2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock
);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node
);
76 EXPORT_PER_CPU_SYMBOL(numa_node
);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
88 int _node_numa_mem_
[MAX_NUMNODES
];
92 * Array of node states.
94 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
95 [N_POSSIBLE
] = NODE_MASK_ALL
,
96 [N_ONLINE
] = { { [0] = 1UL } },
98 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
100 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
102 #ifdef CONFIG_MOVABLE_NODE
103 [N_MEMORY
] = { { [0] = 1UL } },
105 [N_CPU
] = { { [0] = 1UL } },
108 EXPORT_SYMBOL(node_states
);
110 /* Protect totalram_pages and zone->managed_pages */
111 static DEFINE_SPINLOCK(managed_page_count_lock
);
113 unsigned long totalram_pages __read_mostly
;
114 unsigned long totalreserve_pages __read_mostly
;
115 unsigned long totalcma_pages __read_mostly
;
117 * When calculating the number of globally allowed dirty pages, there
118 * is a certain number of per-zone reserves that should not be
119 * considered dirtyable memory. This is the sum of those reserves
120 * over all existing zones that contribute dirtyable memory.
122 unsigned long dirty_balance_reserve __read_mostly
;
124 int percpu_pagelist_fraction
;
125 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
127 #ifdef CONFIG_PM_SLEEP
129 * The following functions are used by the suspend/hibernate code to temporarily
130 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
131 * while devices are suspended. To avoid races with the suspend/hibernate code,
132 * they should always be called with pm_mutex held (gfp_allowed_mask also should
133 * only be modified with pm_mutex held, unless the suspend/hibernate code is
134 * guaranteed not to run in parallel with that modification).
137 static gfp_t saved_gfp_mask
;
139 void pm_restore_gfp_mask(void)
141 WARN_ON(!mutex_is_locked(&pm_mutex
));
142 if (saved_gfp_mask
) {
143 gfp_allowed_mask
= saved_gfp_mask
;
148 void pm_restrict_gfp_mask(void)
150 WARN_ON(!mutex_is_locked(&pm_mutex
));
151 WARN_ON(saved_gfp_mask
);
152 saved_gfp_mask
= gfp_allowed_mask
;
153 gfp_allowed_mask
&= ~GFP_IOFS
;
156 bool pm_suspended_storage(void)
158 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
162 #endif /* CONFIG_PM_SLEEP */
164 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
165 int pageblock_order __read_mostly
;
168 static void __free_pages_ok(struct page
*page
, unsigned int order
);
171 * results with 256, 32 in the lowmem_reserve sysctl:
172 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
173 * 1G machine -> (16M dma, 784M normal, 224M high)
174 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
175 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
176 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
178 * TBD: should special case ZONE_DMA32 machines here - in those we normally
179 * don't need any ZONE_NORMAL reservation
181 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
182 #ifdef CONFIG_ZONE_DMA
185 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 EXPORT_SYMBOL(totalram_pages
);
196 static char * const zone_names
[MAX_NR_ZONES
] = {
197 #ifdef CONFIG_ZONE_DMA
200 #ifdef CONFIG_ZONE_DMA32
204 #ifdef CONFIG_HIGHMEM
210 int min_free_kbytes
= 1024;
211 int user_min_free_kbytes
= -1;
213 static unsigned long __meminitdata nr_kernel_pages
;
214 static unsigned long __meminitdata nr_all_pages
;
215 static unsigned long __meminitdata dma_reserve
;
217 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
218 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
219 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
220 static unsigned long __initdata required_kernelcore
;
221 static unsigned long __initdata required_movablecore
;
222 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
224 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
226 EXPORT_SYMBOL(movable_zone
);
227 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
230 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
231 int nr_online_nodes __read_mostly
= 1;
232 EXPORT_SYMBOL(nr_node_ids
);
233 EXPORT_SYMBOL(nr_online_nodes
);
236 int page_group_by_mobility_disabled __read_mostly
;
238 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
240 if (unlikely(page_group_by_mobility_disabled
&&
241 migratetype
< MIGRATE_PCPTYPES
))
242 migratetype
= MIGRATE_UNMOVABLE
;
244 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
245 PB_migrate
, PB_migrate_end
);
248 #ifdef CONFIG_DEBUG_VM
249 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
253 unsigned long pfn
= page_to_pfn(page
);
254 unsigned long sp
, start_pfn
;
257 seq
= zone_span_seqbegin(zone
);
258 start_pfn
= zone
->zone_start_pfn
;
259 sp
= zone
->spanned_pages
;
260 if (!zone_spans_pfn(zone
, pfn
))
262 } while (zone_span_seqretry(zone
, seq
));
265 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
266 pfn
, zone_to_nid(zone
), zone
->name
,
267 start_pfn
, start_pfn
+ sp
);
272 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
274 if (!pfn_valid_within(page_to_pfn(page
)))
276 if (zone
!= page_zone(page
))
282 * Temporary debugging check for pages not lying within a given zone.
284 static int bad_range(struct zone
*zone
, struct page
*page
)
286 if (page_outside_zone_boundaries(zone
, page
))
288 if (!page_is_consistent(zone
, page
))
294 static inline int bad_range(struct zone
*zone
, struct page
*page
)
300 static void bad_page(struct page
*page
, const char *reason
,
301 unsigned long bad_flags
)
303 static unsigned long resume
;
304 static unsigned long nr_shown
;
305 static unsigned long nr_unshown
;
307 /* Don't complain about poisoned pages */
308 if (PageHWPoison(page
)) {
309 page_mapcount_reset(page
); /* remove PageBuddy */
314 * Allow a burst of 60 reports, then keep quiet for that minute;
315 * or allow a steady drip of one report per second.
317 if (nr_shown
== 60) {
318 if (time_before(jiffies
, resume
)) {
324 "BUG: Bad page state: %lu messages suppressed\n",
331 resume
= jiffies
+ 60 * HZ
;
333 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
334 current
->comm
, page_to_pfn(page
));
335 dump_page_badflags(page
, reason
, bad_flags
);
340 /* Leave bad fields for debug, except PageBuddy could make trouble */
341 page_mapcount_reset(page
); /* remove PageBuddy */
342 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
346 * Higher-order pages are called "compound pages". They are structured thusly:
348 * The first PAGE_SIZE page is called the "head page".
350 * The remaining PAGE_SIZE pages are called "tail pages".
352 * All pages have PG_compound set. All tail pages have their ->first_page
353 * pointing at the head page.
355 * The first tail page's ->lru.next holds the address of the compound page's
356 * put_page() function. Its ->lru.prev holds the order of allocation.
357 * This usage means that zero-order pages may not be compound.
360 static void free_compound_page(struct page
*page
)
362 __free_pages_ok(page
, compound_order(page
));
365 void prep_compound_page(struct page
*page
, unsigned long order
)
368 int nr_pages
= 1 << order
;
370 set_compound_page_dtor(page
, free_compound_page
);
371 set_compound_order(page
, order
);
373 for (i
= 1; i
< nr_pages
; i
++) {
374 struct page
*p
= page
+ i
;
375 set_page_count(p
, 0);
376 p
->first_page
= page
;
377 /* Make sure p->first_page is always valid for PageTail() */
383 static inline void prep_zero_page(struct page
*page
, unsigned int order
,
389 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
390 * and __GFP_HIGHMEM from hard or soft interrupt context.
392 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
393 for (i
= 0; i
< (1 << order
); i
++)
394 clear_highpage(page
+ i
);
397 #ifdef CONFIG_DEBUG_PAGEALLOC
398 unsigned int _debug_guardpage_minorder
;
399 bool _debug_pagealloc_enabled __read_mostly
;
400 bool _debug_guardpage_enabled __read_mostly
;
402 static int __init
early_debug_pagealloc(char *buf
)
407 if (strcmp(buf
, "on") == 0)
408 _debug_pagealloc_enabled
= true;
412 early_param("debug_pagealloc", early_debug_pagealloc
);
414 static bool need_debug_guardpage(void)
416 /* If we don't use debug_pagealloc, we don't need guard page */
417 if (!debug_pagealloc_enabled())
423 static void init_debug_guardpage(void)
425 if (!debug_pagealloc_enabled())
428 _debug_guardpage_enabled
= true;
431 struct page_ext_operations debug_guardpage_ops
= {
432 .need
= need_debug_guardpage
,
433 .init
= init_debug_guardpage
,
436 static int __init
debug_guardpage_minorder_setup(char *buf
)
440 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
441 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
444 _debug_guardpage_minorder
= res
;
445 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
448 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
450 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
451 unsigned int order
, int migratetype
)
453 struct page_ext
*page_ext
;
455 if (!debug_guardpage_enabled())
458 page_ext
= lookup_page_ext(page
);
459 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
461 INIT_LIST_HEAD(&page
->lru
);
462 set_page_private(page
, order
);
463 /* Guard pages are not available for any usage */
464 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
467 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
468 unsigned int order
, int migratetype
)
470 struct page_ext
*page_ext
;
472 if (!debug_guardpage_enabled())
475 page_ext
= lookup_page_ext(page
);
476 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
478 set_page_private(page
, 0);
479 if (!is_migrate_isolate(migratetype
))
480 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
483 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
484 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
485 unsigned int order
, int migratetype
) {}
486 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
487 unsigned int order
, int migratetype
) {}
490 static inline void set_page_order(struct page
*page
, unsigned int order
)
492 set_page_private(page
, order
);
493 __SetPageBuddy(page
);
496 static inline void rmv_page_order(struct page
*page
)
498 __ClearPageBuddy(page
);
499 set_page_private(page
, 0);
503 * This function checks whether a page is free && is the buddy
504 * we can do coalesce a page and its buddy if
505 * (a) the buddy is not in a hole &&
506 * (b) the buddy is in the buddy system &&
507 * (c) a page and its buddy have the same order &&
508 * (d) a page and its buddy are in the same zone.
510 * For recording whether a page is in the buddy system, we set ->_mapcount
511 * PAGE_BUDDY_MAPCOUNT_VALUE.
512 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
513 * serialized by zone->lock.
515 * For recording page's order, we use page_private(page).
517 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
520 if (!pfn_valid_within(page_to_pfn(buddy
)))
523 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
524 if (page_zone_id(page
) != page_zone_id(buddy
))
527 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
532 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
534 * zone check is done late to avoid uselessly
535 * calculating zone/node ids for pages that could
538 if (page_zone_id(page
) != page_zone_id(buddy
))
541 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
549 * Freeing function for a buddy system allocator.
551 * The concept of a buddy system is to maintain direct-mapped table
552 * (containing bit values) for memory blocks of various "orders".
553 * The bottom level table contains the map for the smallest allocatable
554 * units of memory (here, pages), and each level above it describes
555 * pairs of units from the levels below, hence, "buddies".
556 * At a high level, all that happens here is marking the table entry
557 * at the bottom level available, and propagating the changes upward
558 * as necessary, plus some accounting needed to play nicely with other
559 * parts of the VM system.
560 * At each level, we keep a list of pages, which are heads of continuous
561 * free pages of length of (1 << order) and marked with _mapcount
562 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
564 * So when we are allocating or freeing one, we can derive the state of the
565 * other. That is, if we allocate a small block, and both were
566 * free, the remainder of the region must be split into blocks.
567 * If a block is freed, and its buddy is also free, then this
568 * triggers coalescing into a block of larger size.
573 static inline void __free_one_page(struct page
*page
,
575 struct zone
*zone
, unsigned int order
,
578 unsigned long page_idx
;
579 unsigned long combined_idx
;
580 unsigned long uninitialized_var(buddy_idx
);
582 int max_order
= MAX_ORDER
;
584 VM_BUG_ON(!zone_is_initialized(zone
));
585 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
587 VM_BUG_ON(migratetype
== -1);
588 if (is_migrate_isolate(migratetype
)) {
590 * We restrict max order of merging to prevent merge
591 * between freepages on isolate pageblock and normal
592 * pageblock. Without this, pageblock isolation
593 * could cause incorrect freepage accounting.
595 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
597 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
600 page_idx
= pfn
& ((1 << max_order
) - 1);
602 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
603 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
605 while (order
< max_order
- 1) {
606 buddy_idx
= __find_buddy_index(page_idx
, order
);
607 buddy
= page
+ (buddy_idx
- page_idx
);
608 if (!page_is_buddy(page
, buddy
, order
))
611 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
612 * merge with it and move up one order.
614 if (page_is_guard(buddy
)) {
615 clear_page_guard(zone
, buddy
, order
, migratetype
);
617 list_del(&buddy
->lru
);
618 zone
->free_area
[order
].nr_free
--;
619 rmv_page_order(buddy
);
621 combined_idx
= buddy_idx
& page_idx
;
622 page
= page
+ (combined_idx
- page_idx
);
623 page_idx
= combined_idx
;
626 set_page_order(page
, order
);
629 * If this is not the largest possible page, check if the buddy
630 * of the next-highest order is free. If it is, it's possible
631 * that pages are being freed that will coalesce soon. In case,
632 * that is happening, add the free page to the tail of the list
633 * so it's less likely to be used soon and more likely to be merged
634 * as a higher order page
636 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
637 struct page
*higher_page
, *higher_buddy
;
638 combined_idx
= buddy_idx
& page_idx
;
639 higher_page
= page
+ (combined_idx
- page_idx
);
640 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
641 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
642 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
643 list_add_tail(&page
->lru
,
644 &zone
->free_area
[order
].free_list
[migratetype
]);
649 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
651 zone
->free_area
[order
].nr_free
++;
654 static inline int free_pages_check(struct page
*page
)
656 const char *bad_reason
= NULL
;
657 unsigned long bad_flags
= 0;
659 if (unlikely(page_mapcount(page
)))
660 bad_reason
= "nonzero mapcount";
661 if (unlikely(page
->mapping
!= NULL
))
662 bad_reason
= "non-NULL mapping";
663 if (unlikely(atomic_read(&page
->_count
) != 0))
664 bad_reason
= "nonzero _count";
665 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
666 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
667 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
670 if (unlikely(page
->mem_cgroup
))
671 bad_reason
= "page still charged to cgroup";
673 if (unlikely(bad_reason
)) {
674 bad_page(page
, bad_reason
, bad_flags
);
677 page_cpupid_reset_last(page
);
678 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
679 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
684 * Frees a number of pages from the PCP lists
685 * Assumes all pages on list are in same zone, and of same order.
686 * count is the number of pages to free.
688 * If the zone was previously in an "all pages pinned" state then look to
689 * see if this freeing clears that state.
691 * And clear the zone's pages_scanned counter, to hold off the "all pages are
692 * pinned" detection logic.
694 static void free_pcppages_bulk(struct zone
*zone
, int count
,
695 struct per_cpu_pages
*pcp
)
700 unsigned long nr_scanned
;
702 spin_lock(&zone
->lock
);
703 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
705 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
709 struct list_head
*list
;
712 * Remove pages from lists in a round-robin fashion. A
713 * batch_free count is maintained that is incremented when an
714 * empty list is encountered. This is so more pages are freed
715 * off fuller lists instead of spinning excessively around empty
720 if (++migratetype
== MIGRATE_PCPTYPES
)
722 list
= &pcp
->lists
[migratetype
];
723 } while (list_empty(list
));
725 /* This is the only non-empty list. Free them all. */
726 if (batch_free
== MIGRATE_PCPTYPES
)
727 batch_free
= to_free
;
730 int mt
; /* migratetype of the to-be-freed page */
732 page
= list_entry(list
->prev
, struct page
, lru
);
733 /* must delete as __free_one_page list manipulates */
734 list_del(&page
->lru
);
735 mt
= get_freepage_migratetype(page
);
736 if (unlikely(has_isolate_pageblock(zone
)))
737 mt
= get_pageblock_migratetype(page
);
739 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
740 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
741 trace_mm_page_pcpu_drain(page
, 0, mt
);
742 } while (--to_free
&& --batch_free
&& !list_empty(list
));
744 spin_unlock(&zone
->lock
);
747 static void free_one_page(struct zone
*zone
,
748 struct page
*page
, unsigned long pfn
,
752 unsigned long nr_scanned
;
753 spin_lock(&zone
->lock
);
754 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
756 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
758 if (unlikely(has_isolate_pageblock(zone
) ||
759 is_migrate_isolate(migratetype
))) {
760 migratetype
= get_pfnblock_migratetype(page
, pfn
);
762 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
763 spin_unlock(&zone
->lock
);
766 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
768 if (!IS_ENABLED(CONFIG_DEBUG_VM
))
770 if (unlikely(!PageTail(page
))) {
771 bad_page(page
, "PageTail not set", 0);
774 if (unlikely(page
->first_page
!= head_page
)) {
775 bad_page(page
, "first_page not consistent", 0);
781 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
783 bool compound
= PageCompound(page
);
786 VM_BUG_ON_PAGE(PageTail(page
), page
);
787 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
789 trace_mm_page_free(page
, order
);
790 kmemcheck_free_shadow(page
, order
);
791 kasan_free_pages(page
, order
);
794 page
->mapping
= NULL
;
795 bad
+= free_pages_check(page
);
796 for (i
= 1; i
< (1 << order
); i
++) {
798 bad
+= free_tail_pages_check(page
, page
+ i
);
799 bad
+= free_pages_check(page
+ i
);
804 reset_page_owner(page
, order
);
806 if (!PageHighMem(page
)) {
807 debug_check_no_locks_freed(page_address(page
),
809 debug_check_no_obj_freed(page_address(page
),
812 arch_free_page(page
, order
);
813 kernel_map_pages(page
, 1 << order
, 0);
818 static void __free_pages_ok(struct page
*page
, unsigned int order
)
822 unsigned long pfn
= page_to_pfn(page
);
824 if (!free_pages_prepare(page
, order
))
827 migratetype
= get_pfnblock_migratetype(page
, pfn
);
828 local_irq_save(flags
);
829 __count_vm_events(PGFREE
, 1 << order
);
830 set_freepage_migratetype(page
, migratetype
);
831 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
832 local_irq_restore(flags
);
835 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
837 unsigned int nr_pages
= 1 << order
;
838 struct page
*p
= page
;
842 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
844 __ClearPageReserved(p
);
845 set_page_count(p
, 0);
847 __ClearPageReserved(p
);
848 set_page_count(p
, 0);
850 page_zone(page
)->managed_pages
+= nr_pages
;
851 set_page_refcounted(page
);
852 __free_pages(page
, order
);
856 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
857 void __init
init_cma_reserved_pageblock(struct page
*page
)
859 unsigned i
= pageblock_nr_pages
;
860 struct page
*p
= page
;
863 __ClearPageReserved(p
);
864 set_page_count(p
, 0);
867 set_pageblock_migratetype(page
, MIGRATE_CMA
);
869 if (pageblock_order
>= MAX_ORDER
) {
870 i
= pageblock_nr_pages
;
873 set_page_refcounted(p
);
874 __free_pages(p
, MAX_ORDER
- 1);
875 p
+= MAX_ORDER_NR_PAGES
;
876 } while (i
-= MAX_ORDER_NR_PAGES
);
878 set_page_refcounted(page
);
879 __free_pages(page
, pageblock_order
);
882 adjust_managed_page_count(page
, pageblock_nr_pages
);
887 * The order of subdivision here is critical for the IO subsystem.
