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/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_ext.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/prefetch.h>
57 #include <linux/mm_inline.h>
58 #include <linux/migrate.h>
59 #include <linux/page-debug-flags.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
63 #include <asm/sections.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
69 static DEFINE_MUTEX(pcp_batch_high_lock
);
70 #define MIN_PERCPU_PAGELIST_FRACTION (8)
72 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
73 DEFINE_PER_CPU(int, numa_node
);
74 EXPORT_PER_CPU_SYMBOL(numa_node
);
77 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
79 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
80 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
81 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
82 * defined in <linux/topology.h>.
84 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
85 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
86 int _node_numa_mem_
[MAX_NUMNODES
];
90 * Array of node states.
92 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
93 [N_POSSIBLE
] = NODE_MASK_ALL
,
94 [N_ONLINE
] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY
] = { { [0] = 1UL } },
103 [N_CPU
] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states
);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock
);
111 unsigned long totalram_pages __read_mostly
;
112 unsigned long totalreserve_pages __read_mostly
;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly
;
121 int percpu_pagelist_fraction
;
122 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask
;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex
));
139 if (saved_gfp_mask
) {
140 gfp_allowed_mask
= saved_gfp_mask
;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex
));
148 WARN_ON(saved_gfp_mask
);
149 saved_gfp_mask
= gfp_allowed_mask
;
150 gfp_allowed_mask
&= ~GFP_IOFS
;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly
;
165 static void __free_pages_ok(struct page
*page
, unsigned int order
);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages
);
193 static char * const zone_names
[MAX_NR_ZONES
] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes
= 1024;
208 int user_min_free_kbytes
= -1;
210 static unsigned long __meminitdata nr_kernel_pages
;
211 static unsigned long __meminitdata nr_all_pages
;
212 static unsigned long __meminitdata dma_reserve
;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
217 static unsigned long __initdata required_kernelcore
;
218 static unsigned long __initdata required_movablecore
;
219 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone
);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
228 int nr_online_nodes __read_mostly
= 1;
229 EXPORT_SYMBOL(nr_node_ids
);
230 EXPORT_SYMBOL(nr_online_nodes
);
233 int page_group_by_mobility_disabled __read_mostly
;
235 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
237 if (unlikely(page_group_by_mobility_disabled
&&
238 migratetype
< MIGRATE_PCPTYPES
))
239 migratetype
= MIGRATE_UNMOVABLE
;
241 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
242 PB_migrate
, PB_migrate_end
);
245 bool oom_killer_disabled __read_mostly
;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
252 unsigned long pfn
= page_to_pfn(page
);
253 unsigned long sp
, start_pfn
;
256 seq
= zone_span_seqbegin(zone
);
257 start_pfn
= zone
->zone_start_pfn
;
258 sp
= zone
->spanned_pages
;
259 if (!zone_spans_pfn(zone
, pfn
))
261 } while (zone_span_seqretry(zone
, seq
));
264 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
265 pfn
, zone_to_nid(zone
), zone
->name
,
266 start_pfn
, start_pfn
+ sp
);
271 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
273 if (!pfn_valid_within(page_to_pfn(page
)))
275 if (zone
!= page_zone(page
))
281 * Temporary debugging check for pages not lying within a given zone.
283 static int bad_range(struct zone
*zone
, struct page
*page
)
285 if (page_outside_zone_boundaries(zone
, page
))
287 if (!page_is_consistent(zone
, page
))
293 static inline int bad_range(struct zone
*zone
, struct page
*page
)
299 static void bad_page(struct page
*page
, const char *reason
,
300 unsigned long bad_flags
)
302 static unsigned long resume
;
303 static unsigned long nr_shown
;
304 static unsigned long nr_unshown
;
306 /* Don't complain about poisoned pages */
307 if (PageHWPoison(page
)) {
308 page_mapcount_reset(page
); /* remove PageBuddy */
313 * Allow a burst of 60 reports, then keep quiet for that minute;
314 * or allow a steady drip of one report per second.
316 if (nr_shown
== 60) {
317 if (time_before(jiffies
, resume
)) {
323 "BUG: Bad page state: %lu messages suppressed\n",
330 resume
= jiffies
+ 60 * HZ
;
332 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
333 current
->comm
, page_to_pfn(page
));
334 dump_page_badflags(page
, reason
, bad_flags
);
339 /* Leave bad fields for debug, except PageBuddy could make trouble */
340 page_mapcount_reset(page
); /* remove PageBuddy */
341 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
345 * Higher-order pages are called "compound pages". They are structured thusly:
347 * The first PAGE_SIZE page is called the "head page".
349 * The remaining PAGE_SIZE pages are called "tail pages".
351 * All pages have PG_compound set. All tail pages have their ->first_page
352 * pointing at the head page.
354 * The first tail page's ->lru.next holds the address of the compound page's
355 * put_page() function. Its ->lru.prev holds the order of allocation.
356 * This usage means that zero-order pages may not be compound.
359 static void free_compound_page(struct page
*page
)
361 __free_pages_ok(page
, compound_order(page
));
364 void prep_compound_page(struct page
*page
, unsigned long order
)
367 int nr_pages
= 1 << order
;
369 set_compound_page_dtor(page
, free_compound_page
);
370 set_compound_order(page
, order
);
372 for (i
= 1; i
< nr_pages
; i
++) {
373 struct page
*p
= page
+ i
;
374 set_page_count(p
, 0);
375 p
->first_page
= page
;
376 /* Make sure p->first_page is always valid for PageTail() */
382 /* update __split_huge_page_refcount if you change this function */
383 static int destroy_compound_page(struct page
*page
, unsigned long order
)
386 int nr_pages
= 1 << order
;
389 if (unlikely(compound_order(page
) != order
)) {
390 bad_page(page
, "wrong compound order", 0);
394 __ClearPageHead(page
);
396 for (i
= 1; i
< nr_pages
; i
++) {
397 struct page
*p
= page
+ i
;
399 if (unlikely(!PageTail(p
))) {
400 bad_page(page
, "PageTail not set", 0);
402 } else if (unlikely(p
->first_page
!= page
)) {
403 bad_page(page
, "first_page not consistent", 0);
412 static inline void prep_zero_page(struct page
*page
, unsigned int order
,
418 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
419 * and __GFP_HIGHMEM from hard or soft interrupt context.
421 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
422 for (i
= 0; i
< (1 << order
); i
++)
423 clear_highpage(page
+ i
);
426 #ifdef CONFIG_DEBUG_PAGEALLOC
427 unsigned int _debug_guardpage_minorder
;
429 static int __init
debug_guardpage_minorder_setup(char *buf
)
433 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
434 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
437 _debug_guardpage_minorder
= res
;
438 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
441 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
443 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
444 unsigned int order
, int migratetype
)
446 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
447 INIT_LIST_HEAD(&page
->lru
);
448 set_page_private(page
, order
);
449 /* Guard pages are not available for any usage */
450 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
453 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
454 unsigned int order
, int migratetype
)
456 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
457 set_page_private(page
, 0);
458 if (!is_migrate_isolate(migratetype
))
459 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
462 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
463 unsigned int order
, int migratetype
) {}
464 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
465 unsigned int order
, int migratetype
) {}
468 static inline void set_page_order(struct page
*page
, unsigned int order
)
470 set_page_private(page
, order
);
471 __SetPageBuddy(page
);
474 static inline void rmv_page_order(struct page
*page
)
476 __ClearPageBuddy(page
);
477 set_page_private(page
, 0);
481 * This function checks whether a page is free && is the buddy
482 * we can do coalesce a page and its buddy if
483 * (a) the buddy is not in a hole &&
484 * (b) the buddy is in the buddy system &&
485 * (c) a page and its buddy have the same order &&
486 * (d) a page and its buddy are in the same zone.
488 * For recording whether a page is in the buddy system, we set ->_mapcount
489 * PAGE_BUDDY_MAPCOUNT_VALUE.
490 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
491 * serialized by zone->lock.
493 * For recording page's order, we use page_private(page).
495 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
498 if (!pfn_valid_within(page_to_pfn(buddy
)))
501 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
502 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
504 if (page_zone_id(page
) != page_zone_id(buddy
))
510 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
511 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
514 * zone check is done late to avoid uselessly
515 * calculating zone/node ids for pages that could
518 if (page_zone_id(page
) != page_zone_id(buddy
))
527 * Freeing function for a buddy system allocator.
529 * The concept of a buddy system is to maintain direct-mapped table
530 * (containing bit values) for memory blocks of various "orders".
531 * The bottom level table contains the map for the smallest allocatable
532 * units of memory (here, pages), and each level above it describes
533 * pairs of units from the levels below, hence, "buddies".
534 * At a high level, all that happens here is marking the table entry
535 * at the bottom level available, and propagating the changes upward
536 * as necessary, plus some accounting needed to play nicely with other
537 * parts of the VM system.
538 * At each level, we keep a list of pages, which are heads of continuous
539 * free pages of length of (1 << order) and marked with _mapcount
540 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
542 * So when we are allocating or freeing one, we can derive the state of the
543 * other. That is, if we allocate a small block, and both were
544 * free, the remainder of the region must be split into blocks.
545 * If a block is freed, and its buddy is also free, then this
546 * triggers coalescing into a block of larger size.
551 static inline void __free_one_page(struct page
*page
,
553 struct zone
*zone
, unsigned int order
,
556 unsigned long page_idx
;
557 unsigned long combined_idx
;
558 unsigned long uninitialized_var(buddy_idx
);
560 int max_order
= MAX_ORDER
;
562 VM_BUG_ON(!zone_is_initialized(zone
));
564 if (unlikely(PageCompound(page
)))
565 if (unlikely(destroy_compound_page(page
, order
)))
568 VM_BUG_ON(migratetype
== -1);
569 if (is_migrate_isolate(migratetype
)) {
571 * We restrict max order of merging to prevent merge
572 * between freepages on isolate pageblock and normal
573 * pageblock. Without this, pageblock isolation
574 * could cause incorrect freepage accounting.
576 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
578 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
581 page_idx
= pfn
& ((1 << max_order
) - 1);
583 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
584 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
586 while (order
< max_order
- 1) {
587 buddy_idx
= __find_buddy_index(page_idx
, order
);
588 buddy
= page
+ (buddy_idx
- page_idx
);
589 if (!page_is_buddy(page
, buddy
, order
))
592 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
593 * merge with it and move up one order.
