 * (a chicken-egg dilemma).
3908 for_each_possible_cpu(cpu) {
3909 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3911 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3913 * We now know the "local memory node" for each node--
3914 * i.e., the node of the first zone in the generic zonelist.
3915 * Set up numa_mem percpu variable for on-line cpus. During
3916 * boot, only the boot cpu should be on-line; we'll init the
3917 * secondary cpus' numa_mem as they come on-line. During
3918 * node/memory hotplug, we'll fixup all on-line cpus.
3920 if (cpu_online(cpu))
3921 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3928 static noinline void __init
3929 build_all_zonelists_init(void)
3931 __build_all_zonelists(NULL);
3932 mminit_verify_zonelist();
3933 cpuset_init_current_mems_allowed();
3937 * Called with zonelists_mutex held always
3938 * unless system_state == SYSTEM_BOOTING.
3940 * __ref due to (1) call of __meminit annotated setup_zone_pageset
3941 * [we're only called with non-NULL zone through __meminit paths] and
3942 * (2) call of __init annotated helper build_all_zonelists_init
3943 * [protected by SYSTEM_BOOTING].
3945 void __ref build_all_zonelists(pg_data_t *pgdat, struct zone *zone)
3947 set_zonelist_order();
3949 if (system_state == SYSTEM_BOOTING) {
3950 build_all_zonelists_init();
3952 #ifdef CONFIG_MEMORY_HOTPLUG
3954 setup_zone_pageset(zone);
3956 /* we have to stop all cpus to guarantee there is no user
3958 stop_machine(__build_all_zonelists, pgdat, NULL);
3959 /* cpuset refresh routine should be here */
3961 vm_total_pages = nr_free_pagecache_pages();
3963 * Disable grouping by mobility if the number of pages in the
3964 * system is too low to allow the mechanism to work. It would be
3965 * more accurate, but expensive to check per-zone. This check is
3966 * made on memory-hotadd so a system can start with mobility
3967 * disabled and enable it later
3969 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3970 page_group_by_mobility_disabled = 1;
3972 page_group_by_mobility_disabled = 0;
3974 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3975 "Total pages: %ld\n",
3977 zonelist_order_name[current_zonelist_order],
3978 page_group_by_mobility_disabled ? "off" : "on",
3981 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
3986 * Helper functions to size the waitqueue hash table.
3987 * Essentially these want to choose hash table sizes sufficiently
3988 * large so that collisions trying to wait on pages are rare.
3989 * But in fact, the number of active page waitqueues on typical
3990 * systems is ridiculously low, less than 200. So this is even
3991 * conservative, even though it seems large.
3993 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3994 * waitqueues, i.e. the size of the waitq table given the number of pages.
3996 #define PAGES_PER_WAITQUEUE 256
3998 #ifndef CONFIG_MEMORY_HOTPLUG
3999 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4001 unsigned long size = 1;
4003 pages /= PAGES_PER_WAITQUEUE;
4005 while (size < pages)
4009 * Once we have dozens or even hundreds of threads sleeping
4010 * on IO we've got bigger problems than wait queue collision.
4011 * Limit the size of the wait table to a reasonable size.
4013 size = min(size, 4096UL);
4015 return max(size, 4UL);
4019 * A zone's size might be changed by hot-add, so it is not possible to determine
4020 * a suitable size for its wait_table. So we use the maximum size now.
4022 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4024 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4025 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4026 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4028 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4029 * or more by the traditional way. (See above). It equals:
4031 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4032 * ia64(16K page size) : = ( 8G + 4M)byte.
4033 * powerpc (64K page size) : = (32G +16M)byte.
4035 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
4042 * This is an integer logarithm so that shifts can be used later
4043 * to extract the more random high bits from the multiplicative
4044 * hash function before the remainder is taken.
4046 static inline unsigned long wait_table_bits(unsigned long size)
4052 * Check if a pageblock contains reserved pages
4054 static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
4058 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4059 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
4066 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4067 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4068 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4069 * higher will lead to a bigger reserve which will get freed as contiguous
4070 * blocks as reclaim kicks in
4072 static void setup_zone_migrate_reserve(struct zone *zone)
4074 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
4076 unsigned long block_migratetype;
4081 * Get the start pfn, end pfn and the number of blocks to reserve
4082 * We have to be careful to be aligned to pageblock_nr_pages to
4083 * make sure that we always check pfn_valid for the first page in
4086 start_pfn = zone->zone_start_pfn;
4087 end_pfn = zone_end_pfn(zone);
4088 start_pfn = roundup(start_pfn, pageblock_nr_pages);
4089 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
4093 * Reserve blocks are generally in place to help high-order atomic
4094 * allocations that are short-lived. A min_free_kbytes value that
4095 * would result in more than 2 reserve blocks for atomic allocations
4096 * is assumed to be in place to help anti-fragmentation for the
4097 * future allocation of hugepages at runtime.
4099 reserve = min(2, reserve);
4100 old_reserve = zone->nr_migrate_reserve_block;
4102 /* When memory hot-add, we almost always need to do nothing */
4103 if (reserve == old_reserve)
4105 zone->nr_migrate_reserve_block = reserve;
4107 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
4108 if (!pfn_valid(pfn))
4110 page = pfn_to_page(pfn);
4112 /* Watch out for overlapping nodes */
4113 if (page_to_nid(page) != zone_to_nid(zone))
4116 block_migratetype = get_pageblock_migratetype(page);
4118 /* Only test what is necessary when the reserves are not met */
4121 * Blocks with reserved pages will never free, skip
4124 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
4125 if (pageblock_is_reserved(pfn, block_end_pfn))
4128 /* If this block is reserved, account for it */
4129 if (block_migratetype == MIGRATE_RESERVE) {
4134 /* Suitable for reserving if this block is movable */
4135 if (block_migratetype == MIGRATE_MOVABLE) {
4136 set_pageblock_migratetype(page,
4138 move_freepages_block(zone, page,
4143 } else if (!old_reserve) {
4145 * At boot time we don't need to scan the whole zone
4146 * for turning off MIGRATE_RESERVE.
4152 * If the reserve is met and this is a previous reserved block,
4155 if (block_migratetype == MIGRATE_RESERVE) {
4156 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4157 move_freepages_block(zone, page, MIGRATE_MOVABLE);
4163 * Initially all pages are reserved - free ones are freed
4164 * up by free_all_bootmem() once the early boot process is
4165 * done. Non-atomic initialization, single-pass.
4167 void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
4168 unsigned long start_pfn, enum memmap_context context)
4171 unsigned long end_pfn = start_pfn + size;
4175 if (highest_memmap_pfn < end_pfn - 1)
4176 highest_memmap_pfn = end_pfn - 1;
4178 z = &NODE_DATA(nid)->node_zones[zone];
4179 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
4181 * There can be holes in boot-time mem_map[]s
4182 * handed to this function. They do not
4183 * exist on hotplugged memory.
4185 if (context == MEMMAP_EARLY) {
4186 if (!early_pfn_valid(pfn))
4188 if (!early_pfn_in_nid(pfn, nid))
4191 page = pfn_to_page(pfn);
4192 set_page_links(page, zone, nid, pfn);
4193 mminit_verify_page_links(page, zone, nid, pfn);
4194 init_page_count(page);
4195 page_mapcount_reset(page);
4196 page_cpupid_reset_last(page);
4197 SetPageReserved(page);
4199 * Mark the block movable so that blocks are reserved for
4200 * movable at startup. This will force kernel allocations
4201 * to reserve their blocks rather than leaking throughout
4202 * the address space during boot when many long-lived
4203 * kernel allocations are made. Later some blocks near
4204 * the start are marked MIGRATE_RESERVE by
4205 * setup_zone_migrate_reserve()
4207 * bitmap is created for zone's valid pfn range. but memmap
4208 * can be created for invalid pages (for alignment)
4209 * check here not to call set_pageblock_migratetype() against
4212 if ((z->zone_start_pfn <= pfn)
4213 && (pfn < zone_end_pfn(z))
4214 && !(pfn & (pageblock_nr_pages - 1)))
4215 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
4217 INIT_LIST_HEAD(&page->lru);
4218 #ifdef WANT_PAGE_VIRTUAL
4219 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4220 if (!is_highmem_idx(zone))
4221 set_page_address(page, __va(pfn << PAGE_SHIFT));
4226 static void __meminit zone_init_free_lists(struct zone *zone)
4228 unsigned int order, t;
4229 for_each_migratetype_order(order, t) {
4230 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
4231 zone->free_area[order].nr_free = 0;
4235 #ifndef __HAVE_ARCH_MEMMAP_INIT
4236 #define memmap_init(size, nid, zone, start_pfn) \
4237 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4240 static int zone_batchsize(struct zone *zone)
4246 * The per-cpu-pages pools are set to around 1000th of the
4247 * size of the zone. But no more than 1/2 of a meg.
4249 * OK, so we don't know how big the cache is. So guess.
4251 batch = zone->managed_pages / 1024;
4252 if (batch * PAGE_SIZE > 512 * 1024)
4253 batch = (512 * 1024) / PAGE_SIZE;
4254 batch /= 4; /* We effectively *= 4 below */
4259 * Clamp the batch to a 2^n - 1 value. Having a power
4260 * of 2 value was found to be more likely to have
4261 * suboptimal cache aliasing properties in some cases.