888 * Please do not alter this order without good reasons and regression
889 * testing. Specifically, as large blocks of memory are subdivided,
890 * the order in which smaller blocks are delivered depends on the order
891 * they're subdivided in this function. This is the primary factor
892 * influencing the order in which pages are delivered to the IO
893 * subsystem according to empirical testing, and this is also justified
894 * by considering the behavior of a buddy system containing a single
895 * large block of memory acted on by a series of small allocations.
896 * This behavior is a critical factor in sglist merging's success.
900 static inline void expand(struct zone
*zone
, struct page
*page
,
901 int low
, int high
, struct free_area
*area
,
904 unsigned long size
= 1 << high
;
910 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
912 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
913 debug_guardpage_enabled() &&
914 high
< debug_guardpage_minorder()) {
916 * Mark as guard pages (or page), that will allow to
917 * merge back to allocator when buddy will be freed.
918 * Corresponding page table entries will not be touched,
919 * pages will stay not present in virtual address space
921 set_page_guard(zone
, &page
[size
], high
, migratetype
);
924 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
926 set_page_order(&page
[size
], high
);
931 * This page is about to be returned from the page allocator
933 static inline int check_new_page(struct page
*page
)
935 const char *bad_reason
= NULL
;
936 unsigned long bad_flags
= 0;
938 if (unlikely(page_mapcount(page
)))
939 bad_reason
= "nonzero mapcount";
940 if (unlikely(page
->mapping
!= NULL
))
941 bad_reason
= "non-NULL mapping";
942 if (unlikely(atomic_read(&page
->_count
) != 0))
943 bad_reason
= "nonzero _count";
944 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
945 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
946 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
949 if (unlikely(page
->mem_cgroup
))
950 bad_reason
= "page still charged to cgroup";
952 if (unlikely(bad_reason
)) {
953 bad_page(page
, bad_reason
, bad_flags
);
959 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
964 for (i
= 0; i
< (1 << order
); i
++) {
965 struct page
*p
= page
+ i
;
966 if (unlikely(check_new_page(p
)))
970 set_page_private(page
, 0);
971 set_page_refcounted(page
);
973 arch_alloc_page(page
, order
);
974 kernel_map_pages(page
, 1 << order
, 1);
975 kasan_alloc_pages(page
, order
);
977 if (gfp_flags
& __GFP_ZERO
)
978 prep_zero_page(page
, order
, gfp_flags
);
980 if (order
&& (gfp_flags
& __GFP_COMP
))
981 prep_compound_page(page
, order
);
983 set_page_owner(page
, order
, gfp_flags
);
986 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
987 * allocate the page. The expectation is that the caller is taking
988 * steps that will free more memory. The caller should avoid the page
989 * being used for !PFMEMALLOC purposes.
991 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
997 * Go through the free lists for the given migratetype and remove
998 * the smallest available page from the freelists
1001 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1004 unsigned int current_order
;
1005 struct free_area
*area
;
1008 /* Find a page of the appropriate size in the preferred list */
1009 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1010 area
= &(zone
->free_area
[current_order
]);
1011 if (list_empty(&area
->free_list
[migratetype
]))
1014 page
= list_entry(area
->free_list
[migratetype
].next
,
1016 list_del(&page
->lru
);
1017 rmv_page_order(page
);
1019 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1020 set_freepage_migratetype(page
, migratetype
);
1029 * This array describes the order lists are fallen back to when
1030 * the free lists for the desirable migrate type are depleted
1032 static int fallbacks
[MIGRATE_TYPES
][4] = {
1033 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1034 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1036 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
1037 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
1039 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
1041 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
1042 #ifdef CONFIG_MEMORY_ISOLATION
1043 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
1048 * Move the free pages in a range to the free lists of the requested type.
1049 * Note that start_page and end_pages are not aligned on a pageblock
1050 * boundary. If alignment is required, use move_freepages_block()
1052 int move_freepages(struct zone
*zone
,
1053 struct page
*start_page
, struct page
*end_page
,
1057 unsigned long order
;
1058 int pages_moved
= 0;
1060 #ifndef CONFIG_HOLES_IN_ZONE
1062 * page_zone is not safe to call in this context when
1063 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1064 * anyway as we check zone boundaries in move_freepages_block().
1065 * Remove at a later date when no bug reports exist related to
1066 * grouping pages by mobility
1068 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1071 for (page
= start_page
; page
<= end_page
;) {
1072 /* Make sure we are not inadvertently changing nodes */
1073 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1075 if (!pfn_valid_within(page_to_pfn(page
))) {
1080 if (!PageBuddy(page
)) {
1085 order
= page_order(page
);
1086 list_move(&page
->lru
,
1087 &zone
->free_area
[order
].free_list
[migratetype
]);
1088 set_freepage_migratetype(page
, migratetype
);
1090 pages_moved
+= 1 << order
;
1096 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1099 unsigned long start_pfn
, end_pfn
;
1100 struct page
*start_page
, *end_page
;
1102 start_pfn
= page_to_pfn(page
);
1103 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1104 start_page
= pfn_to_page(start_pfn
);
1105 end_page
= start_page
+ pageblock_nr_pages
- 1;
1106 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1108 /* Do not cross zone boundaries */
1109 if (!zone_spans_pfn(zone
, start_pfn
))
1111 if (!zone_spans_pfn(zone
, end_pfn
))
1114 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1117 static void change_pageblock_range(struct page
*pageblock_page
,
1118 int start_order
, int migratetype
)
1120 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1122 while (nr_pageblocks
--) {
1123 set_pageblock_migratetype(pageblock_page
, migratetype
);
1124 pageblock_page
+= pageblock_nr_pages
;
1129 * When we are falling back to another migratetype during allocation, try to
1130 * steal extra free pages from the same pageblocks to satisfy further
1131 * allocations, instead of polluting multiple pageblocks.
1133 * If we are stealing a relatively large buddy page, it is likely there will
1134 * be more free pages in the pageblock, so try to steal them all. For
1135 * reclaimable and unmovable allocations, we steal regardless of page size,
1136 * as fragmentation caused by those allocations polluting movable pageblocks
1137 * is worse than movable allocations stealing from unmovable and reclaimable
1140 * If we claim more than half of the pageblock, change pageblock's migratetype
1143 static void try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1144 int start_type
, int fallback_type
)
1146 int current_order
= page_order(page
);
1148 /* Take ownership for orders >= pageblock_order */
1149 if (current_order
>= pageblock_order
) {
1150 change_pageblock_range(page
, current_order
, start_type
);
1154 if (current_order
>= pageblock_order
/ 2 ||
1155 start_type
== MIGRATE_RECLAIMABLE
||
1156 start_type
== MIGRATE_UNMOVABLE
||
1157 page_group_by_mobility_disabled
) {
1160 pages
= move_freepages_block(zone
, page
, start_type
);
1162 /* Claim the whole block if over half of it is free */
1163 if (pages
>= (1 << (pageblock_order
-1)) ||
1164 page_group_by_mobility_disabled
)
1165 set_pageblock_migratetype(page
, start_type
);
1169 /* Remove an element from the buddy allocator from the fallback list */
1170 static inline struct page
*
1171 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1173 struct free_area
*area
;
1174 unsigned int current_order
;
1177 /* Find the largest possible block of pages in the other list */
1178 for (current_order
= MAX_ORDER
-1;
1179 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1183 int migratetype
= fallbacks
[start_migratetype
][i
];
1184 int buddy_type
= start_migratetype
;
1186 /* MIGRATE_RESERVE handled later if necessary */
1187 if (migratetype
== MIGRATE_RESERVE
)
1190 area
= &(zone
->free_area
[current_order
]);
1191 if (list_empty(&area
->free_list
[migratetype
]))
1194 page
= list_entry(area
->free_list
[migratetype
].next
,
1198 if (!is_migrate_cma(migratetype
)) {
1199 try_to_steal_freepages(zone
, page
,
1204 * When borrowing from MIGRATE_CMA, we need to
1205 * release the excess buddy pages to CMA
1206 * itself, and we do not try to steal extra
1209 buddy_type
= migratetype
;
1212 /* Remove the page from the freelists */
1213 list_del(&page
->lru
);
1214 rmv_page_order(page
);
1216 expand(zone
, page
, order
, current_order
, area
,
1220 * The freepage_migratetype may differ from pageblock's
1221 * migratetype depending on the decisions in
1222 * try_to_steal_freepages(). This is OK as long as it
1223 * does not differ for MIGRATE_CMA pageblocks. For CMA
1224 * we need to make sure unallocated pages flushed from
1225 * pcp lists are returned to the correct freelist.
1227 set_freepage_migratetype(page
, buddy_type
);
1229 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1230 start_migratetype
, migratetype
);
1240 * Do the hard work of removing an element from the buddy allocator.
1241 * Call me with the zone->lock already held.
1243 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1249 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1251 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1252 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1255 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1256 * is used because __rmqueue_smallest is an inline function
1257 * and we want just one call site
1260 migratetype
= MIGRATE_RESERVE
;
1265 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1270 * Obtain a specified number of elements from the buddy allocator, all under
1271 * a single hold of the lock, for efficiency. Add them to the supplied list.
1272 * Returns the number of new pages which were placed at *list.
1274 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1275 unsigned long count
, struct list_head
*list
,
1276 int migratetype
, bool cold
)
1280 spin_lock(&zone
->lock
);
1281 for (i
= 0; i
< count
; ++i
) {
1282 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1283 if (unlikely(page
== NULL
))
1287 * Split buddy pages returned by expand() are received here
1288 * in physical page order. The page is added to the callers and
1289 * list and the list head then moves forward. From the callers
1290 * perspective, the linked list is ordered by page number in
1291 * some conditions. This is useful for IO devices that can
1292 * merge IO requests if the physical pages are ordered
1296 list_add(&page
->lru
, list
);
1298 list_add_tail(&page
->lru
, list
);
1300 if (is_migrate_cma(get_freepage_migratetype(page
)))
1301 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1304 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1305 spin_unlock(&zone
->lock
);
1311 * Called from the vmstat counter updater to drain pagesets of this
1312 * currently executing processor on remote nodes after they have
1315 * Note that this function must be called with the thread pinned to
1316 * a single processor.
1318 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1320 unsigned long flags
;
1321 int to_drain
, batch
;
1323 local_irq_save(flags
);
1324 batch
= ACCESS_ONCE(pcp
->batch
);
1325 to_drain
= min(pcp
->count
, batch
);
1327 free_pcppages_bulk(zone
, to_drain
, pcp
);
1328 pcp
->count
-= to_drain
;
1330 local_irq_restore(flags
);
1335 * Drain pcplists of the indicated processor and zone.
1337 * The processor must either be the current processor and the
1338 * thread pinned to the current processor or a processor that
1341 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1343 unsigned long flags
;
1344 struct per_cpu_pageset
*pset
;
1345 struct per_cpu_pages
*pcp
;
1347 local_irq_save(flags
);
1348 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1352 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1355 local_irq_restore(flags
);
1359 * Drain pcplists of all zones on the indicated processor.
1361 * The processor must either be the current processor and the
1362 * thread pinned to the current processor or a processor that
1365 static void drain_pages(unsigned int cpu
)
1369 for_each_populated_zone(zone
) {
1370 drain_pages_zone(cpu
, zone
);
1375 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1377 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1378 * the single zone's pages.
1380 void drain_local_pages(struct zone
*zone
)
1382 int cpu
= smp_processor_id();
1385 drain_pages_zone(cpu
, zone
);
1391 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1393 * When zone parameter is non-NULL, spill just the single zone's pages.
1395 * Note that this code is protected against sending an IPI to an offline
1396 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1397 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1398 * nothing keeps CPUs from showing up after we populated the cpumask and
1399 * before the call to on_each_cpu_mask().
1401 void drain_all_pages(struct zone
*zone
)
1406 * Allocate in the BSS so we wont require allocation in
1407 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1409 static cpumask_t cpus_with_pcps
;
1412 * We don't care about racing with CPU hotplug event
1413 * as offline notification will cause the notified
1414 * cpu to drain that CPU pcps and on_each_cpu_mask
1415 * disables preemption as part of its processing
1417 for_each_online_cpu(cpu
) {
1418 struct per_cpu_pageset
*pcp
;
1420 bool has_pcps
= false;
1423 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1427 for_each_populated_zone(z
) {
1428 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1429 if (pcp
->pcp
.count
) {
1437 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1439 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1441 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1445 #ifdef CONFIG_HIBERNATION
1447 void mark_free_pages(struct zone
*zone
)
1449 unsigned long pfn
, max_zone_pfn
;
1450 unsigned long flags
;
1451 unsigned int order
, t
;
1452 struct list_head
*curr
;
1454 if (zone_is_empty(zone
))
1457 spin_lock_irqsave(&zone
->lock
, flags
);
1459 max_zone_pfn
= zone_end_pfn(zone
);
1460 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1461 if (pfn_valid(pfn
)) {
1462 struct page
*page
= pfn_to_page(pfn
);
1464 if (!swsusp_page_is_forbidden(page
))
1465 swsusp_unset_page_free(page
);
1468 for_each_migratetype_order(order
, t
) {
1469 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1472 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1473 for (i
= 0; i
< (1UL << order
); i
++)
1474 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1477 spin_unlock_irqrestore(&zone
->lock
, flags
);
1479 #endif /* CONFIG_PM */
1482 * Free a 0-order page
1483 * cold == true ? free a cold page : free a hot page
1485 void free_hot_cold_page(struct page
*page
, bool cold
)
1487 struct zone
*zone
= page_zone(page
);
1488 struct per_cpu_pages
*pcp
;
1489 unsigned long flags
;
1490 unsigned long pfn
= page_to_pfn(page
);
1493 if (!free_pages_prepare(page
, 0))
1496 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1497 set_freepage_migratetype(page
, migratetype
);
1498 local_irq_save(flags
);
1499 __count_vm_event(PGFREE
);
1502 * We only track unmovable, reclaimable and movable on pcp lists.
1503 * Free ISOLATE pages back to the allocator because they are being
1504 * offlined but treat RESERVE as movable pages so we can get those
1505 * areas back if necessary. Otherwise, we may have to free
1506 * excessively into the page allocator
1508 if (migratetype
>= MIGRATE_PCPTYPES
) {
1509 if (unlikely(is_migrate_isolate(migratetype
))) {
1510 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1513 migratetype
= MIGRATE_MOVABLE
;
1516 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1518 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1520 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1522 if (pcp
->count
>= pcp
->high
) {
1523 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1524 free_pcppages_bulk(zone
, batch
, pcp
);
1525 pcp
->count
-= batch
;
1529 local_irq_restore(flags
);
1533 * Free a list of 0-order pages
1535 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1537 struct page
*page
, *next
;
1539 list_for_each_entry_safe(page
, next
, list
, lru
) {
1540 trace_mm_page_free_batched(page
, cold
);
1541 free_hot_cold_page(page
, cold
);
1546 * split_page takes a non-compound higher-order page, and splits it into
1547 * n (1<<order) sub-pages: page[0..n]
1548 * Each sub-page must be freed individually.
1550 * Note: this is probably too low level an operation for use in drivers.
1551 * Please consult with lkml before using this in your driver.
1553 void split_page(struct page
*page
, unsigned int order
)
1557 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1558 VM_BUG_ON_PAGE(!page_count(page
), page
);
1560 #ifdef CONFIG_KMEMCHECK
1562 * Split shadow pages too, because free(page[0]) would
1563 * otherwise free the whole shadow.
1565 if (kmemcheck_page_is_tracked(page
))
1566 split_page(virt_to_page(page
[0].shadow
), order
);
1569 set_page_owner(page
, 0, 0);
1570 for (i
= 1; i
< (1 << order
); i
++) {
1571 set_page_refcounted(page
+ i
);
1572 set_page_owner(page
+ i
, 0, 0);
1575 EXPORT_SYMBOL_GPL(split_page
);
1577 int __isolate_free_page(struct page
*page
, unsigned int order
)
1579 unsigned long watermark
;
1583 BUG_ON(!PageBuddy(page
));
1585 zone
= page_zone(page
);
1586 mt
= get_pageblock_migratetype(page
);
1588 if (!is_migrate_isolate(mt
)) {
1589 /* Obey watermarks as if the page was being allocated */
1590 watermark
= low_wmark_pages(zone
) + (1 << order
);
1591 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1594 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1597 /* Remove page from free list */
1598 list_del(&page
->lru
);
1599 zone
->free_area
[order
].nr_free
--;
1600 rmv_page_order(page
);
1602 /* Set the pageblock if the isolated page is at least a pageblock */
1603 if (order
>= pageblock_order
- 1) {
1604 struct page
*endpage
= page
+ (1 << order
) - 1;
1605 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1606 int mt
= get_pageblock_migratetype(page
);
1607 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1608 set_pageblock_migratetype(page
,
1613 set_page_owner(page
, order
, 0);
1614 return 1UL << order
;
1618 * Similar to split_page except the page is already free. As this is only
1619 * being used for migration, the migratetype of the block also changes.
1620 * As this is called with interrupts disabled, the caller is responsible
1621 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1624 * Note: this is probably too low level an operation for use in drivers.
1625 * Please consult with lkml before using this in your driver.
1627 int split_free_page(struct page
*page
)
1632 order
= page_order(page
);
1634 nr_pages
= __isolate_free_page(page
, order
);
1638 /* Split into individual pages */
1639 set_page_refcounted(page
);
1640 split_page(page
, order
);
1645 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1648 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1649 struct zone
*zone
, unsigned int order
,
1650 gfp_t gfp_flags
, int migratetype
)
1652 unsigned long flags
;
1654 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1656 if (likely(order
== 0)) {
1657 struct per_cpu_pages
*pcp
;
1658 struct list_head
*list
;
1660 local_irq_save(flags
);
1661 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1662 list
= &pcp
->lists
[migratetype
];
1663 if (list_empty(list
)) {
1664 pcp
->count
+= rmqueue_bulk(zone
, 0,
1667 if (unlikely(list_empty(list
)))
1672 page
= list_entry(list
->prev
, struct page
, lru
);
1674 page
= list_entry(list
->next
, struct page
, lru
);
1676 list_del(&page
->lru
);
1679 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1681 * __GFP_NOFAIL is not to be used in new code.