595 if (page_is_guard(buddy
)) {
596 clear_page_guard(zone
, buddy
, order
, migratetype
);
598 list_del(&buddy
->lru
);
599 zone
->free_area
[order
].nr_free
--;
600 rmv_page_order(buddy
);
602 combined_idx
= buddy_idx
& page_idx
;
603 page
= page
+ (combined_idx
- page_idx
);
604 page_idx
= combined_idx
;
607 set_page_order(page
, order
);
610 * If this is not the largest possible page, check if the buddy
611 * of the next-highest order is free. If it is, it's possible
612 * that pages are being freed that will coalesce soon. In case,
613 * that is happening, add the free page to the tail of the list
614 * so it's less likely to be used soon and more likely to be merged
615 * as a higher order page
617 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
618 struct page
*higher_page
, *higher_buddy
;
619 combined_idx
= buddy_idx
& page_idx
;
620 higher_page
= page
+ (combined_idx
- page_idx
);
621 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
622 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
623 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
624 list_add_tail(&page
->lru
,
625 &zone
->free_area
[order
].free_list
[migratetype
]);
630 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
632 zone
->free_area
[order
].nr_free
++;
635 static inline int free_pages_check(struct page
*page
)
637 const char *bad_reason
= NULL
;
638 unsigned long bad_flags
= 0;
640 if (unlikely(page_mapcount(page
)))
641 bad_reason
= "nonzero mapcount";
642 if (unlikely(page
->mapping
!= NULL
))
643 bad_reason
= "non-NULL mapping";
644 if (unlikely(atomic_read(&page
->_count
) != 0))
645 bad_reason
= "nonzero _count";
646 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
647 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
648 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
651 if (unlikely(page
->mem_cgroup
))
652 bad_reason
= "page still charged to cgroup";
654 if (unlikely(bad_reason
)) {
655 bad_page(page
, bad_reason
, bad_flags
);
658 page_cpupid_reset_last(page
);
659 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
660 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
665 * Frees a number of pages from the PCP lists
666 * Assumes all pages on list are in same zone, and of same order.
667 * count is the number of pages to free.
669 * If the zone was previously in an "all pages pinned" state then look to
670 * see if this freeing clears that state.
672 * And clear the zone's pages_scanned counter, to hold off the "all pages are
673 * pinned" detection logic.
675 static void free_pcppages_bulk(struct zone
*zone
, int count
,
676 struct per_cpu_pages
*pcp
)
681 unsigned long nr_scanned
;
683 spin_lock(&zone
->lock
);
684 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
686 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
690 struct list_head
*list
;
693 * Remove pages from lists in a round-robin fashion. A
694 * batch_free count is maintained that is incremented when an
695 * empty list is encountered. This is so more pages are freed
696 * off fuller lists instead of spinning excessively around empty
701 if (++migratetype
== MIGRATE_PCPTYPES
)
703 list
= &pcp
->lists
[migratetype
];
704 } while (list_empty(list
));
706 /* This is the only non-empty list. Free them all. */
707 if (batch_free
== MIGRATE_PCPTYPES
)
708 batch_free
= to_free
;
711 int mt
; /* migratetype of the to-be-freed page */
713 page
= list_entry(list
->prev
, struct page
, lru
);
714 /* must delete as __free_one_page list manipulates */
715 list_del(&page
->lru
);
716 mt
= get_freepage_migratetype(page
);
717 if (unlikely(has_isolate_pageblock(zone
)))
718 mt
= get_pageblock_migratetype(page
);
720 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
721 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
722 trace_mm_page_pcpu_drain(page
, 0, mt
);
723 } while (--to_free
&& --batch_free
&& !list_empty(list
));
725 spin_unlock(&zone
->lock
);
728 static void free_one_page(struct zone
*zone
,
729 struct page
*page
, unsigned long pfn
,
733 unsigned long nr_scanned
;
734 spin_lock(&zone
->lock
);
735 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
737 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
739 if (unlikely(has_isolate_pageblock(zone
) ||
740 is_migrate_isolate(migratetype
))) {
741 migratetype
= get_pfnblock_migratetype(page
, pfn
);
743 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
744 spin_unlock(&zone
->lock
);
747 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
752 VM_BUG_ON_PAGE(PageTail(page
), page
);
753 VM_BUG_ON_PAGE(PageHead(page
) && compound_order(page
) != order
, page
);
755 trace_mm_page_free(page
, order
);
756 kmemcheck_free_shadow(page
, order
);
759 page
->mapping
= NULL
;
760 for (i
= 0; i
< (1 << order
); i
++)
761 bad
+= free_pages_check(page
+ i
);
765 if (!PageHighMem(page
)) {
766 debug_check_no_locks_freed(page_address(page
),
768 debug_check_no_obj_freed(page_address(page
),
771 arch_free_page(page
, order
);
772 kernel_map_pages(page
, 1 << order
, 0);
777 static void __free_pages_ok(struct page
*page
, unsigned int order
)
781 unsigned long pfn
= page_to_pfn(page
);
783 if (!free_pages_prepare(page
, order
))
786 migratetype
= get_pfnblock_migratetype(page
, pfn
);
787 local_irq_save(flags
);
788 __count_vm_events(PGFREE
, 1 << order
);
789 set_freepage_migratetype(page
, migratetype
);
790 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
791 local_irq_restore(flags
);
794 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
796 unsigned int nr_pages
= 1 << order
;
797 struct page
*p
= page
;
801 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
803 __ClearPageReserved(p
);
804 set_page_count(p
, 0);
806 __ClearPageReserved(p
);
807 set_page_count(p
, 0);
809 page_zone(page
)->managed_pages
+= nr_pages
;
810 set_page_refcounted(page
);
811 __free_pages(page
, order
);
815 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
816 void __init
init_cma_reserved_pageblock(struct page
*page
)
818 unsigned i
= pageblock_nr_pages
;
819 struct page
*p
= page
;
822 __ClearPageReserved(p
);
823 set_page_count(p
, 0);
826 set_pageblock_migratetype(page
, MIGRATE_CMA
);
828 if (pageblock_order
>= MAX_ORDER
) {
829 i
= pageblock_nr_pages
;
832 set_page_refcounted(p
);
833 __free_pages(p
, MAX_ORDER
- 1);
834 p
+= MAX_ORDER_NR_PAGES
;
835 } while (i
-= MAX_ORDER_NR_PAGES
);
837 set_page_refcounted(page
);
838 __free_pages(page
, pageblock_order
);
841 adjust_managed_page_count(page
, pageblock_nr_pages
);
846 * The order of subdivision here is critical for the IO subsystem.
847 * Please do not alter this order without good reasons and regression
848 * testing. Specifically, as large blocks of memory are subdivided,
849 * the order in which smaller blocks are delivered depends on the order
850 * they're subdivided in this function. This is the primary factor
851 * influencing the order in which pages are delivered to the IO
852 * subsystem according to empirical testing, and this is also justified
853 * by considering the behavior of a buddy system containing a single
854 * large block of memory acted on by a series of small allocations.
855 * This behavior is a critical factor in sglist merging's success.
859 static inline void expand(struct zone
*zone
, struct page
*page
,
860 int low
, int high
, struct free_area
*area
,
863 unsigned long size
= 1 << high
;
869 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
871 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
872 high
< debug_guardpage_minorder()) {
874 * Mark as guard pages (or page), that will allow to
875 * merge back to allocator when buddy will be freed.
876 * Corresponding page table entries will not be touched,
877 * pages will stay not present in virtual address space
879 set_page_guard(zone
, &page
[size
], high
, migratetype
);
882 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
884 set_page_order(&page
[size
], high
);
889 * This page is about to be returned from the page allocator
891 static inline int check_new_page(struct page
*page
)
893 const char *bad_reason
= NULL
;
894 unsigned long bad_flags
= 0;
896 if (unlikely(page_mapcount(page
)))
897 bad_reason
= "nonzero mapcount";
898 if (unlikely(page
->mapping
!= NULL
))
899 bad_reason
= "non-NULL mapping";
900 if (unlikely(atomic_read(&page
->_count
) != 0))
901 bad_reason
= "nonzero _count";
902 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
903 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
904 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
907 if (unlikely(page
->mem_cgroup
))
908 bad_reason
= "page still charged to cgroup";
910 if (unlikely(bad_reason
)) {
911 bad_page(page
, bad_reason
, bad_flags
);
917 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
)
921 for (i
= 0; i
< (1 << order
); i
++) {
922 struct page
*p
= page
+ i
;
923 if (unlikely(check_new_page(p
)))
927 set_page_private(page
, 0);
928 set_page_refcounted(page
);
930 arch_alloc_page(page
, order
);
931 kernel_map_pages(page
, 1 << order
, 1);
933 if (gfp_flags
& __GFP_ZERO
)
934 prep_zero_page(page
, order
, gfp_flags
);
936 if (order
&& (gfp_flags
& __GFP_COMP
))
937 prep_compound_page(page
, order
);
943 * Go through the free lists for the given migratetype and remove
944 * the smallest available page from the freelists
947 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
950 unsigned int current_order
;
951 struct free_area
*area
;
954 /* Find a page of the appropriate size in the preferred list */
955 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
956 area
= &(zone
->free_area
[current_order
]);
957 if (list_empty(&area
->free_list
[migratetype
]))
960 page
= list_entry(area
->free_list
[migratetype
].next
,
962 list_del(&page
->lru
);
963 rmv_page_order(page
);
965 expand(zone
, page
, order
, current_order
, area
, migratetype
);
966 set_freepage_migratetype(page
, migratetype
);
975 * This array describes the order lists are fallen back to when
976 * the free lists for the desirable migrate type are depleted
978 static int fallbacks
[MIGRATE_TYPES
][4] = {
979 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
980 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
982 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
983 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
985 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
987 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
988 #ifdef CONFIG_MEMORY_ISOLATION
989 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
994 * Move the free pages in a range to the free lists of the requested type.
995 * Note that start_page and end_pages are not aligned on a pageblock
996 * boundary. If alignment is required, use move_freepages_block()
998 int move_freepages(struct zone
*zone
,
999 struct page
*start_page
, struct page
*end_page
,
1003 unsigned long order
;
1004 int pages_moved
= 0;
1006 #ifndef CONFIG_HOLES_IN_ZONE
1008 * page_zone is not safe to call in this context when
1009 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1010 * anyway as we check zone boundaries in move_freepages_block().