4263 * For example if 2 tasks are alternately allocating
4264 * batches of pages, one task can end up with a lot
4265 * of pages of one half of the possible page colors
4266 * and the other with pages of the other colors.
4268 batch = rounddown_pow_of_two(batch + batch/2) - 1;
4273 /* The deferral and batching of frees should be suppressed under NOMMU
4276 * The problem is that NOMMU needs to be able to allocate large chunks
4277 * of contiguous memory as there's no hardware page translation to
4278 * assemble apparent contiguous memory from discontiguous pages.
4280 * Queueing large contiguous runs of pages for batching, however,
4281 * causes the pages to actually be freed in smaller chunks. As there
4282 * can be a significant delay between the individual batches being
4283 * recycled, this leads to the once large chunks of space being
4284 * fragmented and becoming unavailable for high-order allocations.
4291 * pcp->high and pcp->batch values are related and dependent on one another:
4292 * ->batch must never be higher then ->high.
4293 * The following function updates them in a safe manner without read side
4296 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4297 * those fields changing asynchronously (acording the the above rule).
4299 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4300 * outside of boot time (or some other assurance that no concurrent updaters
4303 static void pageset_update(struct per_cpu_pages *pcp, unsigned long high,
4304 unsigned long batch)
4306 /* start with a fail safe value for batch */
4310 /* Update high, then batch, in order */
4317 /* a companion to pageset_set_high() */
4318 static void pageset_set_batch(struct per_cpu_pageset *p, unsigned long batch)
4320 pageset_update(&p->pcp, 6 * batch, max(1UL, 1 * batch));
4323 static void pageset_init(struct per_cpu_pageset *p)
4325 struct per_cpu_pages *pcp;
4328 memset(p, 0, sizeof(*p));
4332 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
4333 INIT_LIST_HEAD(&pcp->lists[migratetype]);
4336 static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
4339 pageset_set_batch(p, batch);
4343 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4344 * to the value high for the pageset p.
4346 static void pageset_set_high(struct per_cpu_pageset *p,
4349 unsigned long batch = max(1UL, high / 4);
4350 if ((high / 4) > (PAGE_SHIFT * 8))
4351 batch = PAGE_SHIFT * 8;
4353 pageset_update(&p->pcp, high, batch);
4356 static void pageset_set_high_and_batch(struct zone *zone,
4357 struct per_cpu_pageset *pcp)
4359 if (percpu_pagelist_fraction)
4360 pageset_set_high(pcp,
4361 (zone->managed_pages /
4362 percpu_pagelist_fraction));
4364 pageset_set_batch(pcp, zone_batchsize(zone));
4367 static void __meminit zone_pageset_init(struct zone *zone, int cpu)
4369 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
4372 pageset_set_high_and_batch(zone, pcp);
4375 static void __meminit setup_zone_pageset(struct zone *zone)
4378 zone->pageset = alloc_percpu(struct per_cpu_pageset);
4379 for_each_possible_cpu(cpu)
4380 zone_pageset_init(zone, cpu);
4384 * Allocate per cpu pagesets and initialize them.
4385 * Before this call only boot pagesets were available.
4387 void __init setup_per_cpu_pageset(void)
4391 for_each_populated_zone(zone)
4392 setup_zone_pageset(zone);
4395 static noinline __init_refok
4396 int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
4402 * The per-page waitqueue mechanism uses hashed waitqueues
4405 zone->wait_table_hash_nr_entries =
4406 wait_table_hash_nr_entries(zone_size_pages);
4407 zone->wait_table_bits =
4408 wait_table_bits(zone->wait_table_hash_nr_entries);
4409 alloc_size = zone->wait_table_hash_nr_entries
4410 * sizeof(wait_queue_head_t);
4412 if (!slab_is_available()) {
4413 zone->wait_table = (wait_queue_head_t *)
4414 memblock_virt_alloc_node_nopanic(
4415 alloc_size, zone->zone_pgdat->node_id);
4418 * This case means that a zone whose size was 0 gets new memory
4419 * via memory hot-add.
4420 * But it may be the case that a new node was hot-added. In
4421 * this case vmalloc() will not be able to use this new node's
4422 * memory - this wait_table must be initialized to use this new
4423 * node itself as well.
4424 * To use this new node's memory, further consideration will be
4427 zone->wait_table = vmalloc(alloc_size);
4429 if (!zone->wait_table)
4432 for (i = 0; i < zone->wait_table_hash_nr_entries; ++i)
4433 init_waitqueue_head(zone->wait_table + i);
4438 static __meminit void zone_pcp_init(struct zone *zone)
4441 * per cpu subsystem is not up at this point. The following code
4442 * relies on the ability of the linker to provide the
4443 * offset of a (static) per cpu variable into the per cpu area.
4445 zone->pageset = &boot_pageset;
4447 if (populated_zone(zone))
4448 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
4449 zone->name, zone->present_pages,
4450 zone_batchsize(zone));
4453 int __meminit init_currently_empty_zone(struct zone *zone,
4454 unsigned long zone_start_pfn,
4456 enum memmap_context context)
4458 struct pglist_data *pgdat = zone->zone_pgdat;
4460 ret = zone_wait_table_init(zone, size);
4463 pgdat->nr_zones = zone_idx(zone) + 1;
4465 zone->zone_start_pfn = zone_start_pfn;
4467 mminit_dprintk(MMINIT_TRACE, "memmap_init",
4468 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4470 (unsigned long)zone_idx(zone),
4471 zone_start_pfn, (zone_start_pfn + size));
4473 zone_init_free_lists(zone);
4478 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4479 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4481 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4483 int __meminit __early_pfn_to_nid(unsigned long pfn)
4485 unsigned long start_pfn, end_pfn;
4488 * NOTE: The following SMP-unsafe globals are only used early in boot
4489 * when the kernel is running single-threaded.
4491 static unsigned long __meminitdata last_start_pfn, last_end_pfn;
4492 static int __meminitdata last_nid;
4494 if (last_start_pfn <= pfn && pfn < last_end_pfn)
4497 nid = memblock_search_pfn_nid(pfn, &start_pfn, &end_pfn);
4499 last_start_pfn = start_pfn;
4500 last_end_pfn = end_pfn;
4506 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4508 int __meminit early_pfn_to_nid(unsigned long pfn)
4512 nid = __early_pfn_to_nid(pfn);
4515 /* just returns 0 */
4519 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4520 bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
4524 nid = __early_pfn_to_nid(pfn);
4525 if (nid >= 0 && nid != node)
4532 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4533 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4534 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4536 * If an architecture guarantees that all ranges registered contain no holes
4537 * and may be freed, this this function may be used instead of calling
4538 * memblock_free_early_nid() manually.
4540 void __init free_bootmem_with_active_regions(int nid, unsigned long max_low_pfn)
4542 unsigned long start_pfn, end_pfn;
4545 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid) {
4546 start_pfn = min(start_pfn, max_low_pfn);
4547 end_pfn = min(end_pfn, max_low_pfn);
4549 if (start_pfn < end_pfn)
4550 memblock_free_early_nid(PFN_PHYS(start_pfn),
4551 (end_pfn - start_pfn) << PAGE_SHIFT,
4557 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4558 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4560 * If an architecture guarantees that all ranges registered contain no holes and may
4561 * be freed, this function may be used instead of calling memory_present() manually.
4563 void __init sparse_memory_present_with_active_regions(int nid)
4565 unsigned long start_pfn, end_pfn;
4568 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, &this_nid)
4569 memory_present(this_nid, start_pfn, end_pfn);
4573 * get_pfn_range_for_nid - Return the start and end page frames for a node
4574 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4575 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4576 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4578 * It returns the start and end page frame of a node based on information
4579 * provided by memblock_set_node(). If called for a node
4580 * with no available memory, a warning is printed and the start and end
4583 void __meminit get_pfn_range_for_nid(unsigned int nid,
4584 unsigned long *start_pfn, unsigned long *end_pfn)
4586 unsigned long this_start_pfn, this_end_pfn;
4592 for_each_mem_pfn_range(i, nid, &this_start_pfn, &this_end_pfn, NULL) {
4593 *start_pfn = min(*start_pfn, this_start_pfn);
4594 *end_pfn = max(*end_pfn, this_end_pfn);
4597 if (*start_pfn == -1UL)
4602 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4603 * assumption is made that zones within a node are ordered in monotonic
4604 * increasing memory addresses so that the "highest" populated zone is used
4606 static void __init find_usable_zone_for_movable(void)
4609 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
4610 if (zone_index == ZONE_MOVABLE)
4613 if (arch_zone_highest_possible_pfn[zone_index] >
4614 arch_zone_lowest_possible_pfn[zone_index])
4618 VM_BUG_ON(zone_index == -1);
4619 movable_zone = zone_index;
4623 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4624 * because it is sized independent of architecture. Unlike the other zones,
4625 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4626 * in each node depending on the size of each node and how evenly kernelcore
4627 * is distributed. This helper function adjusts the zone ranges
4628 * provided by the architecture for a given node by using the end of the
4629 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4630 * zones within a node are in order of monotonic increases memory addresses
4632 static void __meminit adjust_zone_range_for_zone_movable(int nid,
4633 unsigned long zone_type,
4634 unsigned long node_start_pfn,
4635 unsigned long node_end_pfn,
4636 unsigned long *zone_start_pfn,
4637 unsigned long *zone_end_pfn)
4639 /* Only adjust if ZONE_MOVABLE is on this node */
4640 if (zone_movable_pfn[nid]) {
4641 /* Size ZONE_MOVABLE */
4642 if (zone_type == ZONE_MOVABLE) {
4643 *zone_start_pfn = zone_movable_pfn[nid];
4644 *zone_end_pfn = min(node_end_pfn,
4645 arch_zone_highest_possible_pfn[movable_zone]);
4647 /* Adjust for ZONE_MOVABLE starting within this range */
4648 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4649 *zone_end_pfn > zone_movable_pfn[nid]) {
4650 *zone_end_pfn = zone_movable_pfn[nid];
4652 /* Check if this whole range is within ZONE_MOVABLE */
4653 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4654 *zone_start_pfn = *zone_end_pfn;
4659 * Return the number of pages a zone spans in a node, including holes
4660 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4662 static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4663 unsigned long zone_type,
4664 unsigned long node_start_pfn,
4665 unsigned long node_end_pfn,
4666 unsigned long *ignored)
4668 unsigned long zone_start_pfn, zone_end_pfn;
4670 /* Get the start and end of the zone */
4671 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4672 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4673 adjust_zone_range_for_zone_movable(nid, zone_type,
4674 node_start_pfn, node_end_pfn,
4675 &zone_start_pfn, &zone_end_pfn);
4677 /* Check that this node has pages within the zone's required range */
4678 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4681 /* Move the zone boundaries inside the node if necessary */
4682 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4683 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4685 /* Return the spanned pages */
4686 return zone_end_pfn - zone_start_pfn;
4690 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4691 * then all holes in the requested range will be accounted for.