1683 * All __GFP_NOFAIL callers should be fixed so that they
1684 * properly detect and handle allocation failures.
1686 * We most definitely don't want callers attempting to
1687 * allocate greater than order-1 page units with
1690 WARN_ON_ONCE(order
> 1);
1692 spin_lock_irqsave(&zone
->lock
, flags
);
1693 page
= __rmqueue(zone
, order
, migratetype
);
1694 spin_unlock(&zone
->lock
);
1697 __mod_zone_freepage_state(zone
, -(1 << order
),
1698 get_freepage_migratetype(page
));
1701 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1702 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1703 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1704 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1706 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1707 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1708 local_irq_restore(flags
);
1710 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1714 local_irq_restore(flags
);
1718 #ifdef CONFIG_FAIL_PAGE_ALLOC
1721 struct fault_attr attr
;
1723 u32 ignore_gfp_highmem
;
1724 u32 ignore_gfp_wait
;
1726 } fail_page_alloc
= {
1727 .attr
= FAULT_ATTR_INITIALIZER
,
1728 .ignore_gfp_wait
= 1,
1729 .ignore_gfp_highmem
= 1,
1733 static int __init
setup_fail_page_alloc(char *str
)
1735 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1737 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1739 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1741 if (order
< fail_page_alloc
.min_order
)
1743 if (gfp_mask
& __GFP_NOFAIL
)
1745 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1747 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1750 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1753 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1755 static int __init
fail_page_alloc_debugfs(void)
1757 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1760 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1761 &fail_page_alloc
.attr
);
1763 return PTR_ERR(dir
);
1765 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1766 &fail_page_alloc
.ignore_gfp_wait
))
1768 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1769 &fail_page_alloc
.ignore_gfp_highmem
))
1771 if (!debugfs_create_u32("min-order", mode
, dir
,
1772 &fail_page_alloc
.min_order
))
1777 debugfs_remove_recursive(dir
);
1782 late_initcall(fail_page_alloc_debugfs
);
1784 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1786 #else /* CONFIG_FAIL_PAGE_ALLOC */
1788 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1793 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1796 * Return true if free pages are above 'mark'. This takes into account the order
1797 * of the allocation.
1799 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1800 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1803 /* free_pages may go negative - that's OK */
1808 free_pages
-= (1 << order
) - 1;
1809 if (alloc_flags
& ALLOC_HIGH
)
1811 if (alloc_flags
& ALLOC_HARDER
)
1814 /* If allocation can't use CMA areas don't use free CMA pages */
1815 if (!(alloc_flags
& ALLOC_CMA
))
1816 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1819 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1821 for (o
= 0; o
< order
; o
++) {
1822 /* At the next order, this order's pages become unavailable */
1823 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1825 /* Require fewer higher order pages to be free */
1828 if (free_pages
<= min
)
1834 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1835 int classzone_idx
, int alloc_flags
)
1837 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1838 zone_page_state(z
, NR_FREE_PAGES
));
1841 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1842 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1844 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1846 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1847 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1849 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1855 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1856 * skip over zones that are not allowed by the cpuset, or that have
1857 * been recently (in last second) found to be nearly full. See further
1858 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1859 * that have to skip over a lot of full or unallowed zones.
1861 * If the zonelist cache is present in the passed zonelist, then
1862 * returns a pointer to the allowed node mask (either the current
1863 * tasks mems_allowed, or node_states[N_MEMORY].)
1865 * If the zonelist cache is not available for this zonelist, does
1866 * nothing and returns NULL.
1868 * If the fullzones BITMAP in the zonelist cache is stale (more than
1869 * a second since last zap'd) then we zap it out (clear its bits.)
1871 * We hold off even calling zlc_setup, until after we've checked the
1872 * first zone in the zonelist, on the theory that most allocations will
1873 * be satisfied from that first zone, so best to examine that zone as
1874 * quickly as we can.
1876 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1878 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1879 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1881 zlc
= zonelist
->zlcache_ptr
;
1885 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1886 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1887 zlc
->last_full_zap
= jiffies
;
1890 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1891 &cpuset_current_mems_allowed
:
1892 &node_states
[N_MEMORY
];
1893 return allowednodes
;
1897 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1898 * if it is worth looking at further for free memory:
1899 * 1) Check that the zone isn't thought to be full (doesn't have its
1900 * bit set in the zonelist_cache fullzones BITMAP).
1901 * 2) Check that the zones node (obtained from the zonelist_cache
1902 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1903 * Return true (non-zero) if zone is worth looking at further, or
1904 * else return false (zero) if it is not.
1906 * This check -ignores- the distinction between various watermarks,
1907 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1908 * found to be full for any variation of these watermarks, it will
1909 * be considered full for up to one second by all requests, unless
1910 * we are so low on memory on all allowed nodes that we are forced
1911 * into the second scan of the zonelist.
1913 * In the second scan we ignore this zonelist cache and exactly
1914 * apply the watermarks to all zones, even it is slower to do so.
1915 * We are low on memory in the second scan, and should leave no stone
1916 * unturned looking for a free page.
1918 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1919 nodemask_t
*allowednodes
)
1921 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1922 int i
; /* index of *z in zonelist zones */
1923 int n
; /* node that zone *z is on */
1925 zlc
= zonelist
->zlcache_ptr
;
1929 i
= z
- zonelist
->_zonerefs
;
1932 /* This zone is worth trying if it is allowed but not full */
1933 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1937 * Given 'z' scanning a zonelist, set the corresponding bit in
1938 * zlc->fullzones, so that subsequent attempts to allocate a page
1939 * from that zone don't waste time re-examining it.
1941 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1943 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1944 int i
; /* index of *z in zonelist zones */
1946 zlc
= zonelist
->zlcache_ptr
;
1950 i
= z
- zonelist
->_zonerefs
;
1952 set_bit(i
, zlc
->fullzones
);
1956 * clear all zones full, called after direct reclaim makes progress so that
1957 * a zone that was recently full is not skipped over for up to a second
1959 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1961 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1963 zlc
= zonelist
->zlcache_ptr
;
1967 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1970 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1972 return local_zone
->node
== zone
->node
;
1975 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1977 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1981 #else /* CONFIG_NUMA */
1983 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1988 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1989 nodemask_t
*allowednodes
)
1994 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1998 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
2002 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2007 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2012 #endif /* CONFIG_NUMA */
2014 static void reset_alloc_batches(struct zone
*preferred_zone
)
2016 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
2019 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2020 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
2021 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
2022 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2023 } while (zone
++ != preferred_zone
);
2027 * get_page_from_freelist goes through the zonelist trying to allocate
2030 static struct page
*
2031 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2032 const struct alloc_context
*ac
)
2034 struct zonelist
*zonelist
= ac
->zonelist
;
2036 struct page
*page
= NULL
;
2038 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
2039 int zlc_active
= 0; /* set if using zonelist_cache */
2040 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
2041 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
2042 (gfp_mask
& __GFP_WRITE
);
2043 int nr_fair_skipped
= 0;
2044 bool zonelist_rescan
;
2047 zonelist_rescan
= false;
2050 * Scan zonelist, looking for a zone with enough free.
2051 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2053 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2057 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2058 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2060 if (cpusets_enabled() &&
2061 (alloc_flags
& ALLOC_CPUSET
) &&
2062 !cpuset_zone_allowed(zone
, gfp_mask
))
2065 * Distribute pages in proportion to the individual
2066 * zone size to ensure fair page aging. The zone a
2067 * page was allocated in should have no effect on the
2068 * time the page has in memory before being reclaimed.
2070 if (alloc_flags
& ALLOC_FAIR
) {
2071 if (!zone_local(ac
->preferred_zone
, zone
))
2073 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2079 * When allocating a page cache page for writing, we
2080 * want to get it from a zone that is within its dirty
2081 * limit, such that no single zone holds more than its
2082 * proportional share of globally allowed dirty pages.
2083 * The dirty limits take into account the zone's
2084 * lowmem reserves and high watermark so that kswapd
2085 * should be able to balance it without having to
2086 * write pages from its LRU list.
2088 * This may look like it could increase pressure on
2089 * lower zones by failing allocations in higher zones
2090 * before they are full. But the pages that do spill
2091 * over are limited as the lower zones are protected
2092 * by this very same mechanism. It should not become
2093 * a practical burden to them.
2095 * XXX: For now, allow allocations to potentially
2096 * exceed the per-zone dirty limit in the slowpath
2097 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2098 * which is important when on a NUMA setup the allowed
2099 * zones are together not big enough to reach the
2100 * global limit. The proper fix for these situations
2101 * will require awareness of zones in the
2102 * dirty-throttling and the flusher threads.
2104 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2107 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2108 if (!zone_watermark_ok(zone
, order
, mark
,
2109 ac
->classzone_idx
, alloc_flags
)) {
2112 /* Checked here to keep the fast path fast */
2113 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2114 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2117 if (IS_ENABLED(CONFIG_NUMA
) &&
2118 !did_zlc_setup
&& nr_online_nodes
> 1) {
2120 * we do zlc_setup if there are multiple nodes
2121 * and before considering the first zone allowed
2124 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2129 if (zone_reclaim_mode
== 0 ||
2130 !zone_allows_reclaim(ac
->preferred_zone
, zone
))
2131 goto this_zone_full
;
2134 * As we may have just activated ZLC, check if the first
2135 * eligible zone has failed zone_reclaim recently.
2137 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2138 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2141 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2143 case ZONE_RECLAIM_NOSCAN
:
2146 case ZONE_RECLAIM_FULL
:
2147 /* scanned but unreclaimable */
2150 /* did we reclaim enough */
2151 if (zone_watermark_ok(zone
, order
, mark
,
2152 ac
->classzone_idx
, alloc_flags
))
2156 * Failed to reclaim enough to meet watermark.
2157 * Only mark the zone full if checking the min
2158 * watermark or if we failed to reclaim just
2159 * 1<<order pages or else the page allocator
2160 * fastpath will prematurely mark zones full
2161 * when the watermark is between the low and
2164 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2165 ret
== ZONE_RECLAIM_SOME
)
2166 goto this_zone_full
;
2173 page
= buffered_rmqueue(ac
->preferred_zone
, zone
, order
,
2174 gfp_mask
, ac
->migratetype
);
2176 if (prep_new_page(page
, order
, gfp_mask
, alloc_flags
))
2181 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2182 zlc_mark_zone_full(zonelist
, z
);
2186 * The first pass makes sure allocations are spread fairly within the
2187 * local node. However, the local node might have free pages left
2188 * after the fairness batches are exhausted, and remote zones haven't
2189 * even been considered yet. Try once more without fairness, and
2190 * include remote zones now, before entering the slowpath and waking
2191 * kswapd: prefer spilling to a remote zone over swapping locally.
2193 if (alloc_flags
& ALLOC_FAIR
) {
2194 alloc_flags
&= ~ALLOC_FAIR
;
2195 if (nr_fair_skipped
) {
2196 zonelist_rescan
= true;
2197 reset_alloc_batches(ac
->preferred_zone
);
2199 if (nr_online_nodes
> 1)
2200 zonelist_rescan
= true;
2203 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2204 /* Disable zlc cache for second zonelist scan */
2206 zonelist_rescan
= true;
2209 if (zonelist_rescan
)
2216 * Large machines with many possible nodes should not always dump per-node
2217 * meminfo in irq context.
2219 static inline bool should_suppress_show_mem(void)
2224 ret
= in_interrupt();
2229 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2230 DEFAULT_RATELIMIT_INTERVAL
,
2231 DEFAULT_RATELIMIT_BURST
);
2233 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2235 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2237 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2238 debug_guardpage_minorder() > 0)
2242 * This documents exceptions given to allocations in certain
2243 * contexts that are allowed to allocate outside current's set
2246 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2247 if (test_thread_flag(TIF_MEMDIE
) ||
2248 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2249 filter
&= ~SHOW_MEM_FILTER_NODES
;
2250 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2251 filter
&= ~SHOW_MEM_FILTER_NODES
;
2254 struct va_format vaf
;
2257 va_start(args
, fmt
);
2262 pr_warn("%pV", &vaf
);
2267 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2268 current
->comm
, order
, gfp_mask
);
2271 if (!should_suppress_show_mem())
2276 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2277 unsigned long did_some_progress
,
2278 unsigned long pages_reclaimed
)
2280 /* Do not loop if specifically requested */
2281 if (gfp_mask
& __GFP_NORETRY
)
2284 /* Always retry if specifically requested */
2285 if (gfp_mask
& __GFP_NOFAIL
)
2289 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2290 * making forward progress without invoking OOM. Suspend also disables
2291 * storage devices so kswapd will not help. Bail if we are suspending.
2293 if (!did_some_progress
&& pm_suspended_storage())
2297 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2298 * means __GFP_NOFAIL, but that may not be true in other
2301 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2305 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2306 * specified, then we retry until we no longer reclaim any pages
2307 * (above), or we've reclaimed an order of pages at least as
2308 * large as the allocation's order. In both cases, if the
2309 * allocation still fails, we stop retrying.
2311 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2317 static inline struct page
*
2318 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2319 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
2323 *did_some_progress
= 0;
2326 * Acquire the per-zone oom lock for each zone. If that
2327 * fails, somebody else is making progress for us.
2329 if (!oom_zonelist_trylock(ac
->zonelist
, gfp_mask
)) {
2330 *did_some_progress
= 1;
2331 schedule_timeout_uninterruptible(1);
2336 * Go through the zonelist yet one more time, keep very high watermark
2337 * here, this is only to catch a parallel oom killing, we must fail if
2338 * we're still under heavy pressure.
2340 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
2341 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
2345 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2346 /* Coredumps can quickly deplete all memory reserves */
2347 if (current
->flags
& PF_DUMPCORE
)
2349 /* The OOM killer will not help higher order allocs */
2350 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2352 /* The OOM killer does not needlessly kill tasks for lowmem */
2353 if (ac
->high_zoneidx
< ZONE_NORMAL
)
2355 /* The OOM killer does not compensate for light reclaim */
2356 if (!(gfp_mask
& __GFP_FS
))
2359 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2360 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2361 * The caller should handle page allocation failure by itself if
2362 * it specifies __GFP_THISNODE.
2363 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2365 if (gfp_mask
& __GFP_THISNODE
)
2368 /* Exhausted what can be done so it's blamo time */
2369 if (out_of_memory(ac
->zonelist
, gfp_mask
, order
, ac
->nodemask
, false))
2370 *did_some_progress
= 1;
2372 oom_zonelist_unlock(ac
->zonelist
, gfp_mask
);
2376 #ifdef CONFIG_COMPACTION
2377 /* Try memory compaction for high-order allocations before reclaim */
2378 static struct page
*
2379 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2380 int alloc_flags
, const struct alloc_context
*ac
,
2381 enum migrate_mode mode
, int *contended_compaction
,
2382 bool *deferred_compaction
)
2384 unsigned long compact_result
;
2390 current
->flags
|= PF_MEMALLOC
;
2391 compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
2392 mode
, contended_compaction
);
2393 current
->flags
&= ~PF_MEMALLOC
;
2395 switch (compact_result
) {
2396 case COMPACT_DEFERRED
:
2397 *deferred_compaction
= true;
2399 case COMPACT_SKIPPED
:
2406 * At least in one zone compaction wasn't deferred or skipped, so let's
2407 * count a compaction stall
2409 count_vm_event(COMPACTSTALL
);
2411 page
= get_page_from_freelist(gfp_mask
, order
,
2412 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2415 struct zone
*zone
= page_zone(page
);
2417 zone
->compact_blockskip_flush
= false;
2418 compaction_defer_reset(zone
, order
, true);
2419 count_vm_event(COMPACTSUCCESS
);
2424 * It's bad if compaction run occurs and fails. The most likely reason
2425 * is that pages exist, but not enough to satisfy watermarks.
2427 count_vm_event(COMPACTFAIL
);
2434 static inline struct page
*
2435 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2436 int alloc_flags
, const struct alloc_context
*ac
,
2437 enum migrate_mode mode
, int *contended_compaction
,
2438 bool *deferred_compaction
)
2442 #endif /* CONFIG_COMPACTION */
2444 /* Perform direct synchronous page reclaim */
2446 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
2447 const struct alloc_context
*ac
)
2449 struct reclaim_state reclaim_state
;
2454 /* We now go into synchronous reclaim */
2455 cpuset_memory_pressure_bump();
2456 current
->flags
|= PF_MEMALLOC
;
2457 lockdep_set_current_reclaim_state(gfp_mask
);
2458 reclaim_state
.reclaimed_slab
= 0;
2459 current
->reclaim_state
= &reclaim_state
;
2461 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
2464 current
->reclaim_state
= NULL
;
2465 lockdep_clear_current_reclaim_state();
2466 current
->flags
&= ~PF_MEMALLOC
;
2473 /* The really slow allocator path where we enter direct reclaim */
2474 static inline struct page
*
2475 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2476 int alloc_flags
, const struct alloc_context
*ac
,
2477 unsigned long *did_some_progress
)
2479 struct page
*page
= NULL
;
2480 bool drained
= false;
2482 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
2483 if (unlikely(!(*did_some_progress
)))
2486 /* After successful reclaim, reconsider all zones for allocation */
2487 if (IS_ENABLED(CONFIG_NUMA
))
2488 zlc_clear_zones_full(ac
->zonelist
);
2491 page
= get_page_from_freelist(gfp_mask
, order
,
2492 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2495 * If an allocation failed after direct reclaim, it could be because
2496 * pages are pinned on the per-cpu lists. Drain them and try again
2498 if (!page
&& !drained
) {
2499 drain_all_pages(NULL
);
2508 * This is called in the allocator slow-path if the allocation request is of
2509 * sufficient urgency to ignore watermarks and take other desperate measures
2511 static inline struct page
*
2512 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2513 const struct alloc_context
*ac
)
2518 page
= get_page_from_freelist(gfp_mask
, order
,
2519 ALLOC_NO_WATERMARKS
, ac
);
2521 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2522 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
,
2524 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2529 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
2534 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2535 ac
->high_zoneidx
, ac
->nodemask
)
2536 wakeup_kswapd(zone
, order
, zone_idx(ac
->preferred_zone
));
2540 gfp_to_alloc_flags(gfp_t gfp_mask
)
2542 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2543 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2545 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2546 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2549 * The caller may dip into page reserves a bit more if the caller
2550 * cannot run direct reclaim, or if the caller has realtime scheduling
2551 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2552 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2554 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2558 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2559 * if it can't schedule.
2561 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2562 alloc_flags
|= ALLOC_HARDER
;
2564 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2565 * comment for __cpuset_node_allowed().