1011 * Remove at a later date when no bug reports exist related to
1012 * grouping pages by mobility
1014 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1017 for (page
= start_page
; page
<= end_page
;) {
1018 /* Make sure we are not inadvertently changing nodes */
1019 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1021 if (!pfn_valid_within(page_to_pfn(page
))) {
1026 if (!PageBuddy(page
)) {
1031 order
= page_order(page
);
1032 list_move(&page
->lru
,
1033 &zone
->free_area
[order
].free_list
[migratetype
]);
1034 set_freepage_migratetype(page
, migratetype
);
1036 pages_moved
+= 1 << order
;
1042 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1045 unsigned long start_pfn
, end_pfn
;
1046 struct page
*start_page
, *end_page
;
1048 start_pfn
= page_to_pfn(page
);
1049 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1050 start_page
= pfn_to_page(start_pfn
);
1051 end_page
= start_page
+ pageblock_nr_pages
- 1;
1052 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1054 /* Do not cross zone boundaries */
1055 if (!zone_spans_pfn(zone
, start_pfn
))
1057 if (!zone_spans_pfn(zone
, end_pfn
))
1060 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1063 static void change_pageblock_range(struct page
*pageblock_page
,
1064 int start_order
, int migratetype
)
1066 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1068 while (nr_pageblocks
--) {
1069 set_pageblock_migratetype(pageblock_page
, migratetype
);
1070 pageblock_page
+= pageblock_nr_pages
;
1075 * If breaking a large block of pages, move all free pages to the preferred
1076 * allocation list. If falling back for a reclaimable kernel allocation, be
1077 * more aggressive about taking ownership of free pages.
1079 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1080 * nor move CMA pages to different free lists. We don't want unmovable pages
1081 * to be allocated from MIGRATE_CMA areas.
1083 * Returns the new migratetype of the pageblock (or the same old migratetype
1084 * if it was unchanged).
1086 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1087 int start_type
, int fallback_type
)
1089 int current_order
= page_order(page
);
1092 * When borrowing from MIGRATE_CMA, we need to release the excess
1093 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1094 * is set to CMA so it is returned to the correct freelist in case
1095 * the page ends up being not actually allocated from the pcp lists.
1097 if (is_migrate_cma(fallback_type
))
1098 return fallback_type
;
1100 /* Take ownership for orders >= pageblock_order */
1101 if (current_order
>= pageblock_order
) {
1102 change_pageblock_range(page
, current_order
, start_type
);
1106 if (current_order
>= pageblock_order
/ 2 ||
1107 start_type
== MIGRATE_RECLAIMABLE
||
1108 page_group_by_mobility_disabled
) {
1111 pages
= move_freepages_block(zone
, page
, start_type
);
1113 /* Claim the whole block if over half of it is free */
1114 if (pages
>= (1 << (pageblock_order
-1)) ||
1115 page_group_by_mobility_disabled
) {
1117 set_pageblock_migratetype(page
, start_type
);
1123 return fallback_type
;
1126 /* Remove an element from the buddy allocator from the fallback list */
1127 static inline struct page
*
1128 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1130 struct free_area
*area
;
1131 unsigned int current_order
;
1133 int migratetype
, new_type
, i
;
1135 /* Find the largest possible block of pages in the other list */
1136 for (current_order
= MAX_ORDER
-1;
1137 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1140 migratetype
= fallbacks
[start_migratetype
][i
];
1142 /* MIGRATE_RESERVE handled later if necessary */
1143 if (migratetype
== MIGRATE_RESERVE
)
1146 area
= &(zone
->free_area
[current_order
]);
1147 if (list_empty(&area
->free_list
[migratetype
]))
1150 page
= list_entry(area
->free_list
[migratetype
].next
,
1154 new_type
= try_to_steal_freepages(zone
, page
,
1158 /* Remove the page from the freelists */
1159 list_del(&page
->lru
);
1160 rmv_page_order(page
);
1162 expand(zone
, page
, order
, current_order
, area
,
1164 /* The freepage_migratetype may differ from pageblock's
1165 * migratetype depending on the decisions in
1166 * try_to_steal_freepages. This is OK as long as it does
1167 * not differ for MIGRATE_CMA type.
1169 set_freepage_migratetype(page
, new_type
);
1171 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1172 start_migratetype
, migratetype
, new_type
);
1182 * Do the hard work of removing an element from the buddy allocator.
1183 * Call me with the zone->lock already held.
1185 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1191 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1193 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1194 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1197 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1198 * is used because __rmqueue_smallest is an inline function
1199 * and we want just one call site
1202 migratetype
= MIGRATE_RESERVE
;
1207 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1212 * Obtain a specified number of elements from the buddy allocator, all under
1213 * a single hold of the lock, for efficiency. Add them to the supplied list.
1214 * Returns the number of new pages which were placed at *list.
1216 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1217 unsigned long count
, struct list_head
*list
,
1218 int migratetype
, bool cold
)
1222 spin_lock(&zone
->lock
);
1223 for (i
= 0; i
< count
; ++i
) {
1224 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1225 if (unlikely(page
== NULL
))
1229 * Split buddy pages returned by expand() are received here
1230 * in physical page order. The page is added to the callers and
1231 * list and the list head then moves forward. From the callers
1232 * perspective, the linked list is ordered by page number in
1233 * some conditions. This is useful for IO devices that can
1234 * merge IO requests if the physical pages are ordered
1238 list_add(&page
->lru
, list
);
1240 list_add_tail(&page
->lru
, list
);
1242 if (is_migrate_cma(get_freepage_migratetype(page
)))
1243 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1246 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1247 spin_unlock(&zone
->lock
);
1253 * Called from the vmstat counter updater to drain pagesets of this
1254 * currently executing processor on remote nodes after they have
1257 * Note that this function must be called with the thread pinned to
1258 * a single processor.
1260 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1262 unsigned long flags
;
1263 int to_drain
, batch
;
1265 local_irq_save(flags
);
1266 batch
= ACCESS_ONCE(pcp
->batch
);
1267 to_drain
= min(pcp
->count
, batch
);
1269 free_pcppages_bulk(zone
, to_drain
, pcp
);
1270 pcp
->count
-= to_drain
;
1272 local_irq_restore(flags
);
1277 * Drain pcplists of the indicated processor and zone.
1279 * The processor must either be the current processor and the
1280 * thread pinned to the current processor or a processor that
1283 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1285 unsigned long flags
;
1286 struct per_cpu_pageset
*pset
;
1287 struct per_cpu_pages
*pcp
;
1289 local_irq_save(flags
);
1290 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1294 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1297 local_irq_restore(flags
);
1301 * Drain pcplists of all zones on the indicated processor.
1303 * The processor must either be the current processor and the
1304 * thread pinned to the current processor or a processor that
1307 static void drain_pages(unsigned int cpu
)
1311 for_each_populated_zone(zone
) {
1312 drain_pages_zone(cpu
, zone
);
1317 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1319 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1320 * the single zone's pages.
1322 void drain_local_pages(struct zone
*zone
)
1324 int cpu
= smp_processor_id();
1327 drain_pages_zone(cpu
, zone
);
1333 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1335 * When zone parameter is non-NULL, spill just the single zone's pages.
1337 * Note that this code is protected against sending an IPI to an offline
1338 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1339 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1340 * nothing keeps CPUs from showing up after we populated the cpumask and
1341 * before the call to on_each_cpu_mask().
1343 void drain_all_pages(struct zone
*zone
)
1348 * Allocate in the BSS so we wont require allocation in
1349 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1351 static cpumask_t cpus_with_pcps
;
1354 * We don't care about racing with CPU hotplug event
1355 * as offline notification will cause the notified
1356 * cpu to drain that CPU pcps and on_each_cpu_mask
1357 * disables preemption as part of its processing
1359 for_each_online_cpu(cpu
) {
1360 struct per_cpu_pageset
*pcp
;
1362 bool has_pcps
= false;
1365 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1369 for_each_populated_zone(z
) {
1370 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1371 if (pcp
->pcp
.count
) {
1379 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1381 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1383 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1387 #ifdef CONFIG_HIBERNATION
1389 void mark_free_pages(struct zone
*zone
)
1391 unsigned long pfn
, max_zone_pfn
;
1392 unsigned long flags
;
1393 unsigned int order
, t
;
1394 struct list_head
*curr
;
1396 if (zone_is_empty(zone
))
1399 spin_lock_irqsave(&zone
->lock
, flags
);
1401 max_zone_pfn
= zone_end_pfn(zone
);
1402 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1403 if (pfn_valid(pfn
)) {
1404 struct page
*page
= pfn_to_page(pfn
);
1406 if (!swsusp_page_is_forbidden(page
))
1407 swsusp_unset_page_free(page
);
1410 for_each_migratetype_order(order
, t
) {
1411 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1414 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1415 for (i
= 0; i
< (1UL << order
); i
++)
1416 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1419 spin_unlock_irqrestore(&zone
->lock
, flags
);
1421 #endif /* CONFIG_PM */
1424 * Free a 0-order page
1425 * cold == true ? free a cold page : free a hot page
1427 void free_hot_cold_page(struct page
*page
, bool cold
)
1429 struct zone
*zone
= page_zone(page
);
1430 struct per_cpu_pages
*pcp
;
1431 unsigned long flags
;
1432 unsigned long pfn
= page_to_pfn(page
);
1435 if (!free_pages_prepare(page
, 0))
1438 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1439 set_freepage_migratetype(page
, migratetype
);
1440 local_irq_save(flags
);
1441 __count_vm_event(PGFREE
);
1444 * We only track unmovable, reclaimable and movable on pcp lists.
1445 * Free ISOLATE pages back to the allocator because they are being
1446 * offlined but treat RESERVE as movable pages so we can get those
1447 * areas back if necessary. Otherwise, we may have to free
1448 * excessively into the page allocator
1450 if (migratetype
>= MIGRATE_PCPTYPES
) {
1451 if (unlikely(is_migrate_isolate(migratetype
))) {
1452 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1455 migratetype
= MIGRATE_MOVABLE
;
1458 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1460 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1462 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1464 if (pcp
->count
>= pcp
->high
) {
1465 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1466 free_pcppages_bulk(zone
, batch
, pcp
);
1467 pcp
->count
-= batch
;
1471 local_irq_restore(flags
);
1475 * Free a list of 0-order pages
1477 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1479 struct page
*page
, *next
;
1481 list_for_each_entry_safe(page
, next
, list
, lru
) {
1482 trace_mm_page_free_batched(page
, cold
);
1483 free_hot_cold_page(page
, cold
);
1488 * split_page takes a non-compound higher-order page, and splits it into
1489 * n (1<<order) sub-pages: page[0..n]
1490 * Each sub-page must be freed individually.
1492 * Note: this is probably too low level an operation for use in drivers.
1493 * Please consult with lkml before using this in your driver.
1495 void split_page(struct page
*page
, unsigned int order
)
1499 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1500 VM_BUG_ON_PAGE(!page_count(page
), page
);
1502 #ifdef CONFIG_KMEMCHECK
1504 * Split shadow pages too, because free(page[0]) would
1505 * otherwise free the whole shadow.
