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 #ifdef CONFIG_DEBUG_PAGEALLOC
384 unsigned int _debug_guardpage_minorder
;
385 bool _debug_pagealloc_enabled __read_mostly
;
386 bool _debug_guardpage_enabled __read_mostly
;
388 static int __init
early_debug_pagealloc(char *buf
)
393 if (strcmp(buf
, "on") == 0)
394 _debug_pagealloc_enabled
= true;
398 early_param("debug_pagealloc", early_debug_pagealloc
);
400 static bool need_debug_guardpage(void)
402 /* If we don't use debug_pagealloc, we don't need guard page */
403 if (!debug_pagealloc_enabled())
409 static void init_debug_guardpage(void)
411 if (!debug_pagealloc_enabled())
414 _debug_guardpage_enabled
= true;
417 struct page_ext_operations debug_guardpage_ops
= {
418 .need
= need_debug_guardpage
,
419 .init
= init_debug_guardpage
,
422 static int __init
debug_guardpage_minorder_setup(char *buf
)
426 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
427 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
430 _debug_guardpage_minorder
= res
;
431 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
434 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
436 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
437 unsigned int order
, int migratetype
)
439 struct page_ext
*page_ext
;
441 if (!debug_guardpage_enabled())
444 page_ext
= lookup_page_ext(page
);
445 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
447 INIT_LIST_HEAD(&page
->lru
);
448 set_page_private(page
, order
);
449 /* Guard pages are not available for any usage */
450 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
453 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
454 unsigned int order
, int migratetype
)
456 struct page_ext
*page_ext
;
458 if (!debug_guardpage_enabled())
461 page_ext
= lookup_page_ext(page
);
462 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
464 set_page_private(page
, 0);
465 if (!is_migrate_isolate(migratetype
))
466 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
469 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
470 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
471 unsigned int order
, int migratetype
) {}
472 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
473 unsigned int order
, int migratetype
) {}
476 static inline void set_page_order(struct page
*page
, unsigned int order
)
478 set_page_private(page
, order
);
479 __SetPageBuddy(page
);
482 static inline void rmv_page_order(struct page
*page
)
484 __ClearPageBuddy(page
);
485 set_page_private(page
, 0);
489 * This function checks whether a page is free && is the buddy
490 * we can do coalesce a page and its buddy if
491 * (a) the buddy is not in a hole &&
492 * (b) the buddy is in the buddy system &&
493 * (c) a page and its buddy have the same order &&
494 * (d) a page and its buddy are in the same zone.
496 * For recording whether a page is in the buddy system, we set ->_mapcount
497 * PAGE_BUDDY_MAPCOUNT_VALUE.
498 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
499 * serialized by zone->lock.
501 * For recording page's order, we use page_private(page).
503 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
506 if (!pfn_valid_within(page_to_pfn(buddy
)))
509 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
510 if (page_zone_id(page
) != page_zone_id(buddy
))
513 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
518 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
520 * zone check is done late to avoid uselessly
521 * calculating zone/node ids for pages that could
524 if (page_zone_id(page
) != page_zone_id(buddy
))
527 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
535 * Freeing function for a buddy system allocator.
537 * The concept of a buddy system is to maintain direct-mapped table
538 * (containing bit values) for memory blocks of various "orders".
539 * The bottom level table contains the map for the smallest allocatable
540 * units of memory (here, pages), and each level above it describes
541 * pairs of units from the levels below, hence, "buddies".
542 * At a high level, all that happens here is marking the table entry
543 * at the bottom level available, and propagating the changes upward
544 * as necessary, plus some accounting needed to play nicely with other
545 * parts of the VM system.
546 * At each level, we keep a list of pages, which are heads of continuous
547 * free pages of length of (1 << order) and marked with _mapcount
548 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
550 * So when we are allocating or freeing one, we can derive the state of the
551 * other. That is, if we allocate a small block, and both were
552 * free, the remainder of the region must be split into blocks.
553 * If a block is freed, and its buddy is also free, then this
554 * triggers coalescing into a block of larger size.
559 static inline void __free_one_page(struct page
*page
,
561 struct zone
*zone
, unsigned int order
,
564 unsigned long page_idx
;
565 unsigned long combined_idx
;
566 unsigned long uninitialized_var(buddy_idx
);
568 int max_order
= MAX_ORDER
;
570 VM_BUG_ON(!zone_is_initialized(zone
));
571 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
573 VM_BUG_ON(migratetype
== -1);
574 if (is_migrate_isolate(migratetype
)) {
576 * We restrict max order of merging to prevent merge
577 * between freepages on isolate pageblock and normal
578 * pageblock. Without this, pageblock isolation
579 * could cause incorrect freepage accounting.
581 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
583 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
586 page_idx
= pfn
& ((1 << max_order
) - 1);
588 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
589 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
591 while (order
< max_order
- 1) {
592 buddy_idx
= __find_buddy_index(page_idx
, order
);
593 buddy
= page
+ (buddy_idx
- page_idx
);
594 if (!page_is_buddy(page
, buddy
, order
))
597 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
598 * merge with it and move up one order.
600 if (page_is_guard(buddy
)) {
601 clear_page_guard(zone
, buddy
, order
, migratetype
);
603 list_del(&buddy
->lru
);
604 zone
->free_area
[order
].nr_free
--;
605 rmv_page_order(buddy
);
607 combined_idx
= buddy_idx
& page_idx
;
608 page
= page
+ (combined_idx
- page_idx
);
609 page_idx
= combined_idx
;
612 set_page_order(page
, order
);
615 * If this is not the largest possible page, check if the buddy
616 * of the next-highest order is free. If it is, it's possible
617 * that pages are being freed that will coalesce soon. In case,
618 * that is happening, add the free page to the tail of the list
619 * so it's less likely to be used soon and more likely to be merged
620 * as a higher order page
622 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
623 struct page
*higher_page
, *higher_buddy
;
624 combined_idx
= buddy_idx
& page_idx
;
625 higher_page
= page
+ (combined_idx
- page_idx
);
626 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
627 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
628 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
629 list_add_tail(&page
->lru
,
630 &zone
->free_area
[order
].free_list
[migratetype
]);
635 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
637 zone
->free_area
[order
].nr_free
++;
640 static inline int free_pages_check(struct page
*page
)
642 const char *bad_reason
= NULL
;
643 unsigned long bad_flags
= 0;
645 if (unlikely(page_mapcount(page
)))
646 bad_reason
= "nonzero mapcount";
647 if (unlikely(page
->mapping
!= NULL
))
648 bad_reason
= "non-NULL mapping";
649 if (unlikely(atomic_read(&page
->_count
) != 0))
650 bad_reason
= "nonzero _count";
651 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
652 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
653 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
656 if (unlikely(page
->mem_cgroup
))
657 bad_reason
= "page still charged to cgroup";
659 if (unlikely(bad_reason
)) {
660 bad_page(page
, bad_reason
, bad_flags
);
663 page_cpupid_reset_last(page
);
664 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
665 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
670 * Frees a number of pages from the PCP lists
671 * Assumes all pages on list are in same zone, and of same order.
672 * count is the number of pages to free.
674 * If the zone was previously in an "all pages pinned" state then look to
675 * see if this freeing clears that state.
677 * And clear the zone's pages_scanned counter, to hold off the "all pages are
678 * pinned" detection logic.
680 static void free_pcppages_bulk(struct zone
*zone
, int count
,
681 struct per_cpu_pages
*pcp
)
686 unsigned long nr_scanned
;
688 spin_lock(&zone
->lock
);
689 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
691 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
695 struct list_head
*list
;
698 * Remove pages from lists in a round-robin fashion. A
699 * batch_free count is maintained that is incremented when an
700 * empty list is encountered. This is so more pages are freed
701 * off fuller lists instead of spinning excessively around empty
706 if (++migratetype
== MIGRATE_PCPTYPES
)
708 list
= &pcp
->lists
[migratetype
];
709 } while (list_empty(list
));
711 /* This is the only non-empty list. Free them all. */
712 if (batch_free
== MIGRATE_PCPTYPES
)
713 batch_free
= to_free
;
716 int mt
; /* migratetype of the to-be-freed page */
718 page
= list_entry(list
->prev
, struct page
, lru
);
719 /* must delete as __free_one_page list manipulates */
720 list_del(&page
->lru
);
721 mt
= get_freepage_migratetype(page
);
722 if (unlikely(has_isolate_pageblock(zone
)))
723 mt
= get_pageblock_migratetype(page
);
725 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
726 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
727 trace_mm_page_pcpu_drain(page
, 0, mt
);
728 } while (--to_free
&& --batch_free
&& !list_empty(list
));
730 spin_unlock(&zone
->lock
);
733 static void free_one_page(struct zone
*zone
,
734 struct page
*page
, unsigned long pfn
,
738 unsigned long nr_scanned
;
739 spin_lock(&zone
->lock
);
740 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
742 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
744 if (unlikely(has_isolate_pageblock(zone
) ||
745 is_migrate_isolate(migratetype
))) {
746 migratetype
= get_pfnblock_migratetype(page
, pfn
);
748 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
749 spin_unlock(&zone
->lock
);
752 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
754 if (!IS_ENABLED(CONFIG_DEBUG_VM
))
756 if (unlikely(!PageTail(page
))) {
757 bad_page(page
, "PageTail not set", 0);
760 if (unlikely(page
->first_page
!= head_page
)) {
761 bad_page(page
, "first_page not consistent", 0);
767 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
768 unsigned long zone
, int nid
)
770 struct zone
*z
= &NODE_DATA(nid
)->node_zones
[zone
];
772 set_page_links(page
, zone
, nid
, pfn
);
773 mminit_verify_page_links(page
, zone
, nid
, pfn
);
774 init_page_count(page
);
775 page_mapcount_reset(page
);
776 page_cpupid_reset_last(page
);
779 * Mark the block movable so that blocks are reserved for
780 * movable at startup. This will force kernel allocations
781 * to reserve their blocks rather than leaking throughout
782 * the address space during boot when many long-lived
783 * kernel allocations are made. Later some blocks near
784 * the start are marked MIGRATE_RESERVE by
785 * setup_zone_migrate_reserve()
787 * bitmap is created for zone's valid pfn range. but memmap
788 * can be created for invalid pages (for alignment)
789 * check here not to call set_pageblock_migratetype() against
792 if ((z
->zone_start_pfn
<= pfn
)
793 && (pfn
< zone_end_pfn(z
))
794 && !(pfn
& (pageblock_nr_pages
- 1)))
795 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
797 INIT_LIST_HEAD(&page
->lru
);
798 #ifdef WANT_PAGE_VIRTUAL
799 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
800 if (!is_highmem_idx(zone
))
801 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
805 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
808 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
812 * Initialised pages do not have PageReserved set. This function is
813 * called for each range allocated by the bootmem allocator and
814 * marks the pages PageReserved. The remaining valid pages are later
815 * sent to the buddy page allocator.
817 void reserve_bootmem_region(unsigned long start
, unsigned long end
)
819 unsigned long start_pfn
= PFN_DOWN(start
);
820 unsigned long end_pfn
= PFN_UP(end
);
822 for (; start_pfn
< end_pfn
; start_pfn
++)
823 if (pfn_valid(start_pfn
))
824 SetPageReserved(pfn_to_page(start_pfn
));
827 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
829 bool compound
= PageCompound(page
);
832 VM_BUG_ON_PAGE(PageTail(page
), page
);
833 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
835 trace_mm_page_free(page
, order
);
836 kmemcheck_free_shadow(page
, order
);
837 kasan_free_pages(page
, order
);
840 page
->mapping
= NULL
;
841 bad
+= free_pages_check(page
);
842 for (i
= 1; i
< (1 << order
); i
++) {
844 bad
+= free_tail_pages_check(page
, page
+ i
);
845 bad
+= free_pages_check(page
+ i
);
850 reset_page_owner(page
, order
);
852 if (!PageHighMem(page
)) {
853 debug_check_no_locks_freed(page_address(page
),
855 debug_check_no_obj_freed(page_address(page
),
858 arch_free_page(page
, order
);
859 kernel_map_pages(page
, 1 << order
, 0);
864 static void __free_pages_ok(struct page
*page
, unsigned int order
)
868 unsigned long pfn
= page_to_pfn(page
);
870 if (!free_pages_prepare(page
, order
))
873 migratetype
= get_pfnblock_migratetype(page
, pfn
);
874 local_irq_save(flags
);
875 __count_vm_events(PGFREE
, 1 << order
);
876 set_freepage_migratetype(page
, migratetype
);
877 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
878 local_irq_restore(flags
);
881 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
884 unsigned int nr_pages
= 1 << order
;
885 struct page
*p
= page
;
889 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
891 __ClearPageReserved(p
);
892 set_page_count(p
, 0);
894 __ClearPageReserved(p
);
895 set_page_count(p
, 0);
897 page_zone(page
)->managed_pages
+= nr_pages
;
898 set_page_refcounted(page
);
899 __free_pages(page
, order
);
903 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
904 void __init
init_cma_reserved_pageblock(struct page
*page
)
906 unsigned i
= pageblock_nr_pages
;
907 struct page
*p
= page
;
910 __ClearPageReserved(p
);
911 set_page_count(p
, 0);
914 set_pageblock_migratetype(page
, MIGRATE_CMA
);
916 if (pageblock_order
>= MAX_ORDER
) {
917 i
= pageblock_nr_pages
;
920 set_page_refcounted(p
);
921 __free_pages(p
, MAX_ORDER
- 1);
922 p
+= MAX_ORDER_NR_PAGES
;
923 } while (i
-= MAX_ORDER_NR_PAGES
);
925 set_page_refcounted(page
);
926 __free_pages(page
, pageblock_order
);
929 adjust_managed_page_count(page
, pageblock_nr_pages
);
934 * The order of subdivision here is critical for the IO subsystem.
935 * Please do not alter this order without good reasons and regression
936 * testing. Specifically, as large blocks of memory are subdivided,
937 * the order in which smaller blocks are delivered depends on the order
938 * they're subdivided in this function. This is the primary factor
939 * influencing the order in which pages are delivered to the IO
940 * subsystem according to empirical testing, and this is also justified
941 * by considering the behavior of a buddy system containing a single
942 * large block of memory acted on by a series of small allocations.
943 * This behavior is a critical factor in sglist merging's success.
947 static inline void expand(struct zone
*zone
, struct page
*page
,
948 int low
, int high
, struct free_area
*area
,
951 unsigned long size
= 1 << high
;
957 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
959 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
960 debug_guardpage_enabled() &&
961 high
< debug_guardpage_minorder()) {
963 * Mark as guard pages (or page), that will allow to
964 * merge back to allocator when buddy will be freed.
965 * Corresponding page table entries will not be touched,
966 * pages will stay not present in virtual address space
968 set_page_guard(zone
, &page
[size
], high
, migratetype
);
971 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
973 set_page_order(&page
[size
], high
);
978 * This page is about to be returned from the page allocator
980 static inline int check_new_page(struct page
*page
)
982 const char *bad_reason
= NULL
;
983 unsigned long bad_flags
= 0;
985 if (unlikely(page_mapcount(page
)))
986 bad_reason
= "nonzero mapcount";
987 if (unlikely(page
->mapping
!= NULL
))
988 bad_reason
= "non-NULL mapping";
989 if (unlikely(atomic_read(&page
->_count
) != 0))
990 bad_reason
= "nonzero _count";
991 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
992 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
993 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
996 if (unlikely(page
->mem_cgroup
))
997 bad_reason
= "page still charged to cgroup";
999 if (unlikely(bad_reason
)) {
1000 bad_page(page
, bad_reason
, bad_flags
);
1006 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1011 for (i
= 0; i
< (1 << order
); i
++) {
1012 struct page
*p
= page
+ i
;
1013 if (unlikely(check_new_page(p
)))
1017 set_page_private(page
, 0);
1018 set_page_refcounted(page
);
1020 arch_alloc_page(page
, order
);
1021 kernel_map_pages(page
, 1 << order
, 1);
1022 kasan_alloc_pages(page
, order
);
1024 if (gfp_flags
& __GFP_ZERO
)
1025 for (i
= 0; i
< (1 << order
); i
++)
1026 clear_highpage(page
+ i
);
1028 if (order
&& (gfp_flags
& __GFP_COMP
))
1029 prep_compound_page(page
, order
);
1031 set_page_owner(page
, order
, gfp_flags
);
1034 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
1035 * allocate the page. The expectation is that the caller is taking
1036 * steps that will free more memory. The caller should avoid the page
1037 * being used for !PFMEMALLOC purposes.
1039 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
1045 * Go through the free lists for the given migratetype and remove
1046 * the smallest available page from the freelists
1049 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1052 unsigned int current_order
;
1053 struct free_area
*area
;
1056 /* Find a page of the appropriate size in the preferred list */
1057 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1058 area
= &(zone
->free_area
[current_order
]);
1059 if (list_empty(&area
->free_list
[migratetype
]))
1062 page
= list_entry(area
->free_list
[migratetype
].next
,
1064 list_del(&page
->lru
);
1065 rmv_page_order(page
);
1067 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1068 set_freepage_migratetype(page
, migratetype
);
1077 * This array describes the order lists are fallen back to when
1078 * the free lists for the desirable migrate type are depleted
1080 static int fallbacks
[MIGRATE_TYPES
][4] = {
1081 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1082 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1083 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
1085 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
1087 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
1088 #ifdef CONFIG_MEMORY_ISOLATION
1089 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
1094 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1097 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1100 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1101 unsigned int order
) { return NULL
; }
1105 * Move the free pages in a range to the free lists of the requested type.
1106 * Note that start_page and end_pages are not aligned on a pageblock
1107 * boundary. If alignment is required, use move_freepages_block()
1109 int move_freepages(struct zone
*zone
,
1110 struct page
*start_page
, struct page
*end_page
,
1114 unsigned long order
;
1115 int pages_moved
= 0;
1117 #ifndef CONFIG_HOLES_IN_ZONE
1119 * page_zone is not safe to call in this context when
1120 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1121 * anyway as we check zone boundaries in move_freepages_block().