2567 alloc_flags
&= ~ALLOC_CPUSET
;
2568 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2569 alloc_flags
|= ALLOC_HARDER
;
2571 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2572 if (gfp_mask
& __GFP_MEMALLOC
)
2573 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2574 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2575 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2576 else if (!in_interrupt() &&
2577 ((current
->flags
& PF_MEMALLOC
) ||
2578 unlikely(test_thread_flag(TIF_MEMDIE
))))
2579 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2582 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2583 alloc_flags
|= ALLOC_CMA
;
2588 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2590 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2593 static inline struct page
*
2594 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2595 struct alloc_context
*ac
)
2597 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2598 struct page
*page
= NULL
;
2600 unsigned long pages_reclaimed
= 0;
2601 unsigned long did_some_progress
;
2602 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2603 bool deferred_compaction
= false;
2604 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2607 * In the slowpath, we sanity check order to avoid ever trying to
2608 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2609 * be using allocators in order of preference for an area that is
2612 if (order
>= MAX_ORDER
) {
2613 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2618 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2619 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2620 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2621 * using a larger set of nodes after it has established that the
2622 * allowed per node queues are empty and that nodes are
2625 if (IS_ENABLED(CONFIG_NUMA
) &&
2626 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2630 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2631 wake_all_kswapds(order
, ac
);
2634 * OK, we're below the kswapd watermark and have kicked background
2635 * reclaim. Now things get more complex, so set up alloc_flags according
2636 * to how we want to proceed.
2638 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2641 * Find the true preferred zone if the allocation is unconstrained by
2644 if (!(alloc_flags
& ALLOC_CPUSET
) && !ac
->nodemask
) {
2645 struct zoneref
*preferred_zoneref
;
2646 preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
2647 ac
->high_zoneidx
, NULL
, &ac
->preferred_zone
);
2648 ac
->classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2651 /* This is the last chance, in general, before the goto nopage. */
2652 page
= get_page_from_freelist(gfp_mask
, order
,
2653 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2657 /* Allocate without watermarks if the context allows */
2658 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2660 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2661 * the allocation is high priority and these type of
2662 * allocations are system rather than user orientated
2664 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2666 page
= __alloc_pages_high_priority(gfp_mask
, order
, ac
);
2673 /* Atomic allocations - we can't balance anything */
2676 * All existing users of the deprecated __GFP_NOFAIL are
2677 * blockable, so warn of any new users that actually allow this
2678 * type of allocation to fail.
2680 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2684 /* Avoid recursion of direct reclaim */
2685 if (current
->flags
& PF_MEMALLOC
)
2688 /* Avoid allocations with no watermarks from looping endlessly */
2689 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2693 * Try direct compaction. The first pass is asynchronous. Subsequent
2694 * attempts after direct reclaim are synchronous
2696 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
2698 &contended_compaction
,
2699 &deferred_compaction
);
2703 /* Checks for THP-specific high-order allocations */
2704 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2706 * If compaction is deferred for high-order allocations, it is
2707 * because sync compaction recently failed. If this is the case
2708 * and the caller requested a THP allocation, we do not want
2709 * to heavily disrupt the system, so we fail the allocation
2710 * instead of entering direct reclaim.
2712 if (deferred_compaction
)
2716 * In all zones where compaction was attempted (and not
2717 * deferred or skipped), lock contention has been detected.
2718 * For THP allocation we do not want to disrupt the others
2719 * so we fallback to base pages instead.
2721 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2725 * If compaction was aborted due to need_resched(), we do not
2726 * want to further increase allocation latency, unless it is
2727 * khugepaged trying to collapse.
2729 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2730 && !(current
->flags
& PF_KTHREAD
))
2735 * It can become very expensive to allocate transparent hugepages at
2736 * fault, so use asynchronous memory compaction for THP unless it is
2737 * khugepaged trying to collapse.
2739 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2740 (current
->flags
& PF_KTHREAD
))
2741 migration_mode
= MIGRATE_SYNC_LIGHT
;
2743 /* Try direct reclaim and then allocating */
2744 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
2745 &did_some_progress
);
2749 /* Check if we should retry the allocation */
2750 pages_reclaimed
+= did_some_progress
;
2751 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2754 * If we fail to make progress by freeing individual
2755 * pages, but the allocation wants us to keep going,
2756 * start OOM killing tasks.
2758 if (!did_some_progress
) {
2759 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
,
2760 &did_some_progress
);
2763 if (!did_some_progress
)
2766 /* Wait for some write requests to complete then retry */
2767 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2771 * High-order allocations do not necessarily loop after
2772 * direct reclaim and reclaim/compaction depends on compaction
2773 * being called after reclaim so call directly if necessary
2775 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2776 alloc_flags
, ac
, migration_mode
,
2777 &contended_compaction
,
2778 &deferred_compaction
);
2784 warn_alloc_failed(gfp_mask
, order
, NULL
);
2790 * This is the 'heart' of the zoned buddy allocator.
2793 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2794 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2796 struct zoneref
*preferred_zoneref
;
2797 struct page
*page
= NULL
;
2798 unsigned int cpuset_mems_cookie
;
2799 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2800 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
2801 struct alloc_context ac
= {
2802 .high_zoneidx
= gfp_zone(gfp_mask
),
2803 .nodemask
= nodemask
,
2804 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
2807 gfp_mask
&= gfp_allowed_mask
;
2809 lockdep_trace_alloc(gfp_mask
);
2811 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2813 if (should_fail_alloc_page(gfp_mask
, order
))
2817 * Check the zones suitable for the gfp_mask contain at least one
2818 * valid zone. It's possible to have an empty zonelist as a result
2819 * of GFP_THISNODE and a memoryless node
2821 if (unlikely(!zonelist
->_zonerefs
->zone
))
2824 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
2825 alloc_flags
|= ALLOC_CMA
;
2828 cpuset_mems_cookie
= read_mems_allowed_begin();
2830 /* We set it here, as __alloc_pages_slowpath might have changed it */
2831 ac
.zonelist
= zonelist
;
2832 /* The preferred zone is used for statistics later */
2833 preferred_zoneref
= first_zones_zonelist(ac
.zonelist
, ac
.high_zoneidx
,
2834 ac
.nodemask
? : &cpuset_current_mems_allowed
,
2835 &ac
.preferred_zone
);
2836 if (!ac
.preferred_zone
)
2838 ac
.classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2840 /* First allocation attempt */
2841 alloc_mask
= gfp_mask
|__GFP_HARDWALL
;
2842 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
2843 if (unlikely(!page
)) {
2845 * Runtime PM, block IO and its error handling path
2846 * can deadlock because I/O on the device might not
2849 alloc_mask
= memalloc_noio_flags(gfp_mask
);
2851 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
2854 if (kmemcheck_enabled
&& page
)
2855 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2857 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
2861 * When updating a task's mems_allowed, it is possible to race with
2862 * parallel threads in such a way that an allocation can fail while
2863 * the mask is being updated. If a page allocation is about to fail,
2864 * check if the cpuset changed during allocation and if so, retry.
2866 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2871 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2874 * Common helper functions.
2876 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2881 * __get_free_pages() returns a 32-bit address, which cannot represent
2884 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2886 page
= alloc_pages(gfp_mask
, order
);
2889 return (unsigned long) page_address(page
);
2891 EXPORT_SYMBOL(__get_free_pages
);
2893 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2895 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2897 EXPORT_SYMBOL(get_zeroed_page
);
2899 void __free_pages(struct page
*page
, unsigned int order
)
2901 if (put_page_testzero(page
)) {
2903 free_hot_cold_page(page
, false);
2905 __free_pages_ok(page
, order
);
2909 EXPORT_SYMBOL(__free_pages
);
2911 void free_pages(unsigned long addr
, unsigned int order
)
2914 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2915 __free_pages(virt_to_page((void *)addr
), order
);
2919 EXPORT_SYMBOL(free_pages
);
2922 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2923 * of the current memory cgroup.
2925 * It should be used when the caller would like to use kmalloc, but since the
2926 * allocation is large, it has to fall back to the page allocator.
2928 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2931 struct mem_cgroup
*memcg
= NULL
;
2933 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2935 page
= alloc_pages(gfp_mask
, order
);
2936 memcg_kmem_commit_charge(page
, memcg
, order
);
2940 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2943 struct mem_cgroup
*memcg
= NULL
;
2945 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2947 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2948 memcg_kmem_commit_charge(page
, memcg
, order
);
2953 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2956 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2958 memcg_kmem_uncharge_pages(page
, order
);
2959 __free_pages(page
, order
);
2962 void free_kmem_pages(unsigned long addr
, unsigned int order
)
2965 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2966 __free_kmem_pages(virt_to_page((void *)addr
), order
);
2970 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2973 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2974 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2976 split_page(virt_to_page((void *)addr
), order
);
2977 while (used
< alloc_end
) {
2982 return (void *)addr
;
2986 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2987 * @size: the number of bytes to allocate
2988 * @gfp_mask: GFP flags for the allocation
2990 * This function is similar to alloc_pages(), except that it allocates the
2991 * minimum number of pages to satisfy the request. alloc_pages() can only
2992 * allocate memory in power-of-two pages.
2994 * This function is also limited by MAX_ORDER.
2996 * Memory allocated by this function must be released by free_pages_exact().
2998 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3000 unsigned int order
= get_order(size
);
3003 addr
= __get_free_pages(gfp_mask
, order
);
3004 return make_alloc_exact(addr
, order
, size
);
3006 EXPORT_SYMBOL(alloc_pages_exact
);
3009 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3011 * @nid: the preferred node ID where memory should be allocated
3012 * @size: the number of bytes to allocate
3013 * @gfp_mask: GFP flags for the allocation
3015 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3017 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3020 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3022 unsigned order
= get_order(size
);
3023 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3026 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3030 * free_pages_exact - release memory allocated via alloc_pages_exact()
3031 * @virt: the value returned by alloc_pages_exact.
3032 * @size: size of allocation, same value as passed to alloc_pages_exact().
3034 * Release the memory allocated by a previous call to alloc_pages_exact.
3036 void free_pages_exact(void *virt
, size_t size
)
3038 unsigned long addr
= (unsigned long)virt
;
3039 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3041 while (addr
< end
) {
3046 EXPORT_SYMBOL(free_pages_exact
);
3049 * nr_free_zone_pages - count number of pages beyond high watermark
3050 * @offset: The zone index of the highest zone
3052 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3053 * high watermark within all zones at or below a given zone index. For each
3054 * zone, the number of pages is calculated as:
3055 * managed_pages - high_pages
3057 static unsigned long nr_free_zone_pages(int offset
)
3062 /* Just pick one node, since fallback list is circular */
3063 unsigned long sum
= 0;
3065 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3067 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3068 unsigned long size
= zone
->managed_pages
;
3069 unsigned long high
= high_wmark_pages(zone
);
3078 * nr_free_buffer_pages - count number of pages beyond high watermark
3080 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3081 * watermark within ZONE_DMA and ZONE_NORMAL.
3083 unsigned long nr_free_buffer_pages(void)
3085 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3087 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3090 * nr_free_pagecache_pages - count number of pages beyond high watermark
3092 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3093 * high watermark within all zones.
3095 unsigned long nr_free_pagecache_pages(void)
3097 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3100 static inline void show_node(struct zone
*zone
)
3102 if (IS_ENABLED(CONFIG_NUMA
))
3103 printk("Node %d ", zone_to_nid(zone
));
3106 void si_meminfo(struct sysinfo
*val
)
3108 val
->totalram
= totalram_pages
;
3109 val
->sharedram
= global_page_state(NR_SHMEM
);
3110 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3111 val
->bufferram
= nr_blockdev_pages();
3112 val
->totalhigh
= totalhigh_pages
;
3113 val
->freehigh
= nr_free_highpages();
3114 val
->mem_unit
= PAGE_SIZE
;
3117 EXPORT_SYMBOL(si_meminfo
);
3120 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3122 int zone_type
; /* needs to be signed */
3123 unsigned long managed_pages
= 0;
3124 pg_data_t
*pgdat
= NODE_DATA(nid
);
3126 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3127 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3128 val
->totalram
= managed_pages
;
3129 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3130 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3131 #ifdef CONFIG_HIGHMEM
3132 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3133 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3139 val
->mem_unit
= PAGE_SIZE
;
3144 * Determine whether the node should be displayed or not, depending on whether
3145 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3147 bool skip_free_areas_node(unsigned int flags
, int nid
)
3150 unsigned int cpuset_mems_cookie
;
3152 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3156 cpuset_mems_cookie
= read_mems_allowed_begin();
3157 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3158 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3163 #define K(x) ((x) << (PAGE_SHIFT-10))
3165 static void show_migration_types(unsigned char type
)
3167 static const char types
[MIGRATE_TYPES
] = {
3168 [MIGRATE_UNMOVABLE
] = 'U',
3169 [MIGRATE_RECLAIMABLE
] = 'E',
3170 [MIGRATE_MOVABLE
] = 'M',
3171 [MIGRATE_RESERVE
] = 'R',
3173 [MIGRATE_CMA
] = 'C',
3175 #ifdef CONFIG_MEMORY_ISOLATION
3176 [MIGRATE_ISOLATE
] = 'I',
3179 char tmp
[MIGRATE_TYPES
+ 1];
3183 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3184 if (type
& (1 << i
))
3189 printk("(%s) ", tmp
);
3193 * Show free area list (used inside shift_scroll-lock stuff)
3194 * We also calculate the percentage fragmentation. We do this by counting the
3195 * memory on each free list with the exception of the first item on the list.
3196 * Suppresses nodes that are not allowed by current's cpuset if
3197 * SHOW_MEM_FILTER_NODES is passed.
3199 void show_free_areas(unsigned int filter
)
3204 for_each_populated_zone(zone
) {
3205 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3208 printk("%s per-cpu:\n", zone
->name
);
3210 for_each_online_cpu(cpu
) {
3211 struct per_cpu_pageset
*pageset
;
3213 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3215 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3216 cpu
, pageset
->pcp
.high
,
3217 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3221 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3222 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3224 " dirty:%lu writeback:%lu unstable:%lu\n"
3225 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3226 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3228 global_page_state(NR_ACTIVE_ANON
),
3229 global_page_state(NR_INACTIVE_ANON
),
3230 global_page_state(NR_ISOLATED_ANON
),
3231 global_page_state(NR_ACTIVE_FILE
),
3232 global_page_state(NR_INACTIVE_FILE
),
3233 global_page_state(NR_ISOLATED_FILE
),
3234 global_page_state(NR_UNEVICTABLE
),
3235 global_page_state(NR_FILE_DIRTY
),
3236 global_page_state(NR_WRITEBACK
),
3237 global_page_state(NR_UNSTABLE_NFS
),
3238 global_page_state(NR_FREE_PAGES
),
3239 global_page_state(NR_SLAB_RECLAIMABLE
),
3240 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3241 global_page_state(NR_FILE_MAPPED
),
3242 global_page_state(NR_SHMEM
),
3243 global_page_state(NR_PAGETABLE
),
3244 global_page_state(NR_BOUNCE
),
3245 global_page_state(NR_FREE_CMA_PAGES
));
3247 for_each_populated_zone(zone
) {
3250 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3258 " active_anon:%lukB"
3259 " inactive_anon:%lukB"
3260 " active_file:%lukB"
3261 " inactive_file:%lukB"
3262 " unevictable:%lukB"
3263 " isolated(anon):%lukB"
3264 " isolated(file):%lukB"
3272 " slab_reclaimable:%lukB"
3273 " slab_unreclaimable:%lukB"
3274 " kernel_stack:%lukB"
3279 " writeback_tmp:%lukB"
3280 " pages_scanned:%lu"
3281 " all_unreclaimable? %s"
3284 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3285 K(min_wmark_pages(zone
)),
3286 K(low_wmark_pages(zone
)),
3287 K(high_wmark_pages(zone
)),
3288 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3289 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3290 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3291 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3292 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3293 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3294 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3295 K(zone
->present_pages
),
3296 K(zone
->managed_pages
),
3297 K(zone_page_state(zone
, NR_MLOCK
)),
3298 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3299 K(zone_page_state(zone
, NR_WRITEBACK
)),
3300 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3301 K(zone_page_state(zone
, NR_SHMEM
)),
3302 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3303 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3304 zone_page_state(zone
, NR_KERNEL_STACK
) *
3306 K(zone_page_state(zone
, NR_PAGETABLE
)),
3307 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3308 K(zone_page_state(zone
, NR_BOUNCE
)),
3309 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3310 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3311 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3312 (!zone_reclaimable(zone
) ? "yes" : "no")
3314 printk("lowmem_reserve[]:");
3315 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3316 printk(" %ld", zone
->lowmem_reserve
[i
]);
3320 for_each_populated_zone(zone
) {
3321 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3322 unsigned char types
[MAX_ORDER
];
3324 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3327 printk("%s: ", zone
->name
);
3329 spin_lock_irqsave(&zone
->lock
, flags
);
3330 for (order
= 0; order
< MAX_ORDER
; order
++) {
3331 struct free_area
*area
= &zone
->free_area
[order
];
3334 nr
[order
] = area
->nr_free
;
3335 total
+= nr
[order
] << order
;
3338 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3339 if (!list_empty(&area
->free_list
[type
]))
3340 types
[order
] |= 1 << type
;
3343 spin_unlock_irqrestore(&zone
->lock
, flags
);
3344 for (order
= 0; order
< MAX_ORDER
; order
++) {
3345 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3347 show_migration_types(types
[order
]);
3349 printk("= %lukB\n", K(total
));
3352 hugetlb_show_meminfo();
3354 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3356 show_swap_cache_info();
3359 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3361 zoneref
->zone
= zone
;
3362 zoneref
->zone_idx
= zone_idx(zone
);
3366 * Builds allocation fallback zone lists.
3368 * Add all populated zones of a node to the zonelist.
3370 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3374 enum zone_type zone_type
= MAX_NR_ZONES
;
3378 zone
= pgdat
->node_zones
+ zone_type
;
3379 if (populated_zone(zone
)) {
3380 zoneref_set_zone(zone
,
3381 &zonelist
->_zonerefs
[nr_zones
++]);
3382 check_highest_zone(zone_type
);
3384 } while (zone_type
);
3392 * 0 = automatic detection of better ordering.
3393 * 1 = order by ([node] distance, -zonetype)
3394 * 2 = order by (-zonetype, [node] distance)
3396 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3397 * the same zonelist. So only NUMA can configure this param.
3399 #define ZONELIST_ORDER_DEFAULT 0
3400 #define ZONELIST_ORDER_NODE 1
3401 #define ZONELIST_ORDER_ZONE 2
3403 /* zonelist order in the kernel.