1507 if (kmemcheck_page_is_tracked(page
))
1508 split_page(virt_to_page(page
[0].shadow
), order
);
1511 for (i
= 1; i
< (1 << order
); i
++)
1512 set_page_refcounted(page
+ i
);
1514 EXPORT_SYMBOL_GPL(split_page
);
1516 int __isolate_free_page(struct page
*page
, unsigned int order
)
1518 unsigned long watermark
;
1522 BUG_ON(!PageBuddy(page
));
1524 zone
= page_zone(page
);
1525 mt
= get_pageblock_migratetype(page
);
1527 if (!is_migrate_isolate(mt
)) {
1528 /* Obey watermarks as if the page was being allocated */
1529 watermark
= low_wmark_pages(zone
) + (1 << order
);
1530 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1533 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1536 /* Remove page from free list */
1537 list_del(&page
->lru
);
1538 zone
->free_area
[order
].nr_free
--;
1539 rmv_page_order(page
);
1541 /* Set the pageblock if the isolated page is at least a pageblock */
1542 if (order
>= pageblock_order
- 1) {
1543 struct page
*endpage
= page
+ (1 << order
) - 1;
1544 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1545 int mt
= get_pageblock_migratetype(page
);
1546 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1547 set_pageblock_migratetype(page
,
1552 return 1UL << order
;
1556 * Similar to split_page except the page is already free. As this is only
1557 * being used for migration, the migratetype of the block also changes.
1558 * As this is called with interrupts disabled, the caller is responsible
1559 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1562 * Note: this is probably too low level an operation for use in drivers.
1563 * Please consult with lkml before using this in your driver.
1565 int split_free_page(struct page
*page
)
1570 order
= page_order(page
);
1572 nr_pages
= __isolate_free_page(page
, order
);
1576 /* Split into individual pages */
1577 set_page_refcounted(page
);
1578 split_page(page
, order
);
1583 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1584 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1588 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1589 struct zone
*zone
, unsigned int order
,
1590 gfp_t gfp_flags
, int migratetype
)
1592 unsigned long flags
;
1594 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1597 if (likely(order
== 0)) {
1598 struct per_cpu_pages
*pcp
;
1599 struct list_head
*list
;
1601 local_irq_save(flags
);
1602 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1603 list
= &pcp
->lists
[migratetype
];
1604 if (list_empty(list
)) {
1605 pcp
->count
+= rmqueue_bulk(zone
, 0,
1608 if (unlikely(list_empty(list
)))
1613 page
= list_entry(list
->prev
, struct page
, lru
);
1615 page
= list_entry(list
->next
, struct page
, lru
);
1617 list_del(&page
->lru
);
1620 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1622 * __GFP_NOFAIL is not to be used in new code.
1624 * All __GFP_NOFAIL callers should be fixed so that they
1625 * properly detect and handle allocation failures.
1627 * We most definitely don't want callers attempting to
1628 * allocate greater than order-1 page units with
1631 WARN_ON_ONCE(order
> 1);
1633 spin_lock_irqsave(&zone
->lock
, flags
);
1634 page
= __rmqueue(zone
, order
, migratetype
);
1635 spin_unlock(&zone
->lock
);
1638 __mod_zone_freepage_state(zone
, -(1 << order
),
1639 get_freepage_migratetype(page
));
1642 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1643 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1644 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1645 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1647 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1648 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1649 local_irq_restore(flags
);
1651 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1652 if (prep_new_page(page
, order
, gfp_flags
))
1657 local_irq_restore(flags
);
1661 #ifdef CONFIG_FAIL_PAGE_ALLOC
1664 struct fault_attr attr
;
1666 u32 ignore_gfp_highmem
;
1667 u32 ignore_gfp_wait
;
1669 } fail_page_alloc
= {
1670 .attr
= FAULT_ATTR_INITIALIZER
,
1671 .ignore_gfp_wait
= 1,
1672 .ignore_gfp_highmem
= 1,
1676 static int __init
setup_fail_page_alloc(char *str
)
1678 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1680 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1682 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1684 if (order
< fail_page_alloc
.min_order
)
1686 if (gfp_mask
& __GFP_NOFAIL
)
1688 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1690 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1693 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1696 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1698 static int __init
fail_page_alloc_debugfs(void)
1700 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1703 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1704 &fail_page_alloc
.attr
);
1706 return PTR_ERR(dir
);
1708 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1709 &fail_page_alloc
.ignore_gfp_wait
))
1711 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1712 &fail_page_alloc
.ignore_gfp_highmem
))
1714 if (!debugfs_create_u32("min-order", mode
, dir
,
1715 &fail_page_alloc
.min_order
))
1720 debugfs_remove_recursive(dir
);
1725 late_initcall(fail_page_alloc_debugfs
);
1727 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1729 #else /* CONFIG_FAIL_PAGE_ALLOC */
1731 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1736 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1739 * Return true if free pages are above 'mark'. This takes into account the order
1740 * of the allocation.
1742 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1743 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1746 /* free_pages may go negative - that's OK */
1751 free_pages
-= (1 << order
) - 1;
1752 if (alloc_flags
& ALLOC_HIGH
)
1754 if (alloc_flags
& ALLOC_HARDER
)
1757 /* If allocation can't use CMA areas don't use free CMA pages */
1758 if (!(alloc_flags
& ALLOC_CMA
))
1759 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1762 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1764 for (o
= 0; o
< order
; o
++) {
1765 /* At the next order, this order's pages become unavailable */
1766 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1768 /* Require fewer higher order pages to be free */
1771 if (free_pages
<= min
)
1777 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1778 int classzone_idx
, int alloc_flags
)
1780 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1781 zone_page_state(z
, NR_FREE_PAGES
));
1784 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1785 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1787 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1789 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1790 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1792 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1798 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1799 * skip over zones that are not allowed by the cpuset, or that have
1800 * been recently (in last second) found to be nearly full. See further
1801 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1802 * that have to skip over a lot of full or unallowed zones.
1804 * If the zonelist cache is present in the passed zonelist, then
1805 * returns a pointer to the allowed node mask (either the current
1806 * tasks mems_allowed, or node_states[N_MEMORY].)
1808 * If the zonelist cache is not available for this zonelist, does
1809 * nothing and returns NULL.
1811 * If the fullzones BITMAP in the zonelist cache is stale (more than
1812 * a second since last zap'd) then we zap it out (clear its bits.)
1814 * We hold off even calling zlc_setup, until after we've checked the
1815 * first zone in the zonelist, on the theory that most allocations will
1816 * be satisfied from that first zone, so best to examine that zone as
1817 * quickly as we can.
1819 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1821 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1822 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1824 zlc
= zonelist
->zlcache_ptr
;
1828 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1829 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1830 zlc
->last_full_zap
= jiffies
;
1833 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1834 &cpuset_current_mems_allowed
:
1835 &node_states
[N_MEMORY
];
1836 return allowednodes
;
1840 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1841 * if it is worth looking at further for free memory:
1842 * 1) Check that the zone isn't thought to be full (doesn't have its
1843 * bit set in the zonelist_cache fullzones BITMAP).
1844 * 2) Check that the zones node (obtained from the zonelist_cache
1845 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1846 * Return true (non-zero) if zone is worth looking at further, or
1847 * else return false (zero) if it is not.
1849 * This check -ignores- the distinction between various watermarks,
1850 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1851 * found to be full for any variation of these watermarks, it will
1852 * be considered full for up to one second by all requests, unless
1853 * we are so low on memory on all allowed nodes that we are forced
1854 * into the second scan of the zonelist.
1856 * In the second scan we ignore this zonelist cache and exactly
1857 * apply the watermarks to all zones, even it is slower to do so.
1858 * We are low on memory in the second scan, and should leave no stone
1859 * unturned looking for a free page.
1861 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1862 nodemask_t
*allowednodes
)
1864 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1865 int i
; /* index of *z in zonelist zones */
1866 int n
; /* node that zone *z is on */
1868 zlc
= zonelist
->zlcache_ptr
;
1872 i
= z
- zonelist
->_zonerefs
;
1875 /* This zone is worth trying if it is allowed but not full */
1876 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1880 * Given 'z' scanning a zonelist, set the corresponding bit in
1881 * zlc->fullzones, so that subsequent attempts to allocate a page
1882 * from that zone don't waste time re-examining it.
1884 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1886 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1887 int i
; /* index of *z in zonelist zones */
1889 zlc
= zonelist
->zlcache_ptr
;
1893 i
= z
- zonelist
->_zonerefs
;
1895 set_bit(i
, zlc
->fullzones
);
1899 * clear all zones full, called after direct reclaim makes progress so that
1900 * a zone that was recently full is not skipped over for up to a second
1902 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1904 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1906 zlc
= zonelist
->zlcache_ptr
;
1910 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1913 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1915 return local_zone
->node
== zone
->node
;
1918 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1920 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1924 #else /* CONFIG_NUMA */
1926 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1931 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1932 nodemask_t
*allowednodes
)
1937 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1941 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1945 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1950 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1955 #endif /* CONFIG_NUMA */
1957 static void reset_alloc_batches(struct zone
*preferred_zone
)
1959 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
1962 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
1963 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
1964 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
1965 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1966 } while (zone
++ != preferred_zone
);
1970 * get_page_from_freelist goes through the zonelist trying to allocate
1973 static struct page
*
1974 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1975 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1976 struct zone
*preferred_zone
, int classzone_idx
, int migratetype
)
1979 struct page
*page
= NULL
;
1981 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1982 int zlc_active
= 0; /* set if using zonelist_cache */
1983 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1984 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
1985 (gfp_mask
& __GFP_WRITE
);
1986 int nr_fair_skipped
= 0;
1987 bool zonelist_rescan
;
1990 zonelist_rescan
= false;
1993 * Scan zonelist, looking for a zone with enough free.
1994 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
1996 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1997 high_zoneidx
, nodemask
) {
2000 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2001 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2003 if (cpusets_enabled() &&
2004 (alloc_flags
& ALLOC_CPUSET
) &&
2005 !cpuset_zone_allowed(zone
, gfp_mask
))
2008 * Distribute pages in proportion to the individual
2009 * zone size to ensure fair page aging. The zone a
2010 * page was allocated in should have no effect on the
2011 * time the page has in memory before being reclaimed.
2013 if (alloc_flags
& ALLOC_FAIR
) {
2014 if (!zone_local(preferred_zone
, zone
))
2016 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2022 * When allocating a page cache page for writing, we
2023 * want to get it from a zone that is within its dirty
2024 * limit, such that no single zone holds more than its
2025 * proportional share of globally allowed dirty pages.
2026 * The dirty limits take into account the zone's
2027 * lowmem reserves and high watermark so that kswapd
2028 * should be able to balance it without having to
2029 * write pages from its LRU list.
2031 * This may look like it could increase pressure on
2032 * lower zones by failing allocations in higher zones
2033 * before they are full. But the pages that do spill
2034 * over are limited as the lower zones are protected
2035 * by this very same mechanism. It should not become
2036 * a practical burden to them.