4693 unsigned long __meminit __absent_pages_in_range(int nid,
4694 unsigned long range_start_pfn,
4695 unsigned long range_end_pfn)
4697 unsigned long nr_absent = range_end_pfn - range_start_pfn;
4698 unsigned long start_pfn, end_pfn;
4701 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
4702 start_pfn = clamp(start_pfn, range_start_pfn, range_end_pfn);
4703 end_pfn = clamp(end_pfn, range_start_pfn, range_end_pfn);
4704 nr_absent -= end_pfn - start_pfn;
4710 * absent_pages_in_range - Return number of page frames in holes within a range
4711 * @start_pfn: The start PFN to start searching for holes
4712 * @end_pfn: The end PFN to stop searching for holes
4714 * It returns the number of pages frames in memory holes within a range.
4716 unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4717 unsigned long end_pfn)
4719 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4722 /* Return the number of page frames in holes in a zone on a node */
4723 static unsigned long __meminit zone_absent_pages_in_node(int nid,
4724 unsigned long zone_type,
4725 unsigned long node_start_pfn,
4726 unsigned long node_end_pfn,
4727 unsigned long *ignored)
4729 unsigned long zone_low = arch_zone_lowest_possible_pfn[zone_type];
4730 unsigned long zone_high = arch_zone_highest_possible_pfn[zone_type];
4731 unsigned long zone_start_pfn, zone_end_pfn;
4733 zone_start_pfn = clamp(node_start_pfn, zone_low, zone_high);
4734 zone_end_pfn = clamp(node_end_pfn, zone_low, zone_high);
4736 adjust_zone_range_for_zone_movable(nid, zone_type,
4737 node_start_pfn, node_end_pfn,
4738 &zone_start_pfn, &zone_end_pfn);
4739 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4742 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4743 static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4744 unsigned long zone_type,
4745 unsigned long node_start_pfn,
4746 unsigned long node_end_pfn,
4747 unsigned long *zones_size)
4749 return zones_size[zone_type];
4752 static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4753 unsigned long zone_type,
4754 unsigned long node_start_pfn,
4755 unsigned long node_end_pfn,
4756 unsigned long *zholes_size)
4761 return zholes_size[zone_type];
4764 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4766 static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4767 unsigned long node_start_pfn,
4768 unsigned long node_end_pfn,
4769 unsigned long *zones_size,
4770 unsigned long *zholes_size)
4772 unsigned long realtotalpages, totalpages = 0;
4775 for (i = 0; i < MAX_NR_ZONES; i++)
4776 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4780 pgdat->node_spanned_pages = totalpages;
4782 realtotalpages = totalpages;
4783 for (i = 0; i < MAX_NR_ZONES; i++)
4785 zone_absent_pages_in_node(pgdat->node_id, i,
4786 node_start_pfn, node_end_pfn,
4788 pgdat->node_present_pages = realtotalpages;
4789 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4793 #ifndef CONFIG_SPARSEMEM
4795 * Calculate the size of the zone->blockflags rounded to an unsigned long
4796 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4797 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4798 * round what is now in bits to nearest long in bits, then return it in
4801 static unsigned long __init usemap_size(unsigned long zone_start_pfn, unsigned long zonesize)
4803 unsigned long usemapsize;
4805 zonesize += zone_start_pfn & (pageblock_nr_pages-1);
4806 usemapsize = roundup(zonesize, pageblock_nr_pages);
4807 usemapsize = usemapsize >> pageblock_order;
4808 usemapsize *= NR_PAGEBLOCK_BITS;
4809 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4811 return usemapsize / 8;
4814 static void __init setup_usemap(struct pglist_data *pgdat,
4816 unsigned long zone_start_pfn,
4817 unsigned long zonesize)
4819 unsigned long usemapsize = usemap_size(zone_start_pfn, zonesize);
4820 zone->pageblock_flags = NULL;
4822 zone->pageblock_flags =
4823 memblock_virt_alloc_node_nopanic(usemapsize,
4827 static inline void setup_usemap(struct pglist_data *pgdat, struct zone *zone,
4828 unsigned long zone_start_pfn, unsigned long zonesize) {}
4829 #endif /* CONFIG_SPARSEMEM */
4831 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4833 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4834 void __paginginit set_pageblock_order(void)
4838 /* Check that pageblock_nr_pages has not already been setup */
4839 if (pageblock_order)
4842 if (HPAGE_SHIFT > PAGE_SHIFT)
4843 order = HUGETLB_PAGE_ORDER;
4845 order = MAX_ORDER - 1;
4848 * Assume the largest contiguous order of interest is a huge page.
4849 * This value may be variable depending on boot parameters on IA64 and
4852 pageblock_order = order;
4854 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4857 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4858 * is unused as pageblock_order is set at compile-time. See
4859 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4862 void __paginginit set_pageblock_order(void)
4866 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4868 static unsigned long __paginginit calc_memmap_size(unsigned long spanned_pages,
4869 unsigned long present_pages)
4871 unsigned long pages = spanned_pages;
4874 * Provide a more accurate estimation if there are holes within
4875 * the zone and SPARSEMEM is in use. If there are holes within the
4876 * zone, each populated memory region may cost us one or two extra
4877 * memmap pages due to alignment because memmap pages for each
4878 * populated regions may not naturally algined on page boundary.
4879 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4881 if (spanned_pages > present_pages + (present_pages >> 4) &&
4882 IS_ENABLED(CONFIG_SPARSEMEM))
4883 pages = present_pages;
4885 return PAGE_ALIGN(pages * sizeof(struct page)) >> PAGE_SHIFT;
4889 * Set up the zone data structures:
4890 * - mark all pages reserved
4891 * - mark all memory queues empty
4892 * - clear the memory bitmaps
4894 * NOTE: pgdat should get zeroed by caller.
4896 static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4897 unsigned long node_start_pfn, unsigned long node_end_pfn,
4898 unsigned long *zones_size, unsigned long *zholes_size)
4901 int nid = pgdat->node_id;
4902 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4905 pgdat_resize_init(pgdat);
4906 #ifdef CONFIG_NUMA_BALANCING
4907 spin_lock_init(&pgdat->numabalancing_migrate_lock);
4908 pgdat->numabalancing_migrate_nr_pages = 0;
4909 pgdat->numabalancing_migrate_next_window = jiffies;
4911 init_waitqueue_head(&pgdat->kswapd_wait);
4912 init_waitqueue_head(&pgdat->pfmemalloc_wait);
4913 pgdat_page_ext_init(pgdat);
4915 for (j = 0; j < MAX_NR_ZONES; j++) {
4916 struct zone *zone = pgdat->node_zones + j;
4917 unsigned long size, realsize, freesize, memmap_pages;
4919 size = zone_spanned_pages_in_node(nid, j, node_start_pfn,
4920 node_end_pfn, zones_size);
4921 realsize = freesize = size - zone_absent_pages_in_node(nid, j,
4927 * Adjust freesize so that it accounts for how much memory
4928 * is used by this zone for memmap. This affects the watermark
4929 * and per-cpu initialisations
4931 memmap_pages = calc_memmap_size(size, realsize);
4932 if (!is_highmem_idx(j)) {
4933 if (freesize >= memmap_pages) {
4934 freesize -= memmap_pages;
4937 " %s zone: %lu pages used for memmap\n",
4938 zone_names[j], memmap_pages);
4941 " %s zone: %lu pages exceeds freesize %lu\n",
4942 zone_names[j], memmap_pages, freesize);
4945 /* Account for reserved pages */
4946 if (j == 0 && freesize > dma_reserve) {
4947 freesize -= dma_reserve;
4948 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4949 zone_names[0], dma_reserve);
4952 if (!is_highmem_idx(j))
4953 nr_kernel_pages += freesize;
4954 /* Charge for highmem memmap if there are enough kernel pages */
4955 else if (nr_kernel_pages > memmap_pages * 2)
4956 nr_kernel_pages -= memmap_pages;
4957 nr_all_pages += freesize;
4959 zone->spanned_pages = size;
4960 zone->present_pages = realsize;
4962 * Set an approximate value for lowmem here, it will be adjusted
4963 * when the bootmem allocator frees pages into the buddy system.