3404 * set_zonelist_order() will set this to NODE or ZONE.
3406 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3407 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3411 /* The value user specified ....changed by config */
3412 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3413 /* string for sysctl */
3414 #define NUMA_ZONELIST_ORDER_LEN 16
3415 char numa_zonelist_order
[16] = "default";
3418 * interface for configure zonelist ordering.
3419 * command line option "numa_zonelist_order"
3420 * = "[dD]efault - default, automatic configuration.
3421 * = "[nN]ode - order by node locality, then by zone within node
3422 * = "[zZ]one - order by zone, then by locality within zone
3425 static int __parse_numa_zonelist_order(char *s
)
3427 if (*s
== 'd' || *s
== 'D') {
3428 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3429 } else if (*s
== 'n' || *s
== 'N') {
3430 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3431 } else if (*s
== 'z' || *s
== 'Z') {
3432 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3435 "Ignoring invalid numa_zonelist_order value: "
3442 static __init
int setup_numa_zonelist_order(char *s
)
3449 ret
= __parse_numa_zonelist_order(s
);
3451 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3455 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3458 * sysctl handler for numa_zonelist_order
3460 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3461 void __user
*buffer
, size_t *length
,
3464 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3466 static DEFINE_MUTEX(zl_order_mutex
);
3468 mutex_lock(&zl_order_mutex
);
3470 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3474 strcpy(saved_string
, (char *)table
->data
);
3476 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3480 int oldval
= user_zonelist_order
;
3482 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3485 * bogus value. restore saved string
3487 strncpy((char *)table
->data
, saved_string
,
3488 NUMA_ZONELIST_ORDER_LEN
);
3489 user_zonelist_order
= oldval
;
3490 } else if (oldval
!= user_zonelist_order
) {
3491 mutex_lock(&zonelists_mutex
);
3492 build_all_zonelists(NULL
, NULL
);
3493 mutex_unlock(&zonelists_mutex
);
3497 mutex_unlock(&zl_order_mutex
);
3502 #define MAX_NODE_LOAD (nr_online_nodes)
3503 static int node_load
[MAX_NUMNODES
];
3506 * find_next_best_node - find the next node that should appear in a given node's fallback list
3507 * @node: node whose fallback list we're appending
3508 * @used_node_mask: nodemask_t of already used nodes
3510 * We use a number of factors to determine which is the next node that should
3511 * appear on a given node's fallback list. The node should not have appeared
3512 * already in @node's fallback list, and it should be the next closest node
3513 * according to the distance array (which contains arbitrary distance values
3514 * from each node to each node in the system), and should also prefer nodes
3515 * with no CPUs, since presumably they'll have very little allocation pressure
3516 * on them otherwise.
3517 * It returns -1 if no node is found.
3519 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3522 int min_val
= INT_MAX
;
3523 int best_node
= NUMA_NO_NODE
;
3524 const struct cpumask
*tmp
= cpumask_of_node(0);
3526 /* Use the local node if we haven't already */
3527 if (!node_isset(node
, *used_node_mask
)) {
3528 node_set(node
, *used_node_mask
);
3532 for_each_node_state(n
, N_MEMORY
) {
3534 /* Don't want a node to appear more than once */
3535 if (node_isset(n
, *used_node_mask
))
3538 /* Use the distance array to find the distance */
3539 val
= node_distance(node
, n
);
3541 /* Penalize nodes under us ("prefer the next node") */
3544 /* Give preference to headless and unused nodes */
3545 tmp
= cpumask_of_node(n
);
3546 if (!cpumask_empty(tmp
))
3547 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3549 /* Slight preference for less loaded node */
3550 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3551 val
+= node_load
[n
];
3553 if (val
< min_val
) {
3560 node_set(best_node
, *used_node_mask
);
3567 * Build zonelists ordered by node and zones within node.
3568 * This results in maximum locality--normal zone overflows into local
3569 * DMA zone, if any--but risks exhausting DMA zone.
3571 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3574 struct zonelist
*zonelist
;
3576 zonelist
= &pgdat
->node_zonelists
[0];
3577 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3579 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3580 zonelist
->_zonerefs
[j
].zone
= NULL
;
3581 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3585 * Build gfp_thisnode zonelists
3587 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3590 struct zonelist
*zonelist
;
3592 zonelist
= &pgdat
->node_zonelists
[1];
3593 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3594 zonelist
->_zonerefs
[j
].zone
= NULL
;
3595 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3599 * Build zonelists ordered by zone and nodes within zones.
3600 * This results in conserving DMA zone[s] until all Normal memory is
3601 * exhausted, but results in overflowing to remote node while memory
3602 * may still exist in local DMA zone.
3604 static int node_order
[MAX_NUMNODES
];
3606 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3609 int zone_type
; /* needs to be signed */
3611 struct zonelist
*zonelist
;
3613 zonelist
= &pgdat
->node_zonelists
[0];
3615 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3616 for (j
= 0; j
< nr_nodes
; j
++) {
3617 node
= node_order
[j
];
3618 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3619 if (populated_zone(z
)) {
3621 &zonelist
->_zonerefs
[pos
++]);
3622 check_highest_zone(zone_type
);
3626 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3627 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3630 #if defined(CONFIG_64BIT)
3632 * Devices that require DMA32/DMA are relatively rare and do not justify a
3633 * penalty to every machine in case the specialised case applies. Default
3634 * to Node-ordering on 64-bit NUMA machines
3636 static int default_zonelist_order(void)
3638 return ZONELIST_ORDER_NODE
;
3642 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3643 * by the kernel. If processes running on node 0 deplete the low memory zone
3644 * then reclaim will occur more frequency increasing stalls and potentially
3645 * be easier to OOM if a large percentage of the zone is under writeback or
3646 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3647 * Hence, default to zone ordering on 32-bit.
3649 static int default_zonelist_order(void)
3651 return ZONELIST_ORDER_ZONE
;
3653 #endif /* CONFIG_64BIT */
3655 static void set_zonelist_order(void)
3657 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3658 current_zonelist_order
= default_zonelist_order();
3660 current_zonelist_order
= user_zonelist_order
;
3663 static void build_zonelists(pg_data_t
*pgdat
)
3667 nodemask_t used_mask
;
3668 int local_node
, prev_node
;
3669 struct zonelist
*zonelist
;
3670 int order
= current_zonelist_order
;
3672 /* initialize zonelists */
3673 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3674 zonelist
= pgdat
->node_zonelists
+ i
;
3675 zonelist
->_zonerefs
[0].zone
= NULL
;
3676 zonelist
->_zonerefs
[0].zone_idx
= 0;
3679 /* NUMA-aware ordering of nodes */
3680 local_node
= pgdat
->node_id
;
3681 load
= nr_online_nodes
;
3682 prev_node
= local_node
;
3683 nodes_clear(used_mask
);
3685 memset(node_order
, 0, sizeof(node_order
));
3688 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3690 * We don't want to pressure a particular node.
3691 * So adding penalty to the first node in same
3692 * distance group to make it round-robin.
3694 if (node_distance(local_node
, node
) !=
3695 node_distance(local_node
, prev_node
))
3696 node_load
[node
] = load
;
3700 if (order
== ZONELIST_ORDER_NODE
)
3701 build_zonelists_in_node_order(pgdat
, node
);
3703 node_order
[j
++] = node
; /* remember order */
3706 if (order
== ZONELIST_ORDER_ZONE
) {
3707 /* calculate node order -- i.e., DMA last! */
3708 build_zonelists_in_zone_order(pgdat
, j
);
3711 build_thisnode_zonelists(pgdat
);
3714 /* Construct the zonelist performance cache - see further mmzone.h */
3715 static void build_zonelist_cache(pg_data_t
*pgdat
)
3717 struct zonelist
*zonelist
;
3718 struct zonelist_cache
*zlc
;
3721 zonelist
= &pgdat
->node_zonelists
[0];
3722 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3723 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3724 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3725 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3728 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3730 * Return node id of node used for "local" allocations.
3731 * I.e., first node id of first zone in arg node's generic zonelist.
3732 * Used for initializing percpu 'numa_mem', which is used primarily
3733 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3735 int local_memory_node(int node
)
3739 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3740 gfp_zone(GFP_KERNEL
),
3747 #else /* CONFIG_NUMA */
3749 static void set_zonelist_order(void)
3751 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3754 static void build_zonelists(pg_data_t
*pgdat
)
3756 int node
, local_node
;
3758 struct zonelist
*zonelist
;
3760 local_node
= pgdat
->node_id
;
3762 zonelist
= &pgdat
->node_zonelists
[0];
3763 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3766 * Now we build the zonelist so that it contains the zones
3767 * of all the other nodes.
3768 * We don't want to pressure a particular node, so when
3769 * building the zones for node N, we make sure that the
3770 * zones coming right after the local ones are those from
3771 * node N+1 (modulo N)
3773 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3774 if (!node_online(node
))
3776 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3778 for (node
= 0; node
< local_node
; node
++) {
3779 if (!node_online(node
))
3781 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3784 zonelist
->_zonerefs
[j
].zone
= NULL
;
3785 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3788 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3789 static void build_zonelist_cache(pg_data_t
*pgdat
)
3791 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3794 #endif /* CONFIG_NUMA */
3797 * Boot pageset table. One per cpu which is going to be used for all
3798 * zones and all nodes. The parameters will be set in such a way
3799 * that an item put on a list will immediately be handed over to
3800 * the buddy list. This is safe since pageset manipulation is done
3801 * with interrupts disabled.
3803 * The boot_pagesets must be kept even after bootup is complete for
3804 * unused processors and/or zones. They do play a role for bootstrapping
3805 * hotplugged processors.
3807 * zoneinfo_show() and maybe other functions do
3808 * not check if the processor is online before following the pageset pointer.
3809 * Other parts of the kernel may not check if the zone is available.
3811 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3812 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3813 static void setup_zone_pageset(struct zone
*zone
);
3816 * Global mutex to protect against size modification of zonelists
3817 * as well as to serialize pageset setup for the new populated zone.
3819 DEFINE_MUTEX(zonelists_mutex
);
3821 /* return values int ....just for stop_machine() */
3822 static int __build_all_zonelists(void *data
)
3826 pg_data_t
*self
= data
;
3829 memset(node_load
, 0, sizeof(node_load
));
3832 if (self
&& !node_online(self
->node_id
)) {
3833 build_zonelists(self
);
3834 build_zonelist_cache(self
);
3837 for_each_online_node(nid
) {
3838 pg_data_t
*pgdat
= NODE_DATA(nid
);
3840 build_zonelists(pgdat
);
3841 build_zonelist_cache(pgdat
);
3845 * Initialize the boot_pagesets that are going to be used
3846 * for bootstrapping processors. The real pagesets for
3847 * each zone will be allocated later when the per cpu
3848 * allocator is available.
3850 * boot_pagesets are used also for bootstrapping offline
3851 * cpus if the system is already booted because the pagesets
3852 * are needed to initialize allocators on a specific cpu too.
3853 * F.e. the percpu allocator needs the page allocator which
3854 * needs the percpu allocator in order to allocate its pagesets
3855 * (a chicken-egg dilemma).
3857 for_each_possible_cpu(cpu
) {
3858 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3860 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3862 * We now know the "local memory node" for each node--
3863 * i.e., the node of the first zone in the generic zonelist.
3864 * Set up numa_mem percpu variable for on-line cpus. During
3865 * boot, only the boot cpu should be on-line; we'll init the
3866 * secondary cpus' numa_mem as they come on-line. During
3867 * node/memory hotplug, we'll fixup all on-line cpus.
3869 if (cpu_online(cpu
))
3870 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3877 static noinline
void __init
3878 build_all_zonelists_init(void)
3880 __build_all_zonelists(NULL
);
3881 mminit_verify_zonelist();
3882 cpuset_init_current_mems_allowed();
3886 * Called with zonelists_mutex held always
3887 * unless system_state == SYSTEM_BOOTING.
3889 * __ref due to (1) call of __meminit annotated setup_zone_pageset
3890 * [we're only called with non-NULL zone through __meminit paths] and
3891 * (2) call of __init annotated helper build_all_zonelists_init
3892 * [protected by SYSTEM_BOOTING].
3894 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3896 set_zonelist_order();
3898 if (system_state
== SYSTEM_BOOTING
) {
3899 build_all_zonelists_init();
3901 #ifdef CONFIG_MEMORY_HOTPLUG
3903 setup_zone_pageset(zone
);
3905 /* we have to stop all cpus to guarantee there is no user
3907 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3908 /* cpuset refresh routine should be here */
3910 vm_total_pages
= nr_free_pagecache_pages();
3912 * Disable grouping by mobility if the number of pages in the
3913 * system is too low to allow the mechanism to work. It would be
3914 * more accurate, but expensive to check per-zone. This check is
3915 * made on memory-hotadd so a system can start with mobility
3916 * disabled and enable it later
3918 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3919 page_group_by_mobility_disabled
= 1;
3921 page_group_by_mobility_disabled
= 0;
3923 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3924 "Total pages: %ld\n",
3926 zonelist_order_name
[current_zonelist_order
],
3927 page_group_by_mobility_disabled
? "off" : "on",
3930 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
3935 * Helper functions to size the waitqueue hash table.
3936 * Essentially these want to choose hash table sizes sufficiently
3937 * large so that collisions trying to wait on pages are rare.
3938 * But in fact, the number of active page waitqueues on typical
3939 * systems is ridiculously low, less than 200. So this is even
3940 * conservative, even though it seems large.
3942 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3943 * waitqueues, i.e. the size of the waitq table given the number of pages.
3945 #define PAGES_PER_WAITQUEUE 256
3947 #ifndef CONFIG_MEMORY_HOTPLUG
3948 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3950 unsigned long size
= 1;
3952 pages
/= PAGES_PER_WAITQUEUE
;
3954 while (size
< pages
)
3958 * Once we have dozens or even hundreds of threads sleeping
3959 * on IO we've got bigger problems than wait queue collision.
3960 * Limit the size of the wait table to a reasonable size.
3962 size
= min(size
, 4096UL);
3964 return max(size
, 4UL);
3968 * A zone's size might be changed by hot-add, so it is not possible to determine
3969 * a suitable size for its wait_table. So we use the maximum size now.
3971 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3973 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3974 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3975 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3977 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3978 * or more by the traditional way. (See above). It equals:
3980 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3981 * ia64(16K page size) : = ( 8G + 4M)byte.
3982 * powerpc (64K page size) : = (32G +16M)byte.
3984 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3991 * This is an integer logarithm so that shifts can be used later
3992 * to extract the more random high bits from the multiplicative
3993 * hash function before the remainder is taken.
3995 static inline unsigned long wait_table_bits(unsigned long size
)
4001 * Check if a pageblock contains reserved pages
4003 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
4007 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4008 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
4015 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4016 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4017 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4018 * higher will lead to a bigger reserve which will get freed as contiguous
4019 * blocks as reclaim kicks in
4021 static void setup_zone_migrate_reserve(struct zone
*zone
)
4023 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4025 unsigned long block_migratetype
;
4030 * Get the start pfn, end pfn and the number of blocks to reserve
4031 * We have to be careful to be aligned to pageblock_nr_pages to
4032 * make sure that we always check pfn_valid for the first page in
4035 start_pfn
= zone
->zone_start_pfn
;
4036 end_pfn
= zone_end_pfn(zone
);
4037 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4038 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4042 * Reserve blocks are generally in place to help high-order atomic
4043 * allocations that are short-lived. A min_free_kbytes value that
4044 * would result in more than 2 reserve blocks for atomic allocations
4045 * is assumed to be in place to help anti-fragmentation for the
4046 * future allocation of hugepages at runtime.
4048 reserve
= min(2, reserve
);
4049 old_reserve
= zone
->nr_migrate_reserve_block
;
4051 /* When memory hot-add, we almost always need to do nothing */
4052 if (reserve
== old_reserve
)
4054 zone
->nr_migrate_reserve_block
= reserve
;
4056 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4057 if (!pfn_valid(pfn
))
4059 page
= pfn_to_page(pfn
);
4061 /* Watch out for overlapping nodes */
4062 if (page_to_nid(page
) != zone_to_nid(zone
))
4065 block_migratetype
= get_pageblock_migratetype(page
);
4067 /* Only test what is necessary when the reserves are not met */
4070 * Blocks with reserved pages will never free, skip
4073 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4074 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4077 /* If this block is reserved, account for it */
4078 if (block_migratetype
== MIGRATE_RESERVE
) {
4083 /* Suitable for reserving if this block is movable */
4084 if (block_migratetype
== MIGRATE_MOVABLE
) {
4085 set_pageblock_migratetype(page
,
4087 move_freepages_block(zone
, page
,
4092 } else if (!old_reserve
) {
4094 * At boot time we don't need to scan the whole zone
4095 * for turning off MIGRATE_RESERVE.
4101 * If the reserve is met and this is a previous reserved block,
4104 if (block_migratetype
== MIGRATE_RESERVE
) {
4105 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4106 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4112 * Initially all pages are reserved - free ones are freed
4113 * up by free_all_bootmem() once the early boot process is
4114 * done. Non-atomic initialization, single-pass.
4116 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4117 unsigned long start_pfn
, enum memmap_context context
)
4120 unsigned long end_pfn
= start_pfn
+ size
;
4124 if (highest_memmap_pfn
< end_pfn
- 1)
4125 highest_memmap_pfn
= end_pfn
- 1;
4127 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4128 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4130 * There can be holes in boot-time mem_map[]s
4131 * handed to this function. They do not
4132 * exist on hotplugged memory.
4134 if (context
== MEMMAP_EARLY
) {
4135 if (!early_pfn_valid(pfn
))
4137 if (!early_pfn_in_nid(pfn
, nid
))
4140 page
= pfn_to_page(pfn
);
4141 set_page_links(page
, zone
, nid
, pfn
);
4142 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4143 init_page_count(page
);
4144 page_mapcount_reset(page
);
4145 page_cpupid_reset_last(page
);
4146 SetPageReserved(page
);
4148 * Mark the block movable so that blocks are reserved for
4149 * movable at startup. This will force kernel allocations
4150 * to reserve their blocks rather than leaking throughout
4151 * the address space during boot when many long-lived
4152 * kernel allocations are made. Later some blocks near
4153 * the start are marked MIGRATE_RESERVE by
4154 * setup_zone_migrate_reserve()
4156 * bitmap is created for zone's valid pfn range. but memmap
4157 * can be created for invalid pages (for alignment)
4158 * check here not to call set_pageblock_migratetype() against
4161 if ((z
->zone_start_pfn
<= pfn
)
4162 && (pfn
< zone_end_pfn(z
))
4163 && !(pfn
& (pageblock_nr_pages
- 1)))
4164 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4166 INIT_LIST_HEAD(&page
->lru
);
4167 #ifdef WANT_PAGE_VIRTUAL
4168 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4169 if (!is_highmem_idx(zone
))
4170 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4175 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4177 unsigned int order
, t
;
4178 for_each_migratetype_order(order
, t
) {
4179 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4180 zone
->free_area
[order
].nr_free
= 0;
4184 #ifndef __HAVE_ARCH_MEMMAP_INIT
4185 #define memmap_init(size, nid, zone, start_pfn) \
4186 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4189 static int zone_batchsize(struct zone
*zone
)
4195 * The per-cpu-pages pools are set to around 1000th of the
4196 * size of the zone. But no more than 1/2 of a meg.