2038 * XXX: For now, allow allocations to potentially
2039 * exceed the per-zone dirty limit in the slowpath
2040 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2041 * which is important when on a NUMA setup the allowed
2042 * zones are together not big enough to reach the
2043 * global limit. The proper fix for these situations
2044 * will require awareness of zones in the
2045 * dirty-throttling and the flusher threads.
2047 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2050 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2051 if (!zone_watermark_ok(zone
, order
, mark
,
2052 classzone_idx
, alloc_flags
)) {
2055 /* Checked here to keep the fast path fast */
2056 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2057 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2060 if (IS_ENABLED(CONFIG_NUMA
) &&
2061 !did_zlc_setup
&& nr_online_nodes
> 1) {
2063 * we do zlc_setup if there are multiple nodes
2064 * and before considering the first zone allowed
2067 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2072 if (zone_reclaim_mode
== 0 ||
2073 !zone_allows_reclaim(preferred_zone
, zone
))
2074 goto this_zone_full
;
2077 * As we may have just activated ZLC, check if the first
2078 * eligible zone has failed zone_reclaim recently.
2080 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2081 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2084 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2086 case ZONE_RECLAIM_NOSCAN
:
2089 case ZONE_RECLAIM_FULL
:
2090 /* scanned but unreclaimable */
2093 /* did we reclaim enough */
2094 if (zone_watermark_ok(zone
, order
, mark
,
2095 classzone_idx
, alloc_flags
))
2099 * Failed to reclaim enough to meet watermark.
2100 * Only mark the zone full if checking the min
2101 * watermark or if we failed to reclaim just
2102 * 1<<order pages or else the page allocator
2103 * fastpath will prematurely mark zones full
2104 * when the watermark is between the low and
2107 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2108 ret
== ZONE_RECLAIM_SOME
)
2109 goto this_zone_full
;
2116 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2117 gfp_mask
, migratetype
);
2121 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2122 zlc_mark_zone_full(zonelist
, z
);
2127 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2128 * necessary to allocate the page. The expectation is
2129 * that the caller is taking steps that will free more
2130 * memory. The caller should avoid the page being used
2131 * for !PFMEMALLOC purposes.
2133 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2138 * The first pass makes sure allocations are spread fairly within the
2139 * local node. However, the local node might have free pages left
2140 * after the fairness batches are exhausted, and remote zones haven't
2141 * even been considered yet. Try once more without fairness, and
2142 * include remote zones now, before entering the slowpath and waking
2143 * kswapd: prefer spilling to a remote zone over swapping locally.
2145 if (alloc_flags
& ALLOC_FAIR
) {
2146 alloc_flags
&= ~ALLOC_FAIR
;
2147 if (nr_fair_skipped
) {
2148 zonelist_rescan
= true;
2149 reset_alloc_batches(preferred_zone
);
2151 if (nr_online_nodes
> 1)
2152 zonelist_rescan
= true;
2155 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2156 /* Disable zlc cache for second zonelist scan */
2158 zonelist_rescan
= true;
2161 if (zonelist_rescan
)
2168 * Large machines with many possible nodes should not always dump per-node
2169 * meminfo in irq context.
2171 static inline bool should_suppress_show_mem(void)
2176 ret
= in_interrupt();
2181 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2182 DEFAULT_RATELIMIT_INTERVAL
,
2183 DEFAULT_RATELIMIT_BURST
);
2185 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2187 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2189 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2190 debug_guardpage_minorder() > 0)
2194 * This documents exceptions given to allocations in certain
2195 * contexts that are allowed to allocate outside current's set
2198 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2199 if (test_thread_flag(TIF_MEMDIE
) ||
2200 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2201 filter
&= ~SHOW_MEM_FILTER_NODES
;
2202 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2203 filter
&= ~SHOW_MEM_FILTER_NODES
;
2206 struct va_format vaf
;
2209 va_start(args
, fmt
);
2214 pr_warn("%pV", &vaf
);
2219 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2220 current
->comm
, order
, gfp_mask
);
2223 if (!should_suppress_show_mem())
2228 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2229 unsigned long did_some_progress
,
2230 unsigned long pages_reclaimed
)
2232 /* Do not loop if specifically requested */
2233 if (gfp_mask
& __GFP_NORETRY
)
2236 /* Always retry if specifically requested */
2237 if (gfp_mask
& __GFP_NOFAIL
)
2241 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2242 * making forward progress without invoking OOM. Suspend also disables
2243 * storage devices so kswapd will not help. Bail if we are suspending.
2245 if (!did_some_progress
&& pm_suspended_storage())
2249 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2250 * means __GFP_NOFAIL, but that may not be true in other
2253 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2257 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2258 * specified, then we retry until we no longer reclaim any pages
2259 * (above), or we've reclaimed an order of pages at least as
2260 * large as the allocation's order. In both cases, if the
2261 * allocation still fails, we stop retrying.
2263 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2269 static inline struct page
*
2270 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2271 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2272 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2273 int classzone_idx
, int migratetype
)
2277 /* Acquire the per-zone oom lock for each zone */
2278 if (!oom_zonelist_trylock(zonelist
, gfp_mask
)) {
2279 schedule_timeout_uninterruptible(1);
2284 * PM-freezer should be notified that there might be an OOM killer on
2285 * its way to kill and wake somebody up. This is too early and we might
2286 * end up not killing anything but false positives are acceptable.
2287 * See freeze_processes.
2292 * Go through the zonelist yet one more time, keep very high watermark
2293 * here, this is only to catch a parallel oom killing, we must fail if
2294 * we're still under heavy pressure.
2296 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2297 order
, zonelist
, high_zoneidx
,
2298 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2299 preferred_zone
, classzone_idx
, migratetype
);
2303 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2304 /* The OOM killer will not help higher order allocs */
2305 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2307 /* The OOM killer does not needlessly kill tasks for lowmem */
2308 if (high_zoneidx
< ZONE_NORMAL
)
2311 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2312 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2313 * The caller should handle page allocation failure by itself if
2314 * it specifies __GFP_THISNODE.
2315 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2317 if (gfp_mask
& __GFP_THISNODE
)
2320 /* Exhausted what can be done so it's blamo time */
2321 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2324 oom_zonelist_unlock(zonelist
, gfp_mask
);
2328 #ifdef CONFIG_COMPACTION
2329 /* Try memory compaction for high-order allocations before reclaim */
2330 static struct page
*
2331 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2332 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2333 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2334 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2335 int *contended_compaction
, bool *deferred_compaction
)
2337 unsigned long compact_result
;
2343 current
->flags
|= PF_MEMALLOC
;
2344 compact_result
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2346 contended_compaction
,
2347 alloc_flags
, classzone_idx
);
2348 current
->flags
&= ~PF_MEMALLOC
;
2350 switch (compact_result
) {
2351 case COMPACT_DEFERRED
:
2352 *deferred_compaction
= true;
2354 case COMPACT_SKIPPED
:
2361 * At least in one zone compaction wasn't deferred or skipped, so let's
2362 * count a compaction stall
2364 count_vm_event(COMPACTSTALL
);
2366 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2367 order
, zonelist
, high_zoneidx
,
2368 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2369 preferred_zone
, classzone_idx
, migratetype
);
2372 struct zone
*zone
= page_zone(page
);
2374 zone
->compact_blockskip_flush
= false;
2375 compaction_defer_reset(zone
, order
, true);
2376 count_vm_event(COMPACTSUCCESS
);
2381 * It's bad if compaction run occurs and fails. The most likely reason
2382 * is that pages exist, but not enough to satisfy watermarks.
2384 count_vm_event(COMPACTFAIL
);
2391 static inline struct page
*
2392 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2393 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2394 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2395 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2396 int *contended_compaction
, bool *deferred_compaction
)
2400 #endif /* CONFIG_COMPACTION */
2402 /* Perform direct synchronous page reclaim */
2404 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2405 nodemask_t
*nodemask
)
2407 struct reclaim_state reclaim_state
;
2412 /* We now go into synchronous reclaim */
2413 cpuset_memory_pressure_bump();
2414 current
->flags
|= PF_MEMALLOC
;
2415 lockdep_set_current_reclaim_state(gfp_mask
);
2416 reclaim_state
.reclaimed_slab
= 0;
2417 current
->reclaim_state
= &reclaim_state
;
2419 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2421 current
->reclaim_state
= NULL
;
2422 lockdep_clear_current_reclaim_state();
2423 current
->flags
&= ~PF_MEMALLOC
;
2430 /* The really slow allocator path where we enter direct reclaim */
2431 static inline struct page
*
2432 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2433 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2434 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2435 int classzone_idx
, int migratetype
, unsigned long *did_some_progress
)
2437 struct page
*page
= NULL
;
2438 bool drained
= false;
2440 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2442 if (unlikely(!(*did_some_progress
)))
2445 /* After successful reclaim, reconsider all zones for allocation */
2446 if (IS_ENABLED(CONFIG_NUMA
))
2447 zlc_clear_zones_full(zonelist
);
2450 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2451 zonelist
, high_zoneidx
,
2452 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2453 preferred_zone
, classzone_idx
,
2457 * If an allocation failed after direct reclaim, it could be because
2458 * pages are pinned on the per-cpu lists. Drain them and try again
2460 if (!page
&& !drained
) {
2461 drain_all_pages(NULL
);
2470 * This is called in the allocator slow-path if the allocation request is of
2471 * sufficient urgency to ignore watermarks and take other desperate measures
2473 static inline struct page
*
2474 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2475 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2476 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2477 int classzone_idx
, int migratetype
)
2482 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2483 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2484 preferred_zone
, classzone_idx
, migratetype
);
2486 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2487 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2488 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2493 static void wake_all_kswapds(unsigned int order
,
2494 struct zonelist
*zonelist
,
2495 enum zone_type high_zoneidx
,
2496 struct zone
*preferred_zone
,
2497 nodemask_t
*nodemask
)
2502 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2503 high_zoneidx
, nodemask
)
2504 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2508 gfp_to_alloc_flags(gfp_t gfp_mask
)
2510 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2511 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2513 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2514 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2517 * The caller may dip into page reserves a bit more if the caller
2518 * cannot run direct reclaim, or if the caller has realtime scheduling
2519 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2520 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2522 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2526 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2527 * if it can't schedule.
2529 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2530 alloc_flags
|= ALLOC_HARDER
;
2532 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2533 * comment for __cpuset_node_allowed().