1122 * Remove at a later date when no bug reports exist related to
1123 * grouping pages by mobility
1125 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1128 for (page
= start_page
; page
<= end_page
;) {
1129 /* Make sure we are not inadvertently changing nodes */
1130 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1132 if (!pfn_valid_within(page_to_pfn(page
))) {
1137 if (!PageBuddy(page
)) {
1142 order
= page_order(page
);
1143 list_move(&page
->lru
,
1144 &zone
->free_area
[order
].free_list
[migratetype
]);
1145 set_freepage_migratetype(page
, migratetype
);
1147 pages_moved
+= 1 << order
;
1153 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1156 unsigned long start_pfn
, end_pfn
;
1157 struct page
*start_page
, *end_page
;
1159 start_pfn
= page_to_pfn(page
);
1160 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1161 start_page
= pfn_to_page(start_pfn
);
1162 end_page
= start_page
+ pageblock_nr_pages
- 1;
1163 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1165 /* Do not cross zone boundaries */
1166 if (!zone_spans_pfn(zone
, start_pfn
))
1168 if (!zone_spans_pfn(zone
, end_pfn
))
1171 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1174 static void change_pageblock_range(struct page
*pageblock_page
,
1175 int start_order
, int migratetype
)
1177 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1179 while (nr_pageblocks
--) {
1180 set_pageblock_migratetype(pageblock_page
, migratetype
);
1181 pageblock_page
+= pageblock_nr_pages
;
1186 * When we are falling back to another migratetype during allocation, try to
1187 * steal extra free pages from the same pageblocks to satisfy further
1188 * allocations, instead of polluting multiple pageblocks.
1190 * If we are stealing a relatively large buddy page, it is likely there will
1191 * be more free pages in the pageblock, so try to steal them all. For
1192 * reclaimable and unmovable allocations, we steal regardless of page size,
1193 * as fragmentation caused by those allocations polluting movable pageblocks
1194 * is worse than movable allocations stealing from unmovable and reclaimable
1197 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1200 * Leaving this order check is intended, although there is
1201 * relaxed order check in next check. The reason is that
1202 * we can actually steal whole pageblock if this condition met,
1203 * but, below check doesn't guarantee it and that is just heuristic
1204 * so could be changed anytime.
1206 if (order
>= pageblock_order
)
1209 if (order
>= pageblock_order
/ 2 ||
1210 start_mt
== MIGRATE_RECLAIMABLE
||
1211 start_mt
== MIGRATE_UNMOVABLE
||
1212 page_group_by_mobility_disabled
)
1219 * This function implements actual steal behaviour. If order is large enough,
1220 * we can steal whole pageblock. If not, we first move freepages in this
1221 * pageblock and check whether half of pages are moved or not. If half of
1222 * pages are moved, we can change migratetype of pageblock and permanently
1223 * use it's pages as requested migratetype in the future.
1225 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1228 int current_order
= page_order(page
);
1231 /* Take ownership for orders >= pageblock_order */
1232 if (current_order
>= pageblock_order
) {
1233 change_pageblock_range(page
, current_order
, start_type
);
1237 pages
= move_freepages_block(zone
, page
, start_type
);
1239 /* Claim the whole block if over half of it is free */
1240 if (pages
>= (1 << (pageblock_order
-1)) ||
1241 page_group_by_mobility_disabled
)
1242 set_pageblock_migratetype(page
, start_type
);
1246 * Check whether there is a suitable fallback freepage with requested order.
1247 * If only_stealable is true, this function returns fallback_mt only if
1248 * we can steal other freepages all together. This would help to reduce
1249 * fragmentation due to mixed migratetype pages in one pageblock.
1251 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1252 int migratetype
, bool only_stealable
, bool *can_steal
)
1257 if (area
->nr_free
== 0)
1262 fallback_mt
= fallbacks
[migratetype
][i
];
1263 if (fallback_mt
== MIGRATE_RESERVE
)
1266 if (list_empty(&area
->free_list
[fallback_mt
]))
1269 if (can_steal_fallback(order
, migratetype
))
1272 if (!only_stealable
)
1282 /* Remove an element from the buddy allocator from the fallback list */
1283 static inline struct page
*
1284 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1286 struct free_area
*area
;
1287 unsigned int current_order
;
1292 /* Find the largest possible block of pages in the other list */
1293 for (current_order
= MAX_ORDER
-1;
1294 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1296 area
= &(zone
->free_area
[current_order
]);
1297 fallback_mt
= find_suitable_fallback(area
, current_order
,
1298 start_migratetype
, false, &can_steal
);
1299 if (fallback_mt
== -1)
1302 page
= list_entry(area
->free_list
[fallback_mt
].next
,
1305 steal_suitable_fallback(zone
, page
, start_migratetype
);
1307 /* Remove the page from the freelists */
1309 list_del(&page
->lru
);
1310 rmv_page_order(page
);
1312 expand(zone
, page
, order
, current_order
, area
,
1315 * The freepage_migratetype may differ from pageblock's
1316 * migratetype depending on the decisions in
1317 * try_to_steal_freepages(). This is OK as long as it
1318 * does not differ for MIGRATE_CMA pageblocks. For CMA
1319 * we need to make sure unallocated pages flushed from
1320 * pcp lists are returned to the correct freelist.
1322 set_freepage_migratetype(page
, start_migratetype
);
1324 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1325 start_migratetype
, fallback_mt
);
1334 * Do the hard work of removing an element from the buddy allocator.
1335 * Call me with the zone->lock already held.
1337 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1343 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1345 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1346 if (migratetype
== MIGRATE_MOVABLE
)
1347 page
= __rmqueue_cma_fallback(zone
, order
);
1350 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1353 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1354 * is used because __rmqueue_smallest is an inline function
1355 * and we want just one call site
1358 migratetype
= MIGRATE_RESERVE
;
1363 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1368 * Obtain a specified number of elements from the buddy allocator, all under
1369 * a single hold of the lock, for efficiency. Add them to the supplied list.
1370 * Returns the number of new pages which were placed at *list.
1372 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1373 unsigned long count
, struct list_head
*list
,
1374 int migratetype
, bool cold
)
1378 spin_lock(&zone
->lock
);
1379 for (i
= 0; i
< count
; ++i
) {
1380 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1381 if (unlikely(page
== NULL
))
1385 * Split buddy pages returned by expand() are received here
1386 * in physical page order. The page is added to the callers and
1387 * list and the list head then moves forward. From the callers
1388 * perspective, the linked list is ordered by page number in
1389 * some conditions. This is useful for IO devices that can
1390 * merge IO requests if the physical pages are ordered
1394 list_add(&page
->lru
, list
);
1396 list_add_tail(&page
->lru
, list
);
1398 if (is_migrate_cma(get_freepage_migratetype(page
)))
1399 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1402 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1403 spin_unlock(&zone
->lock
);
1409 * Called from the vmstat counter updater to drain pagesets of this
1410 * currently executing processor on remote nodes after they have
1413 * Note that this function must be called with the thread pinned to
1414 * a single processor.
1416 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1418 unsigned long flags
;
1419 int to_drain
, batch
;
1421 local_irq_save(flags
);
1422 batch
= READ_ONCE(pcp
->batch
);
1423 to_drain
= min(pcp
->count
, batch
);
1425 free_pcppages_bulk(zone
, to_drain
, pcp
);
1426 pcp
->count
-= to_drain
;
1428 local_irq_restore(flags
);
1433 * Drain pcplists of the indicated processor and zone.
1435 * The processor must either be the current processor and the
1436 * thread pinned to the current processor or a processor that
1439 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1441 unsigned long flags
;
1442 struct per_cpu_pageset
*pset
;
1443 struct per_cpu_pages
*pcp
;
1445 local_irq_save(flags
);
1446 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1450 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1453 local_irq_restore(flags
);
1457 * Drain pcplists of all zones on the indicated processor.
1459 * The processor must either be the current processor and the
1460 * thread pinned to the current processor or a processor that
1463 static void drain_pages(unsigned int cpu
)
1467 for_each_populated_zone(zone
) {
1468 drain_pages_zone(cpu
, zone
);
1473 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1475 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1476 * the single zone's pages.
1478 void drain_local_pages(struct zone
*zone
)
1480 int cpu
= smp_processor_id();
1483 drain_pages_zone(cpu
, zone
);
1489 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1491 * When zone parameter is non-NULL, spill just the single zone's pages.
1493 * Note that this code is protected against sending an IPI to an offline
1494 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1495 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1496 * nothing keeps CPUs from showing up after we populated the cpumask and
1497 * before the call to on_each_cpu_mask().
1499 void drain_all_pages(struct zone
*zone
)
1504 * Allocate in the BSS so we wont require allocation in
1505 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1507 static cpumask_t cpus_with_pcps
;
1510 * We don't care about racing with CPU hotplug event
1511 * as offline notification will cause the notified
1512 * cpu to drain that CPU pcps and on_each_cpu_mask
1513 * disables preemption as part of its processing
1515 for_each_online_cpu(cpu
) {
1516 struct per_cpu_pageset
*pcp
;
1518 bool has_pcps
= false;
1521 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1525 for_each_populated_zone(z
) {
1526 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1527 if (pcp
->pcp
.count
) {
1535 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1537 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1539 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1543 #ifdef CONFIG_HIBERNATION
1545 void mark_free_pages(struct zone
*zone
)
1547 unsigned long pfn
, max_zone_pfn
;
1548 unsigned long flags
;
1549 unsigned int order
, t
;
1550 struct list_head
*curr
;
1552 if (zone_is_empty(zone
))
1555 spin_lock_irqsave(&zone
->lock
, flags
);
1557 max_zone_pfn
= zone_end_pfn(zone
);
1558 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1559 if (pfn_valid(pfn
)) {
1560 struct page
*page
= pfn_to_page(pfn
);
1562 if (!swsusp_page_is_forbidden(page
))
1563 swsusp_unset_page_free(page
);
1566 for_each_migratetype_order(order
, t
) {
1567 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1570 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1571 for (i
= 0; i
< (1UL << order
); i
++)
1572 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1575 spin_unlock_irqrestore(&zone
->lock
, flags
);
1577 #endif /* CONFIG_PM */
1580 * Free a 0-order page
1581 * cold == true ? free a cold page : free a hot page
1583 void free_hot_cold_page(struct page
*page
, bool cold
)
1585 struct zone
*zone
= page_zone(page
);
1586 struct per_cpu_pages
*pcp
;
1587 unsigned long flags
;
1588 unsigned long pfn
= page_to_pfn(page
);
1591 if (!free_pages_prepare(page
, 0))
1594 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1595 set_freepage_migratetype(page
, migratetype
);
1596 local_irq_save(flags
);
1597 __count_vm_event(PGFREE
);
1600 * We only track unmovable, reclaimable and movable on pcp lists.
1601 * Free ISOLATE pages back to the allocator because they are being
1602 * offlined but treat RESERVE as movable pages so we can get those
1603 * areas back if necessary. Otherwise, we may have to free
1604 * excessively into the page allocator
1606 if (migratetype
>= MIGRATE_PCPTYPES
) {
1607 if (unlikely(is_migrate_isolate(migratetype
))) {
1608 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1611 migratetype
= MIGRATE_MOVABLE
;
1614 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1616 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1618 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1620 if (pcp
->count
>= pcp
->high
) {
1621 unsigned long batch
= READ_ONCE(pcp
->batch
);
1622 free_pcppages_bulk(zone
, batch
, pcp
);
1623 pcp
->count
-= batch
;
1627 local_irq_restore(flags
);
1631 * Free a list of 0-order pages
1633 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1635 struct page
*page
, *next
;
1637 list_for_each_entry_safe(page
, next
, list
, lru
) {
1638 trace_mm_page_free_batched(page
, cold
);
1639 free_hot_cold_page(page
, cold
);
1644 * split_page takes a non-compound higher-order page, and splits it into
1645 * n (1<<order) sub-pages: page[0..n]
1646 * Each sub-page must be freed individually.
1648 * Note: this is probably too low level an operation for use in drivers.
1649 * Please consult with lkml before using this in your driver.
1651 void split_page(struct page
*page
, unsigned int order
)
1655 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1656 VM_BUG_ON_PAGE(!page_count(page
), page
);
1658 #ifdef CONFIG_KMEMCHECK
1660 * Split shadow pages too, because free(page[0]) would
1661 * otherwise free the whole shadow.
1663 if (kmemcheck_page_is_tracked(page
))
1664 split_page(virt_to_page(page
[0].shadow
), order
);
1667 set_page_owner(page
, 0, 0);
1668 for (i
= 1; i
< (1 << order
); i
++) {
1669 set_page_refcounted(page
+ i
);
1670 set_page_owner(page
+ i
, 0, 0);
1673 EXPORT_SYMBOL_GPL(split_page
);
1675 int __isolate_free_page(struct page
*page
, unsigned int order
)
1677 unsigned long watermark
;
1681 BUG_ON(!PageBuddy(page
));
1683 zone
= page_zone(page
);
1684 mt
= get_pageblock_migratetype(page
);
1686 if (!is_migrate_isolate(mt
)) {
1687 /* Obey watermarks as if the page was being allocated */
1688 watermark
= low_wmark_pages(zone
) + (1 << order
);
1689 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1692 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1695 /* Remove page from free list */
1696 list_del(&page
->lru
);
1697 zone
->free_area
[order
].nr_free
--;
1698 rmv_page_order(page
);
1700 /* Set the pageblock if the isolated page is at least a pageblock */
1701 if (order
>= pageblock_order
- 1) {
1702 struct page
*endpage
= page
+ (1 << order
) - 1;
1703 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1704 int mt
= get_pageblock_migratetype(page
);
1705 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1706 set_pageblock_migratetype(page
,
1711 set_page_owner(page
, order
, 0);
1712 return 1UL << order
;
1716 * Similar to split_page except the page is already free. As this is only
1717 * being used for migration, the migratetype of the block also changes.
1718 * As this is called with interrupts disabled, the caller is responsible
1719 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1722 * Note: this is probably too low level an operation for use in drivers.
1723 * Please consult with lkml before using this in your driver.
1725 int split_free_page(struct page
*page
)
1730 order
= page_order(page
);
1732 nr_pages
= __isolate_free_page(page
, order
);
1736 /* Split into individual pages */
1737 set_page_refcounted(page
);
1738 split_page(page
, order
);
1743 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1746 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1747 struct zone
*zone
, unsigned int order
,
1748 gfp_t gfp_flags
, int migratetype
)
1750 unsigned long flags
;
1752 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1754 if (likely(order
== 0)) {
1755 struct per_cpu_pages
*pcp
;
1756 struct list_head
*list
;
1758 local_irq_save(flags
);
1759 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1760 list
= &pcp
->lists
[migratetype
];
1761 if (list_empty(list
)) {
1762 pcp
->count
+= rmqueue_bulk(zone
, 0,
1765 if (unlikely(list_empty(list
)))
1770 page
= list_entry(list
->prev
, struct page
, lru
);
1772 page
= list_entry(list
->next
, struct page
, lru
);
1774 list_del(&page
->lru
);
1777 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1779 * __GFP_NOFAIL is not to be used in new code.
1781 * All __GFP_NOFAIL callers should be fixed so that they
1782 * properly detect and handle allocation failures.
1784 * We most definitely don't want callers attempting to
1785 * allocate greater than order-1 page units with
1788 WARN_ON_ONCE(order
> 1);
1790 spin_lock_irqsave(&zone
->lock
, flags
);
1791 page
= __rmqueue(zone
, order
, migratetype
);
1792 spin_unlock(&zone
->lock
);
1795 __mod_zone_freepage_state(zone
, -(1 << order
),
1796 get_freepage_migratetype(page
));
1799 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1800 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1801 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1802 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1804 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1805 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1806 local_irq_restore(flags
);
1808 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1812 local_irq_restore(flags
);
1816 #ifdef CONFIG_FAIL_PAGE_ALLOC
1819 struct fault_attr attr
;
1821 u32 ignore_gfp_highmem
;
1822 u32 ignore_gfp_wait
;
1824 } fail_page_alloc
= {
1825 .attr
= FAULT_ATTR_INITIALIZER
,
1826 .ignore_gfp_wait
= 1,
1827 .ignore_gfp_highmem
= 1,
1831 static int __init
setup_fail_page_alloc(char *str
)
1833 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1835 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1837 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1839 if (order
< fail_page_alloc
.min_order
)
1841 if (gfp_mask
& __GFP_NOFAIL
)
1843 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1845 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1848 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1851 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1853 static int __init
fail_page_alloc_debugfs(void)
1855 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1858 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1859 &fail_page_alloc
.attr
);
1861 return PTR_ERR(dir
);
1863 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1864 &fail_page_alloc
.ignore_gfp_wait
))
1866 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1867 &fail_page_alloc
.ignore_gfp_highmem
))
1869 if (!debugfs_create_u32("min-order", mode
, dir
,
1870 &fail_page_alloc
.min_order
))
1875 debugfs_remove_recursive(dir
);
1880 late_initcall(fail_page_alloc_debugfs
);
1882 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1884 #else /* CONFIG_FAIL_PAGE_ALLOC */
1886 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1891 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1894 * Return true if free pages are above 'mark'. This takes into account the order
1895 * of the allocation.
1897 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1898 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1901 /* free_pages may go negative - that's OK */
1906 free_pages
-= (1 << order
) - 1;
1907 if (alloc_flags
& ALLOC_HIGH
)
1909 if (alloc_flags
& ALLOC_HARDER
)
1912 /* If allocation can't use CMA areas don't use free CMA pages */
1913 if (!(alloc_flags
& ALLOC_CMA
))
1914 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1917 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1919 for (o
= 0; o
< order
; o
++) {
1920 /* At the next order, this order's pages become unavailable */
1921 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1923 /* Require fewer higher order pages to be free */
1926 if (free_pages
<= min
)
1932 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1933 int classzone_idx
, int alloc_flags
)
1935 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1936 zone_page_state(z
, NR_FREE_PAGES
));
1939 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1940 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1942 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1944 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1945 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1947 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1953 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1954 * skip over zones that are not allowed by the cpuset, or that have
1955 * been recently (in last second) found to be nearly full. See further
1956 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1957 * that have to skip over a lot of full or unallowed zones.
1959 * If the zonelist cache is present in the passed zonelist, then
1960 * returns a pointer to the allowed node mask (either the current
1961 * tasks mems_allowed, or node_states[N_MEMORY].)
1963 * If the zonelist cache is not available for this zonelist, does
1964 * nothing and returns NULL.