4964 * And all highmem pages will be managed by the buddy system.
4966 zone->managed_pages = is_highmem_idx(j) ? realsize : freesize;
4969 zone->min_unmapped_pages = (freesize*sysctl_min_unmapped_ratio)
4971 zone->min_slab_pages = (freesize * sysctl_min_slab_ratio) / 100;
4973 zone->name = zone_names[j];
4974 spin_lock_init(&zone->lock);
4975 spin_lock_init(&zone->lru_lock);
4976 zone_seqlock_init(zone);
4977 zone->zone_pgdat = pgdat;
4978 zone_pcp_init(zone);
4980 /* For bootup, initialized properly in watermark setup */
4981 mod_zone_page_state(zone, NR_ALLOC_BATCH, zone->managed_pages);
4983 lruvec_init(&zone->lruvec);
4987 set_pageblock_order();
4988 setup_usemap(pgdat, zone, zone_start_pfn, size);
4989 ret = init_currently_empty_zone(zone, zone_start_pfn,
4990 size, MEMMAP_EARLY);
4992 memmap_init(size, nid, j, zone_start_pfn);
4993 zone_start_pfn += size;
4997 static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4999 /* Skip empty nodes */
5000 if (!pgdat->node_spanned_pages)
5003 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5004 /* ia64 gets its own node_mem_map, before this, without bootmem */
5005 if (!pgdat->node_mem_map) {
5006 unsigned long size, start, end;
5010 * The zone's endpoints aren't required to be MAX_ORDER
5011 * aligned but the node_mem_map endpoints must be in order
5012 * for the buddy allocator to function correctly.
5014 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
5015 end = pgdat_end_pfn(pgdat);
5016 end = ALIGN(end, MAX_ORDER_NR_PAGES);
5017 size = (end - start) * sizeof(struct page);
5018 map = alloc_remap(pgdat->node_id, size);
5020 map = memblock_virt_alloc_node_nopanic(size,
5022 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
5024 #ifndef CONFIG_NEED_MULTIPLE_NODES
5026 * With no DISCONTIG, the global mem_map is just set as node 0's
5028 if (pgdat == NODE_DATA(0)) {
5029 mem_map = NODE_DATA(0)->node_mem_map;
5030 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5031 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
5032 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
5033 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5036 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5039 void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
5040 unsigned long node_start_pfn, unsigned long *zholes_size)
5042 pg_data_t *pgdat = NODE_DATA(nid);
5043 unsigned long start_pfn = 0;
5044 unsigned long end_pfn = 0;
5046 /* pg_data_t should be reset to zero when it's allocated */
5047 WARN_ON(pgdat->nr_zones || pgdat->classzone_idx);
5049 pgdat->node_id = nid;
5050 pgdat->node_start_pfn = node_start_pfn;
5051 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5052 get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
5053 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid,
5054 (u64)start_pfn << PAGE_SHIFT, ((u64)end_pfn << PAGE_SHIFT) - 1);
5056 calculate_node_totalpages(pgdat, start_pfn, end_pfn,
5057 zones_size, zholes_size);
5059 alloc_node_mem_map(pgdat);
5060 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5061 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5062 nid, (unsigned long)pgdat,
5063 (unsigned long)pgdat->node_mem_map);
5066 free_area_init_core(pgdat, start_pfn, end_pfn,
5067 zones_size, zholes_size);
5070 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5072 #if MAX_NUMNODES > 1
5074 * Figure out the number of possible node ids.
5076 void __init setup_nr_node_ids(void)
5079 unsigned int highest = 0;
5081 for_each_node_mask(node, node_possible_map)
5083 nr_node_ids = highest + 1;
5088 * node_map_pfn_alignment - determine the maximum internode alignment
5090 * This function should be called after node map is populated and sorted.
5091 * It calculates the maximum power of two alignment which can distinguish
5094 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5095 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5096 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5097 * shifted, 1GiB is enough and this function will indicate so.
5099 * This is used to test whether pfn -> nid mapping of the chosen memory
5100 * model has fine enough granularity to avoid incorrect mapping for the
5101 * populated node map.
5103 * Returns the determined alignment in pfn's. 0 if there is no alignment
5104 * requirement (single node).
5106 unsigned long __init node_map_pfn_alignment(void)
5108 unsigned long accl_mask = 0, last_end = 0;
5109 unsigned long start, end, mask;
5113 for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid) {
5114 if (!start || last_nid < 0 || last_nid == nid) {
5121 * Start with a mask granular enough to pin-point to the
5122 * start pfn and tick off bits one-by-one until it becomes
5123 * too coarse to separate the current node from the last.
5125 mask = ~((1 << __ffs(start)) - 1);
5126 while (mask && last_end <= (start & (mask << 1)))
5129 /* accumulate all internode masks */
5133 /* convert mask to number of pages */
5134 return ~accl_mask + 1;
5137 /* Find the lowest pfn for a node */
5138 static unsigned long __init find_min_pfn_for_node(int nid)
5140 unsigned long min_pfn = ULONG_MAX;
5141 unsigned long start_pfn;
5144 for_each_mem_pfn_range(i, nid, &start_pfn, NULL, NULL)
5145 min_pfn = min(min_pfn, start_pfn);
5147 if (min_pfn == ULONG_MAX) {
5149 "Could not find start_pfn for node %d\n", nid);
5157 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5159 * It returns the minimum PFN based on information provided via
5160 * memblock_set_node().
5162 unsigned long __init find_min_pfn_with_active_regions(void)
5164 return find_min_pfn_for_node(MAX_NUMNODES);
5168 * early_calculate_totalpages()
5169 * Sum pages in active regions for movable zone.
5170 * Populate N_MEMORY for calculating usable_nodes.
5172 static unsigned long __init early_calculate_totalpages(void)
5174 unsigned long totalpages = 0;
5175 unsigned long start_pfn, end_pfn;
5178 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid) {
5179 unsigned long pages = end_pfn - start_pfn;
5181 totalpages += pages;
5183 node_set_state(nid, N_MEMORY);
5189 * Find the PFN the Movable zone begins in each node. Kernel memory
5190 * is spread evenly between nodes as long as the nodes have enough
5191 * memory. When they don't, some nodes will have more kernelcore than
5194 static void __init find_zone_movable_pfns_for_nodes(void)
5197 unsigned long usable_startpfn;
5198 unsigned long kernelcore_node, kernelcore_remaining;
5199 /* save the state before borrow the nodemask */
5200 nodemask_t saved_node_state = node_states[N_MEMORY];
5201 unsigned long totalpages = early_calculate_totalpages();
5202 int usable_nodes = nodes_weight(node_states[N_MEMORY]);
5203 struct memblock_region *r;
5205 /* Need to find movable_zone earlier when movable_node is specified. */
5206 find_usable_zone_for_movable();
5209 * If movable_node is specified, ignore kernelcore and movablecore
5212 if (movable_node_is_enabled()) {
5213 for_each_memblock(memory, r) {
5214 if (!memblock_is_hotpluggable(r))
5219 usable_startpfn = PFN_DOWN(r->base);
5220 zone_movable_pfn[nid] = zone_movable_pfn[nid] ?
5221 min(usable_startpfn, zone_movable_pfn[nid]) :
5229 * If movablecore=nn[KMG] was specified, calculate what size of
5230 * kernelcore that corresponds so that memory usable for
5231 * any allocation type is evenly spread. If both kernelcore
5232 * and movablecore are specified, then the value of kernelcore
5233 * will be used for required_kernelcore if it's greater than
5234 * what movablecore would have allowed.