4198 * OK, so we don't know how big the cache is. So guess.
4200 batch
= zone
->managed_pages
/ 1024;
4201 if (batch
* PAGE_SIZE
> 512 * 1024)
4202 batch
= (512 * 1024) / PAGE_SIZE
;
4203 batch
/= 4; /* We effectively *= 4 below */
4208 * Clamp the batch to a 2^n - 1 value. Having a power
4209 * of 2 value was found to be more likely to have
4210 * suboptimal cache aliasing properties in some cases.
4212 * For example if 2 tasks are alternately allocating
4213 * batches of pages, one task can end up with a lot
4214 * of pages of one half of the possible page colors
4215 * and the other with pages of the other colors.
4217 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4222 /* The deferral and batching of frees should be suppressed under NOMMU
4225 * The problem is that NOMMU needs to be able to allocate large chunks
4226 * of contiguous memory as there's no hardware page translation to
4227 * assemble apparent contiguous memory from discontiguous pages.
4229 * Queueing large contiguous runs of pages for batching, however,
4230 * causes the pages to actually be freed in smaller chunks. As there
4231 * can be a significant delay between the individual batches being
4232 * recycled, this leads to the once large chunks of space being
4233 * fragmented and becoming unavailable for high-order allocations.
4240 * pcp->high and pcp->batch values are related and dependent on one another:
4241 * ->batch must never be higher then ->high.
4242 * The following function updates them in a safe manner without read side
4245 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4246 * those fields changing asynchronously (acording the the above rule).
4248 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4249 * outside of boot time (or some other assurance that no concurrent updaters
4252 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4253 unsigned long batch
)
4255 /* start with a fail safe value for batch */
4259 /* Update high, then batch, in order */
4266 /* a companion to pageset_set_high() */
4267 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4269 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4272 static void pageset_init(struct per_cpu_pageset
*p
)
4274 struct per_cpu_pages
*pcp
;
4277 memset(p
, 0, sizeof(*p
));
4281 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4282 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4285 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4288 pageset_set_batch(p
, batch
);
4292 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4293 * to the value high for the pageset p.
4295 static void pageset_set_high(struct per_cpu_pageset
*p
,
4298 unsigned long batch
= max(1UL, high
/ 4);
4299 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4300 batch
= PAGE_SHIFT
* 8;
4302 pageset_update(&p
->pcp
, high
, batch
);
4305 static void pageset_set_high_and_batch(struct zone
*zone
,
4306 struct per_cpu_pageset
*pcp
)
4308 if (percpu_pagelist_fraction
)
4309 pageset_set_high(pcp
,
4310 (zone
->managed_pages
/
4311 percpu_pagelist_fraction
));
4313 pageset_set_batch(pcp
, zone_batchsize(zone
));
4316 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4318 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4321 pageset_set_high_and_batch(zone
, pcp
);
4324 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4327 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4328 for_each_possible_cpu(cpu
)
4329 zone_pageset_init(zone
, cpu
);
4333 * Allocate per cpu pagesets and initialize them.
4334 * Before this call only boot pagesets were available.
4336 void __init
setup_per_cpu_pageset(void)
4340 for_each_populated_zone(zone
)
4341 setup_zone_pageset(zone
);
4344 static noinline __init_refok
4345 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4351 * The per-page waitqueue mechanism uses hashed waitqueues
4354 zone
->wait_table_hash_nr_entries
=
4355 wait_table_hash_nr_entries(zone_size_pages
);
4356 zone
->wait_table_bits
=
4357 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4358 alloc_size
= zone
->wait_table_hash_nr_entries
4359 * sizeof(wait_queue_head_t
);
4361 if (!slab_is_available()) {
4362 zone
->wait_table
= (wait_queue_head_t
*)
4363 memblock_virt_alloc_node_nopanic(
4364 alloc_size
, zone
->zone_pgdat
->node_id
);
4367 * This case means that a zone whose size was 0 gets new memory
4368 * via memory hot-add.
4369 * But it may be the case that a new node was hot-added. In
4370 * this case vmalloc() will not be able to use this new node's
4371 * memory - this wait_table must be initialized to use this new
4372 * node itself as well.
4373 * To use this new node's memory, further consideration will be
4376 zone
->wait_table
= vmalloc(alloc_size
);
4378 if (!zone
->wait_table
)
4381 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4382 init_waitqueue_head(zone
->wait_table
+ i
);
4387 static __meminit
void zone_pcp_init(struct zone
*zone
)
4390 * per cpu subsystem is not up at this point. The following code
4391 * relies on the ability of the linker to provide the
4392 * offset of a (static) per cpu variable into the per cpu area.
4394 zone
->pageset
= &boot_pageset
;
4396 if (populated_zone(zone
))
4397 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4398 zone
->name
, zone
->present_pages
,
4399 zone_batchsize(zone
));
4402 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4403 unsigned long zone_start_pfn
,
4405 enum memmap_context context
)
4407 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4409 ret
= zone_wait_table_init(zone
, size
);
4412 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4414 zone
->zone_start_pfn
= zone_start_pfn
;
4416 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4417 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4419 (unsigned long)zone_idx(zone
),
4420 zone_start_pfn
, (zone_start_pfn
+ size
));
4422 zone_init_free_lists(zone
);
4427 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4428 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4430 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4432 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4434 unsigned long start_pfn
, end_pfn
;
4437 * NOTE: The following SMP-unsafe globals are only used early in boot
4438 * when the kernel is running single-threaded.
4440 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4441 static int __meminitdata last_nid
;
4443 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4446 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4448 last_start_pfn
= start_pfn
;
4449 last_end_pfn
= end_pfn
;
4455 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4457 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4461 nid
= __early_pfn_to_nid(pfn
);
4464 /* just returns 0 */
4468 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4469 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4473 nid
= __early_pfn_to_nid(pfn
);
4474 if (nid
>= 0 && nid
!= node
)
4481 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4482 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4483 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4485 * If an architecture guarantees that all ranges registered contain no holes
4486 * and may be freed, this this function may be used instead of calling
4487 * memblock_free_early_nid() manually.
4489 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4491 unsigned long start_pfn
, end_pfn
;
4494 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4495 start_pfn
= min(start_pfn
, max_low_pfn
);
4496 end_pfn
= min(end_pfn
, max_low_pfn
);
4498 if (start_pfn
< end_pfn
)
4499 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4500 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4506 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4507 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4509 * If an architecture guarantees that all ranges registered contain no holes and may
4510 * be freed, this function may be used instead of calling memory_present() manually.
4512 void __init
sparse_memory_present_with_active_regions(int nid
)
4514 unsigned long start_pfn
, end_pfn
;
4517 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4518 memory_present(this_nid
, start_pfn
, end_pfn
);
4522 * get_pfn_range_for_nid - Return the start and end page frames for a node
4523 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4524 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4525 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4527 * It returns the start and end page frame of a node based on information
4528 * provided by memblock_set_node(). If called for a node
4529 * with no available memory, a warning is printed and the start and end
4532 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4533 unsigned long *start_pfn
, unsigned long *end_pfn
)
4535 unsigned long this_start_pfn
, this_end_pfn
;
4541 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4542 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4543 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4546 if (*start_pfn
== -1UL)
4551 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4552 * assumption is made that zones within a node are ordered in monotonic
4553 * increasing memory addresses so that the "highest" populated zone is used
4555 static void __init
find_usable_zone_for_movable(void)
4558 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4559 if (zone_index
== ZONE_MOVABLE
)
4562 if (arch_zone_highest_possible_pfn
[zone_index
] >
4563 arch_zone_lowest_possible_pfn
[zone_index
])
4567 VM_BUG_ON(zone_index
== -1);
4568 movable_zone
= zone_index
;
4572 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4573 * because it is sized independent of architecture. Unlike the other zones,
4574 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4575 * in each node depending on the size of each node and how evenly kernelcore
4576 * is distributed. This helper function adjusts the zone ranges
4577 * provided by the architecture for a given node by using the end of the
4578 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4579 * zones within a node are in order of monotonic increases memory addresses
4581 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4582 unsigned long zone_type
,
4583 unsigned long node_start_pfn
,
4584 unsigned long node_end_pfn
,
4585 unsigned long *zone_start_pfn
,
4586 unsigned long *zone_end_pfn
)
4588 /* Only adjust if ZONE_MOVABLE is on this node */
4589 if (zone_movable_pfn
[nid
]) {
4590 /* Size ZONE_MOVABLE */
4591 if (zone_type
== ZONE_MOVABLE
) {
4592 *zone_start_pfn
= zone_movable_pfn
[nid
];
4593 *zone_end_pfn
= min(node_end_pfn
,
4594 arch_zone_highest_possible_pfn
[movable_zone
]);
4596 /* Adjust for ZONE_MOVABLE starting within this range */
4597 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4598 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4599 *zone_end_pfn
= zone_movable_pfn
[nid
];
4601 /* Check if this whole range is within ZONE_MOVABLE */
4602 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4603 *zone_start_pfn
= *zone_end_pfn
;
4608 * Return the number of pages a zone spans in a node, including holes
4609 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4611 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4612 unsigned long zone_type
,
4613 unsigned long node_start_pfn
,
4614 unsigned long node_end_pfn
,
4615 unsigned long *ignored
)
4617 unsigned long zone_start_pfn
, zone_end_pfn
;
4619 /* Get the start and end of the zone */
4620 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4621 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4622 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4623 node_start_pfn
, node_end_pfn
,
4624 &zone_start_pfn
, &zone_end_pfn
);
4626 /* Check that this node has pages within the zone's required range */
4627 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4630 /* Move the zone boundaries inside the node if necessary */
4631 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4632 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4634 /* Return the spanned pages */
4635 return zone_end_pfn
- zone_start_pfn
;
4639 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4640 * then all holes in the requested range will be accounted for.
4642 unsigned long __meminit
__absent_pages_in_range(int nid
,
4643 unsigned long range_start_pfn
,
4644 unsigned long range_end_pfn
)
4646 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4647 unsigned long start_pfn
, end_pfn
;
4650 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4651 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4652 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4653 nr_absent
-= end_pfn
- start_pfn
;
4659 * absent_pages_in_range - Return number of page frames in holes within a range
4660 * @start_pfn: The start PFN to start searching for holes
4661 * @end_pfn: The end PFN to stop searching for holes
4663 * It returns the number of pages frames in memory holes within a range.
4665 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4666 unsigned long end_pfn
)
4668 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4671 /* Return the number of page frames in holes in a zone on a node */
4672 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4673 unsigned long zone_type
,
4674 unsigned long node_start_pfn
,
4675 unsigned long node_end_pfn
,
4676 unsigned long *ignored
)
4678 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4679 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4680 unsigned long zone_start_pfn
, zone_end_pfn
;
4682 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4683 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4685 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4686 node_start_pfn
, node_end_pfn
,
4687 &zone_start_pfn
, &zone_end_pfn
);
4688 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4691 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4692 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4693 unsigned long zone_type
,
4694 unsigned long node_start_pfn
,
4695 unsigned long node_end_pfn
,
4696 unsigned long *zones_size
)
4698 return zones_size
[zone_type
];
4701 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4702 unsigned long zone_type
,
4703 unsigned long node_start_pfn
,
4704 unsigned long node_end_pfn
,
4705 unsigned long *zholes_size
)
4710 return zholes_size
[zone_type
];
4713 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4715 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4716 unsigned long node_start_pfn
,
4717 unsigned long node_end_pfn
,
4718 unsigned long *zones_size
,
4719 unsigned long *zholes_size
)
4721 unsigned long realtotalpages
, totalpages
= 0;
4724 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4725 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4729 pgdat
->node_spanned_pages
= totalpages
;
4731 realtotalpages
= totalpages
;
4732 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4734 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4735 node_start_pfn
, node_end_pfn
,
4737 pgdat
->node_present_pages
= realtotalpages
;
4738 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4742 #ifndef CONFIG_SPARSEMEM
4744 * Calculate the size of the zone->blockflags rounded to an unsigned long
4745 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4746 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4747 * round what is now in bits to nearest long in bits, then return it in
4750 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4752 unsigned long usemapsize
;
4754 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4755 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4756 usemapsize
= usemapsize
>> pageblock_order
;
4757 usemapsize
*= NR_PAGEBLOCK_BITS
;
4758 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4760 return usemapsize
/ 8;
4763 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4765 unsigned long zone_start_pfn
,
4766 unsigned long zonesize
)
4768 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4769 zone
->pageblock_flags
= NULL
;
4771 zone
->pageblock_flags
=
4772 memblock_virt_alloc_node_nopanic(usemapsize
,
4776 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4777 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4778 #endif /* CONFIG_SPARSEMEM */
4780 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4782 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4783 void __paginginit
set_pageblock_order(void)
4787 /* Check that pageblock_nr_pages has not already been setup */
4788 if (pageblock_order
)
4791 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4792 order
= HUGETLB_PAGE_ORDER
;
4794 order
= MAX_ORDER
- 1;
4797 * Assume the largest contiguous order of interest is a huge page.
4798 * This value may be variable depending on boot parameters on IA64 and
4801 pageblock_order
= order
;
4803 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4806 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4807 * is unused as pageblock_order is set at compile-time. See
4808 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4811 void __paginginit
set_pageblock_order(void)
4815 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4817 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4818 unsigned long present_pages
)
4820 unsigned long pages
= spanned_pages
;
4823 * Provide a more accurate estimation if there are holes within
4824 * the zone and SPARSEMEM is in use. If there are holes within the
4825 * zone, each populated memory region may cost us one or two extra
4826 * memmap pages due to alignment because memmap pages for each
4827 * populated regions may not naturally algined on page boundary.
4828 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4830 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4831 IS_ENABLED(CONFIG_SPARSEMEM
))
4832 pages
= present_pages
;
4834 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4838 * Set up the zone data structures:
4839 * - mark all pages reserved
4840 * - mark all memory queues empty
4841 * - clear the memory bitmaps
4843 * NOTE: pgdat should get zeroed by caller.
4845 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4846 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4847 unsigned long *zones_size
, unsigned long *zholes_size
)
4850 int nid
= pgdat
->node_id
;
4851 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4854 pgdat_resize_init(pgdat
);
4855 #ifdef CONFIG_NUMA_BALANCING
4856 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4857 pgdat
->numabalancing_migrate_nr_pages
= 0;
4858 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4860 init_waitqueue_head(&pgdat
->kswapd_wait
);
4861 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4862 pgdat_page_ext_init(pgdat
);
4864 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4865 struct zone
*zone
= pgdat
->node_zones
+ j
;
4866 unsigned long size
, realsize
, freesize
, memmap_pages
;
4868 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4869 node_end_pfn
, zones_size
);
4870 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4876 * Adjust freesize so that it accounts for how much memory
4877 * is used by this zone for memmap. This affects the watermark
4878 * and per-cpu initialisations
4880 memmap_pages
= calc_memmap_size(size
, realsize
);
4881 if (!is_highmem_idx(j
)) {
4882 if (freesize
>= memmap_pages
) {
4883 freesize
-= memmap_pages
;
4886 " %s zone: %lu pages used for memmap\n",
4887 zone_names
[j
], memmap_pages
);
4890 " %s zone: %lu pages exceeds freesize %lu\n",
4891 zone_names
[j
], memmap_pages
, freesize
);
4894 /* Account for reserved pages */
4895 if (j
== 0 && freesize
> dma_reserve
) {
4896 freesize
-= dma_reserve
;
4897 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4898 zone_names
[0], dma_reserve
);
4901 if (!is_highmem_idx(j
))
4902 nr_kernel_pages
+= freesize
;
4903 /* Charge for highmem memmap if there are enough kernel pages */
4904 else if (nr_kernel_pages
> memmap_pages
* 2)
4905 nr_kernel_pages
-= memmap_pages
;
4906 nr_all_pages
+= freesize
;
4908 zone
->spanned_pages
= size
;
4909 zone
->present_pages
= realsize
;
4911 * Set an approximate value for lowmem here, it will be adjusted
4912 * when the bootmem allocator frees pages into the buddy system.
4913 * And all highmem pages will be managed by the buddy system.
4915 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4918 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4920 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4922 zone
->name
= zone_names
[j
];
4923 spin_lock_init(&zone
->lock
);
4924 spin_lock_init(&zone
->lru_lock
);
4925 zone_seqlock_init(zone
);
4926 zone
->zone_pgdat
= pgdat
;
4927 zone_pcp_init(zone
);
4929 /* For bootup, initialized properly in watermark setup */
4930 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4932 lruvec_init(&zone
->lruvec
);
4936 set_pageblock_order();
4937 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4938 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4939 size
, MEMMAP_EARLY
);
4941 memmap_init(size
, nid
, j
, zone_start_pfn
);
4942 zone_start_pfn
+= size
;
4946 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4948 /* Skip empty nodes */
4949 if (!pgdat
->node_spanned_pages
)
4952 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4953 /* ia64 gets its own node_mem_map, before this, without bootmem */
4954 if (!pgdat
->node_mem_map
) {
4955 unsigned long size
, start
, end
;
4959 * The zone's endpoints aren't required to be MAX_ORDER
4960 * aligned but the node_mem_map endpoints must be in order
4961 * for the buddy allocator to function correctly.
4963 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4964 end
= pgdat_end_pfn(pgdat
);
4965 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4966 size
= (end
- start
) * sizeof(struct page
);
4967 map
= alloc_remap(pgdat
->node_id
, size
);
4969 map
= memblock_virt_alloc_node_nopanic(size
,
4971 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4973 #ifndef CONFIG_NEED_MULTIPLE_NODES
4975 * With no DISCONTIG, the global mem_map is just set as node 0's
4977 if (pgdat
== NODE_DATA(0)) {
4978 mem_map
= NODE_DATA(0)->node_mem_map
;
4979 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4980 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4981 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4982 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4985 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4988 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4989 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4991 pg_data_t
*pgdat
= NODE_DATA(nid
);
4992 unsigned long start_pfn
= 0;
4993 unsigned long end_pfn
= 0;
4995 /* pg_data_t should be reset to zero when it's allocated */
4996 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4998 pgdat
->node_id
= nid
;
4999 pgdat
->node_start_pfn
= node_start_pfn
;
5000 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5001 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5002 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5003 (u64
)start_pfn
<< PAGE_SHIFT
, ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5005 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5006 zones_size
, zholes_size
);
5008 alloc_node_mem_map(pgdat
);
5009 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5010 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5011 nid
, (unsigned long)pgdat
,
5012 (unsigned long)pgdat
->node_mem_map
);
5015 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
5016 zones_size
, zholes_size
);
5019 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5021 #if MAX_NUMNODES > 1
5023 * Figure out the number of possible node ids.