1966 * If the fullzones BITMAP in the zonelist cache is stale (more than
1967 * a second since last zap'd) then we zap it out (clear its bits.)
1969 * We hold off even calling zlc_setup, until after we've checked the
1970 * first zone in the zonelist, on the theory that most allocations will
1971 * be satisfied from that first zone, so best to examine that zone as
1972 * quickly as we can.
1974 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1976 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1977 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1979 zlc
= zonelist
->zlcache_ptr
;
1983 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1984 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1985 zlc
->last_full_zap
= jiffies
;
1988 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1989 &cpuset_current_mems_allowed
:
1990 &node_states
[N_MEMORY
];
1991 return allowednodes
;
1995 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1996 * if it is worth looking at further for free memory:
1997 * 1) Check that the zone isn't thought to be full (doesn't have its
1998 * bit set in the zonelist_cache fullzones BITMAP).
1999 * 2) Check that the zones node (obtained from the zonelist_cache
2000 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2001 * Return true (non-zero) if zone is worth looking at further, or
2002 * else return false (zero) if it is not.
2004 * This check -ignores- the distinction between various watermarks,
2005 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2006 * found to be full for any variation of these watermarks, it will
2007 * be considered full for up to one second by all requests, unless
2008 * we are so low on memory on all allowed nodes that we are forced
2009 * into the second scan of the zonelist.
2011 * In the second scan we ignore this zonelist cache and exactly
2012 * apply the watermarks to all zones, even it is slower to do so.
2013 * We are low on memory in the second scan, and should leave no stone
2014 * unturned looking for a free page.
2016 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
2017 nodemask_t
*allowednodes
)
2019 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2020 int i
; /* index of *z in zonelist zones */
2021 int n
; /* node that zone *z is on */
2023 zlc
= zonelist
->zlcache_ptr
;
2027 i
= z
- zonelist
->_zonerefs
;
2030 /* This zone is worth trying if it is allowed but not full */
2031 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
2035 * Given 'z' scanning a zonelist, set the corresponding bit in
2036 * zlc->fullzones, so that subsequent attempts to allocate a page
2037 * from that zone don't waste time re-examining it.
2039 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
2041 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2042 int i
; /* index of *z in zonelist zones */
2044 zlc
= zonelist
->zlcache_ptr
;
2048 i
= z
- zonelist
->_zonerefs
;
2050 set_bit(i
, zlc
->fullzones
);
2054 * clear all zones full, called after direct reclaim makes progress so that
2055 * a zone that was recently full is not skipped over for up to a second
2057 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
2059 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2061 zlc
= zonelist
->zlcache_ptr
;
2065 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2068 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2070 return local_zone
->node
== zone
->node
;
2073 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2075 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2079 #else /* CONFIG_NUMA */
2081 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
2086 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
2087 nodemask_t
*allowednodes
)
2092 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
2096 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
2100 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2105 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2110 #endif /* CONFIG_NUMA */
2112 static void reset_alloc_batches(struct zone
*preferred_zone
)
2114 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
2117 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2118 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
2119 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
2120 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2121 } while (zone
++ != preferred_zone
);
2125 * get_page_from_freelist goes through the zonelist trying to allocate
2128 static struct page
*
2129 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2130 const struct alloc_context
*ac
)
2132 struct zonelist
*zonelist
= ac
->zonelist
;
2134 struct page
*page
= NULL
;
2136 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
2137 int zlc_active
= 0; /* set if using zonelist_cache */
2138 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
2139 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
2140 (gfp_mask
& __GFP_WRITE
);
2141 int nr_fair_skipped
= 0;
2142 bool zonelist_rescan
;
2145 zonelist_rescan
= false;
2148 * Scan zonelist, looking for a zone with enough free.
2149 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2151 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2155 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2156 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2158 if (cpusets_enabled() &&
2159 (alloc_flags
& ALLOC_CPUSET
) &&
2160 !cpuset_zone_allowed(zone
, gfp_mask
))
2163 * Distribute pages in proportion to the individual
2164 * zone size to ensure fair page aging. The zone a
2165 * page was allocated in should have no effect on the
2166 * time the page has in memory before being reclaimed.
2168 if (alloc_flags
& ALLOC_FAIR
) {
2169 if (!zone_local(ac
->preferred_zone
, zone
))
2171 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2177 * When allocating a page cache page for writing, we
2178 * want to get it from a zone that is within its dirty
2179 * limit, such that no single zone holds more than its
2180 * proportional share of globally allowed dirty pages.
2181 * The dirty limits take into account the zone's
2182 * lowmem reserves and high watermark so that kswapd
2183 * should be able to balance it without having to
2184 * write pages from its LRU list.
2186 * This may look like it could increase pressure on
2187 * lower zones by failing allocations in higher zones
2188 * before they are full. But the pages that do spill
2189 * over are limited as the lower zones are protected
2190 * by this very same mechanism. It should not become
2191 * a practical burden to them.
2193 * XXX: For now, allow allocations to potentially
2194 * exceed the per-zone dirty limit in the slowpath
2195 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2196 * which is important when on a NUMA setup the allowed
2197 * zones are together not big enough to reach the
2198 * global limit. The proper fix for these situations
2199 * will require awareness of zones in the
2200 * dirty-throttling and the flusher threads.
2202 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2205 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2206 if (!zone_watermark_ok(zone
, order
, mark
,
2207 ac
->classzone_idx
, alloc_flags
)) {
2210 /* Checked here to keep the fast path fast */
2211 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2212 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2215 if (IS_ENABLED(CONFIG_NUMA
) &&
2216 !did_zlc_setup
&& nr_online_nodes
> 1) {
2218 * we do zlc_setup if there are multiple nodes
2219 * and before considering the first zone allowed
2222 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2227 if (zone_reclaim_mode
== 0 ||
2228 !zone_allows_reclaim(ac
->preferred_zone
, zone
))
2229 goto this_zone_full
;
2232 * As we may have just activated ZLC, check if the first
2233 * eligible zone has failed zone_reclaim recently.
2235 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2236 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2239 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2241 case ZONE_RECLAIM_NOSCAN
:
2244 case ZONE_RECLAIM_FULL
:
2245 /* scanned but unreclaimable */
2248 /* did we reclaim enough */
2249 if (zone_watermark_ok(zone
, order
, mark
,
2250 ac
->classzone_idx
, alloc_flags
))
2254 * Failed to reclaim enough to meet watermark.
2255 * Only mark the zone full if checking the min
2256 * watermark or if we failed to reclaim just
2257 * 1<<order pages or else the page allocator
2258 * fastpath will prematurely mark zones full
2259 * when the watermark is between the low and
2262 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2263 ret
== ZONE_RECLAIM_SOME
)
2264 goto this_zone_full
;
2271 page
= buffered_rmqueue(ac
->preferred_zone
, zone
, order
,
2272 gfp_mask
, ac
->migratetype
);
2274 if (prep_new_page(page
, order
, gfp_mask
, alloc_flags
))
2279 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2280 zlc_mark_zone_full(zonelist
, z
);
2284 * The first pass makes sure allocations are spread fairly within the
2285 * local node. However, the local node might have free pages left
2286 * after the fairness batches are exhausted, and remote zones haven't
2287 * even been considered yet. Try once more without fairness, and
2288 * include remote zones now, before entering the slowpath and waking
2289 * kswapd: prefer spilling to a remote zone over swapping locally.
2291 if (alloc_flags
& ALLOC_FAIR
) {
2292 alloc_flags
&= ~ALLOC_FAIR
;
2293 if (nr_fair_skipped
) {
2294 zonelist_rescan
= true;
2295 reset_alloc_batches(ac
->preferred_zone
);
2297 if (nr_online_nodes
> 1)
2298 zonelist_rescan
= true;
2301 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2302 /* Disable zlc cache for second zonelist scan */
2304 zonelist_rescan
= true;
2307 if (zonelist_rescan
)
2314 * Large machines with many possible nodes should not always dump per-node
2315 * meminfo in irq context.
2317 static inline bool should_suppress_show_mem(void)
2322 ret
= in_interrupt();
2327 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2328 DEFAULT_RATELIMIT_INTERVAL
,
2329 DEFAULT_RATELIMIT_BURST
);
2331 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2333 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2335 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2336 debug_guardpage_minorder() > 0)
2340 * This documents exceptions given to allocations in certain
2341 * contexts that are allowed to allocate outside current's set
2344 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2345 if (test_thread_flag(TIF_MEMDIE
) ||
2346 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2347 filter
&= ~SHOW_MEM_FILTER_NODES
;
2348 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2349 filter
&= ~SHOW_MEM_FILTER_NODES
;
2352 struct va_format vaf
;
2355 va_start(args
, fmt
);
2360 pr_warn("%pV", &vaf
);
2365 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2366 current
->comm
, order
, gfp_mask
);
2369 if (!should_suppress_show_mem())
2373 static inline struct page
*
2374 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2375 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
2379 *did_some_progress
= 0;
2382 * Acquire the oom lock. If that fails, somebody else is
2383 * making progress for us.
2385 if (!mutex_trylock(&oom_lock
)) {
2386 *did_some_progress
= 1;
2387 schedule_timeout_uninterruptible(1);
2392 * Go through the zonelist yet one more time, keep very high watermark
2393 * here, this is only to catch a parallel oom killing, we must fail if
2394 * we're still under heavy pressure.
2396 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
2397 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
2401 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2402 /* Coredumps can quickly deplete all memory reserves */
2403 if (current
->flags
& PF_DUMPCORE
)
2405 /* The OOM killer will not help higher order allocs */
2406 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2408 /* The OOM killer does not needlessly kill tasks for lowmem */
2409 if (ac
->high_zoneidx
< ZONE_NORMAL
)
2411 /* The OOM killer does not compensate for IO-less reclaim */
2412 if (!(gfp_mask
& __GFP_FS
)) {
2414 * XXX: Page reclaim didn't yield anything,
2415 * and the OOM killer can't be invoked, but
2416 * keep looping as per tradition.
2418 *did_some_progress
= 1;
2421 if (pm_suspended_storage())
2423 /* The OOM killer may not free memory on a specific node */
2424 if (gfp_mask
& __GFP_THISNODE
)
2427 /* Exhausted what can be done so it's blamo time */
2428 if (out_of_memory(ac
->zonelist
, gfp_mask
, order
, ac
->nodemask
, false)
2429 || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
))
2430 *did_some_progress
= 1;
2432 mutex_unlock(&oom_lock
);
2436 #ifdef CONFIG_COMPACTION
2437 /* Try memory compaction for high-order allocations before reclaim */
2438 static struct page
*
2439 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2440 int alloc_flags
, const struct alloc_context
*ac
,
2441 enum migrate_mode mode
, int *contended_compaction
,
2442 bool *deferred_compaction
)
2444 unsigned long compact_result
;
2450 current
->flags
|= PF_MEMALLOC
;
2451 compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
2452 mode
, contended_compaction
);
2453 current
->flags
&= ~PF_MEMALLOC
;
2455 switch (compact_result
) {
2456 case COMPACT_DEFERRED
:
2457 *deferred_compaction
= true;
2459 case COMPACT_SKIPPED
:
2466 * At least in one zone compaction wasn't deferred or skipped, so let's
2467 * count a compaction stall
2469 count_vm_event(COMPACTSTALL
);
2471 page
= get_page_from_freelist(gfp_mask
, order
,
2472 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2475 struct zone
*zone
= page_zone(page
);
2477 zone
->compact_blockskip_flush
= false;
2478 compaction_defer_reset(zone
, order
, true);
2479 count_vm_event(COMPACTSUCCESS
);
2484 * It's bad if compaction run occurs and fails. The most likely reason
2485 * is that pages exist, but not enough to satisfy watermarks.
2487 count_vm_event(COMPACTFAIL
);
2494 static inline struct page
*
2495 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2496 int alloc_flags
, const struct alloc_context
*ac
,
2497 enum migrate_mode mode
, int *contended_compaction
,
2498 bool *deferred_compaction
)
2502 #endif /* CONFIG_COMPACTION */
2504 /* Perform direct synchronous page reclaim */
2506 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
2507 const struct alloc_context
*ac
)
2509 struct reclaim_state reclaim_state
;
2514 /* We now go into synchronous reclaim */
2515 cpuset_memory_pressure_bump();
2516 current
->flags
|= PF_MEMALLOC
;
2517 lockdep_set_current_reclaim_state(gfp_mask
);
2518 reclaim_state
.reclaimed_slab
= 0;
2519 current
->reclaim_state
= &reclaim_state
;
2521 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
2524 current
->reclaim_state
= NULL
;
2525 lockdep_clear_current_reclaim_state();
2526 current
->flags
&= ~PF_MEMALLOC
;
2533 /* The really slow allocator path where we enter direct reclaim */
2534 static inline struct page
*
2535 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2536 int alloc_flags
, const struct alloc_context
*ac
,
2537 unsigned long *did_some_progress
)
2539 struct page
*page
= NULL
;
2540 bool drained
= false;
2542 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
2543 if (unlikely(!(*did_some_progress
)))
2546 /* After successful reclaim, reconsider all zones for allocation */
2547 if (IS_ENABLED(CONFIG_NUMA
))
2548 zlc_clear_zones_full(ac
->zonelist
);
2551 page
= get_page_from_freelist(gfp_mask
, order
,
2552 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2555 * If an allocation failed after direct reclaim, it could be because
2556 * pages are pinned on the per-cpu lists. Drain them and try again
2558 if (!page
&& !drained
) {
2559 drain_all_pages(NULL
);
2568 * This is called in the allocator slow-path if the allocation request is of
2569 * sufficient urgency to ignore watermarks and take other desperate measures
2571 static inline struct page
*
2572 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2573 const struct alloc_context
*ac
)
2578 page
= get_page_from_freelist(gfp_mask
, order
,
2579 ALLOC_NO_WATERMARKS
, ac
);
2581 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2582 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
,
2584 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2589 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
2594 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2595 ac
->high_zoneidx
, ac
->nodemask
)
2596 wakeup_kswapd(zone
, order
, zone_idx(ac
->preferred_zone
));
2600 gfp_to_alloc_flags(gfp_t gfp_mask
)
2602 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2603 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2605 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2606 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2609 * The caller may dip into page reserves a bit more if the caller
2610 * cannot run direct reclaim, or if the caller has realtime scheduling
2611 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2612 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2614 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2618 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2619 * if it can't schedule.
2621 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2622 alloc_flags
|= ALLOC_HARDER
;
2624 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2625 * comment for __cpuset_node_allowed().
2627 alloc_flags
&= ~ALLOC_CPUSET
;
2628 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2629 alloc_flags
|= ALLOC_HARDER
;
2631 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2632 if (gfp_mask
& __GFP_MEMALLOC
)
2633 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2634 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2635 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2636 else if (!in_interrupt() &&
2637 ((current
->flags
& PF_MEMALLOC
) ||
2638 unlikely(test_thread_flag(TIF_MEMDIE
))))
2639 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2642 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2643 alloc_flags
|= ALLOC_CMA
;
2648 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2650 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2653 static inline struct page
*
2654 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2655 struct alloc_context
*ac
)
2657 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2658 struct page
*page
= NULL
;
2660 unsigned long pages_reclaimed
= 0;
2661 unsigned long did_some_progress
;
2662 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2663 bool deferred_compaction
= false;
2664 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2667 * In the slowpath, we sanity check order to avoid ever trying to
2668 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2669 * be using allocators in order of preference for an area that is
2672 if (order
>= MAX_ORDER
) {
2673 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2678 * If this allocation cannot block and it is for a specific node, then
2679 * fail early. There's no need to wakeup kswapd or retry for a
2680 * speculative node-specific allocation.
2682 if (IS_ENABLED(CONFIG_NUMA
) && (gfp_mask
& __GFP_THISNODE
) && !wait
)
2686 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2687 wake_all_kswapds(order
, ac
);
2690 * OK, we're below the kswapd watermark and have kicked background
2691 * reclaim. Now things get more complex, so set up alloc_flags according
2692 * to how we want to proceed.
2694 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2697 * Find the true preferred zone if the allocation is unconstrained by
2700 if (!(alloc_flags
& ALLOC_CPUSET
) && !ac
->nodemask
) {
2701 struct zoneref
*preferred_zoneref
;
2702 preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
2703 ac
->high_zoneidx
, NULL
, &ac
->preferred_zone
);
2704 ac
->classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2707 /* This is the last chance, in general, before the goto nopage. */
2708 page
= get_page_from_freelist(gfp_mask
, order
,
2709 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2713 /* Allocate without watermarks if the context allows */
2714 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2716 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2717 * the allocation is high priority and these type of
2718 * allocations are system rather than user orientated
2720 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2722 page
= __alloc_pages_high_priority(gfp_mask
, order
, ac
);
2729 /* Atomic allocations - we can't balance anything */
2732 * All existing users of the deprecated __GFP_NOFAIL are
2733 * blockable, so warn of any new users that actually allow this
2734 * type of allocation to fail.
2736 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2740 /* Avoid recursion of direct reclaim */
2741 if (current
->flags
& PF_MEMALLOC
)
2744 /* Avoid allocations with no watermarks from looping endlessly */
2745 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2749 * Try direct compaction. The first pass is asynchronous. Subsequent
2750 * attempts after direct reclaim are synchronous
2752 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
2754 &contended_compaction
,
2755 &deferred_compaction
);
2759 /* Checks for THP-specific high-order allocations */
2760 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2762 * If compaction is deferred for high-order allocations, it is
2763 * because sync compaction recently failed. If this is the case
2764 * and the caller requested a THP allocation, we do not want
2765 * to heavily disrupt the system, so we fail the allocation
2766 * instead of entering direct reclaim.
2768 if (deferred_compaction
)
2772 * In all zones where compaction was attempted (and not
2773 * deferred or skipped), lock contention has been detected.
2774 * For THP allocation we do not want to disrupt the others
2775 * so we fallback to base pages instead.
2777 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2781 * If compaction was aborted due to need_resched(), we do not
2782 * want to further increase allocation latency, unless it is
2783 * khugepaged trying to collapse.
2785 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2786 && !(current
->flags
& PF_KTHREAD
))
2791 * It can become very expensive to allocate transparent hugepages at
2792 * fault, so use asynchronous memory compaction for THP unless it is
2793 * khugepaged trying to collapse.