5236 if (required_movablecore) {
5237 unsigned long corepages;
5240 * Round-up so that ZONE_MOVABLE is at least as large as what
5241 * was requested by the user
5243 required_movablecore =
5244 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
5245 corepages = totalpages - required_movablecore;
5247 required_kernelcore = max(required_kernelcore, corepages);
5250 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5251 if (!required_kernelcore)
5254 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5255 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
5258 /* Spread kernelcore memory as evenly as possible throughout nodes */
5259 kernelcore_node = required_kernelcore / usable_nodes;
5260 for_each_node_state(nid, N_MEMORY) {
5261 unsigned long start_pfn, end_pfn;
5264 * Recalculate kernelcore_node if the division per node
5265 * now exceeds what is necessary to satisfy the requested
5266 * amount of memory for the kernel
5268 if (required_kernelcore < kernelcore_node)
5269 kernelcore_node = required_kernelcore / usable_nodes;
5272 * As the map is walked, we track how much memory is usable
5273 * by the kernel using kernelcore_remaining. When it is
5274 * 0, the rest of the node is usable by ZONE_MOVABLE
5276 kernelcore_remaining = kernelcore_node;
5278 /* Go through each range of PFNs within this node */
5279 for_each_mem_pfn_range(i, nid, &start_pfn, &end_pfn, NULL) {
5280 unsigned long size_pages;
5282 start_pfn = max(start_pfn, zone_movable_pfn[nid]);
5283 if (start_pfn >= end_pfn)
5286 /* Account for what is only usable for kernelcore */
5287 if (start_pfn < usable_startpfn) {
5288 unsigned long kernel_pages;
5289 kernel_pages = min(end_pfn, usable_startpfn)
5292 kernelcore_remaining -= min(kernel_pages,
5293 kernelcore_remaining);
5294 required_kernelcore -= min(kernel_pages,
5295 required_kernelcore);
5297 /* Continue if range is now fully accounted */
5298 if (end_pfn <= usable_startpfn) {
5301 * Push zone_movable_pfn to the end so
5302 * that if we have to rebalance
5303 * kernelcore across nodes, we will
5304 * not double account here
5306 zone_movable_pfn[nid] = end_pfn;
5309 start_pfn = usable_startpfn;
5313 * The usable PFN range for ZONE_MOVABLE is from
5314 * start_pfn->end_pfn. Calculate size_pages as the
5315 * number of pages used as kernelcore
5317 size_pages = end_pfn - start_pfn;
5318 if (size_pages > kernelcore_remaining)
5319 size_pages = kernelcore_remaining;
5320 zone_movable_pfn[nid] = start_pfn + size_pages;
5323 * Some kernelcore has been met, update counts and
5324 * break if the kernelcore for this node has been
5327 required_kernelcore -= min(required_kernelcore,
5329 kernelcore_remaining -= size_pages;
5330 if (!kernelcore_remaining)
5336 * If there is still required_kernelcore, we do another pass with one
5337 * less node in the count. This will push zone_movable_pfn[nid] further
5338 * along on the nodes that still have memory until kernelcore is
5342 if (usable_nodes && required_kernelcore > usable_nodes)
5346 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5347 for (nid = 0; nid < MAX_NUMNODES; nid++)
5348 zone_movable_pfn[nid] =
5349 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
5352 /* restore the node_state */
5353 node_states[N_MEMORY] = saved_node_state;
5356 /* Any regular or high memory on that node ? */
5357 static void check_for_memory(pg_data_t *pgdat, int nid)
5359 enum zone_type zone_type;
5361 if (N_MEMORY == N_NORMAL_MEMORY)
5364 for (zone_type = 0; zone_type <= ZONE_MOVABLE - 1; zone_type++) {
5365 struct zone *zone = &pgdat->node_zones[zone_type];
5366 if (populated_zone(zone)) {
5367 node_set_state(nid, N_HIGH_MEMORY);
5368 if (N_NORMAL_MEMORY != N_HIGH_MEMORY &&
5369 zone_type <= ZONE_NORMAL)
5370 node_set_state(nid, N_NORMAL_MEMORY);
5377 * free_area_init_nodes - Initialise all pg_data_t and zone data
5378 * @max_zone_pfn: an array of max PFNs for each zone
5380 * This will call free_area_init_node() for each active node in the system.
5381 * Using the page ranges provided by memblock_set_node(), the size of each
5382 * zone in each node and their holes is calculated. If the maximum PFN
5383 * between two adjacent zones match, it is assumed that the zone is empty.
5384 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5385 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5386 * starts where the previous one ended. For example, ZONE_DMA32 starts
5387 * at arch_max_dma_pfn.
5389 void __init free_area_init_nodes(unsigned long *max_zone_pfn)
5391 unsigned long start_pfn, end_pfn;
5394 /* Record where the zone boundaries are */
5395 memset(arch_zone_lowest_possible_pfn, 0,
5396 sizeof(arch_zone_lowest_possible_pfn));
5397 memset(arch_zone_highest_possible_pfn, 0,
5398 sizeof(arch_zone_highest_possible_pfn));
5399 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
5400 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
5401 for (i = 1; i < MAX_NR_ZONES; i++) {
5402 if (i == ZONE_MOVABLE)
5404 arch_zone_lowest_possible_pfn[i] =
5405 arch_zone_highest_possible_pfn[i-1];
5406 arch_zone_highest_possible_pfn[i] =
5407 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
5409 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
5410 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
5412 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5413 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
5414 find_zone_movable_pfns_for_nodes();
5416 /* Print out the zone ranges */
5417 pr_info("Zone ranges:\n");
5418 for (i = 0; i < MAX_NR_ZONES; i++) {
5419 if (i == ZONE_MOVABLE)
5421 pr_info(" %-8s ", zone_names[i]);
5422 if (arch_zone_lowest_possible_pfn[i] ==
5423 arch_zone_highest_possible_pfn[i])
5426 pr_cont("[mem %#018Lx-%#018Lx]\n",
5427 (u64)arch_zone_lowest_possible_pfn[i]
5429 ((u64)arch_zone_highest_possible_pfn[i]
5430 << PAGE_SHIFT) - 1);
5433 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5434 pr_info("Movable zone start for each node\n");
5435 for (i = 0; i < MAX_NUMNODES; i++) {
5436 if (zone_movable_pfn[i])
5437 pr_info(" Node %d: %#018Lx\n", i,
5438 (u64)zone_movable_pfn[i] << PAGE_SHIFT);
5441 /* Print out the early node map */
5442 pr_info("Early memory node ranges\n");
5443 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, &nid)
5444 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid,
5445 (u64)start_pfn << PAGE_SHIFT,
5446 ((u64)end_pfn << PAGE_SHIFT) - 1);
5448 /* Initialise every node */
5449 mminit_verify_pageflags_layout();
5450 setup_nr_node_ids();
5451 for_each_online_node(nid) {
5452 pg_data_t *pgdat = NODE_DATA(nid);
5453 free_area_init_node(nid, NULL,
5454 find_min_pfn_for_node(nid), NULL);
5456 /* Any memory on that node */
5457 if (pgdat->node_present_pages)
5458 node_set_state(nid, N_MEMORY);
5459 check_for_memory(pgdat, nid);
5463 static int __init cmdline_parse_core(char *p, unsigned long *core)
5465 unsigned long long coremem;
5469 coremem = memparse(p, &p);
5470 *core = coremem >> PAGE_SHIFT;
5472 /* Paranoid check that UL is enough for the coremem value */
5473 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
5479 * kernelcore=size sets the amount of memory for use for allocations that
5480 * cannot be reclaimed or migrated.
5482 static int __init cmdline_parse_kernelcore(char *p)
5484 return cmdline_parse_core(p, &required_kernelcore);
5488 * movablecore=size sets the amount of memory for use for allocations that
5489 * can be reclaimed or migrated.
5491 static int __init cmdline_parse_movablecore(char *p)
5493 return cmdline_parse_core(p, &required_movablecore);
5496 early_param("kernelcore", cmdline_parse_kernelcore);
5497 early_param("movablecore", cmdline_parse_movablecore);
5499 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5501 void adjust_managed_page_count(struct page *page, long count)
5503 spin_lock(&managed_page_count_lock);
5504 page_zone(page)->managed_pages += count;
5505 totalram_pages += count;
5506 #ifdef CONFIG_HIGHMEM
5507 if (PageHighMem(page))
5508 totalhigh_pages += count;
5510 spin_unlock(&managed_page_count_lock);
5512 EXPORT_SYMBOL(adjust_managed_page_count);
5514 unsigned long free_reserved_area(void *start, void *end, int poison, char *s)
5517 unsigned long pages = 0;
5519 start = (void *)PAGE_ALIGN((unsigned long)start);
5520 end = (void *)((unsigned long)end & PAGE_MASK);
5521 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5522 if ((unsigned int)poison <= 0xFF)
5523 memset(pos, poison, PAGE_SIZE);
5524 free_reserved_page(virt_to_page(pos));
5528 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5529 s, pages << (PAGE_SHIFT - 10), start, end);
5533 EXPORT_SYMBOL(free_reserved_area);
5535 #ifdef CONFIG_HIGHMEM
5536 void free_highmem_page(struct page *page)
5538 __free_reserved_page(page);
5540 page_zone(page)->managed_pages++;
5546 void __init mem_init_print_info(const char *str)
5548 unsigned long physpages, codesize, datasize, rosize, bss_size;
5549 unsigned long init_code_size, init_data_size;
5551 physpages = get_num_physpages();
5552 codesize = _etext - _stext;
5553 datasize = _edata - _sdata;
5554 rosize = __end_rodata - __start_rodata;
5555 bss_size = __bss_stop - __bss_start;
5556 init_data_size = __init_end - __init_begin;
5557 init_code_size = _einittext - _sinittext;
5560 * Detect special cases and adjust section sizes accordingly:
5561 * 1) .init.* may be embedded into .data sections
5562 * 2) .init.text.* may be out of [__init_begin, __init_end],
5563 * please refer to arch/tile/kernel/vmlinux.lds.S.
5564 * 3) .rodata.* may be embedded into .text or .data sections.