2535 alloc_flags
&= ~ALLOC_CPUSET
;
2536 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2537 alloc_flags
|= ALLOC_HARDER
;
2539 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2540 if (gfp_mask
& __GFP_MEMALLOC
)
2541 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2542 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2543 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2544 else if (!in_interrupt() &&
2545 ((current
->flags
& PF_MEMALLOC
) ||
2546 unlikely(test_thread_flag(TIF_MEMDIE
))))
2547 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2550 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2551 alloc_flags
|= ALLOC_CMA
;
2556 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2558 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2561 static inline struct page
*
2562 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2563 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2564 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2565 int classzone_idx
, int migratetype
)
2567 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2568 struct page
*page
= NULL
;
2570 unsigned long pages_reclaimed
= 0;
2571 unsigned long did_some_progress
;
2572 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2573 bool deferred_compaction
= false;
2574 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2577 * In the slowpath, we sanity check order to avoid ever trying to
2578 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2579 * be using allocators in order of preference for an area that is
2582 if (order
>= MAX_ORDER
) {
2583 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2588 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2589 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2590 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2591 * using a larger set of nodes after it has established that the
2592 * allowed per node queues are empty and that nodes are
2595 if (IS_ENABLED(CONFIG_NUMA
) &&
2596 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2600 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2601 wake_all_kswapds(order
, zonelist
, high_zoneidx
,
2602 preferred_zone
, nodemask
);
2605 * OK, we're below the kswapd watermark and have kicked background
2606 * reclaim. Now things get more complex, so set up alloc_flags according
2607 * to how we want to proceed.
2609 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2612 * Find the true preferred zone if the allocation is unconstrained by
2615 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
) {
2616 struct zoneref
*preferred_zoneref
;
2617 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2618 NULL
, &preferred_zone
);
2619 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2623 /* This is the last chance, in general, before the goto nopage. */
2624 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2625 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2626 preferred_zone
, classzone_idx
, migratetype
);
2630 /* Allocate without watermarks if the context allows */
2631 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2633 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2634 * the allocation is high priority and these type of
2635 * allocations are system rather than user orientated
2637 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2639 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2640 zonelist
, high_zoneidx
, nodemask
,
2641 preferred_zone
, classzone_idx
, migratetype
);
2647 /* Atomic allocations - we can't balance anything */
2650 * All existing users of the deprecated __GFP_NOFAIL are
2651 * blockable, so warn of any new users that actually allow this
2652 * type of allocation to fail.
2654 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2658 /* Avoid recursion of direct reclaim */
2659 if (current
->flags
& PF_MEMALLOC
)
2662 /* Avoid allocations with no watermarks from looping endlessly */
2663 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2667 * Try direct compaction. The first pass is asynchronous. Subsequent
2668 * attempts after direct reclaim are synchronous
2670 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2671 high_zoneidx
, nodemask
, alloc_flags
,
2673 classzone_idx
, migratetype
,
2674 migration_mode
, &contended_compaction
,
2675 &deferred_compaction
);
2679 /* Checks for THP-specific high-order allocations */
2680 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2682 * If compaction is deferred for high-order allocations, it is
2683 * because sync compaction recently failed. If this is the case
2684 * and the caller requested a THP allocation, we do not want
2685 * to heavily disrupt the system, so we fail the allocation
2686 * instead of entering direct reclaim.
2688 if (deferred_compaction
)
2692 * In all zones where compaction was attempted (and not
2693 * deferred or skipped), lock contention has been detected.
2694 * For THP allocation we do not want to disrupt the others
2695 * so we fallback to base pages instead.
2697 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2701 * If compaction was aborted due to need_resched(), we do not
2702 * want to further increase allocation latency, unless it is
2703 * khugepaged trying to collapse.
2705 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2706 && !(current
->flags
& PF_KTHREAD
))
2711 * It can become very expensive to allocate transparent hugepages at
2712 * fault, so use asynchronous memory compaction for THP unless it is
2713 * khugepaged trying to collapse.
2715 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2716 (current
->flags
& PF_KTHREAD
))
2717 migration_mode
= MIGRATE_SYNC_LIGHT
;
2719 /* Try direct reclaim and then allocating */
2720 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2721 zonelist
, high_zoneidx
,
2723 alloc_flags
, preferred_zone
,
2724 classzone_idx
, migratetype
,
2725 &did_some_progress
);
2730 * If we failed to make any progress reclaiming, then we are
2731 * running out of options and have to consider going OOM
2733 if (!did_some_progress
) {
2734 if (oom_gfp_allowed(gfp_mask
)) {
2735 if (oom_killer_disabled
)
2737 /* Coredumps can quickly deplete all memory reserves */
2738 if ((current
->flags
& PF_DUMPCORE
) &&
2739 !(gfp_mask
& __GFP_NOFAIL
))
2741 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2742 zonelist
, high_zoneidx
,
2743 nodemask
, preferred_zone
,
2744 classzone_idx
, migratetype
);
2748 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2750 * The oom killer is not called for high-order
2751 * allocations that may fail, so if no progress
2752 * is being made, there are no other options and
2753 * retrying is unlikely to help.
2755 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2758 * The oom killer is not called for lowmem
2759 * allocations to prevent needlessly killing
2762 if (high_zoneidx
< ZONE_NORMAL
)
2770 /* Check if we should retry the allocation */
2771 pages_reclaimed
+= did_some_progress
;
2772 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2774 /* Wait for some write requests to complete then retry */
2775 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2779 * High-order allocations do not necessarily loop after
2780 * direct reclaim and reclaim/compaction depends on compaction
2781 * being called after reclaim so call directly if necessary
2783 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2784 high_zoneidx
, nodemask
, alloc_flags
,
2786 classzone_idx
, migratetype
,
2787 migration_mode
, &contended_compaction
,
2788 &deferred_compaction
);
2794 warn_alloc_failed(gfp_mask
, order
, NULL
);
2797 if (kmemcheck_enabled
)
2798 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2804 * This is the 'heart' of the zoned buddy allocator.
2807 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2808 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2810 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2811 struct zone
*preferred_zone
;
2812 struct zoneref
*preferred_zoneref
;
2813 struct page
*page
= NULL
;
2814 int migratetype
= gfpflags_to_migratetype(gfp_mask
);
2815 unsigned int cpuset_mems_cookie
;
2816 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2819 gfp_mask
&= gfp_allowed_mask
;
2821 lockdep_trace_alloc(gfp_mask
);
2823 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2825 if (should_fail_alloc_page(gfp_mask
, order
))
2829 * Check the zones suitable for the gfp_mask contain at least one
2830 * valid zone. It's possible to have an empty zonelist as a result
2831 * of GFP_THISNODE and a memoryless node
2833 if (unlikely(!zonelist
->_zonerefs
->zone
))
2836 if (IS_ENABLED(CONFIG_CMA
) && migratetype
== MIGRATE_MOVABLE
)
2837 alloc_flags
|= ALLOC_CMA
;
2840 cpuset_mems_cookie
= read_mems_allowed_begin();
2842 /* The preferred zone is used for statistics later */
2843 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2844 nodemask
? : &cpuset_current_mems_allowed
,
2846 if (!preferred_zone
)
2848 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2850 /* First allocation attempt */
2851 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2852 zonelist
, high_zoneidx
, alloc_flags
,
2853 preferred_zone
, classzone_idx
, migratetype
);
2854 if (unlikely(!page
)) {
2856 * Runtime PM, block IO and its error handling path
2857 * can deadlock because I/O on the device might not
2860 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2861 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2862 zonelist
, high_zoneidx
, nodemask
,
2863 preferred_zone
, classzone_idx
, migratetype
);
2866 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2870 * When updating a task's mems_allowed, it is possible to race with
2871 * parallel threads in such a way that an allocation can fail while
2872 * the mask is being updated. If a page allocation is about to fail,
2873 * check if the cpuset changed during allocation and if so, retry.
2875 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2880 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2883 * Common helper functions.
2885 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2890 * __get_free_pages() returns a 32-bit address, which cannot represent
2893 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2895 page
= alloc_pages(gfp_mask
, order
);
2898 return (unsigned long) page_address(page
);
2900 EXPORT_SYMBOL(__get_free_pages
);
2902 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2904 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2906 EXPORT_SYMBOL(get_zeroed_page
);
2908 void __free_pages(struct page
*page
, unsigned int order
)
2910 if (put_page_testzero(page
)) {
2912 free_hot_cold_page(page
, false);
2914 __free_pages_ok(page
, order
);
2918 EXPORT_SYMBOL(__free_pages
);
2920 void free_pages(unsigned long addr
, unsigned int order
)
2923 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2924 __free_pages(virt_to_page((void *)addr
), order
);
2928 EXPORT_SYMBOL(free_pages
);
2931 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2932 * of the current memory cgroup.
2934 * It should be used when the caller would like to use kmalloc, but since the
2935 * allocation is large, it has to fall back to the page allocator.
2937 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2940 struct mem_cgroup
*memcg
= NULL
;
2942 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2944 page
= alloc_pages(gfp_mask
, order
);
2945 memcg_kmem_commit_charge(page
, memcg
, order
);
2949 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2952 struct mem_cgroup
*memcg
= NULL
;
2954 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2956 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2957 memcg_kmem_commit_charge(page
, memcg
, order
);
2962 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2965 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2967 memcg_kmem_uncharge_pages(page
, order
);
2968 __free_pages(page
, order
);
2971 void free_kmem_pages(unsigned long addr
, unsigned int order
)
2974 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2975 __free_kmem_pages(virt_to_page((void *)addr
), order
);
2979 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2982 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2983 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2985 split_page(virt_to_page((void *)addr
), order
);
2986 while (used
< alloc_end
) {
2991 return (void *)addr
;
2995 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2996 * @size: the number of bytes to allocate
2997 * @gfp_mask: GFP flags for the allocation
2999 * This function is similar to alloc_pages(), except that it allocates the
3000 * minimum number of pages to satisfy the request. alloc_pages() can only
3001 * allocate memory in power-of-two pages.
3003 * This function is also limited by MAX_ORDER.
3005 * Memory allocated by this function must be released by free_pages_exact().
3007 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3009 unsigned int order
= get_order(size
);
3012 addr
= __get_free_pages(gfp_mask
, order
);
3013 return make_alloc_exact(addr
, order
, size
);
3015 EXPORT_SYMBOL(alloc_pages_exact
);
3018 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3020 * @nid: the preferred node ID where memory should be allocated
3021 * @size: the number of bytes to allocate
3022 * @gfp_mask: GFP flags for the allocation
3024 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3026 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3029 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3031 unsigned order
= get_order(size
);
3032 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3035 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3039 * free_pages_exact - release memory allocated via alloc_pages_exact()
3040 * @virt: the value returned by alloc_pages_exact.
3041 * @size: size of allocation, same value as passed to alloc_pages_exact().
3043 * Release the memory allocated by a previous call to alloc_pages_exact.