2795 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2796 (current
->flags
& PF_KTHREAD
))
2797 migration_mode
= MIGRATE_SYNC_LIGHT
;
2799 /* Try direct reclaim and then allocating */
2800 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
2801 &did_some_progress
);
2805 /* Do not loop if specifically requested */
2806 if (gfp_mask
& __GFP_NORETRY
)
2809 /* Keep reclaiming pages as long as there is reasonable progress */
2810 pages_reclaimed
+= did_some_progress
;
2811 if ((did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
) ||
2812 ((gfp_mask
& __GFP_REPEAT
) && pages_reclaimed
< (1 << order
))) {
2813 /* Wait for some write requests to complete then retry */
2814 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2818 /* Reclaim has failed us, start killing things */
2819 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
2823 /* Retry as long as the OOM killer is making progress */
2824 if (did_some_progress
)
2829 * High-order allocations do not necessarily loop after
2830 * direct reclaim and reclaim/compaction depends on compaction
2831 * being called after reclaim so call directly if necessary
2833 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
,
2835 &contended_compaction
,
2836 &deferred_compaction
);
2840 warn_alloc_failed(gfp_mask
, order
, NULL
);
2846 * This is the 'heart' of the zoned buddy allocator.
2849 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2850 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2852 struct zoneref
*preferred_zoneref
;
2853 struct page
*page
= NULL
;
2854 unsigned int cpuset_mems_cookie
;
2855 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2856 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
2857 struct alloc_context ac
= {
2858 .high_zoneidx
= gfp_zone(gfp_mask
),
2859 .nodemask
= nodemask
,
2860 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
2863 gfp_mask
&= gfp_allowed_mask
;
2865 lockdep_trace_alloc(gfp_mask
);
2867 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2869 if (should_fail_alloc_page(gfp_mask
, order
))
2873 * Check the zones suitable for the gfp_mask contain at least one
2874 * valid zone. It's possible to have an empty zonelist as a result
2875 * of __GFP_THISNODE and a memoryless node
2877 if (unlikely(!zonelist
->_zonerefs
->zone
))
2880 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
2881 alloc_flags
|= ALLOC_CMA
;
2884 cpuset_mems_cookie
= read_mems_allowed_begin();
2886 /* We set it here, as __alloc_pages_slowpath might have changed it */
2887 ac
.zonelist
= zonelist
;
2888 /* The preferred zone is used for statistics later */
2889 preferred_zoneref
= first_zones_zonelist(ac
.zonelist
, ac
.high_zoneidx
,
2890 ac
.nodemask
? : &cpuset_current_mems_allowed
,
2891 &ac
.preferred_zone
);
2892 if (!ac
.preferred_zone
)
2894 ac
.classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2896 /* First allocation attempt */
2897 alloc_mask
= gfp_mask
|__GFP_HARDWALL
;
2898 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
2899 if (unlikely(!page
)) {
2901 * Runtime PM, block IO and its error handling path
2902 * can deadlock because I/O on the device might not
2905 alloc_mask
= memalloc_noio_flags(gfp_mask
);
2907 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
2910 if (kmemcheck_enabled
&& page
)
2911 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2913 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
2917 * When updating a task's mems_allowed, it is possible to race with
2918 * parallel threads in such a way that an allocation can fail while
2919 * the mask is being updated. If a page allocation is about to fail,
2920 * check if the cpuset changed during allocation and if so, retry.
2922 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2927 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2930 * Common helper functions.
2932 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2937 * __get_free_pages() returns a 32-bit address, which cannot represent
2940 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2942 page
= alloc_pages(gfp_mask
, order
);
2945 return (unsigned long) page_address(page
);
2947 EXPORT_SYMBOL(__get_free_pages
);
2949 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2951 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2953 EXPORT_SYMBOL(get_zeroed_page
);
2955 void __free_pages(struct page
*page
, unsigned int order
)
2957 if (put_page_testzero(page
)) {
2959 free_hot_cold_page(page
, false);
2961 __free_pages_ok(page
, order
);
2965 EXPORT_SYMBOL(__free_pages
);
2967 void free_pages(unsigned long addr
, unsigned int order
)
2970 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2971 __free_pages(virt_to_page((void *)addr
), order
);
2975 EXPORT_SYMBOL(free_pages
);
2979 * An arbitrary-length arbitrary-offset area of memory which resides
2980 * within a 0 or higher order page. Multiple fragments within that page
2981 * are individually refcounted, in the page's reference counter.
2983 * The page_frag functions below provide a simple allocation framework for
2984 * page fragments. This is used by the network stack and network device
2985 * drivers to provide a backing region of memory for use as either an
2986 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
2988 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
2991 struct page
*page
= NULL
;
2992 gfp_t gfp
= gfp_mask
;
2994 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
2995 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
2997 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
2998 PAGE_FRAG_CACHE_MAX_ORDER
);
2999 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3001 if (unlikely(!page
))
3002 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3004 nc
->va
= page
? page_address(page
) : NULL
;
3009 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3010 unsigned int fragsz
, gfp_t gfp_mask
)
3012 unsigned int size
= PAGE_SIZE
;
3016 if (unlikely(!nc
->va
)) {
3018 page
= __page_frag_refill(nc
, gfp_mask
);
3022 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3023 /* if size can vary use size else just use PAGE_SIZE */
3026 /* Even if we own the page, we do not use atomic_set().
3027 * This would break get_page_unless_zero() users.
3029 atomic_add(size
- 1, &page
->_count
);
3031 /* reset page count bias and offset to start of new frag */
3032 nc
->pfmemalloc
= page
->pfmemalloc
;
3033 nc
->pagecnt_bias
= size
;
3037 offset
= nc
->offset
- fragsz
;
3038 if (unlikely(offset
< 0)) {
3039 page
= virt_to_page(nc
->va
);
3041 if (!atomic_sub_and_test(nc
->pagecnt_bias
, &page
->_count
))
3044 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3045 /* if size can vary use size else just use PAGE_SIZE */
3048 /* OK, page count is 0, we can safely set it */
3049 atomic_set(&page
->_count
, size
);
3051 /* reset page count bias and offset to start of new frag */
3052 nc
->pagecnt_bias
= size
;
3053 offset
= size
- fragsz
;
3057 nc
->offset
= offset
;
3059 return nc
->va
+ offset
;
3061 EXPORT_SYMBOL(__alloc_page_frag
);
3064 * Frees a page fragment allocated out of either a compound or order 0 page.
3066 void __free_page_frag(void *addr
)
3068 struct page
*page
= virt_to_head_page(addr
);
3070 if (unlikely(put_page_testzero(page
)))
3071 __free_pages_ok(page
, compound_order(page
));
3073 EXPORT_SYMBOL(__free_page_frag
);
3076 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3077 * of the current memory cgroup.
3079 * It should be used when the caller would like to use kmalloc, but since the
3080 * allocation is large, it has to fall back to the page allocator.
3082 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
3085 struct mem_cgroup
*memcg
= NULL
;
3087 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
3089 page
= alloc_pages(gfp_mask
, order
);
3090 memcg_kmem_commit_charge(page
, memcg
, order
);
3094 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
3097 struct mem_cgroup
*memcg
= NULL
;
3099 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
3101 page
= alloc_pages_node(nid
, gfp_mask
, order
);
3102 memcg_kmem_commit_charge(page
, memcg
, order
);
3107 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3110 void __free_kmem_pages(struct page
*page
, unsigned int order
)
3112 memcg_kmem_uncharge_pages(page
, order
);
3113 __free_pages(page
, order
);
3116 void free_kmem_pages(unsigned long addr
, unsigned int order
)
3119 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3120 __free_kmem_pages(virt_to_page((void *)addr
), order
);
3124 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
3127 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3128 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3130 split_page(virt_to_page((void *)addr
), order
);
3131 while (used
< alloc_end
) {
3136 return (void *)addr
;
3140 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3141 * @size: the number of bytes to allocate
3142 * @gfp_mask: GFP flags for the allocation
3144 * This function is similar to alloc_pages(), except that it allocates the
3145 * minimum number of pages to satisfy the request. alloc_pages() can only
3146 * allocate memory in power-of-two pages.
3148 * This function is also limited by MAX_ORDER.
3150 * Memory allocated by this function must be released by free_pages_exact().
3152 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3154 unsigned int order
= get_order(size
);
3157 addr
= __get_free_pages(gfp_mask
, order
);
3158 return make_alloc_exact(addr
, order
, size
);
3160 EXPORT_SYMBOL(alloc_pages_exact
);
3163 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3165 * @nid: the preferred node ID where memory should be allocated
3166 * @size: the number of bytes to allocate
3167 * @gfp_mask: GFP flags for the allocation
3169 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3171 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3174 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3176 unsigned order
= get_order(size
);
3177 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3180 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3184 * free_pages_exact - release memory allocated via alloc_pages_exact()
3185 * @virt: the value returned by alloc_pages_exact.
3186 * @size: size of allocation, same value as passed to alloc_pages_exact().
3188 * Release the memory allocated by a previous call to alloc_pages_exact.
3190 void free_pages_exact(void *virt
, size_t size
)
3192 unsigned long addr
= (unsigned long)virt
;
3193 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3195 while (addr
< end
) {
3200 EXPORT_SYMBOL(free_pages_exact
);
3203 * nr_free_zone_pages - count number of pages beyond high watermark
3204 * @offset: The zone index of the highest zone
3206 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3207 * high watermark within all zones at or below a given zone index. For each
3208 * zone, the number of pages is calculated as:
3209 * managed_pages - high_pages
3211 static unsigned long nr_free_zone_pages(int offset
)
3216 /* Just pick one node, since fallback list is circular */
3217 unsigned long sum
= 0;
3219 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3221 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3222 unsigned long size
= zone
->managed_pages
;
3223 unsigned long high
= high_wmark_pages(zone
);
3232 * nr_free_buffer_pages - count number of pages beyond high watermark
3234 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3235 * watermark within ZONE_DMA and ZONE_NORMAL.
3237 unsigned long nr_free_buffer_pages(void)
3239 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3241 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3244 * nr_free_pagecache_pages - count number of pages beyond high watermark
3246 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3247 * high watermark within all zones.
3249 unsigned long nr_free_pagecache_pages(void)
3251 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3254 static inline void show_node(struct zone
*zone
)
3256 if (IS_ENABLED(CONFIG_NUMA
))
3257 printk("Node %d ", zone_to_nid(zone
));
3260 void si_meminfo(struct sysinfo
*val
)
3262 val
->totalram
= totalram_pages
;
3263 val
->sharedram
= global_page_state(NR_SHMEM
);
3264 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3265 val
->bufferram
= nr_blockdev_pages();
3266 val
->totalhigh
= totalhigh_pages
;
3267 val
->freehigh
= nr_free_highpages();
3268 val
->mem_unit
= PAGE_SIZE
;
3271 EXPORT_SYMBOL(si_meminfo
);
3274 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3276 int zone_type
; /* needs to be signed */
3277 unsigned long managed_pages
= 0;
3278 pg_data_t
*pgdat
= NODE_DATA(nid
);
3280 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3281 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3282 val
->totalram
= managed_pages
;
3283 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3284 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3285 #ifdef CONFIG_HIGHMEM
3286 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3287 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3293 val
->mem_unit
= PAGE_SIZE
;
3298 * Determine whether the node should be displayed or not, depending on whether
3299 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3301 bool skip_free_areas_node(unsigned int flags
, int nid
)
3304 unsigned int cpuset_mems_cookie
;
3306 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3310 cpuset_mems_cookie
= read_mems_allowed_begin();
3311 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3312 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3317 #define K(x) ((x) << (PAGE_SHIFT-10))
3319 static void show_migration_types(unsigned char type
)
3321 static const char types
[MIGRATE_TYPES
] = {
3322 [MIGRATE_UNMOVABLE
] = 'U',
3323 [MIGRATE_RECLAIMABLE
] = 'E',
3324 [MIGRATE_MOVABLE
] = 'M',
3325 [MIGRATE_RESERVE
] = 'R',
3327 [MIGRATE_CMA
] = 'C',
3329 #ifdef CONFIG_MEMORY_ISOLATION
3330 [MIGRATE_ISOLATE
] = 'I',
3333 char tmp
[MIGRATE_TYPES
+ 1];
3337 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3338 if (type
& (1 << i
))
3343 printk("(%s) ", tmp
);
3347 * Show free area list (used inside shift_scroll-lock stuff)
3348 * We also calculate the percentage fragmentation. We do this by counting the
3349 * memory on each free list with the exception of the first item on the list.
3352 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3355 void show_free_areas(unsigned int filter
)
3357 unsigned long free_pcp
= 0;
3361 for_each_populated_zone(zone
) {
3362 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3365 for_each_online_cpu(cpu
)
3366 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
3369 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3370 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3371 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3372 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3373 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3374 " free:%lu free_pcp:%lu free_cma:%lu\n",
3375 global_page_state(NR_ACTIVE_ANON
),
3376 global_page_state(NR_INACTIVE_ANON
),
3377 global_page_state(NR_ISOLATED_ANON
),
3378 global_page_state(NR_ACTIVE_FILE
),
3379 global_page_state(NR_INACTIVE_FILE
),
3380 global_page_state(NR_ISOLATED_FILE
),
3381 global_page_state(NR_UNEVICTABLE
),
3382 global_page_state(NR_FILE_DIRTY
),
3383 global_page_state(NR_WRITEBACK
),
3384 global_page_state(NR_UNSTABLE_NFS
),
3385 global_page_state(NR_SLAB_RECLAIMABLE
),
3386 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3387 global_page_state(NR_FILE_MAPPED
),
3388 global_page_state(NR_SHMEM
),
3389 global_page_state(NR_PAGETABLE
),
3390 global_page_state(NR_BOUNCE
),
3391 global_page_state(NR_FREE_PAGES
),
3393 global_page_state(NR_FREE_CMA_PAGES
));
3395 for_each_populated_zone(zone
) {
3398 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3402 for_each_online_cpu(cpu
)
3403 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
3411 " active_anon:%lukB"
3412 " inactive_anon:%lukB"
3413 " active_file:%lukB"
3414 " inactive_file:%lukB"
3415 " unevictable:%lukB"
3416 " isolated(anon):%lukB"
3417 " isolated(file):%lukB"
3425 " slab_reclaimable:%lukB"
3426 " slab_unreclaimable:%lukB"
3427 " kernel_stack:%lukB"
3434 " writeback_tmp:%lukB"
3435 " pages_scanned:%lu"
3436 " all_unreclaimable? %s"
3439 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3440 K(min_wmark_pages(zone
)),
3441 K(low_wmark_pages(zone
)),
3442 K(high_wmark_pages(zone
)),
3443 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3444 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3445 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3446 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3447 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3448 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3449 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3450 K(zone
->present_pages
),
3451 K(zone
->managed_pages
),
3452 K(zone_page_state(zone
, NR_MLOCK
)),
3453 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3454 K(zone_page_state(zone
, NR_WRITEBACK
)),
3455 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3456 K(zone_page_state(zone
, NR_SHMEM
)),
3457 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3458 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3459 zone_page_state(zone
, NR_KERNEL_STACK
) *
3461 K(zone_page_state(zone
, NR_PAGETABLE
)),
3462 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3463 K(zone_page_state(zone
, NR_BOUNCE
)),
3465 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
3466 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3467 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3468 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3469 (!zone_reclaimable(zone
) ? "yes" : "no")
3471 printk("lowmem_reserve[]:");
3472 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3473 printk(" %ld", zone
->lowmem_reserve
[i
]);
3477 for_each_populated_zone(zone
) {
3478 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3479 unsigned char types
[MAX_ORDER
];
3481 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3484 printk("%s: ", zone
->name
);
3486 spin_lock_irqsave(&zone
->lock
, flags
);
3487 for (order
= 0; order
< MAX_ORDER
; order
++) {
3488 struct free_area
*area
= &zone
->free_area
[order
];
3491 nr
[order
] = area
->nr_free
;
3492 total
+= nr
[order
] << order
;
3495 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3496 if (!list_empty(&area
->free_list
[type
]))
3497 types
[order
] |= 1 << type
;
3500 spin_unlock_irqrestore(&zone
->lock
, flags
);
3501 for (order
= 0; order
< MAX_ORDER
; order
++) {
3502 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3504 show_migration_types(types
[order
]);
3506 printk("= %lukB\n", K(total
));
3509 hugetlb_show_meminfo();
3511 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3513 show_swap_cache_info();
3516 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3518 zoneref
->zone
= zone
;
3519 zoneref
->zone_idx
= zone_idx(zone
);
3523 * Builds allocation fallback zone lists.
3525 * Add all populated zones of a node to the zonelist.
3527 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3531 enum zone_type zone_type
= MAX_NR_ZONES
;
3535 zone
= pgdat
->node_zones
+ zone_type
;
3536 if (populated_zone(zone
)) {
3537 zoneref_set_zone(zone
,
3538 &zonelist
->_zonerefs
[nr_zones
++]);
3539 check_highest_zone(zone_type
);
3541 } while (zone_type
);
3549 * 0 = automatic detection of better ordering.
3550 * 1 = order by ([node] distance, -zonetype)
3551 * 2 = order by (-zonetype, [node] distance)
3553 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3554 * the same zonelist. So only NUMA can configure this param.
3556 #define ZONELIST_ORDER_DEFAULT 0
3557 #define ZONELIST_ORDER_NODE 1
3558 #define ZONELIST_ORDER_ZONE 2
3560 /* zonelist order in the kernel.
3561 * set_zonelist_order() will set this to NODE or ZONE.
3563 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3564 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3568 /* The value user specified ....changed by config */
3569 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3570 /* string for sysctl */
3571 #define NUMA_ZONELIST_ORDER_LEN 16
3572 char numa_zonelist_order
[16] = "default";
3575 * interface for configure zonelist ordering.
3576 * command line option "numa_zonelist_order"
3577 * = "[dD]efault - default, automatic configuration.