5025 void __init
setup_nr_node_ids(void)
5028 unsigned int highest
= 0;
5030 for_each_node_mask(node
, node_possible_map
)
5032 nr_node_ids
= highest
+ 1;
5037 * node_map_pfn_alignment - determine the maximum internode alignment
5039 * This function should be called after node map is populated and sorted.
5040 * It calculates the maximum power of two alignment which can distinguish
5043 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5044 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5045 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5046 * shifted, 1GiB is enough and this function will indicate so.
5048 * This is used to test whether pfn -> nid mapping of the chosen memory
5049 * model has fine enough granularity to avoid incorrect mapping for the
5050 * populated node map.
5052 * Returns the determined alignment in pfn's. 0 if there is no alignment
5053 * requirement (single node).
5055 unsigned long __init
node_map_pfn_alignment(void)
5057 unsigned long accl_mask
= 0, last_end
= 0;
5058 unsigned long start
, end
, mask
;
5062 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5063 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5070 * Start with a mask granular enough to pin-point to the
5071 * start pfn and tick off bits one-by-one until it becomes
5072 * too coarse to separate the current node from the last.
5074 mask
= ~((1 << __ffs(start
)) - 1);
5075 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5078 /* accumulate all internode masks */
5082 /* convert mask to number of pages */
5083 return ~accl_mask
+ 1;
5086 /* Find the lowest pfn for a node */
5087 static unsigned long __init
find_min_pfn_for_node(int nid
)
5089 unsigned long min_pfn
= ULONG_MAX
;
5090 unsigned long start_pfn
;
5093 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5094 min_pfn
= min(min_pfn
, start_pfn
);
5096 if (min_pfn
== ULONG_MAX
) {
5098 "Could not find start_pfn for node %d\n", nid
);
5106 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5108 * It returns the minimum PFN based on information provided via
5109 * memblock_set_node().
5111 unsigned long __init
find_min_pfn_with_active_regions(void)
5113 return find_min_pfn_for_node(MAX_NUMNODES
);
5117 * early_calculate_totalpages()
5118 * Sum pages in active regions for movable zone.
5119 * Populate N_MEMORY for calculating usable_nodes.
5121 static unsigned long __init
early_calculate_totalpages(void)
5123 unsigned long totalpages
= 0;
5124 unsigned long start_pfn
, end_pfn
;
5127 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5128 unsigned long pages
= end_pfn
- start_pfn
;
5130 totalpages
+= pages
;
5132 node_set_state(nid
, N_MEMORY
);
5138 * Find the PFN the Movable zone begins in each node. Kernel memory
5139 * is spread evenly between nodes as long as the nodes have enough
5140 * memory. When they don't, some nodes will have more kernelcore than
5143 static void __init
find_zone_movable_pfns_for_nodes(void)
5146 unsigned long usable_startpfn
;
5147 unsigned long kernelcore_node
, kernelcore_remaining
;
5148 /* save the state before borrow the nodemask */
5149 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5150 unsigned long totalpages
= early_calculate_totalpages();
5151 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5152 struct memblock_region
*r
;
5154 /* Need to find movable_zone earlier when movable_node is specified. */
5155 find_usable_zone_for_movable();
5158 * If movable_node is specified, ignore kernelcore and movablecore
5161 if (movable_node_is_enabled()) {
5162 for_each_memblock(memory
, r
) {
5163 if (!memblock_is_hotpluggable(r
))
5168 usable_startpfn
= PFN_DOWN(r
->base
);
5169 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5170 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5178 * If movablecore=nn[KMG] was specified, calculate what size of
5179 * kernelcore that corresponds so that memory usable for
5180 * any allocation type is evenly spread. If both kernelcore
5181 * and movablecore are specified, then the value of kernelcore
5182 * will be used for required_kernelcore if it's greater than
5183 * what movablecore would have allowed.
5185 if (required_movablecore
) {
5186 unsigned long corepages
;
5189 * Round-up so that ZONE_MOVABLE is at least as large as what
5190 * was requested by the user
5192 required_movablecore
=
5193 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5194 corepages
= totalpages
- required_movablecore
;
5196 required_kernelcore
= max(required_kernelcore
, corepages
);
5199 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5200 if (!required_kernelcore
)
5203 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5204 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5207 /* Spread kernelcore memory as evenly as possible throughout nodes */
5208 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5209 for_each_node_state(nid
, N_MEMORY
) {
5210 unsigned long start_pfn
, end_pfn
;
5213 * Recalculate kernelcore_node if the division per node
5214 * now exceeds what is necessary to satisfy the requested
5215 * amount of memory for the kernel
5217 if (required_kernelcore
< kernelcore_node
)
5218 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5221 * As the map is walked, we track how much memory is usable
5222 * by the kernel using kernelcore_remaining. When it is
5223 * 0, the rest of the node is usable by ZONE_MOVABLE
5225 kernelcore_remaining
= kernelcore_node
;
5227 /* Go through each range of PFNs within this node */
5228 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5229 unsigned long size_pages
;
5231 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5232 if (start_pfn
>= end_pfn
)
5235 /* Account for what is only usable for kernelcore */
5236 if (start_pfn
< usable_startpfn
) {
5237 unsigned long kernel_pages
;
5238 kernel_pages
= min(end_pfn
, usable_startpfn
)
5241 kernelcore_remaining
-= min(kernel_pages
,
5242 kernelcore_remaining
);
5243 required_kernelcore
-= min(kernel_pages
,
5244 required_kernelcore
);
5246 /* Continue if range is now fully accounted */
5247 if (end_pfn
<= usable_startpfn
) {
5250 * Push zone_movable_pfn to the end so
5251 * that if we have to rebalance
5252 * kernelcore across nodes, we will
5253 * not double account here
5255 zone_movable_pfn
[nid
] = end_pfn
;
5258 start_pfn
= usable_startpfn
;
5262 * The usable PFN range for ZONE_MOVABLE is from
5263 * start_pfn->end_pfn. Calculate size_pages as the
5264 * number of pages used as kernelcore
5266 size_pages
= end_pfn
- start_pfn
;
5267 if (size_pages
> kernelcore_remaining
)
5268 size_pages
= kernelcore_remaining
;
5269 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5272 * Some kernelcore has been met, update counts and
5273 * break if the kernelcore for this node has been
5276 required_kernelcore
-= min(required_kernelcore
,
5278 kernelcore_remaining
-= size_pages
;
5279 if (!kernelcore_remaining
)
5285 * If there is still required_kernelcore, we do another pass with one
5286 * less node in the count. This will push zone_movable_pfn[nid] further
5287 * along on the nodes that still have memory until kernelcore is
5291 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5295 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5296 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5297 zone_movable_pfn
[nid
] =
5298 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5301 /* restore the node_state */
5302 node_states
[N_MEMORY
] = saved_node_state
;
5305 /* Any regular or high memory on that node ? */
5306 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5308 enum zone_type zone_type
;
5310 if (N_MEMORY
== N_NORMAL_MEMORY
)
5313 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5314 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5315 if (populated_zone(zone
)) {
5316 node_set_state(nid
, N_HIGH_MEMORY
);
5317 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5318 zone_type
<= ZONE_NORMAL
)
5319 node_set_state(nid
, N_NORMAL_MEMORY
);
5326 * free_area_init_nodes - Initialise all pg_data_t and zone data
5327 * @max_zone_pfn: an array of max PFNs for each zone
5329 * This will call free_area_init_node() for each active node in the system.
5330 * Using the page ranges provided by memblock_set_node(), the size of each
5331 * zone in each node and their holes is calculated. If the maximum PFN
5332 * between two adjacent zones match, it is assumed that the zone is empty.
5333 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5334 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5335 * starts where the previous one ended. For example, ZONE_DMA32 starts
5336 * at arch_max_dma_pfn.
5338 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5340 unsigned long start_pfn
, end_pfn
;
5343 /* Record where the zone boundaries are */
5344 memset(arch_zone_lowest_possible_pfn
, 0,
5345 sizeof(arch_zone_lowest_possible_pfn
));
5346 memset(arch_zone_highest_possible_pfn
, 0,
5347 sizeof(arch_zone_highest_possible_pfn
));
5348 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5349 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5350 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5351 if (i
== ZONE_MOVABLE
)
5353 arch_zone_lowest_possible_pfn
[i
] =
5354 arch_zone_highest_possible_pfn
[i
-1];
5355 arch_zone_highest_possible_pfn
[i
] =
5356 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5358 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5359 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5361 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5362 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5363 find_zone_movable_pfns_for_nodes();
5365 /* Print out the zone ranges */
5366 pr_info("Zone ranges:\n");
5367 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5368 if (i
== ZONE_MOVABLE
)
5370 pr_info(" %-8s ", zone_names
[i
]);
5371 if (arch_zone_lowest_possible_pfn
[i
] ==
5372 arch_zone_highest_possible_pfn
[i
])
5375 pr_cont("[mem %#018Lx-%#018Lx]\n",
5376 (u64
)arch_zone_lowest_possible_pfn
[i
]
5378 ((u64
)arch_zone_highest_possible_pfn
[i
]
5379 << PAGE_SHIFT
) - 1);
5382 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5383 pr_info("Movable zone start for each node\n");
5384 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5385 if (zone_movable_pfn
[i
])
5386 pr_info(" Node %d: %#018Lx\n", i
,
5387 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
5390 /* Print out the early node map */
5391 pr_info("Early memory node ranges\n");
5392 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5393 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
5394 (u64
)start_pfn
<< PAGE_SHIFT
,
5395 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5397 /* Initialise every node */
5398 mminit_verify_pageflags_layout();
5399 setup_nr_node_ids();
5400 for_each_online_node(nid
) {
5401 pg_data_t
*pgdat
= NODE_DATA(nid
);
5402 free_area_init_node(nid
, NULL
,
5403 find_min_pfn_for_node(nid
), NULL
);
5405 /* Any memory on that node */
5406 if (pgdat
->node_present_pages
)
5407 node_set_state(nid
, N_MEMORY
);
5408 check_for_memory(pgdat
, nid
);
5412 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5414 unsigned long long coremem
;
5418 coremem
= memparse(p
, &p
);
5419 *core
= coremem
>> PAGE_SHIFT
;
5421 /* Paranoid check that UL is enough for the coremem value */
5422 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5428 * kernelcore=size sets the amount of memory for use for allocations that
5429 * cannot be reclaimed or migrated.
5431 static int __init
cmdline_parse_kernelcore(char *p
)
5433 return cmdline_parse_core(p
, &required_kernelcore
);
5437 * movablecore=size sets the amount of memory for use for allocations that
5438 * can be reclaimed or migrated.
5440 static int __init
cmdline_parse_movablecore(char *p
)
5442 return cmdline_parse_core(p
, &required_movablecore
);
5445 early_param("kernelcore", cmdline_parse_kernelcore
);
5446 early_param("movablecore", cmdline_parse_movablecore
);
5448 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5450 void adjust_managed_page_count(struct page
*page
, long count
)
5452 spin_lock(&managed_page_count_lock
);
5453 page_zone(page
)->managed_pages
+= count
;
5454 totalram_pages
+= count
;
5455 #ifdef CONFIG_HIGHMEM
5456 if (PageHighMem(page
))
5457 totalhigh_pages
+= count
;
5459 spin_unlock(&managed_page_count_lock
);
5461 EXPORT_SYMBOL(adjust_managed_page_count
);
5463 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5466 unsigned long pages
= 0;
5468 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5469 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5470 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5471 if ((unsigned int)poison
<= 0xFF)
5472 memset(pos
, poison
, PAGE_SIZE
);
5473 free_reserved_page(virt_to_page(pos
));
5477 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5478 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5482 EXPORT_SYMBOL(free_reserved_area
);
5484 #ifdef CONFIG_HIGHMEM
5485 void free_highmem_page(struct page
*page
)
5487 __free_reserved_page(page
);
5489 page_zone(page
)->managed_pages
++;
5495 void __init
mem_init_print_info(const char *str
)
5497 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5498 unsigned long init_code_size
, init_data_size
;
5500 physpages
= get_num_physpages();
5501 codesize
= _etext
- _stext
;
5502 datasize
= _edata
- _sdata
;
5503 rosize
= __end_rodata
- __start_rodata
;
5504 bss_size
= __bss_stop
- __bss_start
;
5505 init_data_size
= __init_end
- __init_begin
;
5506 init_code_size
= _einittext
- _sinittext
;
5509 * Detect special cases and adjust section sizes accordingly:
5510 * 1) .init.* may be embedded into .data sections
5511 * 2) .init.text.* may be out of [__init_begin, __init_end],
5512 * please refer to arch/tile/kernel/vmlinux.lds.S.
5513 * 3) .rodata.* may be embedded into .text or .data sections.
5515 #define adj_init_size(start, end, size, pos, adj) \
5517 if (start <= pos && pos < end && size > adj) \
5521 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5522 _sinittext
, init_code_size
);
5523 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5524 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5525 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5526 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5528 #undef adj_init_size
5530 pr_info("Memory: %luK/%luK available "
5531 "(%luK kernel code, %luK rwdata, %luK rodata, "
5532 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5533 #ifdef CONFIG_HIGHMEM
5537 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5538 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5539 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5540 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
-10),
5541 totalcma_pages
<< (PAGE_SHIFT
-10),
5542 #ifdef CONFIG_HIGHMEM
5543 totalhigh_pages
<< (PAGE_SHIFT
-10),
5545 str
? ", " : "", str
? str
: "");
5549 * set_dma_reserve - set the specified number of pages reserved in the first zone
5550 * @new_dma_reserve: The number of pages to mark reserved
5552 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5553 * In the DMA zone, a significant percentage may be consumed by kernel image
5554 * and other unfreeable allocations which can skew the watermarks badly. This
5555 * function may optionally be used to account for unfreeable pages in the
5556 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5557 * smaller per-cpu batchsize.
5559 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5561 dma_reserve
= new_dma_reserve
;
5564 void __init
free_area_init(unsigned long *zones_size
)
5566 free_area_init_node(0, zones_size
,
5567 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5570 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5571 unsigned long action
, void *hcpu
)
5573 int cpu
= (unsigned long)hcpu
;
5575 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5576 lru_add_drain_cpu(cpu
);
5580 * Spill the event counters of the dead processor
5581 * into the current processors event counters.
5582 * This artificially elevates the count of the current
5585 vm_events_fold_cpu(cpu
);
5588 * Zero the differential counters of the dead processor
5589 * so that the vm statistics are consistent.
5591 * This is only okay since the processor is dead and cannot
5592 * race with what we are doing.
5594 cpu_vm_stats_fold(cpu
);
5599 void __init
page_alloc_init(void)
5601 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5605 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5606 * or min_free_kbytes changes.
5608 static void calculate_totalreserve_pages(void)
5610 struct pglist_data
*pgdat
;
5611 unsigned long reserve_pages
= 0;
5612 enum zone_type i
, j
;
5614 for_each_online_pgdat(pgdat
) {
5615 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5616 struct zone
*zone
= pgdat
->node_zones
+ i
;
5619 /* Find valid and maximum lowmem_reserve in the zone */
5620 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5621 if (zone
->lowmem_reserve
[j
] > max
)
5622 max
= zone
->lowmem_reserve
[j
];
5625 /* we treat the high watermark as reserved pages. */
5626 max
+= high_wmark_pages(zone
);
5628 if (max
> zone
->managed_pages
)
5629 max
= zone
->managed_pages
;
5630 reserve_pages
+= max
;
5632 * Lowmem reserves are not available to
5633 * GFP_HIGHUSER page cache allocations and
5634 * kswapd tries to balance zones to their high
5635 * watermark. As a result, neither should be
5636 * regarded as dirtyable memory, to prevent a
5637 * situation where reclaim has to clean pages
5638 * in order to balance the zones.
5640 zone
->dirty_balance_reserve
= max
;
5643 dirty_balance_reserve
= reserve_pages
;
5644 totalreserve_pages
= reserve_pages
;
5648 * setup_per_zone_lowmem_reserve - called whenever
5649 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5650 * has a correct pages reserved value, so an adequate number of
5651 * pages are left in the zone after a successful __alloc_pages().
5653 static void setup_per_zone_lowmem_reserve(void)
5655 struct pglist_data
*pgdat
;
5656 enum zone_type j
, idx
;
5658 for_each_online_pgdat(pgdat
) {
5659 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5660 struct zone
*zone
= pgdat
->node_zones
+ j
;
5661 unsigned long managed_pages
= zone
->managed_pages
;
5663 zone
->lowmem_reserve
[j
] = 0;
5667 struct zone
*lower_zone
;
5671 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5672 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5674 lower_zone
= pgdat
->node_zones
+ idx
;
5675 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5676 sysctl_lowmem_reserve_ratio
[idx
];
5677 managed_pages
+= lower_zone
->managed_pages
;
5682 /* update totalreserve_pages */
5683 calculate_totalreserve_pages();
5686 static void __setup_per_zone_wmarks(void)
5688 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5689 unsigned long lowmem_pages
= 0;
5691 unsigned long flags
;
5693 /* Calculate total number of !ZONE_HIGHMEM pages */
5694 for_each_zone(zone
) {
5695 if (!is_highmem(zone
))
5696 lowmem_pages
+= zone
->managed_pages
;
5699 for_each_zone(zone
) {
5702 spin_lock_irqsave(&zone
->lock
, flags
);
5703 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5704 do_div(tmp
, lowmem_pages
);
5705 if (is_highmem(zone
)) {
5707 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5708 * need highmem pages, so cap pages_min to a small
5711 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5712 * deltas controls asynch page reclaim, and so should
5713 * not be capped for highmem.
5715 unsigned long min_pages
;
5717 min_pages
= zone
->managed_pages
/ 1024;
5718 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5719 zone
->watermark
[WMARK_MIN
] = min_pages
;
5722 * If it's a lowmem zone, reserve a number of pages
5723 * proportionate to the zone's size.