5566 #define adj_init_size(start, end, size, pos, adj) \
5568 if (start <= pos && pos < end && size > adj) \
5572 adj_init_size(__init_begin, __init_end, init_data_size,
5573 _sinittext, init_code_size);
5574 adj_init_size(_stext, _etext, codesize, _sinittext, init_code_size);
5575 adj_init_size(_sdata, _edata, datasize, __init_begin, init_data_size);
5576 adj_init_size(_stext, _etext, codesize, __start_rodata, rosize);
5577 adj_init_size(_sdata, _edata, datasize, __start_rodata, rosize);
5579 #undef adj_init_size
5581 pr_info("Memory: %luK/%luK available "
5582 "(%luK kernel code, %luK rwdata, %luK rodata, "
5583 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5584 #ifdef CONFIG_HIGHMEM
5588 nr_free_pages() << (PAGE_SHIFT-10), physpages << (PAGE_SHIFT-10),
5589 codesize >> 10, datasize >> 10, rosize >> 10,
5590 (init_data_size + init_code_size) >> 10, bss_size >> 10,
5591 (physpages - totalram_pages - totalcma_pages) << (PAGE_SHIFT-10),
5592 totalcma_pages << (PAGE_SHIFT-10),
5593 #ifdef CONFIG_HIGHMEM
5594 totalhigh_pages << (PAGE_SHIFT-10),
5596 str ? ", " : "", str ? str : "");
5600 * set_dma_reserve - set the specified number of pages reserved in the first zone
5601 * @new_dma_reserve: The number of pages to mark reserved
5603 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5604 * In the DMA zone, a significant percentage may be consumed by kernel image
5605 * and other unfreeable allocations which can skew the watermarks badly. This
5606 * function may optionally be used to account for unfreeable pages in the
5607 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5608 * smaller per-cpu batchsize.
5610 void __init set_dma_reserve(unsigned long new_dma_reserve)
5612 dma_reserve = new_dma_reserve;
5615 void __init free_area_init(unsigned long *zones_size)
5617 free_area_init_node(0, zones_size,
5618 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5621 static int page_alloc_cpu_notify(struct notifier_block *self,
5622 unsigned long action, void *hcpu)
5624 int cpu = (unsigned long)hcpu;
5626 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5627 lru_add_drain_cpu(cpu);
5631 * Spill the event counters of the dead processor
5632 * into the current processors event counters.
5633 * This artificially elevates the count of the current
5636 vm_events_fold_cpu(cpu);
5639 * Zero the differential counters of the dead processor
5640 * so that the vm statistics are consistent.
5642 * This is only okay since the processor is dead and cannot
5643 * race with what we are doing.
5645 cpu_vm_stats_fold(cpu);
5650 void __init page_alloc_init(void)
5652 hotcpu_notifier(page_alloc_cpu_notify, 0);
5656 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5657 * or min_free_kbytes changes.
5659 static void calculate_totalreserve_pages(void)
5661 struct pglist_data *pgdat;
5662 unsigned long reserve_pages = 0;
5663 enum zone_type i, j;
5665 for_each_online_pgdat(pgdat) {
5666 for (i = 0; i < MAX_NR_ZONES; i++) {
5667 struct zone *zone = pgdat->node_zones + i;
5670 /* Find valid and maximum lowmem_reserve in the zone */
5671 for (j = i; j < MAX_NR_ZONES; j++) {
5672 if (zone->lowmem_reserve[j] > max)
5673 max = zone->lowmem_reserve[j];
5676 /* we treat the high watermark as reserved pages. */
5677 max += high_wmark_pages(zone);
5679 if (max > zone->managed_pages)
5680 max = zone->managed_pages;
5681 reserve_pages += max;
5683 * Lowmem reserves are not available to
5684 * GFP_HIGHUSER page cache allocations and
5685 * kswapd tries to balance zones to their high
5686 * watermark. As a result, neither should be
5687 * regarded as dirtyable memory, to prevent a
5688 * situation where reclaim has to clean pages
5689 * in order to balance the zones.
5691 zone->dirty_balance_reserve = max;
5694 dirty_balance_reserve = reserve_pages;
5695 totalreserve_pages = reserve_pages;
5699 * setup_per_zone_lowmem_reserve - called whenever
5700 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5701 * has a correct pages reserved value, so an adequate number of
5702 * pages are left in the zone after a successful __alloc_pages().
5704 static void setup_per_zone_lowmem_reserve(void)
5706 struct pglist_data *pgdat;
5707 enum zone_type j, idx;
5709 for_each_online_pgdat(pgdat) {
5710 for (j = 0; j < MAX_NR_ZONES; j++) {
5711 struct zone *zone = pgdat->node_zones + j;
5712 unsigned long managed_pages = zone->managed_pages;
5714 zone->lowmem_reserve[j] = 0;
5718 struct zone *lower_zone;
5722 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5723 sysctl_lowmem_reserve_ratio[idx] = 1;
5725 lower_zone = pgdat->node_zones + idx;
5726 lower_zone->lowmem_reserve[j] = managed_pages /
5727 sysctl_lowmem_reserve_ratio[idx];
5728 managed_pages += lower_zone->managed_pages;
5733 /* update totalreserve_pages */
5734 calculate_totalreserve_pages();
5737 static void __setup_per_zone_wmarks(void)
5739 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5740 unsigned long lowmem_pages = 0;
5742 unsigned long flags;
5744 /* Calculate total number of !ZONE_HIGHMEM pages */
5745 for_each_zone(zone) {
5746 if (!is_highmem(zone))
5747 lowmem_pages += zone->managed_pages;
5750 for_each_zone(zone) {
5753 spin_lock_irqsave(&zone->lock, flags);
5754 tmp = (u64)pages_min * zone->managed_pages;
5755 do_div(tmp, lowmem_pages);
5756 if (is_highmem(zone)) {
5758 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5759 * need highmem pages, so cap pages_min to a small
5762 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5763 * deltas control asynch page reclaim, and so should
5764 * not be capped for highmem.
5766 unsigned long min_pages;
5768 min_pages = zone->managed_pages / 1024;
5769 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5770 zone->watermark[WMARK_MIN] = min_pages;
5773 * If it's a lowmem zone, reserve a number of pages
5774 * proportionate to the zone's size.
5776 zone->watermark[WMARK_MIN] = tmp;
5779 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5780 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5782 __mod_zone_page_state(zone, NR_ALLOC_BATCH,
5783 high_wmark_pages(zone) - low_wmark_pages(zone) -
5784 atomic_long_read(&zone->vm_stat[NR_ALLOC_BATCH]));
5786 setup_zone_migrate_reserve(zone);
5787 spin_unlock_irqrestore(&zone->lock, flags);
5790 /* update totalreserve_pages */
5791 calculate_totalreserve_pages();
5795 * setup_per_zone_wmarks - called when min_free_kbytes changes
5796 * or when memory is hot-{added|removed}
5798 * Ensures that the watermark[min,low,high] values for each zone are set
5799 * correctly with respect to min_free_kbytes.
5801 void setup_per_zone_wmarks(void)
5803 mutex_lock(&zonelists_mutex);
5804 __setup_per_zone_wmarks();
5805 mutex_unlock(&zonelists_mutex);
5809 * The inactive anon list should be small enough that the VM never has to
5810 * do too much work, but large enough that each inactive page has a chance
5811 * to be referenced again before it is swapped out.
5813 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5814 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5815 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5816 * the anonymous pages are kept on the inactive list.
5819 * memory ratio inactive anon
5820 * -------------------------------------
5829 static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5831 unsigned int gb, ratio;
5833 /* Zone size in gigabytes */
5834 gb = zone->managed_pages >> (30 - PAGE_SHIFT);
5836 ratio = int_sqrt(10 * gb);
5840 zone->inactive_ratio = ratio;
5843 static void __meminit setup_per_zone_inactive_ratio(void)
5848 calculate_zone_inactive_ratio(zone);
5852 * Initialise min_free_kbytes.
5854 * For small machines we want it small (128k min). For large machines
5855 * we want it large (64MB max). But it is not linear, because network
5856 * bandwidth does not increase linearly with machine size. We use
5858 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5859 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5875 int __meminit init_per_zone_wmark_min(void)
5877 unsigned long lowmem_kbytes;
5878 int new_min_free_kbytes;
5880 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5881 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5883 if (new_min_free_kbytes > user_min_free_kbytes) {
5884 min_free_kbytes = new_min_free_kbytes;
5885 if (min_free_kbytes < 128)
5886 min_free_kbytes = 128;
5887 if (min_free_kbytes > 65536)
5888 min_free_kbytes = 65536;
5890 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5891 new_min_free_kbytes, user_min_free_kbytes);
5893 setup_per_zone_wmarks();
5894 refresh_zone_stat_thresholds();
5895 setup_per_zone_lowmem_reserve();
5896 setup_per_zone_inactive_ratio();
5899 module_init(init_per_zone_wmark_min)
5902 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5903 * that we can call two helper functions whenever min_free_kbytes
5906 int min_free_kbytes_sysctl_handler(struct ctl_table *table, int write,
5907 void __user *buffer, size_t *length, loff_t *ppos)
5911 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5916 user_min_free_kbytes = min_free_kbytes;
5917 setup_per_zone_wmarks();
5923 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table *table, int write,
5924 void __user *buffer, size_t *length, loff_t *ppos)
5929 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5934 zone->min_unmapped_pages = (zone->managed_pages *
5935 sysctl_min_unmapped_ratio) / 100;
5939 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table *table, int write,
5940 void __user *buffer, size_t *length, loff_t *ppos)
5945 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5950 zone->min_slab_pages = (zone->managed_pages *
5951 sysctl_min_slab_ratio) / 100;
5957 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5958 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5959 * whenever sysctl_lowmem_reserve_ratio changes.