3578 * = "[nN]ode - order by node locality, then by zone within node
3579 * = "[zZ]one - order by zone, then by locality within zone
3582 static int __parse_numa_zonelist_order(char *s
)
3584 if (*s
== 'd' || *s
== 'D') {
3585 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3586 } else if (*s
== 'n' || *s
== 'N') {
3587 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3588 } else if (*s
== 'z' || *s
== 'Z') {
3589 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3592 "Ignoring invalid numa_zonelist_order value: "
3599 static __init
int setup_numa_zonelist_order(char *s
)
3606 ret
= __parse_numa_zonelist_order(s
);
3608 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3612 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3615 * sysctl handler for numa_zonelist_order
3617 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3618 void __user
*buffer
, size_t *length
,
3621 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3623 static DEFINE_MUTEX(zl_order_mutex
);
3625 mutex_lock(&zl_order_mutex
);
3627 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3631 strcpy(saved_string
, (char *)table
->data
);
3633 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3637 int oldval
= user_zonelist_order
;
3639 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3642 * bogus value. restore saved string
3644 strncpy((char *)table
->data
, saved_string
,
3645 NUMA_ZONELIST_ORDER_LEN
);
3646 user_zonelist_order
= oldval
;
3647 } else if (oldval
!= user_zonelist_order
) {
3648 mutex_lock(&zonelists_mutex
);
3649 build_all_zonelists(NULL
, NULL
);
3650 mutex_unlock(&zonelists_mutex
);
3654 mutex_unlock(&zl_order_mutex
);
3659 #define MAX_NODE_LOAD (nr_online_nodes)
3660 static int node_load
[MAX_NUMNODES
];
3663 * find_next_best_node - find the next node that should appear in a given node's fallback list
3664 * @node: node whose fallback list we're appending
3665 * @used_node_mask: nodemask_t of already used nodes
3667 * We use a number of factors to determine which is the next node that should
3668 * appear on a given node's fallback list. The node should not have appeared
3669 * already in @node's fallback list, and it should be the next closest node
3670 * according to the distance array (which contains arbitrary distance values
3671 * from each node to each node in the system), and should also prefer nodes
3672 * with no CPUs, since presumably they'll have very little allocation pressure
3673 * on them otherwise.
3674 * It returns -1 if no node is found.
3676 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3679 int min_val
= INT_MAX
;
3680 int best_node
= NUMA_NO_NODE
;
3681 const struct cpumask
*tmp
= cpumask_of_node(0);
3683 /* Use the local node if we haven't already */
3684 if (!node_isset(node
, *used_node_mask
)) {
3685 node_set(node
, *used_node_mask
);
3689 for_each_node_state(n
, N_MEMORY
) {
3691 /* Don't want a node to appear more than once */
3692 if (node_isset(n
, *used_node_mask
))
3695 /* Use the distance array to find the distance */
3696 val
= node_distance(node
, n
);
3698 /* Penalize nodes under us ("prefer the next node") */
3701 /* Give preference to headless and unused nodes */
3702 tmp
= cpumask_of_node(n
);
3703 if (!cpumask_empty(tmp
))
3704 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3706 /* Slight preference for less loaded node */
3707 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3708 val
+= node_load
[n
];
3710 if (val
< min_val
) {
3717 node_set(best_node
, *used_node_mask
);
3724 * Build zonelists ordered by node and zones within node.
3725 * This results in maximum locality--normal zone overflows into local
3726 * DMA zone, if any--but risks exhausting DMA zone.
3728 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3731 struct zonelist
*zonelist
;
3733 zonelist
= &pgdat
->node_zonelists
[0];
3734 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3736 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3737 zonelist
->_zonerefs
[j
].zone
= NULL
;
3738 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3742 * Build gfp_thisnode zonelists
3744 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3747 struct zonelist
*zonelist
;
3749 zonelist
= &pgdat
->node_zonelists
[1];
3750 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3751 zonelist
->_zonerefs
[j
].zone
= NULL
;
3752 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3756 * Build zonelists ordered by zone and nodes within zones.
3757 * This results in conserving DMA zone[s] until all Normal memory is
3758 * exhausted, but results in overflowing to remote node while memory
3759 * may still exist in local DMA zone.
3761 static int node_order
[MAX_NUMNODES
];
3763 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3766 int zone_type
; /* needs to be signed */
3768 struct zonelist
*zonelist
;
3770 zonelist
= &pgdat
->node_zonelists
[0];
3772 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3773 for (j
= 0; j
< nr_nodes
; j
++) {
3774 node
= node_order
[j
];
3775 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3776 if (populated_zone(z
)) {
3778 &zonelist
->_zonerefs
[pos
++]);
3779 check_highest_zone(zone_type
);
3783 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3784 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3787 #if defined(CONFIG_64BIT)
3789 * Devices that require DMA32/DMA are relatively rare and do not justify a
3790 * penalty to every machine in case the specialised case applies. Default
3791 * to Node-ordering on 64-bit NUMA machines
3793 static int default_zonelist_order(void)
3795 return ZONELIST_ORDER_NODE
;
3799 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3800 * by the kernel. If processes running on node 0 deplete the low memory zone
3801 * then reclaim will occur more frequency increasing stalls and potentially
3802 * be easier to OOM if a large percentage of the zone is under writeback or
3803 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3804 * Hence, default to zone ordering on 32-bit.
3806 static int default_zonelist_order(void)
3808 return ZONELIST_ORDER_ZONE
;
3810 #endif /* CONFIG_64BIT */
3812 static void set_zonelist_order(void)
3814 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3815 current_zonelist_order
= default_zonelist_order();
3817 current_zonelist_order
= user_zonelist_order
;
3820 static void build_zonelists(pg_data_t
*pgdat
)
3824 nodemask_t used_mask
;
3825 int local_node
, prev_node
;
3826 struct zonelist
*zonelist
;
3827 int order
= current_zonelist_order
;
3829 /* initialize zonelists */
3830 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3831 zonelist
= pgdat
->node_zonelists
+ i
;
3832 zonelist
->_zonerefs
[0].zone
= NULL
;
3833 zonelist
->_zonerefs
[0].zone_idx
= 0;
3836 /* NUMA-aware ordering of nodes */
3837 local_node
= pgdat
->node_id
;
3838 load
= nr_online_nodes
;
3839 prev_node
= local_node
;
3840 nodes_clear(used_mask
);
3842 memset(node_order
, 0, sizeof(node_order
));
3845 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3847 * We don't want to pressure a particular node.
3848 * So adding penalty to the first node in same
3849 * distance group to make it round-robin.
3851 if (node_distance(local_node
, node
) !=
3852 node_distance(local_node
, prev_node
))
3853 node_load
[node
] = load
;
3857 if (order
== ZONELIST_ORDER_NODE
)
3858 build_zonelists_in_node_order(pgdat
, node
);
3860 node_order
[j
++] = node
; /* remember order */
3863 if (order
== ZONELIST_ORDER_ZONE
) {
3864 /* calculate node order -- i.e., DMA last! */
3865 build_zonelists_in_zone_order(pgdat
, j
);
3868 build_thisnode_zonelists(pgdat
);
3871 /* Construct the zonelist performance cache - see further mmzone.h */
3872 static void build_zonelist_cache(pg_data_t
*pgdat
)
3874 struct zonelist
*zonelist
;
3875 struct zonelist_cache
*zlc
;
3878 zonelist
= &pgdat
->node_zonelists
[0];
3879 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3880 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3881 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3882 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3885 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3887 * Return node id of node used for "local" allocations.
3888 * I.e., first node id of first zone in arg node's generic zonelist.
3889 * Used for initializing percpu 'numa_mem', which is used primarily
3890 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3892 int local_memory_node(int node
)
3896 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3897 gfp_zone(GFP_KERNEL
),
3904 #else /* CONFIG_NUMA */
3906 static void set_zonelist_order(void)
3908 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3911 static void build_zonelists(pg_data_t
*pgdat
)
3913 int node
, local_node
;
3915 struct zonelist
*zonelist
;
3917 local_node
= pgdat
->node_id
;
3919 zonelist
= &pgdat
->node_zonelists
[0];
3920 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3923 * Now we build the zonelist so that it contains the zones
3924 * of all the other nodes.
3925 * We don't want to pressure a particular node, so when
3926 * building the zones for node N, we make sure that the
3927 * zones coming right after the local ones are those from
3928 * node N+1 (modulo N)
3930 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3931 if (!node_online(node
))
3933 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3935 for (node
= 0; node
< local_node
; node
++) {
3936 if (!node_online(node
))
3938 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3941 zonelist
->_zonerefs
[j
].zone
= NULL
;
3942 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3945 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3946 static void build_zonelist_cache(pg_data_t
*pgdat
)
3948 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3951 #endif /* CONFIG_NUMA */
3954 * Boot pageset table. One per cpu which is going to be used for all
3955 * zones and all nodes. The parameters will be set in such a way
3956 * that an item put on a list will immediately be handed over to
3957 * the buddy list. This is safe since pageset manipulation is done
3958 * with interrupts disabled.
3960 * The boot_pagesets must be kept even after bootup is complete for
3961 * unused processors and/or zones. They do play a role for bootstrapping
3962 * hotplugged processors.
3964 * zoneinfo_show() and maybe other functions do
3965 * not check if the processor is online before following the pageset pointer.
3966 * Other parts of the kernel may not check if the zone is available.
3968 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3969 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3970 static void setup_zone_pageset(struct zone
*zone
);
3973 * Global mutex to protect against size modification of zonelists
3974 * as well as to serialize pageset setup for the new populated zone.
3976 DEFINE_MUTEX(zonelists_mutex
);
3978 /* return values int ....just for stop_machine() */
3979 static int __build_all_zonelists(void *data
)
3983 pg_data_t
*self
= data
;
3986 memset(node_load
, 0, sizeof(node_load
));
3989 if (self
&& !node_online(self
->node_id
)) {
3990 build_zonelists(self
);
3991 build_zonelist_cache(self
);
3994 for_each_online_node(nid
) {
3995 pg_data_t
*pgdat
= NODE_DATA(nid
);
3997 build_zonelists(pgdat
);
3998 build_zonelist_cache(pgdat
);
4002 * Initialize the boot_pagesets that are going to be used
4003 * for bootstrapping processors. The real pagesets for
4004 * each zone will be allocated later when the per cpu
4005 * allocator is available.
4007 * boot_pagesets are used also for bootstrapping offline
4008 * cpus if the system is already booted because the pagesets
4009 * are needed to initialize allocators on a specific cpu too.
4010 * F.e. the percpu allocator needs the page allocator which
4011 * needs the percpu allocator in order to allocate its pagesets
4012 * (a chicken-egg dilemma).
4014 for_each_possible_cpu(cpu
) {
4015 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4017 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4019 * We now know the "local memory node" for each node--
4020 * i.e., the node of the first zone in the generic zonelist.
4021 * Set up numa_mem percpu variable for on-line cpus. During
4022 * boot, only the boot cpu should be on-line; we'll init the
4023 * secondary cpus' numa_mem as they come on-line. During
4024 * node/memory hotplug, we'll fixup all on-line cpus.
4026 if (cpu_online(cpu
))
4027 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4034 static noinline
void __init
4035 build_all_zonelists_init(void)
4037 __build_all_zonelists(NULL
);
4038 mminit_verify_zonelist();
4039 cpuset_init_current_mems_allowed();
4043 * Called with zonelists_mutex held always
4044 * unless system_state == SYSTEM_BOOTING.
4046 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4047 * [we're only called with non-NULL zone through __meminit paths] and
4048 * (2) call of __init annotated helper build_all_zonelists_init
4049 * [protected by SYSTEM_BOOTING].
4051 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4053 set_zonelist_order();
4055 if (system_state
== SYSTEM_BOOTING
) {
4056 build_all_zonelists_init();
4058 #ifdef CONFIG_MEMORY_HOTPLUG
4060 setup_zone_pageset(zone
);
4062 /* we have to stop all cpus to guarantee there is no user
4064 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4065 /* cpuset refresh routine should be here */
4067 vm_total_pages
= nr_free_pagecache_pages();
4069 * Disable grouping by mobility if the number of pages in the
4070 * system is too low to allow the mechanism to work. It would be
4071 * more accurate, but expensive to check per-zone. This check is
4072 * made on memory-hotadd so a system can start with mobility
4073 * disabled and enable it later
4075 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4076 page_group_by_mobility_disabled
= 1;
4078 page_group_by_mobility_disabled
= 0;
4080 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4081 "Total pages: %ld\n",
4083 zonelist_order_name
[current_zonelist_order
],
4084 page_group_by_mobility_disabled
? "off" : "on",
4087 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4092 * Helper functions to size the waitqueue hash table.
4093 * Essentially these want to choose hash table sizes sufficiently
4094 * large so that collisions trying to wait on pages are rare.
4095 * But in fact, the number of active page waitqueues on typical
4096 * systems is ridiculously low, less than 200. So this is even
4097 * conservative, even though it seems large.
4099 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4100 * waitqueues, i.e. the size of the waitq table given the number of pages.
4102 #define PAGES_PER_WAITQUEUE 256
4104 #ifndef CONFIG_MEMORY_HOTPLUG
4105 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4107 unsigned long size
= 1;
4109 pages
/= PAGES_PER_WAITQUEUE
;
4111 while (size
< pages
)
4115 * Once we have dozens or even hundreds of threads sleeping
4116 * on IO we've got bigger problems than wait queue collision.
4117 * Limit the size of the wait table to a reasonable size.
4119 size
= min(size
, 4096UL);
4121 return max(size
, 4UL);
4125 * A zone's size might be changed by hot-add, so it is not possible to determine
4126 * a suitable size for its wait_table. So we use the maximum size now.
4128 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4130 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4131 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4132 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4134 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4135 * or more by the traditional way. (See above). It equals:
4137 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4138 * ia64(16K page size) : = ( 8G + 4M)byte.
4139 * powerpc (64K page size) : = (32G +16M)byte.
4141 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4148 * This is an integer logarithm so that shifts can be used later
4149 * to extract the more random high bits from the multiplicative
4150 * hash function before the remainder is taken.
4152 static inline unsigned long wait_table_bits(unsigned long size
)
4158 * Check if a pageblock contains reserved pages
4160 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
4164 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4165 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
4172 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4173 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4174 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4175 * higher will lead to a bigger reserve which will get freed as contiguous
4176 * blocks as reclaim kicks in
4178 static void setup_zone_migrate_reserve(struct zone
*zone
)
4180 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4182 unsigned long block_migratetype
;
4187 * Get the start pfn, end pfn and the number of blocks to reserve
4188 * We have to be careful to be aligned to pageblock_nr_pages to
4189 * make sure that we always check pfn_valid for the first page in
4192 start_pfn
= zone
->zone_start_pfn
;
4193 end_pfn
= zone_end_pfn(zone
);
4194 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4195 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4199 * Reserve blocks are generally in place to help high-order atomic
4200 * allocations that are short-lived. A min_free_kbytes value that
4201 * would result in more than 2 reserve blocks for atomic allocations
4202 * is assumed to be in place to help anti-fragmentation for the
4203 * future allocation of hugepages at runtime.
4205 reserve
= min(2, reserve
);
4206 old_reserve
= zone
->nr_migrate_reserve_block
;
4208 /* When memory hot-add, we almost always need to do nothing */
4209 if (reserve
== old_reserve
)
4211 zone
->nr_migrate_reserve_block
= reserve
;
4213 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4214 if (!pfn_valid(pfn
))
4216 page
= pfn_to_page(pfn
);
4218 /* Watch out for overlapping nodes */
4219 if (page_to_nid(page
) != zone_to_nid(zone
))
4222 block_migratetype
= get_pageblock_migratetype(page
);
4224 /* Only test what is necessary when the reserves are not met */
4227 * Blocks with reserved pages will never free, skip
4230 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4231 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4234 /* If this block is reserved, account for it */
4235 if (block_migratetype
== MIGRATE_RESERVE
) {
4240 /* Suitable for reserving if this block is movable */
4241 if (block_migratetype
== MIGRATE_MOVABLE
) {
4242 set_pageblock_migratetype(page
,
4244 move_freepages_block(zone
, page
,
4249 } else if (!old_reserve
) {
4251 * At boot time we don't need to scan the whole zone
4252 * for turning off MIGRATE_RESERVE.
4258 * If the reserve is met and this is a previous reserved block,
4261 if (block_migratetype
== MIGRATE_RESERVE
) {
4262 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4263 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4269 * Initially all pages are reserved - free ones are freed
4270 * up by free_all_bootmem() once the early boot process is
4271 * done. Non-atomic initialization, single-pass.
4273 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4274 unsigned long start_pfn
, enum memmap_context context
)
4276 unsigned long end_pfn
= start_pfn
+ size
;
4280 if (highest_memmap_pfn
< end_pfn
- 1)
4281 highest_memmap_pfn
= end_pfn
- 1;
4283 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4284 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4286 * There can be holes in boot-time mem_map[]s
4287 * handed to this function. They do not
4288 * exist on hotplugged memory.
4290 if (context
== MEMMAP_EARLY
) {
4291 if (!early_pfn_valid(pfn
))
4293 if (!early_pfn_in_nid(pfn
, nid
))
4296 __init_single_pfn(pfn
, zone
, nid
);
4300 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4302 unsigned int order
, t
;
4303 for_each_migratetype_order(order
, t
) {
4304 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4305 zone
->free_area
[order
].nr_free
= 0;
4309 #ifndef __HAVE_ARCH_MEMMAP_INIT
4310 #define memmap_init(size, nid, zone, start_pfn) \
4311 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4314 static int zone_batchsize(struct zone
*zone
)
4320 * The per-cpu-pages pools are set to around 1000th of the
4321 * size of the zone. But no more than 1/2 of a meg.
4323 * OK, so we don't know how big the cache is. So guess.
4325 batch
= zone
->managed_pages
/ 1024;
4326 if (batch
* PAGE_SIZE
> 512 * 1024)
4327 batch
= (512 * 1024) / PAGE_SIZE
;
4328 batch
/= 4; /* We effectively *= 4 below */
4333 * Clamp the batch to a 2^n - 1 value. Having a power
4334 * of 2 value was found to be more likely to have
4335 * suboptimal cache aliasing properties in some cases.
4337 * For example if 2 tasks are alternately allocating
4338 * batches of pages, one task can end up with a lot
4339 * of pages of one half of the possible page colors
4340 * and the other with pages of the other colors.
4342 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4347 /* The deferral and batching of frees should be suppressed under NOMMU
4350 * The problem is that NOMMU needs to be able to allocate large chunks
4351 * of contiguous memory as there's no hardware page translation to
4352 * assemble apparent contiguous memory from discontiguous pages.