3045 void free_pages_exact(void *virt
, size_t size
)
3047 unsigned long addr
= (unsigned long)virt
;
3048 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3050 while (addr
< end
) {
3055 EXPORT_SYMBOL(free_pages_exact
);
3058 * nr_free_zone_pages - count number of pages beyond high watermark
3059 * @offset: The zone index of the highest zone
3061 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3062 * high watermark within all zones at or below a given zone index. For each
3063 * zone, the number of pages is calculated as:
3064 * managed_pages - high_pages
3066 static unsigned long nr_free_zone_pages(int offset
)
3071 /* Just pick one node, since fallback list is circular */
3072 unsigned long sum
= 0;
3074 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3076 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3077 unsigned long size
= zone
->managed_pages
;
3078 unsigned long high
= high_wmark_pages(zone
);
3087 * nr_free_buffer_pages - count number of pages beyond high watermark
3089 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3090 * watermark within ZONE_DMA and ZONE_NORMAL.
3092 unsigned long nr_free_buffer_pages(void)
3094 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3096 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3099 * nr_free_pagecache_pages - count number of pages beyond high watermark
3101 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3102 * high watermark within all zones.
3104 unsigned long nr_free_pagecache_pages(void)
3106 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3109 static inline void show_node(struct zone
*zone
)
3111 if (IS_ENABLED(CONFIG_NUMA
))
3112 printk("Node %d ", zone_to_nid(zone
));
3115 void si_meminfo(struct sysinfo
*val
)
3117 val
->totalram
= totalram_pages
;
3118 val
->sharedram
= global_page_state(NR_SHMEM
);
3119 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3120 val
->bufferram
= nr_blockdev_pages();
3121 val
->totalhigh
= totalhigh_pages
;
3122 val
->freehigh
= nr_free_highpages();
3123 val
->mem_unit
= PAGE_SIZE
;
3126 EXPORT_SYMBOL(si_meminfo
);
3129 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3131 int zone_type
; /* needs to be signed */
3132 unsigned long managed_pages
= 0;
3133 pg_data_t
*pgdat
= NODE_DATA(nid
);
3135 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3136 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3137 val
->totalram
= managed_pages
;
3138 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3139 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3140 #ifdef CONFIG_HIGHMEM
3141 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3142 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3148 val
->mem_unit
= PAGE_SIZE
;
3153 * Determine whether the node should be displayed or not, depending on whether
3154 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3156 bool skip_free_areas_node(unsigned int flags
, int nid
)
3159 unsigned int cpuset_mems_cookie
;
3161 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3165 cpuset_mems_cookie
= read_mems_allowed_begin();
3166 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3167 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3172 #define K(x) ((x) << (PAGE_SHIFT-10))
3174 static void show_migration_types(unsigned char type
)
3176 static const char types
[MIGRATE_TYPES
] = {
3177 [MIGRATE_UNMOVABLE
] = 'U',
3178 [MIGRATE_RECLAIMABLE
] = 'E',
3179 [MIGRATE_MOVABLE
] = 'M',
3180 [MIGRATE_RESERVE
] = 'R',
3182 [MIGRATE_CMA
] = 'C',
3184 #ifdef CONFIG_MEMORY_ISOLATION
3185 [MIGRATE_ISOLATE
] = 'I',
3188 char tmp
[MIGRATE_TYPES
+ 1];
3192 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3193 if (type
& (1 << i
))
3198 printk("(%s) ", tmp
);
3202 * Show free area list (used inside shift_scroll-lock stuff)
3203 * We also calculate the percentage fragmentation. We do this by counting the
3204 * memory on each free list with the exception of the first item on the list.
3205 * Suppresses nodes that are not allowed by current's cpuset if
3206 * SHOW_MEM_FILTER_NODES is passed.
3208 void show_free_areas(unsigned int filter
)
3213 for_each_populated_zone(zone
) {
3214 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3217 printk("%s per-cpu:\n", zone
->name
);
3219 for_each_online_cpu(cpu
) {
3220 struct per_cpu_pageset
*pageset
;
3222 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3224 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3225 cpu
, pageset
->pcp
.high
,
3226 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3230 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3231 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3233 " dirty:%lu writeback:%lu unstable:%lu\n"
3234 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3235 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3237 global_page_state(NR_ACTIVE_ANON
),
3238 global_page_state(NR_INACTIVE_ANON
),
3239 global_page_state(NR_ISOLATED_ANON
),
3240 global_page_state(NR_ACTIVE_FILE
),
3241 global_page_state(NR_INACTIVE_FILE
),
3242 global_page_state(NR_ISOLATED_FILE
),
3243 global_page_state(NR_UNEVICTABLE
),
3244 global_page_state(NR_FILE_DIRTY
),
3245 global_page_state(NR_WRITEBACK
),
3246 global_page_state(NR_UNSTABLE_NFS
),
3247 global_page_state(NR_FREE_PAGES
),
3248 global_page_state(NR_SLAB_RECLAIMABLE
),
3249 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3250 global_page_state(NR_FILE_MAPPED
),
3251 global_page_state(NR_SHMEM
),
3252 global_page_state(NR_PAGETABLE
),
3253 global_page_state(NR_BOUNCE
),
3254 global_page_state(NR_FREE_CMA_PAGES
));
3256 for_each_populated_zone(zone
) {
3259 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3267 " active_anon:%lukB"
3268 " inactive_anon:%lukB"
3269 " active_file:%lukB"
3270 " inactive_file:%lukB"
3271 " unevictable:%lukB"
3272 " isolated(anon):%lukB"
3273 " isolated(file):%lukB"
3281 " slab_reclaimable:%lukB"
3282 " slab_unreclaimable:%lukB"
3283 " kernel_stack:%lukB"
3288 " writeback_tmp:%lukB"
3289 " pages_scanned:%lu"
3290 " all_unreclaimable? %s"
3293 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3294 K(min_wmark_pages(zone
)),
3295 K(low_wmark_pages(zone
)),
3296 K(high_wmark_pages(zone
)),
3297 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3298 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3299 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3300 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3301 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3302 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3303 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3304 K(zone
->present_pages
),
3305 K(zone
->managed_pages
),
3306 K(zone_page_state(zone
, NR_MLOCK
)),
3307 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3308 K(zone_page_state(zone
, NR_WRITEBACK
)),
3309 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3310 K(zone_page_state(zone
, NR_SHMEM
)),
3311 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3312 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3313 zone_page_state(zone
, NR_KERNEL_STACK
) *
3315 K(zone_page_state(zone
, NR_PAGETABLE
)),
3316 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3317 K(zone_page_state(zone
, NR_BOUNCE
)),
3318 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3319 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3320 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3321 (!zone_reclaimable(zone
) ? "yes" : "no")
3323 printk("lowmem_reserve[]:");
3324 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3325 printk(" %ld", zone
->lowmem_reserve
[i
]);
3329 for_each_populated_zone(zone
) {
3330 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3331 unsigned char types
[MAX_ORDER
];
3333 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3336 printk("%s: ", zone
->name
);
3338 spin_lock_irqsave(&zone
->lock
, flags
);
3339 for (order
= 0; order
< MAX_ORDER
; order
++) {
3340 struct free_area
*area
= &zone
->free_area
[order
];
3343 nr
[order
] = area
->nr_free
;
3344 total
+= nr
[order
] << order
;
3347 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3348 if (!list_empty(&area
->free_list
[type
]))
3349 types
[order
] |= 1 << type
;
3352 spin_unlock_irqrestore(&zone
->lock
, flags
);
3353 for (order
= 0; order
< MAX_ORDER
; order
++) {
3354 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3356 show_migration_types(types
[order
]);
3358 printk("= %lukB\n", K(total
));
3361 hugetlb_show_meminfo();
3363 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3365 show_swap_cache_info();
3368 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3370 zoneref
->zone
= zone
;
3371 zoneref
->zone_idx
= zone_idx(zone
);
3375 * Builds allocation fallback zone lists.
3377 * Add all populated zones of a node to the zonelist.
3379 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3383 enum zone_type zone_type
= MAX_NR_ZONES
;
3387 zone
= pgdat
->node_zones
+ zone_type
;
3388 if (populated_zone(zone
)) {
3389 zoneref_set_zone(zone
,
3390 &zonelist
->_zonerefs
[nr_zones
++]);
3391 check_highest_zone(zone_type
);
3393 } while (zone_type
);
3401 * 0 = automatic detection of better ordering.
3402 * 1 = order by ([node] distance, -zonetype)
3403 * 2 = order by (-zonetype, [node] distance)
3405 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3406 * the same zonelist. So only NUMA can configure this param.
3408 #define ZONELIST_ORDER_DEFAULT 0
3409 #define ZONELIST_ORDER_NODE 1
3410 #define ZONELIST_ORDER_ZONE 2
3412 /* zonelist order in the kernel.
3413 * set_zonelist_order() will set this to NODE or ZONE.
3415 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3416 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3420 /* The value user specified ....changed by config */
3421 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3422 /* string for sysctl */
3423 #define NUMA_ZONELIST_ORDER_LEN 16
3424 char numa_zonelist_order
[16] = "default";
3427 * interface for configure zonelist ordering.
3428 * command line option "numa_zonelist_order"
3429 * = "[dD]efault - default, automatic configuration.
3430 * = "[nN]ode - order by node locality, then by zone within node
3431 * = "[zZ]one - order by zone, then by locality within zone
3434 static int __parse_numa_zonelist_order(char *s
)
3436 if (*s
== 'd' || *s
== 'D') {
3437 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3438 } else if (*s
== 'n' || *s
== 'N') {
3439 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3440 } else if (*s
== 'z' || *s
== 'Z') {
3441 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3444 "Ignoring invalid numa_zonelist_order value: "
3451 static __init
int setup_numa_zonelist_order(char *s
)
3458 ret
= __parse_numa_zonelist_order(s
);
3460 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3464 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3467 * sysctl handler for numa_zonelist_order
3469 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3470 void __user
*buffer
, size_t *length
,
3473 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3475 static DEFINE_MUTEX(zl_order_mutex
);
3477 mutex_lock(&zl_order_mutex
);
3479 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3483 strcpy(saved_string
, (char *)table
->data
);
3485 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3489 int oldval
= user_zonelist_order
;
3491 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3494 * bogus value. restore saved string
3496 strncpy((char *)table
->data
, saved_string
,
3497 NUMA_ZONELIST_ORDER_LEN
);
3498 user_zonelist_order
= oldval
;
3499 } else if (oldval
!= user_zonelist_order
) {
3500 mutex_lock(&zonelists_mutex
);
3501 build_all_zonelists(NULL
, NULL
);
3502 mutex_unlock(&zonelists_mutex
);
3506 mutex_unlock(&zl_order_mutex
);
3511 #define MAX_NODE_LOAD (nr_online_nodes)
3512 static int node_load
[MAX_NUMNODES
];
3515 * find_next_best_node - find the next node that should appear in a given node's fallback list
3516 * @node: node whose fallback list we're appending
3517 * @used_node_mask: nodemask_t of already used nodes
3519 * We use a number of factors to determine which is the next node that should
3520 * appear on a given node's fallback list. The node should not have appeared
3521 * already in @node's fallback list, and it should be the next closest node
3522 * according to the distance array (which contains arbitrary distance values
3523 * from each node to each node in the system), and should also prefer nodes
3524 * with no CPUs, since presumably they'll have very little allocation pressure
3525 * on them otherwise.