5961 * The reserve ratio obviously has absolutely no relation with the
5962 * minimum watermarks. The lowmem reserve ratio can only make sense
5963 * if in function of the boot time zone sizes.
5965 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table *table, int write,
5966 void __user *buffer, size_t *length, loff_t *ppos)
5968 proc_dointvec_minmax(table, write, buffer, length, ppos);
5969 setup_per_zone_lowmem_reserve();
5974 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5975 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5976 * pagelist can have before it gets flushed back to buddy allocator.
5978 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table *table, int write,
5979 void __user *buffer, size_t *length, loff_t *ppos)
5982 int old_percpu_pagelist_fraction;
5985 mutex_lock(&pcp_batch_high_lock);
5986 old_percpu_pagelist_fraction = percpu_pagelist_fraction;
5988 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5989 if (!write || ret < 0)
5992 /* Sanity checking to avoid pcp imbalance */
5993 if (percpu_pagelist_fraction &&
5994 percpu_pagelist_fraction < MIN_PERCPU_PAGELIST_FRACTION) {
5995 percpu_pagelist_fraction = old_percpu_pagelist_fraction;
6001 if (percpu_pagelist_fraction == old_percpu_pagelist_fraction)
6004 for_each_populated_zone(zone) {
6007 for_each_possible_cpu(cpu)
6008 pageset_set_high_and_batch(zone,
6009 per_cpu_ptr(zone->pageset, cpu));
6012 mutex_unlock(&pcp_batch_high_lock);
6016 int hashdist = HASHDIST_DEFAULT;
6019 static int __init set_hashdist(char *str)
6023 hashdist = simple_strtoul(str, &str, 0);
6026 __setup("hashdist=", set_hashdist);
6030 * allocate a large system hash table from bootmem
6031 * - it is assumed that the hash table must contain an exact power-of-2
6032 * quantity of entries
6033 * - limit is the number of hash buckets, not the total allocation size
6035 void *__init alloc_large_system_hash(const char *tablename,
6036 unsigned long bucketsize,
6037 unsigned long numentries,
6040 unsigned int *_hash_shift,
6041 unsigned int *_hash_mask,
6042 unsigned long low_limit,
6043 unsigned long high_limit)
6045 unsigned long long max = high_limit;
6046 unsigned long log2qty, size;
6049 /* allow the kernel cmdline to have a say */
6051 /* round applicable memory size up to nearest megabyte */
6052 numentries = nr_kernel_pages;
6054 /* It isn't necessary when PAGE_SIZE >= 1MB */
6055 if (PAGE_SHIFT < 20)
6056 numentries = round_up(numentries, (1<<20)/PAGE_SIZE);
6058 /* limit to 1 bucket per 2^scale bytes of low memory */
6059 if (scale > PAGE_SHIFT)
6060 numentries >>= (scale - PAGE_SHIFT);
6062 numentries <<= (PAGE_SHIFT - scale);
6064 /* Make sure we've got at least a 0-order allocation.. */
6065 if (unlikely(flags & HASH_SMALL)) {
6066 /* Makes no sense without HASH_EARLY */
6067 WARN_ON(!(flags & HASH_EARLY));
6068 if (!(numentries >> *_hash_shift)) {
6069 numentries = 1UL << *_hash_shift;
6070 BUG_ON(!numentries);
6072 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
6073 numentries = PAGE_SIZE / bucketsize;
6075 numentries = roundup_pow_of_two(numentries);
6077 /* limit allocation size to 1/16 total memory by default */
6079 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
6080 do_div(max, bucketsize);
6082 max = min(max, 0x80000000ULL);
6084 if (numentries < low_limit)
6085 numentries = low_limit;
6086 if (numentries > max)
6089 log2qty = ilog2(numentries);
6092 size = bucketsize << log2qty;
6093 if (flags & HASH_EARLY)
6094 table = memblock_virt_alloc_nopanic(size, 0);
6096 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
6099 * If bucketsize is not a power-of-two, we may free
6100 * some pages at the end of hash table which
6101 * alloc_pages_exact() automatically does
6103 if (get_order(size) < MAX_ORDER) {
6104 table = alloc_pages_exact(size, GFP_ATOMIC);
6105 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
6108 } while (!table && size > PAGE_SIZE && --log2qty);
6111 panic("Failed to allocate %s hash table\n", tablename);
6113 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
6116 ilog2(size) - PAGE_SHIFT,
6120 *_hash_shift = log2qty;
6122 *_hash_mask = (1 << log2qty) - 1;
6127 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6128 static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
6131 #ifdef CONFIG_SPARSEMEM
6132 return __pfn_to_section(pfn)->pageblock_flags;
6134 return zone->pageblock_flags;
6135 #endif /* CONFIG_SPARSEMEM */
6138 static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
6140 #ifdef CONFIG_SPARSEMEM
6141 pfn &= (PAGES_PER_SECTION-1);
6142 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6144 pfn = pfn - round_down(zone->zone_start_pfn, pageblock_nr_pages);
6145 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
6146 #endif /* CONFIG_SPARSEMEM */
6150 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6151 * @page: The page within the block of interest
6152 * @pfn: The target page frame number
6153 * @end_bitidx: The last bit of interest to retrieve
6154 * @mask: mask of bits that the caller is interested in
6156 * Return: pageblock_bits flags
6158 unsigned long get_pfnblock_flags_mask(struct page *page, unsigned long pfn,
6159 unsigned long end_bitidx,
6163 unsigned long *bitmap;
6164 unsigned long bitidx, word_bitidx;
6167 zone = page_zone(page);
6168 bitmap = get_pageblock_bitmap(zone, pfn);
6169 bitidx = pfn_to_bitidx(zone, pfn);
6170 word_bitidx = bitidx / BITS_PER_LONG;
6171 bitidx &= (BITS_PER_LONG-1);
6173 word = bitmap[word_bitidx];
6174 bitidx += end_bitidx;
6175 return (word >> (BITS_PER_LONG - bitidx - 1)) & mask;
6179 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6180 * @page: The page within the block of interest
6181 * @flags: The flags to set
6182 * @pfn: The target page frame number
6183 * @end_bitidx: The last bit of interest
6184 * @mask: mask of bits that the caller is interested in
6186 void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
6188 unsigned long end_bitidx,
6192 unsigned long *bitmap;
6193 unsigned long bitidx, word_bitidx;
6194 unsigned long old_word, word;
6196 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
6198 zone = page_zone(page);
6199 bitmap = get_pageblock_bitmap(zone, pfn);
6200 bitidx = pfn_to_bitidx(zone, pfn);
6201 word_bitidx = bitidx / BITS_PER_LONG;
6202 bitidx &= (BITS_PER_LONG-1);
6204 VM_BUG_ON_PAGE(!zone_spans_pfn(zone, pfn), page);
6206 bitidx += end_bitidx;
6207 mask <<= (BITS_PER_LONG - bitidx - 1);
6208 flags <<= (BITS_PER_LONG - bitidx - 1);
6210 word = READ_ONCE(bitmap[word_bitidx]);
6212 old_word = cmpxchg(&bitmap[word_bitidx], word, (word & ~mask) | flags);
6213 if (word == old_word)
6220 * This function checks whether pageblock includes unmovable pages or not.
6221 * If @count is not zero, it is okay to include less @count unmovable pages
6223 * PageLRU check without isolation or lru_lock could race so that
6224 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6225 * expect this function should be exact.
6227 bool has_unmovable_pages(struct zone *zone, struct page *page, int count,
6228 bool skip_hwpoisoned_pages)
6230 unsigned long pfn, iter, found;
6234 * For avoiding noise data, lru_add_drain_all() should be called
6235 * If ZONE_MOVABLE, the zone never contains unmovable pages
6237 if (zone_idx(zone) == ZONE_MOVABLE)
6239 mt = get_pageblock_migratetype(page);
6240 if (mt == MIGRATE_MOVABLE || is_migrate_cma(mt))
6243 pfn = page_to_pfn(page);
6244 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
6245 unsigned long check = pfn + iter;
6247 if (!pfn_valid_within(check))
6250 page = pfn_to_page(check);
6253 * Hugepages are not in LRU lists, but they're movable.
6254 * We need not scan over tail pages bacause we don't
6255 * handle each tail page individually in migration.
6257 if (PageHuge(page)) {
6258 iter = round_up(iter + 1, 1<<compound_order(page)) - 1;
6263 * We can't use page_count without pin a page
6264 * because another CPU can free compound page.
6265 * This check already skips compound tails of THP
6266 * because their page->_count is zero at all time.
6268 if (!atomic_read(&page->_count)) {
6269 if (PageBuddy(page))
6270 iter += (1 << page_order(page)) - 1;
6275 * The HWPoisoned page may be not in buddy system, and
6276 * page_count() is not 0.
6278 if (skip_hwpoisoned_pages && PageHWPoison(page))
6284 * If there are RECLAIMABLE pages, we need to check
6285 * it. But now, memory offline itself doesn't call
6286 * shrink_node_slabs() and it still to be fixed.