4354 * Queueing large contiguous runs of pages for batching, however,
4355 * causes the pages to actually be freed in smaller chunks. As there
4356 * can be a significant delay between the individual batches being
4357 * recycled, this leads to the once large chunks of space being
4358 * fragmented and becoming unavailable for high-order allocations.
4365 * pcp->high and pcp->batch values are related and dependent on one another:
4366 * ->batch must never be higher then ->high.
4367 * The following function updates them in a safe manner without read side
4370 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4371 * those fields changing asynchronously (acording the the above rule).
4373 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4374 * outside of boot time (or some other assurance that no concurrent updaters
4377 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4378 unsigned long batch
)
4380 /* start with a fail safe value for batch */
4384 /* Update high, then batch, in order */
4391 /* a companion to pageset_set_high() */
4392 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4394 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4397 static void pageset_init(struct per_cpu_pageset
*p
)
4399 struct per_cpu_pages
*pcp
;
4402 memset(p
, 0, sizeof(*p
));
4406 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4407 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4410 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4413 pageset_set_batch(p
, batch
);
4417 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4418 * to the value high for the pageset p.
4420 static void pageset_set_high(struct per_cpu_pageset
*p
,
4423 unsigned long batch
= max(1UL, high
/ 4);
4424 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4425 batch
= PAGE_SHIFT
* 8;
4427 pageset_update(&p
->pcp
, high
, batch
);
4430 static void pageset_set_high_and_batch(struct zone
*zone
,
4431 struct per_cpu_pageset
*pcp
)
4433 if (percpu_pagelist_fraction
)
4434 pageset_set_high(pcp
,
4435 (zone
->managed_pages
/
4436 percpu_pagelist_fraction
));
4438 pageset_set_batch(pcp
, zone_batchsize(zone
));
4441 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4443 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4446 pageset_set_high_and_batch(zone
, pcp
);
4449 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4452 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4453 for_each_possible_cpu(cpu
)
4454 zone_pageset_init(zone
, cpu
);
4458 * Allocate per cpu pagesets and initialize them.
4459 * Before this call only boot pagesets were available.
4461 void __init
setup_per_cpu_pageset(void)
4465 for_each_populated_zone(zone
)
4466 setup_zone_pageset(zone
);
4469 static noinline __init_refok
4470 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4476 * The per-page waitqueue mechanism uses hashed waitqueues
4479 zone
->wait_table_hash_nr_entries
=
4480 wait_table_hash_nr_entries(zone_size_pages
);
4481 zone
->wait_table_bits
=
4482 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4483 alloc_size
= zone
->wait_table_hash_nr_entries
4484 * sizeof(wait_queue_head_t
);
4486 if (!slab_is_available()) {
4487 zone
->wait_table
= (wait_queue_head_t
*)
4488 memblock_virt_alloc_node_nopanic(
4489 alloc_size
, zone
->zone_pgdat
->node_id
);
4492 * This case means that a zone whose size was 0 gets new memory
4493 * via memory hot-add.
4494 * But it may be the case that a new node was hot-added. In
4495 * this case vmalloc() will not be able to use this new node's
4496 * memory - this wait_table must be initialized to use this new
4497 * node itself as well.
4498 * To use this new node's memory, further consideration will be
4501 zone
->wait_table
= vmalloc(alloc_size
);
4503 if (!zone
->wait_table
)
4506 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4507 init_waitqueue_head(zone
->wait_table
+ i
);
4512 static __meminit
void zone_pcp_init(struct zone
*zone
)
4515 * per cpu subsystem is not up at this point. The following code
4516 * relies on the ability of the linker to provide the
4517 * offset of a (static) per cpu variable into the per cpu area.
4519 zone
->pageset
= &boot_pageset
;
4521 if (populated_zone(zone
))
4522 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4523 zone
->name
, zone
->present_pages
,
4524 zone_batchsize(zone
));
4527 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4528 unsigned long zone_start_pfn
,
4530 enum memmap_context context
)
4532 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4534 ret
= zone_wait_table_init(zone
, size
);
4537 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4539 zone
->zone_start_pfn
= zone_start_pfn
;
4541 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4542 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4544 (unsigned long)zone_idx(zone
),
4545 zone_start_pfn
, (zone_start_pfn
+ size
));
4547 zone_init_free_lists(zone
);
4552 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4553 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4555 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4557 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4559 unsigned long start_pfn
, end_pfn
;
4562 * NOTE: The following SMP-unsafe globals are only used early in boot
4563 * when the kernel is running single-threaded.
4565 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4566 static int __meminitdata last_nid
;
4568 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4571 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4573 last_start_pfn
= start_pfn
;
4574 last_end_pfn
= end_pfn
;
4580 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4582 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4586 nid
= __early_pfn_to_nid(pfn
);
4589 /* just returns 0 */
4593 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4594 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4598 nid
= __early_pfn_to_nid(pfn
);
4599 if (nid
>= 0 && nid
!= node
)
4606 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4607 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4608 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4610 * If an architecture guarantees that all ranges registered contain no holes
4611 * and may be freed, this this function may be used instead of calling
4612 * memblock_free_early_nid() manually.
4614 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4616 unsigned long start_pfn
, end_pfn
;
4619 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4620 start_pfn
= min(start_pfn
, max_low_pfn
);
4621 end_pfn
= min(end_pfn
, max_low_pfn
);
4623 if (start_pfn
< end_pfn
)
4624 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4625 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4631 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4632 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4634 * If an architecture guarantees that all ranges registered contain no holes and may
4635 * be freed, this function may be used instead of calling memory_present() manually.
4637 void __init
sparse_memory_present_with_active_regions(int nid
)
4639 unsigned long start_pfn
, end_pfn
;
4642 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4643 memory_present(this_nid
, start_pfn
, end_pfn
);
4647 * get_pfn_range_for_nid - Return the start and end page frames for a node
4648 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4649 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4650 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4652 * It returns the start and end page frame of a node based on information
4653 * provided by memblock_set_node(). If called for a node
4654 * with no available memory, a warning is printed and the start and end
4657 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4658 unsigned long *start_pfn
, unsigned long *end_pfn
)
4660 unsigned long this_start_pfn
, this_end_pfn
;
4666 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4667 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4668 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4671 if (*start_pfn
== -1UL)
4676 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4677 * assumption is made that zones within a node are ordered in monotonic
4678 * increasing memory addresses so that the "highest" populated zone is used
4680 static void __init
find_usable_zone_for_movable(void)
4683 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4684 if (zone_index
== ZONE_MOVABLE
)
4687 if (arch_zone_highest_possible_pfn
[zone_index
] >
4688 arch_zone_lowest_possible_pfn
[zone_index
])
4692 VM_BUG_ON(zone_index
== -1);
4693 movable_zone
= zone_index
;
4697 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4698 * because it is sized independent of architecture. Unlike the other zones,
4699 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4700 * in each node depending on the size of each node and how evenly kernelcore
4701 * is distributed. This helper function adjusts the zone ranges
4702 * provided by the architecture for a given node by using the end of the
4703 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4704 * zones within a node are in order of monotonic increases memory addresses
4706 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4707 unsigned long zone_type
,
4708 unsigned long node_start_pfn
,
4709 unsigned long node_end_pfn
,
4710 unsigned long *zone_start_pfn
,
4711 unsigned long *zone_end_pfn
)
4713 /* Only adjust if ZONE_MOVABLE is on this node */
4714 if (zone_movable_pfn
[nid
]) {
4715 /* Size ZONE_MOVABLE */
4716 if (zone_type
== ZONE_MOVABLE
) {
4717 *zone_start_pfn
= zone_movable_pfn
[nid
];
4718 *zone_end_pfn
= min(node_end_pfn
,
4719 arch_zone_highest_possible_pfn
[movable_zone
]);
4721 /* Adjust for ZONE_MOVABLE starting within this range */
4722 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4723 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4724 *zone_end_pfn
= zone_movable_pfn
[nid
];
4726 /* Check if this whole range is within ZONE_MOVABLE */
4727 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4728 *zone_start_pfn
= *zone_end_pfn
;
4733 * Return the number of pages a zone spans in a node, including holes
4734 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4736 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4737 unsigned long zone_type
,
4738 unsigned long node_start_pfn
,
4739 unsigned long node_end_pfn
,
4740 unsigned long *ignored
)
4742 unsigned long zone_start_pfn
, zone_end_pfn
;
4744 /* Get the start and end of the zone */
4745 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4746 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4747 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4748 node_start_pfn
, node_end_pfn
,
4749 &zone_start_pfn
, &zone_end_pfn
);
4751 /* Check that this node has pages within the zone's required range */
4752 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4755 /* Move the zone boundaries inside the node if necessary */
4756 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4757 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4759 /* Return the spanned pages */
4760 return zone_end_pfn
- zone_start_pfn
;
4764 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4765 * then all holes in the requested range will be accounted for.
4767 unsigned long __meminit
__absent_pages_in_range(int nid
,
4768 unsigned long range_start_pfn
,
4769 unsigned long range_end_pfn
)
4771 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4772 unsigned long start_pfn
, end_pfn
;
4775 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4776 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4777 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4778 nr_absent
-= end_pfn
- start_pfn
;
4784 * absent_pages_in_range - Return number of page frames in holes within a range
4785 * @start_pfn: The start PFN to start searching for holes
4786 * @end_pfn: The end PFN to stop searching for holes
4788 * It returns the number of pages frames in memory holes within a range.
4790 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4791 unsigned long end_pfn
)
4793 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4796 /* Return the number of page frames in holes in a zone on a node */
4797 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4798 unsigned long zone_type
,
4799 unsigned long node_start_pfn
,
4800 unsigned long node_end_pfn
,
4801 unsigned long *ignored
)
4803 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4804 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4805 unsigned long zone_start_pfn
, zone_end_pfn
;
4807 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4808 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4810 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4811 node_start_pfn
, node_end_pfn
,
4812 &zone_start_pfn
, &zone_end_pfn
);
4813 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4816 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4817 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4818 unsigned long zone_type
,
4819 unsigned long node_start_pfn
,
4820 unsigned long node_end_pfn
,
4821 unsigned long *zones_size
)
4823 return zones_size
[zone_type
];
4826 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4827 unsigned long zone_type
,
4828 unsigned long node_start_pfn
,
4829 unsigned long node_end_pfn
,
4830 unsigned long *zholes_size
)
4835 return zholes_size
[zone_type
];
4838 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4840 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4841 unsigned long node_start_pfn
,
4842 unsigned long node_end_pfn
,
4843 unsigned long *zones_size
,
4844 unsigned long *zholes_size
)
4846 unsigned long realtotalpages
= 0, totalpages
= 0;
4849 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4850 struct zone
*zone
= pgdat
->node_zones
+ i
;
4851 unsigned long size
, real_size
;
4853 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4857 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
4858 node_start_pfn
, node_end_pfn
,
4860 zone
->spanned_pages
= size
;
4861 zone
->present_pages
= real_size
;
4864 realtotalpages
+= real_size
;
4867 pgdat
->node_spanned_pages
= totalpages
;
4868 pgdat
->node_present_pages
= realtotalpages
;
4869 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4873 #ifndef CONFIG_SPARSEMEM
4875 * Calculate the size of the zone->blockflags rounded to an unsigned long
4876 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4877 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4878 * round what is now in bits to nearest long in bits, then return it in
4881 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4883 unsigned long usemapsize
;
4885 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4886 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4887 usemapsize
= usemapsize
>> pageblock_order
;
4888 usemapsize
*= NR_PAGEBLOCK_BITS
;
4889 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4891 return usemapsize
/ 8;
4894 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4896 unsigned long zone_start_pfn
,
4897 unsigned long zonesize
)
4899 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4900 zone
->pageblock_flags
= NULL
;
4902 zone
->pageblock_flags
=
4903 memblock_virt_alloc_node_nopanic(usemapsize
,
4907 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4908 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4909 #endif /* CONFIG_SPARSEMEM */
4911 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4913 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4914 void __paginginit
set_pageblock_order(void)
4918 /* Check that pageblock_nr_pages has not already been setup */
4919 if (pageblock_order
)
4922 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4923 order
= HUGETLB_PAGE_ORDER
;
4925 order
= MAX_ORDER
- 1;
4928 * Assume the largest contiguous order of interest is a huge page.
4929 * This value may be variable depending on boot parameters on IA64 and
4932 pageblock_order
= order
;
4934 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4937 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4938 * is unused as pageblock_order is set at compile-time. See
4939 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4942 void __paginginit
set_pageblock_order(void)
4946 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4948 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4949 unsigned long present_pages
)
4951 unsigned long pages
= spanned_pages
;
4954 * Provide a more accurate estimation if there are holes within
4955 * the zone and SPARSEMEM is in use. If there are holes within the
4956 * zone, each populated memory region may cost us one or two extra
4957 * memmap pages due to alignment because memmap pages for each
4958 * populated regions may not naturally algined on page boundary.
4959 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4961 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4962 IS_ENABLED(CONFIG_SPARSEMEM
))
4963 pages
= present_pages
;
4965 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4969 * Set up the zone data structures:
4970 * - mark all pages reserved
4971 * - mark all memory queues empty
4972 * - clear the memory bitmaps
4974 * NOTE: pgdat should get zeroed by caller.
4976 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4977 unsigned long node_start_pfn
, unsigned long node_end_pfn
)
4980 int nid
= pgdat
->node_id
;
4981 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4984 pgdat_resize_init(pgdat
);
4985 #ifdef CONFIG_NUMA_BALANCING
4986 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4987 pgdat
->numabalancing_migrate_nr_pages
= 0;
4988 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4990 init_waitqueue_head(&pgdat
->kswapd_wait
);
4991 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4992 pgdat_page_ext_init(pgdat
);
4994 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4995 struct zone
*zone
= pgdat
->node_zones
+ j
;
4996 unsigned long size
, realsize
, freesize
, memmap_pages
;
4998 size
= zone
->spanned_pages
;
4999 realsize
= freesize
= zone
->present_pages
;
5002 * Adjust freesize so that it accounts for how much memory
5003 * is used by this zone for memmap. This affects the watermark
5004 * and per-cpu initialisations
5006 memmap_pages
= calc_memmap_size(size
, realsize
);
5007 if (!is_highmem_idx(j
)) {
5008 if (freesize
>= memmap_pages
) {
5009 freesize
-= memmap_pages
;
5012 " %s zone: %lu pages used for memmap\n",
5013 zone_names
[j
], memmap_pages
);
5016 " %s zone: %lu pages exceeds freesize %lu\n",
5017 zone_names
[j
], memmap_pages
, freesize
);
5020 /* Account for reserved pages */
5021 if (j
== 0 && freesize
> dma_reserve
) {
5022 freesize
-= dma_reserve
;
5023 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5024 zone_names
[0], dma_reserve
);
5027 if (!is_highmem_idx(j
))
5028 nr_kernel_pages
+= freesize
;
5029 /* Charge for highmem memmap if there are enough kernel pages */
5030 else if (nr_kernel_pages
> memmap_pages
* 2)
5031 nr_kernel_pages
-= memmap_pages
;
5032 nr_all_pages
+= freesize
;
5035 * Set an approximate value for lowmem here, it will be adjusted
5036 * when the bootmem allocator frees pages into the buddy system.
5037 * And all highmem pages will be managed by the buddy system.
5039 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5042 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
5044 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
5046 zone
->name
= zone_names
[j
];
5047 spin_lock_init(&zone
->lock
);
5048 spin_lock_init(&zone
->lru_lock
);
5049 zone_seqlock_init(zone
);
5050 zone
->zone_pgdat
= pgdat
;
5051 zone_pcp_init(zone
);
5053 /* For bootup, initialized properly in watermark setup */
5054 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
5056 lruvec_init(&zone
->lruvec
);
5060 set_pageblock_order();
5061 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5062 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
5063 size
, MEMMAP_EARLY
);
5065 memmap_init(size
, nid
, j
, zone_start_pfn
);
5066 zone_start_pfn
+= size
;
5070 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
5072 /* Skip empty nodes */
5073 if (!pgdat
->node_spanned_pages
)
5076 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5077 /* ia64 gets its own node_mem_map, before this, without bootmem */
5078 if (!pgdat
->node_mem_map
) {
5079 unsigned long size
, start
, end
;
5083 * The zone's endpoints aren't required to be MAX_ORDER
5084 * aligned but the node_mem_map endpoints must be in order
5085 * for the buddy allocator to function correctly.
5087 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5088 end
= pgdat_end_pfn(pgdat
);
5089 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5090 size
= (end
- start
) * sizeof(struct page
);
5091 map
= alloc_remap(pgdat
->node_id
, size
);
5093 map
= memblock_virt_alloc_node_nopanic(size
,
5095 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
5097 #ifndef CONFIG_NEED_MULTIPLE_NODES
5099 * With no DISCONTIG, the global mem_map is just set as node 0's
5101 if (pgdat
== NODE_DATA(0)) {
5102 mem_map
= NODE_DATA(0)->node_mem_map
;
5103 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5104 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5105 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
5106 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5109 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5112 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5113 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5115 pg_data_t
*pgdat
= NODE_DATA(nid
);
5116 unsigned long start_pfn
= 0;
5117 unsigned long end_pfn
= 0;
5119 /* pg_data_t should be reset to zero when it's allocated */
5120 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
5122 pgdat
->node_id
= nid
;
5123 pgdat
->node_start_pfn
= node_start_pfn
;
5124 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5125 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5126 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5127 (u64
)start_pfn
<< PAGE_SHIFT
, ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5129 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5130 zones_size
, zholes_size
);
5132 alloc_node_mem_map(pgdat
);
5133 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5134 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5135 nid
, (unsigned long)pgdat
,
5136 (unsigned long)pgdat
->node_mem_map
);
5139 free_area_init_core(pgdat
, start_pfn
, end_pfn
);
5142 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5144 #if MAX_NUMNODES > 1
5146 * Figure out the number of possible node ids.
5148 void __init
setup_nr_node_ids(void)
5151 unsigned int highest
= 0;
5153 for_each_node_mask(node
, node_possible_map
)
5155 nr_node_ids
= highest
+ 1;
5160 * node_map_pfn_alignment - determine the maximum internode alignment
5162 * This function should be called after node map is populated and sorted.