5725 zone
->watermark
[WMARK_MIN
] = tmp
;
5728 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5729 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5731 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5732 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5733 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5735 setup_zone_migrate_reserve(zone
);
5736 spin_unlock_irqrestore(&zone
->lock
, flags
);
5739 /* update totalreserve_pages */
5740 calculate_totalreserve_pages();
5744 * setup_per_zone_wmarks - called when min_free_kbytes changes
5745 * or when memory is hot-{added|removed}
5747 * Ensures that the watermark[min,low,high] values for each zone are set
5748 * correctly with respect to min_free_kbytes.
5750 void setup_per_zone_wmarks(void)
5752 mutex_lock(&zonelists_mutex
);
5753 __setup_per_zone_wmarks();
5754 mutex_unlock(&zonelists_mutex
);
5758 * The inactive anon list should be small enough that the VM never has to
5759 * do too much work, but large enough that each inactive page has a chance
5760 * to be referenced again before it is swapped out.
5762 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5763 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5764 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5765 * the anonymous pages are kept on the inactive list.
5768 * memory ratio inactive anon
5769 * -------------------------------------
5778 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5780 unsigned int gb
, ratio
;
5782 /* Zone size in gigabytes */
5783 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5785 ratio
= int_sqrt(10 * gb
);
5789 zone
->inactive_ratio
= ratio
;
5792 static void __meminit
setup_per_zone_inactive_ratio(void)
5797 calculate_zone_inactive_ratio(zone
);
5801 * Initialise min_free_kbytes.
5803 * For small machines we want it small (128k min). For large machines
5804 * we want it large (64MB max). But it is not linear, because network
5805 * bandwidth does not increase linearly with machine size. We use
5807 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5808 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5824 int __meminit
init_per_zone_wmark_min(void)
5826 unsigned long lowmem_kbytes
;
5827 int new_min_free_kbytes
;
5829 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5830 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5832 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5833 min_free_kbytes
= new_min_free_kbytes
;
5834 if (min_free_kbytes
< 128)
5835 min_free_kbytes
= 128;
5836 if (min_free_kbytes
> 65536)
5837 min_free_kbytes
= 65536;
5839 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5840 new_min_free_kbytes
, user_min_free_kbytes
);
5842 setup_per_zone_wmarks();
5843 refresh_zone_stat_thresholds();
5844 setup_per_zone_lowmem_reserve();
5845 setup_per_zone_inactive_ratio();
5848 module_init(init_per_zone_wmark_min
)
5851 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5852 * that we can call two helper functions whenever min_free_kbytes
5855 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5856 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5860 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5865 user_min_free_kbytes
= min_free_kbytes
;
5866 setup_per_zone_wmarks();
5872 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5873 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5878 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5883 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5884 sysctl_min_unmapped_ratio
) / 100;
5888 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5889 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5894 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5899 zone
->min_slab_pages
= (zone
->managed_pages
*
5900 sysctl_min_slab_ratio
) / 100;
5906 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5907 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5908 * whenever sysctl_lowmem_reserve_ratio changes.
5910 * The reserve ratio obviously has absolutely no relation with the
5911 * minimum watermarks. The lowmem reserve ratio can only make sense
5912 * if in function of the boot time zone sizes.
5914 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5915 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5917 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5918 setup_per_zone_lowmem_reserve();
5923 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5924 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5925 * pagelist can have before it gets flushed back to buddy allocator.
5927 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5928 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5931 int old_percpu_pagelist_fraction
;
5934 mutex_lock(&pcp_batch_high_lock
);
5935 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5937 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5938 if (!write
|| ret
< 0)
5941 /* Sanity checking to avoid pcp imbalance */
5942 if (percpu_pagelist_fraction
&&
5943 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5944 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5950 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
5953 for_each_populated_zone(zone
) {
5956 for_each_possible_cpu(cpu
)
5957 pageset_set_high_and_batch(zone
,
5958 per_cpu_ptr(zone
->pageset
, cpu
));
5961 mutex_unlock(&pcp_batch_high_lock
);
5965 int hashdist
= HASHDIST_DEFAULT
;
5968 static int __init
set_hashdist(char *str
)
5972 hashdist
= simple_strtoul(str
, &str
, 0);
5975 __setup("hashdist=", set_hashdist
);
5979 * allocate a large system hash table from bootmem
5980 * - it is assumed that the hash table must contain an exact power-of-2
5981 * quantity of entries
5982 * - limit is the number of hash buckets, not the total allocation size
5984 void *__init
alloc_large_system_hash(const char *tablename
,
5985 unsigned long bucketsize
,
5986 unsigned long numentries
,
5989 unsigned int *_hash_shift
,
5990 unsigned int *_hash_mask
,
5991 unsigned long low_limit
,
5992 unsigned long high_limit
)
5994 unsigned long long max
= high_limit
;
5995 unsigned long log2qty
, size
;
5998 /* allow the kernel cmdline to have a say */
6000 /* round applicable memory size up to nearest megabyte */
6001 numentries
= nr_kernel_pages
;
6003 /* It isn't necessary when PAGE_SIZE >= 1MB */
6004 if (PAGE_SHIFT
< 20)
6005 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6007 /* limit to 1 bucket per 2^scale bytes of low memory */
6008 if (scale
> PAGE_SHIFT
)
6009 numentries
>>= (scale
- PAGE_SHIFT
);
6011 numentries
<<= (PAGE_SHIFT
- scale
);
6013 /* Make sure we've got at least a 0-order allocation.. */
6014 if (unlikely(flags
& HASH_SMALL
)) {
6015 /* Makes no sense without HASH_EARLY */
6016 WARN_ON(!(flags
& HASH_EARLY
));
6017 if (!(numentries
>> *_hash_shift
)) {
6018 numentries
= 1UL << *_hash_shift
;
6019 BUG_ON(!numentries
);
6021 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6022 numentries
= PAGE_SIZE
/ bucketsize
;
6024 numentries
= roundup_pow_of_two(numentries
);
6026 /* limit allocation size to 1/16 total memory by default */
6028 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6029 do_div(max
, bucketsize
);
6031 max
= min(max
, 0x80000000ULL
);
6033 if (numentries
< low_limit
)
6034 numentries
= low_limit
;
6035 if (numentries
> max
)
6038 log2qty
= ilog2(numentries
);
6041 size
= bucketsize
<< log2qty
;
6042 if (flags
& HASH_EARLY
)
6043 table
= memblock_virt_alloc_nopanic(size
, 0);
6045 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6048 * If bucketsize is not a power-of-two, we may free
6049 * some pages at the end of hash table which
6050 * alloc_pages_exact() automatically does
6052 if (get_order(size
) < MAX_ORDER
) {
6053 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6054 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6057 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6060 panic("Failed to allocate %s hash table\n", tablename
);
6062 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6065 ilog2(size
) - PAGE_SHIFT
,
6069 *_hash_shift
= log2qty
;
6071 *_hash_mask
= (1 << log2qty
) - 1;
6076 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6077 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6080 #ifdef CONFIG_SPARSEMEM
6081 return __pfn_to_section(pfn
)->pageblock_flags
;
6083 return zone
->pageblock_flags
;
6084 #endif /* CONFIG_SPARSEMEM */
6087 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6089 #ifdef CONFIG_SPARSEMEM
6090 pfn
&= (PAGES_PER_SECTION
-1);
6091 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6093 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6094 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6095 #endif /* CONFIG_SPARSEMEM */
6099 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6100 * @page: The page within the block of interest
6101 * @pfn: The target page frame number
6102 * @end_bitidx: The last bit of interest to retrieve
6103 * @mask: mask of bits that the caller is interested in
6105 * Return: pageblock_bits flags
6107 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6108 unsigned long end_bitidx
,
6112 unsigned long *bitmap
;
6113 unsigned long bitidx
, word_bitidx
;
6116 zone
= page_zone(page
);
6117 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6118 bitidx
= pfn_to_bitidx(zone
, pfn
);
6119 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6120 bitidx
&= (BITS_PER_LONG
-1);
6122 word
= bitmap
[word_bitidx
];
6123 bitidx
+= end_bitidx
;
6124 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6128 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6129 * @page: The page within the block of interest
6130 * @flags: The flags to set
6131 * @pfn: The target page frame number
6132 * @end_bitidx: The last bit of interest
6133 * @mask: mask of bits that the caller is interested in
6135 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6137 unsigned long end_bitidx
,
6141 unsigned long *bitmap
;
6142 unsigned long bitidx
, word_bitidx
;
6143 unsigned long old_word
, word
;
6145 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6147 zone
= page_zone(page
);
6148 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6149 bitidx
= pfn_to_bitidx(zone
, pfn
);
6150 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6151 bitidx
&= (BITS_PER_LONG
-1);
6153 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6155 bitidx
+= end_bitidx
;
6156 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6157 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6159 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6161 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6162 if (word
== old_word
)
6169 * This function checks whether pageblock includes unmovable pages or not.
6170 * If @count is not zero, it is okay to include less @count unmovable pages
6172 * PageLRU check without isolation or lru_lock could race so that
6173 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6174 * expect this function should be exact.
6176 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6177 bool skip_hwpoisoned_pages
)
6179 unsigned long pfn
, iter
, found
;
6183 * For avoiding noise data, lru_add_drain_all() should be called
6184 * If ZONE_MOVABLE, the zone never contains unmovable pages
6186 if (zone_idx(zone
) == ZONE_MOVABLE
)
6188 mt
= get_pageblock_migratetype(page
);
6189 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6192 pfn
= page_to_pfn(page
);
6193 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6194 unsigned long check
= pfn
+ iter
;
6196 if (!pfn_valid_within(check
))
6199 page
= pfn_to_page(check
);
6202 * Hugepages are not in LRU lists, but they're movable.
6203 * We need not scan over tail pages bacause we don't
6204 * handle each tail page individually in migration.
6206 if (PageHuge(page
)) {
6207 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6212 * We can't use page_count without pin a page
6213 * because another CPU can free compound page.
6214 * This check already skips compound tails of THP
6215 * because their page->_count is zero at all time.
6217 if (!atomic_read(&page
->_count
)) {
6218 if (PageBuddy(page
))
6219 iter
+= (1 << page_order(page
)) - 1;
6224 * The HWPoisoned page may be not in buddy system, and
6225 * page_count() is not 0.
6227 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6233 * If there are RECLAIMABLE pages, we need to check
6234 * it. But now, memory offline itself doesn't call
6235 * shrink_node_slabs() and it still to be fixed.
6238 * If the page is not RAM, page_count()should be 0.
6239 * we don't need more check. This is an _used_ not-movable page.
6241 * The problematic thing here is PG_reserved pages. PG_reserved
6242 * is set to both of a memory hole page and a _used_ kernel
6251 bool is_pageblock_removable_nolock(struct page
*page
)
6257 * We have to be careful here because we are iterating over memory
6258 * sections which are not zone aware so we might end up outside of
6259 * the zone but still within the section.
6260 * We have to take care about the node as well. If the node is offline
6261 * its NODE_DATA will be NULL - see page_zone.
6263 if (!node_online(page_to_nid(page
)))
6266 zone
= page_zone(page
);
6267 pfn
= page_to_pfn(page
);
6268 if (!zone_spans_pfn(zone
, pfn
))
6271 return !has_unmovable_pages(zone
, page
, 0, true);
6276 static unsigned long pfn_max_align_down(unsigned long pfn
)
6278 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6279 pageblock_nr_pages
) - 1);
6282 static unsigned long pfn_max_align_up(unsigned long pfn
)
6284 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6285 pageblock_nr_pages
));
6288 /* [start, end) must belong to a single zone. */
6289 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6290 unsigned long start
, unsigned long end
)
6292 /* This function is based on compact_zone() from compaction.c. */
6293 unsigned long nr_reclaimed
;
6294 unsigned long pfn
= start
;
6295 unsigned int tries
= 0;
6300 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6301 if (fatal_signal_pending(current
)) {
6306 if (list_empty(&cc
->migratepages
)) {
6307 cc
->nr_migratepages
= 0;
6308 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6314 } else if (++tries
== 5) {
6315 ret
= ret
< 0 ? ret
: -EBUSY
;
6319 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6321 cc
->nr_migratepages
-= nr_reclaimed
;
6323 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6324 NULL
, 0, cc
->mode
, MR_CMA
);
6327 putback_movable_pages(&cc
->migratepages
);
6334 * alloc_contig_range() -- tries to allocate given range of pages
6335 * @start: start PFN to allocate
6336 * @end: one-past-the-last PFN to allocate
6337 * @migratetype: migratetype of the underlaying pageblocks (either
6338 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6339 * in range must have the same migratetype and it must
6340 * be either of the two.
6342 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6343 * aligned, however it's the caller's responsibility to guarantee that
6344 * we are the only thread that changes migrate type of pageblocks the
6347 * The PFN range must belong to a single zone.
6349 * Returns zero on success or negative error code. On success all
6350 * pages which PFN is in [start, end) are allocated for the caller and
6351 * need to be freed with free_contig_range().
6353 int alloc_contig_range(unsigned long start
, unsigned long end
,
6354 unsigned migratetype
)
6356 unsigned long outer_start
, outer_end
;
6359 struct compact_control cc
= {
6360 .nr_migratepages
= 0,
6362 .zone
= page_zone(pfn_to_page(start
)),
6363 .mode
= MIGRATE_SYNC
,
6364 .ignore_skip_hint
= true,
6366 INIT_LIST_HEAD(&cc
.migratepages
);
6369 * What we do here is we mark all pageblocks in range as
6370 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6371 * have different sizes, and due to the way page allocator
6372 * work, we align the range to biggest of the two pages so
6373 * that page allocator won't try to merge buddies from
6374 * different pageblocks and change MIGRATE_ISOLATE to some
6375 * other migration type.
6377 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6378 * migrate the pages from an unaligned range (ie. pages that
6379 * we are interested in). This will put all the pages in
6380 * range back to page allocator as MIGRATE_ISOLATE.
6382 * When this is done, we take the pages in range from page
6383 * allocator removing them from the buddy system. This way
6384 * page allocator will never consider using them.
6386 * This lets us mark the pageblocks back as
6387 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6388 * aligned range but not in the unaligned, original range are
6389 * put back to page allocator so that buddy can use them.
6392 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6393 pfn_max_align_up(end
), migratetype
,
6398 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6403 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6404 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6405 * more, all pages in [start, end) are free in page allocator.
6406 * What we are going to do is to allocate all pages from
6407 * [start, end) (that is remove them from page allocator).
6409 * The only problem is that pages at the beginning and at the
6410 * end of interesting range may be not aligned with pages that
6411 * page allocator holds, ie. they can be part of higher order
6412 * pages. Because of this, we reserve the bigger range and
6413 * once this is done free the pages we are not interested in.
6415 * We don't have to hold zone->lock here because the pages are
6416 * isolated thus they won't get removed from buddy.
6419 lru_add_drain_all();
6420 drain_all_pages(cc
.zone
);
6423 outer_start
= start
;
6424 while (!PageBuddy(pfn_to_page(outer_start
))) {
6425 if (++order
>= MAX_ORDER
) {
6429 outer_start
&= ~0UL << order
;
6432 /* Make sure the range is really isolated. */
6433 if (test_pages_isolated(outer_start
, end
, false)) {
6434 pr_info("%s: [%lx, %lx) PFNs busy\n",
6435 __func__
, outer_start
, end
);
6440 /* Grab isolated pages from freelists. */
6441 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6447 /* Free head and tail (if any) */
6448 if (start
!= outer_start
)
6449 free_contig_range(outer_start
, start
- outer_start
);
6450 if (end
!= outer_end
)
6451 free_contig_range(end
, outer_end
- end
);
6454 undo_isolate_page_range(pfn_max_align_down(start
),
6455 pfn_max_align_up(end
), migratetype
);
6459 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6461 unsigned int count
= 0;
6463 for (; nr_pages
--; pfn
++) {
6464 struct page
*page
= pfn_to_page(pfn
);
6466 count
+= page_count(page
) != 1;
6469 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6473 #ifdef CONFIG_MEMORY_HOTPLUG
6475 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6476 * page high values need to be recalulated.
6478 void __meminit
zone_pcp_update(struct zone
*zone
)
6481 mutex_lock(&pcp_batch_high_lock
);
6482 for_each_possible_cpu(cpu
)
6483 pageset_set_high_and_batch(zone
,
6484 per_cpu_ptr(zone
->pageset
, cpu
));
6485 mutex_unlock(&pcp_batch_high_lock
);
6489 void zone_pcp_reset(struct zone
*zone
)
6491 unsigned long flags
;
6493 struct per_cpu_pageset
*pset
;
6495 /* avoid races with drain_pages() */
6496 local_irq_save(flags
);
6497 if (zone
->pageset
!= &boot_pageset
) {
6498 for_each_online_cpu(cpu
) {
6499 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6500 drain_zonestat(zone
, pset
);
6502 free_percpu(zone
->pageset
);
6503 zone
->pageset
= &boot_pageset
;
6505 local_irq_restore(flags
);
6508 #ifdef CONFIG_MEMORY_HOTREMOVE
6510 * All pages in the range must be isolated before calling this.
6513 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6517 unsigned int order
, i
;
6519 unsigned long flags
;
6520 /* find the first valid pfn */
6521 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6526 zone
= page_zone(pfn_to_page(pfn
));
6527 spin_lock_irqsave(&zone
->lock
, flags
);
6529 while (pfn
< end_pfn
) {
6530 if (!pfn_valid(pfn
)) {
6534 page
= pfn_to_page(pfn
);
6536 * The HWPoisoned page may be not in buddy system, and
6537 * page_count() is not 0.
6539 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6541 SetPageReserved(page
);
6545 BUG_ON(page_count(page
));
6546 BUG_ON(!PageBuddy(page
));
6547 order
= page_order(page
);
6548 #ifdef CONFIG_DEBUG_VM
6549 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6550 pfn
, 1 << order
, end_pfn
);
6552 list_del(&page
->lru
);
6553 rmv_page_order(page
);
6554 zone
->free_area
[order
].nr_free
--;
6555 for (i
= 0; i
< (1 << order
); i
++)
6556 SetPageReserved((page
+i
));
6557 pfn
+= (1 << order
);
6559 spin_unlock_irqrestore(&zone
->lock
, flags
);
6563 #ifdef CONFIG_MEMORY_FAILURE
6564 bool is_free_buddy_page(struct page
*page
)
6566 struct zone
*zone
= page_zone(page
);
6567 unsigned long pfn
= page_to_pfn(page
);
6568 unsigned long flags
;
6571 spin_lock_irqsave(&zone
->lock
, flags
);
6572 for (order
= 0; order
< MAX_ORDER
; order
++) {
6573 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6575 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6578 spin_unlock_irqrestore(&zone
->lock
, flags
);
6580 return order
< MAX_ORDER
;