6289 * If the page is not RAM, page_count()should be 0.
6290 * we don't need more check. This is an _used_ not-movable page.
6292 * The problematic thing here is PG_reserved pages. PG_reserved
6293 * is set to both of a memory hole page and a _used_ kernel
6302 bool is_pageblock_removable_nolock(struct page *page)
6308 * We have to be careful here because we are iterating over memory
6309 * sections which are not zone aware so we might end up outside of
6310 * the zone but still within the section.
6311 * We have to take care about the node as well. If the node is offline
6312 * its NODE_DATA will be NULL - see page_zone.
6314 if (!node_online(page_to_nid(page)))
6317 zone = page_zone(page);
6318 pfn = page_to_pfn(page);
6319 if (!zone_spans_pfn(zone, pfn))
6322 return !has_unmovable_pages(zone, page, 0, true);
6327 static unsigned long pfn_max_align_down(unsigned long pfn)
6329 return pfn & ~(max_t(unsigned long, MAX_ORDER_NR_PAGES,
6330 pageblock_nr_pages) - 1);
6333 static unsigned long pfn_max_align_up(unsigned long pfn)
6335 return ALIGN(pfn, max_t(unsigned long, MAX_ORDER_NR_PAGES,
6336 pageblock_nr_pages));
6339 /* [start, end) must belong to a single zone. */
6340 static int __alloc_contig_migrate_range(struct compact_control *cc,
6341 unsigned long start, unsigned long end)
6343 /* This function is based on compact_zone() from compaction.c. */
6344 unsigned long nr_reclaimed;
6345 unsigned long pfn = start;
6346 unsigned int tries = 0;
6351 while (pfn < end || !list_empty(&cc->migratepages)) {
6352 if (fatal_signal_pending(current)) {
6357 if (list_empty(&cc->migratepages)) {
6358 cc->nr_migratepages = 0;
6359 pfn = isolate_migratepages_range(cc, pfn, end);
6365 } else if (++tries == 5) {
6366 ret = ret < 0 ? ret : -EBUSY;
6370 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6372 cc->nr_migratepages -= nr_reclaimed;
6374 ret = migrate_pages(&cc->migratepages, alloc_migrate_target,
6375 NULL, 0, cc->mode, MR_CMA);
6378 putback_movable_pages(&cc->migratepages);
6385 * alloc_contig_range() -- tries to allocate given range of pages
6386 * @start: start PFN to allocate
6387 * @end: one-past-the-last PFN to allocate
6388 * @migratetype: migratetype of the underlaying pageblocks (either
6389 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6390 * in range must have the same migratetype and it must
6391 * be either of the two.
6393 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6394 * aligned, however it's the caller's responsibility to guarantee that
6395 * we are the only thread that changes migrate type of pageblocks the
6398 * The PFN range must belong to a single zone.
6400 * Returns zero on success or negative error code. On success all
6401 * pages which PFN is in [start, end) are allocated for the caller and
6402 * need to be freed with free_contig_range().
6404 int alloc_contig_range(unsigned long start, unsigned long end,
6405 unsigned migratetype)
6407 unsigned long outer_start, outer_end;
6410 struct compact_control cc = {
6411 .nr_migratepages = 0,
6413 .zone = page_zone(pfn_to_page(start)),
6414 .mode = MIGRATE_SYNC,
6415 .ignore_skip_hint = true,
6417 INIT_LIST_HEAD(&cc.migratepages);
6420 * What we do here is we mark all pageblocks in range as
6421 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6422 * have different sizes, and due to the way page allocator
6423 * work, we align the range to biggest of the two pages so
6424 * that page allocator won't try to merge buddies from
6425 * different pageblocks and change MIGRATE_ISOLATE to some
6426 * other migration type.
6428 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6429 * migrate the pages from an unaligned range (ie. pages that
6430 * we are interested in). This will put all the pages in
6431 * range back to page allocator as MIGRATE_ISOLATE.
6433 * When this is done, we take the pages in range from page
6434 * allocator removing them from the buddy system. This way
6435 * page allocator will never consider using them.
6437 * This lets us mark the pageblocks back as
6438 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6439 * aligned range but not in the unaligned, original range are
6440 * put back to page allocator so that buddy can use them.
6443 ret = start_isolate_page_range(pfn_max_align_down(start),
6444 pfn_max_align_up(end), migratetype,
6449 ret = __alloc_contig_migrate_range(&cc, start, end);
6454 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6455 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6456 * more, all pages in [start, end) are free in page allocator.
6457 * What we are going to do is to allocate all pages from
6458 * [start, end) (that is remove them from page allocator).
6460 * The only problem is that pages at the beginning and at the
6461 * end of interesting range may be not aligned with pages that
6462 * page allocator holds, ie. they can be part of higher order
6463 * pages. Because of this, we reserve the bigger range and
6464 * once this is done free the pages we are not interested in.
6466 * We don't have to hold zone->lock here because the pages are
6467 * isolated thus they won't get removed from buddy.
6470 lru_add_drain_all();
6471 drain_all_pages(cc.zone);
6474 outer_start = start;
6475 while (!PageBuddy(pfn_to_page(outer_start))) {
6476 if (++order >= MAX_ORDER) {
6480 outer_start &= ~0UL << order;
6483 /* Make sure the range is really isolated. */
6484 if (test_pages_isolated(outer_start, end, false)) {
6485 pr_info("%s: [%lx, %lx) PFNs busy\n",
6486 __func__, outer_start, end);
6491 /* Grab isolated pages from freelists. */
6492 outer_end = isolate_freepages_range(&cc, outer_start, end);
6498 /* Free head and tail (if any) */
6499 if (start != outer_start)
6500 free_contig_range(outer_start, start - outer_start);
6501 if (end != outer_end)
6502 free_contig_range(end, outer_end - end);
6505 undo_isolate_page_range(pfn_max_align_down(start),
6506 pfn_max_align_up(end), migratetype);
6510 void free_contig_range(unsigned long pfn, unsigned nr_pages)
6512 unsigned int count = 0;
6514 for (; nr_pages--; pfn++) {
6515 struct page *page = pfn_to_page(pfn);
6517 count += page_count(page) != 1;
6520 WARN(count != 0, "%d pages are still in use!\n", count);
6524 #ifdef CONFIG_MEMORY_HOTPLUG
6526 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6527 * page high values need to be recalulated.
6529 void __meminit zone_pcp_update(struct zone *zone)
6532 mutex_lock(&pcp_batch_high_lock);
6533 for_each_possible_cpu(cpu)
6534 pageset_set_high_and_batch(zone,
6535 per_cpu_ptr(zone->pageset, cpu));
6536 mutex_unlock(&pcp_batch_high_lock);
6540 void zone_pcp_reset(struct zone *zone)
6542 unsigned long flags;
6544 struct per_cpu_pageset *pset;
6546 /* avoid races with drain_pages() */
6547 local_irq_save(flags);
6548 if (zone->pageset != &boot_pageset) {
6549 for_each_online_cpu(cpu) {
6550 pset = per_cpu_ptr(zone->pageset, cpu);
6551 drain_zonestat(zone, pset);
6553 free_percpu(zone->pageset);
6554 zone->pageset = &boot_pageset;
6556 local_irq_restore(flags);
6559 #ifdef CONFIG_MEMORY_HOTREMOVE
6561 * All pages in the range must be isolated before calling this.
6564 __offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
6568 unsigned int order, i;
6570 unsigned long flags;
6571 /* find the first valid pfn */
6572 for (pfn = start_pfn; pfn < end_pfn; pfn++)
6577 zone = page_zone(pfn_to_page(pfn));
6578 spin_lock_irqsave(&zone->lock, flags);
6580 while (pfn < end_pfn) {
6581 if (!pfn_valid(pfn)) {
6585 page = pfn_to_page(pfn);
6587 * The HWPoisoned page may be not in buddy system, and
6588 * page_count() is not 0.
6590 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6592 SetPageReserved(page);
6596 BUG_ON(page_count(page));
6597 BUG_ON(!PageBuddy(page));
6598 order = page_order(page);
6599 #ifdef CONFIG_DEBUG_VM
6600 printk(KERN_INFO "remove from free list %lx %d %lx\n",
6601 pfn, 1 << order, end_pfn);
6603 list_del(&page->lru);
6604 rmv_page_order(page);
6605 zone->free_area[order].nr_free--;
6606 for (i = 0; i < (1 << order); i++)
6607 SetPageReserved((page+i));
6608 pfn += (1 << order);
6610 spin_unlock_irqrestore(&zone->lock, flags);
6614 #ifdef CONFIG_MEMORY_FAILURE
6615 bool is_free_buddy_page(struct page *page)
6617 struct zone *zone = page_zone(page);
6618 unsigned long pfn = page_to_pfn(page);
6619 unsigned long flags;
6622 spin_lock_irqsave(&zone->lock, flags);
6623 for (order = 0; order < MAX_ORDER; order++) {
6624 struct page *page_head = page - (pfn & ((1 << order) - 1));
6626 if (PageBuddy(page_head) && page_order(page_head) >= order)
6629 spin_unlock_irqrestore(&zone->lock, flags);
6631 return order < MAX_ORDER;