5163 * It calculates the maximum power of two alignment which can distinguish
5166 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5167 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5168 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5169 * shifted, 1GiB is enough and this function will indicate so.
5171 * This is used to test whether pfn -> nid mapping of the chosen memory
5172 * model has fine enough granularity to avoid incorrect mapping for the
5173 * populated node map.
5175 * Returns the determined alignment in pfn's. 0 if there is no alignment
5176 * requirement (single node).
5178 unsigned long __init
node_map_pfn_alignment(void)
5180 unsigned long accl_mask
= 0, last_end
= 0;
5181 unsigned long start
, end
, mask
;
5185 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5186 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5193 * Start with a mask granular enough to pin-point to the
5194 * start pfn and tick off bits one-by-one until it becomes
5195 * too coarse to separate the current node from the last.
5197 mask
= ~((1 << __ffs(start
)) - 1);
5198 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5201 /* accumulate all internode masks */
5205 /* convert mask to number of pages */
5206 return ~accl_mask
+ 1;
5209 /* Find the lowest pfn for a node */
5210 static unsigned long __init
find_min_pfn_for_node(int nid
)
5212 unsigned long min_pfn
= ULONG_MAX
;
5213 unsigned long start_pfn
;
5216 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5217 min_pfn
= min(min_pfn
, start_pfn
);
5219 if (min_pfn
== ULONG_MAX
) {
5221 "Could not find start_pfn for node %d\n", nid
);
5229 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5231 * It returns the minimum PFN based on information provided via
5232 * memblock_set_node().
5234 unsigned long __init
find_min_pfn_with_active_regions(void)
5236 return find_min_pfn_for_node(MAX_NUMNODES
);
5240 * early_calculate_totalpages()
5241 * Sum pages in active regions for movable zone.
5242 * Populate N_MEMORY for calculating usable_nodes.
5244 static unsigned long __init
early_calculate_totalpages(void)
5246 unsigned long totalpages
= 0;
5247 unsigned long start_pfn
, end_pfn
;
5250 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5251 unsigned long pages
= end_pfn
- start_pfn
;
5253 totalpages
+= pages
;
5255 node_set_state(nid
, N_MEMORY
);
5261 * Find the PFN the Movable zone begins in each node. Kernel memory
5262 * is spread evenly between nodes as long as the nodes have enough
5263 * memory. When they don't, some nodes will have more kernelcore than
5266 static void __init
find_zone_movable_pfns_for_nodes(void)
5269 unsigned long usable_startpfn
;
5270 unsigned long kernelcore_node
, kernelcore_remaining
;
5271 /* save the state before borrow the nodemask */
5272 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5273 unsigned long totalpages
= early_calculate_totalpages();
5274 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5275 struct memblock_region
*r
;
5277 /* Need to find movable_zone earlier when movable_node is specified. */
5278 find_usable_zone_for_movable();
5281 * If movable_node is specified, ignore kernelcore and movablecore
5284 if (movable_node_is_enabled()) {
5285 for_each_memblock(memory
, r
) {
5286 if (!memblock_is_hotpluggable(r
))
5291 usable_startpfn
= PFN_DOWN(r
->base
);
5292 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5293 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5301 * If movablecore=nn[KMG] was specified, calculate what size of
5302 * kernelcore that corresponds so that memory usable for
5303 * any allocation type is evenly spread. If both kernelcore
5304 * and movablecore are specified, then the value of kernelcore
5305 * will be used for required_kernelcore if it's greater than
5306 * what movablecore would have allowed.
5308 if (required_movablecore
) {
5309 unsigned long corepages
;
5312 * Round-up so that ZONE_MOVABLE is at least as large as what
5313 * was requested by the user
5315 required_movablecore
=
5316 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5317 corepages
= totalpages
- required_movablecore
;
5319 required_kernelcore
= max(required_kernelcore
, corepages
);
5322 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5323 if (!required_kernelcore
)
5326 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5327 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5330 /* Spread kernelcore memory as evenly as possible throughout nodes */
5331 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5332 for_each_node_state(nid
, N_MEMORY
) {
5333 unsigned long start_pfn
, end_pfn
;
5336 * Recalculate kernelcore_node if the division per node
5337 * now exceeds what is necessary to satisfy the requested
5338 * amount of memory for the kernel
5340 if (required_kernelcore
< kernelcore_node
)
5341 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5344 * As the map is walked, we track how much memory is usable
5345 * by the kernel using kernelcore_remaining. When it is
5346 * 0, the rest of the node is usable by ZONE_MOVABLE
5348 kernelcore_remaining
= kernelcore_node
;
5350 /* Go through each range of PFNs within this node */
5351 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5352 unsigned long size_pages
;
5354 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5355 if (start_pfn
>= end_pfn
)
5358 /* Account for what is only usable for kernelcore */
5359 if (start_pfn
< usable_startpfn
) {
5360 unsigned long kernel_pages
;
5361 kernel_pages
= min(end_pfn
, usable_startpfn
)
5364 kernelcore_remaining
-= min(kernel_pages
,
5365 kernelcore_remaining
);
5366 required_kernelcore
-= min(kernel_pages
,
5367 required_kernelcore
);
5369 /* Continue if range is now fully accounted */
5370 if (end_pfn
<= usable_startpfn
) {
5373 * Push zone_movable_pfn to the end so
5374 * that if we have to rebalance
5375 * kernelcore across nodes, we will
5376 * not double account here
5378 zone_movable_pfn
[nid
] = end_pfn
;
5381 start_pfn
= usable_startpfn
;
5385 * The usable PFN range for ZONE_MOVABLE is from
5386 * start_pfn->end_pfn. Calculate size_pages as the
5387 * number of pages used as kernelcore
5389 size_pages
= end_pfn
- start_pfn
;
5390 if (size_pages
> kernelcore_remaining
)
5391 size_pages
= kernelcore_remaining
;
5392 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5395 * Some kernelcore has been met, update counts and
5396 * break if the kernelcore for this node has been
5399 required_kernelcore
-= min(required_kernelcore
,
5401 kernelcore_remaining
-= size_pages
;
5402 if (!kernelcore_remaining
)
5408 * If there is still required_kernelcore, we do another pass with one
5409 * less node in the count. This will push zone_movable_pfn[nid] further
5410 * along on the nodes that still have memory until kernelcore is
5414 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5418 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5419 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5420 zone_movable_pfn
[nid
] =
5421 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5424 /* restore the node_state */
5425 node_states
[N_MEMORY
] = saved_node_state
;
5428 /* Any regular or high memory on that node ? */
5429 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5431 enum zone_type zone_type
;
5433 if (N_MEMORY
== N_NORMAL_MEMORY
)
5436 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5437 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5438 if (populated_zone(zone
)) {
5439 node_set_state(nid
, N_HIGH_MEMORY
);
5440 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5441 zone_type
<= ZONE_NORMAL
)
5442 node_set_state(nid
, N_NORMAL_MEMORY
);
5449 * free_area_init_nodes - Initialise all pg_data_t and zone data
5450 * @max_zone_pfn: an array of max PFNs for each zone
5452 * This will call free_area_init_node() for each active node in the system.
5453 * Using the page ranges provided by memblock_set_node(), the size of each
5454 * zone in each node and their holes is calculated. If the maximum PFN
5455 * between two adjacent zones match, it is assumed that the zone is empty.
5456 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5457 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5458 * starts where the previous one ended. For example, ZONE_DMA32 starts
5459 * at arch_max_dma_pfn.
5461 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5463 unsigned long start_pfn
, end_pfn
;
5466 /* Record where the zone boundaries are */
5467 memset(arch_zone_lowest_possible_pfn
, 0,
5468 sizeof(arch_zone_lowest_possible_pfn
));
5469 memset(arch_zone_highest_possible_pfn
, 0,
5470 sizeof(arch_zone_highest_possible_pfn
));
5471 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5472 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5473 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5474 if (i
== ZONE_MOVABLE
)
5476 arch_zone_lowest_possible_pfn
[i
] =
5477 arch_zone_highest_possible_pfn
[i
-1];
5478 arch_zone_highest_possible_pfn
[i
] =
5479 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5481 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5482 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5484 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5485 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5486 find_zone_movable_pfns_for_nodes();
5488 /* Print out the zone ranges */
5489 pr_info("Zone ranges:\n");
5490 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5491 if (i
== ZONE_MOVABLE
)
5493 pr_info(" %-8s ", zone_names
[i
]);
5494 if (arch_zone_lowest_possible_pfn
[i
] ==
5495 arch_zone_highest_possible_pfn
[i
])
5498 pr_cont("[mem %#018Lx-%#018Lx]\n",
5499 (u64
)arch_zone_lowest_possible_pfn
[i
]
5501 ((u64
)arch_zone_highest_possible_pfn
[i
]
5502 << PAGE_SHIFT
) - 1);
5505 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5506 pr_info("Movable zone start for each node\n");
5507 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5508 if (zone_movable_pfn
[i
])
5509 pr_info(" Node %d: %#018Lx\n", i
,
5510 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
5513 /* Print out the early node map */
5514 pr_info("Early memory node ranges\n");
5515 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5516 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
5517 (u64
)start_pfn
<< PAGE_SHIFT
,
5518 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5520 /* Initialise every node */
5521 mminit_verify_pageflags_layout();
5522 setup_nr_node_ids();
5523 for_each_online_node(nid
) {
5524 pg_data_t
*pgdat
= NODE_DATA(nid
);
5525 free_area_init_node(nid
, NULL
,
5526 find_min_pfn_for_node(nid
), NULL
);
5528 /* Any memory on that node */
5529 if (pgdat
->node_present_pages
)
5530 node_set_state(nid
, N_MEMORY
);
5531 check_for_memory(pgdat
, nid
);
5535 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5537 unsigned long long coremem
;
5541 coremem
= memparse(p
, &p
);
5542 *core
= coremem
>> PAGE_SHIFT
;
5544 /* Paranoid check that UL is enough for the coremem value */
5545 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5551 * kernelcore=size sets the amount of memory for use for allocations that
5552 * cannot be reclaimed or migrated.
5554 static int __init
cmdline_parse_kernelcore(char *p
)
5556 return cmdline_parse_core(p
, &required_kernelcore
);
5560 * movablecore=size sets the amount of memory for use for allocations that
5561 * can be reclaimed or migrated.
5563 static int __init
cmdline_parse_movablecore(char *p
)
5565 return cmdline_parse_core(p
, &required_movablecore
);
5568 early_param("kernelcore", cmdline_parse_kernelcore
);
5569 early_param("movablecore", cmdline_parse_movablecore
);
5571 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5573 void adjust_managed_page_count(struct page
*page
, long count
)
5575 spin_lock(&managed_page_count_lock
);
5576 page_zone(page
)->managed_pages
+= count
;
5577 totalram_pages
+= count
;
5578 #ifdef CONFIG_HIGHMEM
5579 if (PageHighMem(page
))
5580 totalhigh_pages
+= count
;
5582 spin_unlock(&managed_page_count_lock
);
5584 EXPORT_SYMBOL(adjust_managed_page_count
);
5586 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5589 unsigned long pages
= 0;
5591 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5592 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5593 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5594 if ((unsigned int)poison
<= 0xFF)
5595 memset(pos
, poison
, PAGE_SIZE
);
5596 free_reserved_page(virt_to_page(pos
));
5600 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5601 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5605 EXPORT_SYMBOL(free_reserved_area
);
5607 #ifdef CONFIG_HIGHMEM
5608 void free_highmem_page(struct page
*page
)
5610 __free_reserved_page(page
);
5612 page_zone(page
)->managed_pages
++;
5618 void __init
mem_init_print_info(const char *str
)
5620 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5621 unsigned long init_code_size
, init_data_size
;
5623 physpages
= get_num_physpages();
5624 codesize
= _etext
- _stext
;
5625 datasize
= _edata
- _sdata
;
5626 rosize
= __end_rodata
- __start_rodata
;
5627 bss_size
= __bss_stop
- __bss_start
;
5628 init_data_size
= __init_end
- __init_begin
;
5629 init_code_size
= _einittext
- _sinittext
;
5632 * Detect special cases and adjust section sizes accordingly:
5633 * 1) .init.* may be embedded into .data sections
5634 * 2) .init.text.* may be out of [__init_begin, __init_end],
5635 * please refer to arch/tile/kernel/vmlinux.lds.S.
5636 * 3) .rodata.* may be embedded into .text or .data sections.
5638 #define adj_init_size(start, end, size, pos, adj) \
5640 if (start <= pos && pos < end && size > adj) \
5644 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5645 _sinittext
, init_code_size
);
5646 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5647 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5648 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5649 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5651 #undef adj_init_size
5653 pr_info("Memory: %luK/%luK available "
5654 "(%luK kernel code, %luK rwdata, %luK rodata, "
5655 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5656 #ifdef CONFIG_HIGHMEM
5660 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5661 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5662 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5663 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
-10),
5664 totalcma_pages
<< (PAGE_SHIFT
-10),
5665 #ifdef CONFIG_HIGHMEM
5666 totalhigh_pages
<< (PAGE_SHIFT
-10),
5668 str
? ", " : "", str
? str
: "");
5672 * set_dma_reserve - set the specified number of pages reserved in the first zone
5673 * @new_dma_reserve: The number of pages to mark reserved
5675 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5676 * In the DMA zone, a significant percentage may be consumed by kernel image
5677 * and other unfreeable allocations which can skew the watermarks badly. This
5678 * function may optionally be used to account for unfreeable pages in the
5679 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5680 * smaller per-cpu batchsize.
5682 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5684 dma_reserve
= new_dma_reserve
;
5687 void __init
free_area_init(unsigned long *zones_size
)
5689 free_area_init_node(0, zones_size
,
5690 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5693 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5694 unsigned long action
, void *hcpu
)
5696 int cpu
= (unsigned long)hcpu
;
5698 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5699 lru_add_drain_cpu(cpu
);
5703 * Spill the event counters of the dead processor
5704 * into the current processors event counters.
5705 * This artificially elevates the count of the current
5708 vm_events_fold_cpu(cpu
);
5711 * Zero the differential counters of the dead processor
5712 * so that the vm statistics are consistent.
5714 * This is only okay since the processor is dead and cannot
5715 * race with what we are doing.
5717 cpu_vm_stats_fold(cpu
);
5722 void __init
page_alloc_init(void)
5724 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5728 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5729 * or min_free_kbytes changes.
5731 static void calculate_totalreserve_pages(void)
5733 struct pglist_data
*pgdat
;
5734 unsigned long reserve_pages
= 0;
5735 enum zone_type i
, j
;
5737 for_each_online_pgdat(pgdat
) {
5738 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5739 struct zone
*zone
= pgdat
->node_zones
+ i
;
5742 /* Find valid and maximum lowmem_reserve in the zone */
5743 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5744 if (zone
->lowmem_reserve
[j
] > max
)
5745 max
= zone
->lowmem_reserve
[j
];
5748 /* we treat the high watermark as reserved pages. */
5749 max
+= high_wmark_pages(zone
);
5751 if (max
> zone
->managed_pages
)
5752 max
= zone
->managed_pages
;
5753 reserve_pages
+= max
;
5755 * Lowmem reserves are not available to
5756 * GFP_HIGHUSER page cache allocations and
5757 * kswapd tries to balance zones to their high
5758 * watermark. As a result, neither should be
5759 * regarded as dirtyable memory, to prevent a
5760 * situation where reclaim has to clean pages
5761 * in order to balance the zones.
5763 zone
->dirty_balance_reserve
= max
;
5766 dirty_balance_reserve
= reserve_pages
;
5767 totalreserve_pages
= reserve_pages
;
5771 * setup_per_zone_lowmem_reserve - called whenever
5772 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5773 * has a correct pages reserved value, so an adequate number of
5774 * pages are left in the zone after a successful __alloc_pages().
5776 static void setup_per_zone_lowmem_reserve(void)
5778 struct pglist_data
*pgdat
;
5779 enum zone_type j
, idx
;
5781 for_each_online_pgdat(pgdat
) {
5782 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5783 struct zone
*zone
= pgdat
->node_zones
+ j
;
5784 unsigned long managed_pages
= zone
->managed_pages
;
5786 zone
->lowmem_reserve
[j
] = 0;
5790 struct zone
*lower_zone
;
5794 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5795 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5797 lower_zone
= pgdat
->node_zones
+ idx
;
5798 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5799 sysctl_lowmem_reserve_ratio
[idx
];
5800 managed_pages
+= lower_zone
->managed_pages
;
5805 /* update totalreserve_pages */
5806 calculate_totalreserve_pages();
5809 static void __setup_per_zone_wmarks(void)
5811 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5812 unsigned long lowmem_pages
= 0;
5814 unsigned long flags
;
5816 /* Calculate total number of !ZONE_HIGHMEM pages */
5817 for_each_zone(zone
) {
5818 if (!is_highmem(zone
))
5819 lowmem_pages
+= zone
->managed_pages
;
5822 for_each_zone(zone
) {
5825 spin_lock_irqsave(&zone
->lock
, flags
);
5826 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5827 do_div(tmp
, lowmem_pages
);
5828 if (is_highmem(zone
)) {
5830 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5831 * need highmem pages, so cap pages_min to a small
5834 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5835 * deltas control asynch page reclaim, and so should
5836 * not be capped for highmem.
5838 unsigned long min_pages
;
5840 min_pages
= zone
->managed_pages
/ 1024;
5841 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5842 zone
->watermark
[WMARK_MIN
] = min_pages
;
5845 * If it's a lowmem zone, reserve a number of pages
5846 * proportionate to the zone's size.
5848 zone
->watermark
[WMARK_MIN
] = tmp
;
5851 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5852 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5854 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5855 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5856 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5858 setup_zone_migrate_reserve(zone
);
5859 spin_unlock_irqrestore(&zone
->lock
, flags
);
5862 /* update totalreserve_pages */
5863 calculate_totalreserve_pages();
5867 * setup_per_zone_wmarks - called when min_free_kbytes changes
5868 * or when memory is hot-{added|removed}
5870 * Ensures that the watermark[min,low,high] values for each zone are set
5871 * correctly with respect to min_free_kbytes.
5873 void setup_per_zone_wmarks(void)
5875 mutex_lock(&zonelists_mutex
);
5876 __setup_per_zone_wmarks();
5877 mutex_unlock(&zonelists_mutex
);
5881 * The inactive anon list should be small enough that the VM never has to
5882 * do too much work, but large enough that each inactive page has a chance
5883 * to be referenced again before it is swapped out.
5885 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5886 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5887 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5888 * the anonymous pages are kept on the inactive list.
5891 * memory ratio inactive anon
5892 * -------------------------------------
5901 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5903 unsigned int gb
, ratio
;
5905 /* Zone size in gigabytes */
5906 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5908 ratio
= int_sqrt(10 * gb
);
5912 zone
->inactive_ratio
= ratio
;
5915 static void __meminit
setup_per_zone_inactive_ratio(void)
5920 calculate_zone_inactive_ratio(zone
);
5924 * Initialise min_free_kbytes.
5926 * For small machines we want it small (128k min). For large machines
5927 * we want it large (64MB max). But it is not linear, because network
5928 * bandwidth does not increase linearly with machine size. We use
5930 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5931 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5947 int __meminit
init_per_zone_wmark_min(void)
5949 unsigned long lowmem_kbytes
;
5950 int new_min_free_kbytes
;
5952 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5953 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5955 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5956 min_free_kbytes
= new_min_free_kbytes
;
5957 if (min_free_kbytes
< 128)
5958 min_free_kbytes
= 128;
5959 if (min_free_kbytes
> 65536)
5960 min_free_kbytes
= 65536;
5962 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5963 new_min_free_kbytes
, user_min_free_kbytes
);
5965 setup_per_zone_wmarks();
5966 refresh_zone_stat_thresholds();
5967 setup_per_zone_lowmem_reserve();
5968 setup_per_zone_inactive_ratio();
5971 module_init(init_per_zone_wmark_min
)
5974 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5975 * that we can call two helper functions whenever min_free_kbytes
5978 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5979 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5983 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5988 user_min_free_kbytes
= min_free_kbytes
;
5989 setup_per_zone_wmarks();
5995 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5996 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6001 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6006 zone
->min_unmapped_pages
= (zone
->managed_pages
*
6007 sysctl_min_unmapped_ratio
) / 100;
6011 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6012 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6017 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6022 zone
->min_slab_pages
= (zone
->managed_pages
*
6023 sysctl_min_slab_ratio
) / 100;
6029 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6030 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6031 * whenever sysctl_lowmem_reserve_ratio changes.
6033 * The reserve ratio obviously has absolutely no relation with the
6034 * minimum watermarks. The lowmem reserve ratio can only make sense
6035 * if in function of the boot time zone sizes.
6037 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6038 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6040 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6041 setup_per_zone_lowmem_reserve();
6046 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6047 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6048 * pagelist can have before it gets flushed back to buddy allocator.
6050 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6051 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6054 int old_percpu_pagelist_fraction
;
6057 mutex_lock(&pcp_batch_high_lock
);
6058 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6060 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6061 if (!write
|| ret
< 0)
6064 /* Sanity checking to avoid pcp imbalance */
6065 if (percpu_pagelist_fraction
&&
6066 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6067 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6073 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6076 for_each_populated_zone(zone
) {
6079 for_each_possible_cpu(cpu
)
6080 pageset_set_high_and_batch(zone
,
6081 per_cpu_ptr(zone
->pageset
, cpu
));
6084 mutex_unlock(&pcp_batch_high_lock
);
6089 int hashdist
= HASHDIST_DEFAULT
;
6091 static int __init
set_hashdist(char *str
)
6095 hashdist
= simple_strtoul(str
, &str
, 0);
6098 __setup("hashdist=", set_hashdist
);
6102 * allocate a large system hash table from bootmem
6103 * - it is assumed that the hash table must contain an exact power-of-2
6104 * quantity of entries
6105 * - limit is the number of hash buckets, not the total allocation size
6107 void *__init
alloc_large_system_hash(const char *tablename
,
6108 unsigned long bucketsize
,
6109 unsigned long numentries
,
6112 unsigned int *_hash_shift
,
6113 unsigned int *_hash_mask
,
6114 unsigned long low_limit
,
6115 unsigned long high_limit
)
6117 unsigned long long max
= high_limit
;
6118 unsigned long log2qty
, size
;
6121 /* allow the kernel cmdline to have a say */
6123 /* round applicable memory size up to nearest megabyte */
6124 numentries
= nr_kernel_pages
;
6126 /* It isn't necessary when PAGE_SIZE >= 1MB */
6127 if (PAGE_SHIFT
< 20)
6128 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6130 /* limit to 1 bucket per 2^scale bytes of low memory */
6131 if (scale
> PAGE_SHIFT
)
6132 numentries
>>= (scale
- PAGE_SHIFT
);
6134 numentries
<<= (PAGE_SHIFT
- scale
);
6136 /* Make sure we've got at least a 0-order allocation.. */
6137 if (unlikely(flags
& HASH_SMALL
)) {
6138 /* Makes no sense without HASH_EARLY */
6139 WARN_ON(!(flags
& HASH_EARLY
));
6140 if (!(numentries
>> *_hash_shift
)) {
6141 numentries
= 1UL << *_hash_shift
;
6142 BUG_ON(!numentries
);
6144 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6145 numentries
= PAGE_SIZE
/ bucketsize
;
6147 numentries
= roundup_pow_of_two(numentries
);
6149 /* limit allocation size to 1/16 total memory by default */
6151 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6152 do_div(max
, bucketsize
);
6154 max
= min(max
, 0x80000000ULL
);
6156 if (numentries
< low_limit
)
6157 numentries
= low_limit
;
6158 if (numentries
> max
)
6161 log2qty
= ilog2(numentries
);
6164 size
= bucketsize
<< log2qty
;
6165 if (flags
& HASH_EARLY
)
6166 table
= memblock_virt_alloc_nopanic(size
, 0);
6168 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6171 * If bucketsize is not a power-of-two, we may free
6172 * some pages at the end of hash table which
6173 * alloc_pages_exact() automatically does
6175 if (get_order(size
) < MAX_ORDER
) {
6176 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6177 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6180 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6183 panic("Failed to allocate %s hash table\n", tablename
);
6185 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6188 ilog2(size
) - PAGE_SHIFT
,
6192 *_hash_shift
= log2qty
;
6194 *_hash_mask
= (1 << log2qty
) - 1;
6199 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6200 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6203 #ifdef CONFIG_SPARSEMEM
6204 return __pfn_to_section(pfn
)->pageblock_flags
;
6206 return zone
->pageblock_flags
;
6207 #endif /* CONFIG_SPARSEMEM */
6210 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6212 #ifdef CONFIG_SPARSEMEM
6213 pfn
&= (PAGES_PER_SECTION
-1);
6214 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6216 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6217 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6218 #endif /* CONFIG_SPARSEMEM */
6222 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6223 * @page: The page within the block of interest
6224 * @pfn: The target page frame number
6225 * @end_bitidx: The last bit of interest to retrieve
6226 * @mask: mask of bits that the caller is interested in
6228 * Return: pageblock_bits flags
6230 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6231 unsigned long end_bitidx
,
6235 unsigned long *bitmap
;
6236 unsigned long bitidx
, word_bitidx
;
6239 zone
= page_zone(page
);
6240 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6241 bitidx
= pfn_to_bitidx(zone
, pfn
);
6242 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6243 bitidx
&= (BITS_PER_LONG
-1);
6245 word
= bitmap
[word_bitidx
];
6246 bitidx
+= end_bitidx
;
6247 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6251 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6252 * @page: The page within the block of interest
6253 * @flags: The flags to set
6254 * @pfn: The target page frame number
6255 * @end_bitidx: The last bit of interest
6256 * @mask: mask of bits that the caller is interested in
6258 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6260 unsigned long end_bitidx
,
6264 unsigned long *bitmap
;
6265 unsigned long bitidx
, word_bitidx
;
6266 unsigned long old_word
, word
;
6268 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6270 zone
= page_zone(page
);
6271 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6272 bitidx
= pfn_to_bitidx(zone
, pfn
);
6273 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6274 bitidx
&= (BITS_PER_LONG
-1);
6276 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6278 bitidx
+= end_bitidx
;
6279 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6280 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6282 word
= READ_ONCE(bitmap
[word_bitidx
]);
6284 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6285 if (word
== old_word
)
6292 * This function checks whether pageblock includes unmovable pages or not.
6293 * If @count is not zero, it is okay to include less @count unmovable pages
6295 * PageLRU check without isolation or lru_lock could race so that
6296 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6297 * expect this function should be exact.
6299 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6300 bool skip_hwpoisoned_pages
)
6302 unsigned long pfn
, iter
, found
;
6306 * For avoiding noise data, lru_add_drain_all() should be called
6307 * If ZONE_MOVABLE, the zone never contains unmovable pages
6309 if (zone_idx(zone
) == ZONE_MOVABLE
)
6311 mt
= get_pageblock_migratetype(page
);
6312 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6315 pfn
= page_to_pfn(page
);
6316 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6317 unsigned long check
= pfn
+ iter
;
6319 if (!pfn_valid_within(check
))
6322 page
= pfn_to_page(check
);
6325 * Hugepages are not in LRU lists, but they're movable.
6326 * We need not scan over tail pages bacause we don't
6327 * handle each tail page individually in migration.
6329 if (PageHuge(page
)) {
6330 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6335 * We can't use page_count without pin a page
6336 * because another CPU can free compound page.
6337 * This check already skips compound tails of THP
6338 * because their page->_count is zero at all time.
6340 if (!atomic_read(&page
->_count
)) {
6341 if (PageBuddy(page
))
6342 iter
+= (1 << page_order(page
)) - 1;
6347 * The HWPoisoned page may be not in buddy system, and
6348 * page_count() is not 0.
6350 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6356 * If there are RECLAIMABLE pages, we need to check
6357 * it. But now, memory offline itself doesn't call
6358 * shrink_node_slabs() and it still to be fixed.
6361 * If the page is not RAM, page_count()should be 0.
6362 * we don't need more check. This is an _used_ not-movable page.
6364 * The problematic thing here is PG_reserved pages. PG_reserved
6365 * is set to both of a memory hole page and a _used_ kernel
6374 bool is_pageblock_removable_nolock(struct page
*page
)
6380 * We have to be careful here because we are iterating over memory
6381 * sections which are not zone aware so we might end up outside of
6382 * the zone but still within the section.
6383 * We have to take care about the node as well. If the node is offline
6384 * its NODE_DATA will be NULL - see page_zone.
6386 if (!node_online(page_to_nid(page
)))
6389 zone
= page_zone(page
);
6390 pfn
= page_to_pfn(page
);
6391 if (!zone_spans_pfn(zone
, pfn
))
6394 return !has_unmovable_pages(zone
, page
, 0, true);
6399 static unsigned long pfn_max_align_down(unsigned long pfn
)
6401 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6402 pageblock_nr_pages
) - 1);
6405 static unsigned long pfn_max_align_up(unsigned long pfn
)
6407 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6408 pageblock_nr_pages
));
6411 /* [start, end) must belong to a single zone. */
6412 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6413 unsigned long start
, unsigned long end
)
6415 /* This function is based on compact_zone() from compaction.c. */
6416 unsigned long nr_reclaimed
;
6417 unsigned long pfn
= start
;
6418 unsigned int tries
= 0;
6423 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6424 if (fatal_signal_pending(current
)) {
6429 if (list_empty(&cc
->migratepages
)) {
6430 cc
->nr_migratepages
= 0;
6431 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6437 } else if (++tries
== 5) {
6438 ret
= ret
< 0 ? ret
: -EBUSY
;
6442 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6444 cc
->nr_migratepages
-= nr_reclaimed
;
6446 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6447 NULL
, 0, cc
->mode
, MR_CMA
);
6450 putback_movable_pages(&cc
->migratepages
);
6457 * alloc_contig_range() -- tries to allocate given range of pages
6458 * @start: start PFN to allocate
6459 * @end: one-past-the-last PFN to allocate
6460 * @migratetype: migratetype of the underlaying pageblocks (either
6461 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6462 * in range must have the same migratetype and it must
6463 * be either of the two.
6465 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6466 * aligned, however it's the caller's responsibility to guarantee that
6467 * we are the only thread that changes migrate type of pageblocks the
6470 * The PFN range must belong to a single zone.
6472 * Returns zero on success or negative error code. On success all
6473 * pages which PFN is in [start, end) are allocated for the caller and
6474 * need to be freed with free_contig_range().
6476 int alloc_contig_range(unsigned long start
, unsigned long end
,
6477 unsigned migratetype
)
6479 unsigned long outer_start
, outer_end
;
6482 struct compact_control cc
= {
6483 .nr_migratepages
= 0,
6485 .zone
= page_zone(pfn_to_page(start
)),
6486 .mode
= MIGRATE_SYNC
,
6487 .ignore_skip_hint
= true,
6489 INIT_LIST_HEAD(&cc
.migratepages
);
6492 * What we do here is we mark all pageblocks in range as
6493 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6494 * have different sizes, and due to the way page allocator
6495 * work, we align the range to biggest of the two pages so
6496 * that page allocator won't try to merge buddies from
6497 * different pageblocks and change MIGRATE_ISOLATE to some
6498 * other migration type.
6500 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6501 * migrate the pages from an unaligned range (ie. pages that
6502 * we are interested in). This will put all the pages in
6503 * range back to page allocator as MIGRATE_ISOLATE.
6505 * When this is done, we take the pages in range from page
6506 * allocator removing them from the buddy system. This way
6507 * page allocator will never consider using them.
6509 * This lets us mark the pageblocks back as
6510 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6511 * aligned range but not in the unaligned, original range are
6512 * put back to page allocator so that buddy can use them.
6515 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6516 pfn_max_align_up(end
), migratetype
,
6521 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6526 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6527 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6528 * more, all pages in [start, end) are free in page allocator.
6529 * What we are going to do is to allocate all pages from
6530 * [start, end) (that is remove them from page allocator).
6532 * The only problem is that pages at the beginning and at the
6533 * end of interesting range may be not aligned with pages that
6534 * page allocator holds, ie. they can be part of higher order
6535 * pages. Because of this, we reserve the bigger range and
6536 * once this is done free the pages we are not interested in.
6538 * We don't have to hold zone->lock here because the pages are
6539 * isolated thus they won't get removed from buddy.
6542 lru_add_drain_all();
6543 drain_all_pages(cc
.zone
);
6546 outer_start
= start
;
6547 while (!PageBuddy(pfn_to_page(outer_start
))) {
6548 if (++order
>= MAX_ORDER
) {
6552 outer_start
&= ~0UL << order
;
6555 /* Make sure the range is really isolated. */
6556 if (test_pages_isolated(outer_start
, end
, false)) {
6557 pr_info("%s: [%lx, %lx) PFNs busy\n",
6558 __func__
, outer_start
, end
);
6563 /* Grab isolated pages from freelists. */
6564 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6570 /* Free head and tail (if any) */
6571 if (start
!= outer_start
)
6572 free_contig_range(outer_start
, start
- outer_start
);
6573 if (end
!= outer_end
)
6574 free_contig_range(end
, outer_end
- end
);
6577 undo_isolate_page_range(pfn_max_align_down(start
),
6578 pfn_max_align_up(end
), migratetype
);
6582 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6584 unsigned int count
= 0;
6586 for (; nr_pages
--; pfn
++) {
6587 struct page
*page
= pfn_to_page(pfn
);
6589 count
+= page_count(page
) != 1;
6592 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6596 #ifdef CONFIG_MEMORY_HOTPLUG
6598 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6599 * page high values need to be recalulated.
6601 void __meminit
zone_pcp_update(struct zone
*zone
)
6604 mutex_lock(&pcp_batch_high_lock
);
6605 for_each_possible_cpu(cpu
)
6606 pageset_set_high_and_batch(zone
,
6607 per_cpu_ptr(zone
->pageset
, cpu
));
6608 mutex_unlock(&pcp_batch_high_lock
);
6612 void zone_pcp_reset(struct zone
*zone
)
6614 unsigned long flags
;
6616 struct per_cpu_pageset
*pset
;
6618 /* avoid races with drain_pages() */
6619 local_irq_save(flags
);
6620 if (zone
->pageset
!= &boot_pageset
) {
6621 for_each_online_cpu(cpu
) {
6622 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6623 drain_zonestat(zone
, pset
);
6625 free_percpu(zone
->pageset
);
6626 zone
->pageset
= &boot_pageset
;
6628 local_irq_restore(flags
);
6631 #ifdef CONFIG_MEMORY_HOTREMOVE
6633 * All pages in the range must be isolated before calling this.
6636 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6640 unsigned int order
, i
;
6642 unsigned long flags
;
6643 /* find the first valid pfn */
6644 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6649 zone
= page_zone(pfn_to_page(pfn
));
6650 spin_lock_irqsave(&zone
->lock
, flags
);
6652 while (pfn
< end_pfn
) {
6653 if (!pfn_valid(pfn
)) {
6657 page
= pfn_to_page(pfn
);
6659 * The HWPoisoned page may be not in buddy system, and
6660 * page_count() is not 0.
6662 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6664 SetPageReserved(page
);
6668 BUG_ON(page_count(page
));
6669 BUG_ON(!PageBuddy(page
));
6670 order
= page_order(page
);
6671 #ifdef CONFIG_DEBUG_VM
6672 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6673 pfn
, 1 << order
, end_pfn
);
6675 list_del(&page
->lru
);
6676 rmv_page_order(page
);
6677 zone
->free_area
[order
].nr_free
--;
6678 for (i
= 0; i
< (1 << order
); i
++)
6679 SetPageReserved((page
+i
));
6680 pfn
+= (1 << order
);
6682 spin_unlock_irqrestore(&zone
->lock
, flags
);
6686 #ifdef CONFIG_MEMORY_FAILURE
6687 bool is_free_buddy_page(struct page
*page
)
6689 struct zone
*zone
= page_zone(page
);
6690 unsigned long pfn
= page_to_pfn(page
);
6691 unsigned long flags
;
6694 spin_lock_irqsave(&zone
->lock
, flags
);
6695 for (order
= 0; order
< MAX_ORDER
; order
++) {
6696 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6698 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6701 spin_unlock_irqrestore(&zone
->lock
, flags
);
6703 return order
< MAX_ORDER
;