3526 * It returns -1 if no node is found.
3528 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3531 int min_val
= INT_MAX
;
3532 int best_node
= NUMA_NO_NODE
;
3533 const struct cpumask
*tmp
= cpumask_of_node(0);
3535 /* Use the local node if we haven't already */
3536 if (!node_isset(node
, *used_node_mask
)) {
3537 node_set(node
, *used_node_mask
);
3541 for_each_node_state(n
, N_MEMORY
) {
3543 /* Don't want a node to appear more than once */
3544 if (node_isset(n
, *used_node_mask
))
3547 /* Use the distance array to find the distance */
3548 val
= node_distance(node
, n
);
3550 /* Penalize nodes under us ("prefer the next node") */
3553 /* Give preference to headless and unused nodes */
3554 tmp
= cpumask_of_node(n
);
3555 if (!cpumask_empty(tmp
))
3556 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3558 /* Slight preference for less loaded node */
3559 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3560 val
+= node_load
[n
];
3562 if (val
< min_val
) {
3569 node_set(best_node
, *used_node_mask
);
3576 * Build zonelists ordered by node and zones within node.
3577 * This results in maximum locality--normal zone overflows into local
3578 * DMA zone, if any--but risks exhausting DMA zone.
3580 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3583 struct zonelist
*zonelist
;
3585 zonelist
= &pgdat
->node_zonelists
[0];
3586 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3588 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3589 zonelist
->_zonerefs
[j
].zone
= NULL
;
3590 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3594 * Build gfp_thisnode zonelists
3596 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3599 struct zonelist
*zonelist
;
3601 zonelist
= &pgdat
->node_zonelists
[1];
3602 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3603 zonelist
->_zonerefs
[j
].zone
= NULL
;
3604 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3608 * Build zonelists ordered by zone and nodes within zones.
3609 * This results in conserving DMA zone[s] until all Normal memory is
3610 * exhausted, but results in overflowing to remote node while memory
3611 * may still exist in local DMA zone.
3613 static int node_order
[MAX_NUMNODES
];
3615 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3618 int zone_type
; /* needs to be signed */
3620 struct zonelist
*zonelist
;
3622 zonelist
= &pgdat
->node_zonelists
[0];
3624 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3625 for (j
= 0; j
< nr_nodes
; j
++) {
3626 node
= node_order
[j
];
3627 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3628 if (populated_zone(z
)) {
3630 &zonelist
->_zonerefs
[pos
++]);
3631 check_highest_zone(zone_type
);
3635 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3636 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3639 #if defined(CONFIG_64BIT)
3641 * Devices that require DMA32/DMA are relatively rare and do not justify a
3642 * penalty to every machine in case the specialised case applies. Default
3643 * to Node-ordering on 64-bit NUMA machines
3645 static int default_zonelist_order(void)
3647 return ZONELIST_ORDER_NODE
;
3651 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3652 * by the kernel. If processes running on node 0 deplete the low memory zone
3653 * then reclaim will occur more frequency increasing stalls and potentially
3654 * be easier to OOM if a large percentage of the zone is under writeback or
3655 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3656 * Hence, default to zone ordering on 32-bit.
3658 static int default_zonelist_order(void)
3660 return ZONELIST_ORDER_ZONE
;
3662 #endif /* CONFIG_64BIT */
3664 static void set_zonelist_order(void)
3666 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3667 current_zonelist_order
= default_zonelist_order();
3669 current_zonelist_order
= user_zonelist_order
;
3672 static void build_zonelists(pg_data_t
*pgdat
)
3676 nodemask_t used_mask
;
3677 int local_node
, prev_node
;
3678 struct zonelist
*zonelist
;
3679 int order
= current_zonelist_order
;
3681 /* initialize zonelists */
3682 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3683 zonelist
= pgdat
->node_zonelists
+ i
;
3684 zonelist
->_zonerefs
[0].zone
= NULL
;
3685 zonelist
->_zonerefs
[0].zone_idx
= 0;
3688 /* NUMA-aware ordering of nodes */
3689 local_node
= pgdat
->node_id
;
3690 load
= nr_online_nodes
;
3691 prev_node
= local_node
;
3692 nodes_clear(used_mask
);
3694 memset(node_order
, 0, sizeof(node_order
));
3697 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3699 * We don't want to pressure a particular node.
3700 * So adding penalty to the first node in same
3701 * distance group to make it round-robin.
3703 if (node_distance(local_node
, node
) !=
3704 node_distance(local_node
, prev_node
))
3705 node_load
[node
] = load
;
3709 if (order
== ZONELIST_ORDER_NODE
)
3710 build_zonelists_in_node_order(pgdat
, node
);
3712 node_order
[j
++] = node
; /* remember order */
3715 if (order
== ZONELIST_ORDER_ZONE
) {
3716 /* calculate node order -- i.e., DMA last! */
3717 build_zonelists_in_zone_order(pgdat
, j
);
3720 build_thisnode_zonelists(pgdat
);
3723 /* Construct the zonelist performance cache - see further mmzone.h */
3724 static void build_zonelist_cache(pg_data_t
*pgdat
)
3726 struct zonelist
*zonelist
;
3727 struct zonelist_cache
*zlc
;
3730 zonelist
= &pgdat
->node_zonelists
[0];
3731 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3732 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3733 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3734 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3737 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3739 * Return node id of node used for "local" allocations.
3740 * I.e., first node id of first zone in arg node's generic zonelist.
3741 * Used for initializing percpu 'numa_mem', which is used primarily
3742 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3744 int local_memory_node(int node
)
3748 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3749 gfp_zone(GFP_KERNEL
),
3756 #else /* CONFIG_NUMA */
3758 static void set_zonelist_order(void)
3760 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3763 static void build_zonelists(pg_data_t
*pgdat
)
3765 int node
, local_node
;
3767 struct zonelist
*zonelist
;
3769 local_node
= pgdat
->node_id
;
3771 zonelist
= &pgdat
->node_zonelists
[0];
3772 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3775 * Now we build the zonelist so that it contains the zones
3776 * of all the other nodes.
3777 * We don't want to pressure a particular node, so when
3778 * building the zones for node N, we make sure that the
3779 * zones coming right after the local ones are those from
3780 * node N+1 (modulo N)
3782 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3783 if (!node_online(node
))
3785 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3787 for (node
= 0; node
< local_node
; node
++) {
3788 if (!node_online(node
))
3790 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3793 zonelist
->_zonerefs
[j
].zone
= NULL
;
3794 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3797 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3798 static void build_zonelist_cache(pg_data_t
*pgdat
)
3800 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3803 #endif /* CONFIG_NUMA */
3806 * Boot pageset table. One per cpu which is going to be used for all
3807 * zones and all nodes. The parameters will be set in such a way
3808 * that an item put on a list will immediately be handed over to
3809 * the buddy list. This is safe since pageset manipulation is done
3810 * with interrupts disabled.
3812 * The boot_pagesets must be kept even after bootup is complete for
3813 * unused processors and/or zones. They do play a role for bootstrapping
3814 * hotplugged processors.
3816 * zoneinfo_show() and maybe other functions do
3817 * not check if the processor is online before following the pageset pointer.
3818 * Other parts of the kernel may not check if the zone is available.
3820 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3821 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3822 static void setup_zone_pageset(struct zone
*zone
);
3825 * Global mutex to protect against size modification of zonelists
3826 * as well as to serialize pageset setup for the new populated zone.
3828 DEFINE_MUTEX(zonelists_mutex
);
3830 /* return values int ....just for stop_machine() */
3831 static int __build_all_zonelists(void *data
)
3835 pg_data_t
*self
= data
;
3838 memset(node_load
, 0, sizeof(node_load
));
3841 if (self
&& !node_online(self
->node_id
)) {
3842 build_zonelists(self
);
3843 build_zonelist_cache(self
);
3846 for_each_online_node(nid
) {
3847 pg_data_t
*pgdat
= NODE_DATA(nid
);
3849 build_zonelists(pgdat
);
3850 build_zonelist_cache(pgdat
);
3854 * Initialize the boot_pagesets that are going to be used
3855 * for bootstrapping processors. The real pagesets for
3856 * each zone will be allocated later when the per cpu
3857 * allocator is available.
3859 * boot_pagesets are used also for bootstrapping offline
3860 * cpus if the system is already booted because the pagesets
3861 * are needed to initialize allocators on a specific cpu too.
3862 * F.e. the percpu allocator needs the page allocator which
3863 * needs the percpu allocator in order to allocate its pagesets
3864 * (a chicken-egg dilemma).
3866 for_each_possible_cpu(cpu
) {
3867 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3869 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3871 * We now know the "local memory node" for each node--
3872 * i.e., the node of the first zone in the generic zonelist.
3873 * Set up numa_mem percpu variable for on-line cpus. During
3874 * boot, only the boot cpu should be on-line; we'll init the
3875 * secondary cpus' numa_mem as they come on-line. During
3876 * node/memory hotplug, we'll fixup all on-line cpus.
3878 if (cpu_online(cpu
))
3879 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3887 * Called with zonelists_mutex held always
3888 * unless system_state == SYSTEM_BOOTING.
3890 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3892 set_zonelist_order();
3894 if (system_state
== SYSTEM_BOOTING
) {
3895 __build_all_zonelists(NULL
);
3896 mminit_verify_zonelist();
3897 cpuset_init_current_mems_allowed();
3899 #ifdef CONFIG_MEMORY_HOTPLUG
3901 setup_zone_pageset(zone
);
3903 /* we have to stop all cpus to guarantee there is no user
3905 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3906 /* cpuset refresh routine should be here */
3908 vm_total_pages
= nr_free_pagecache_pages();
3910 * Disable grouping by mobility if the number of pages in the
3911 * system is too low to allow the mechanism to work. It would be
3912 * more accurate, but expensive to check per-zone. This check is
3913 * made on memory-hotadd so a system can start with mobility
3914 * disabled and enable it later
3916 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3917 page_group_by_mobility_disabled
= 1;
3919 page_group_by_mobility_disabled
= 0;
3921 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3922 "Total pages: %ld\n",
3924 zonelist_order_name
[current_zonelist_order
],
3925 page_group_by_mobility_disabled
? "off" : "on",
3928 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
3933 * Helper functions to size the waitqueue hash table.
3934 * Essentially these want to choose hash table sizes sufficiently
3935 * large so that collisions trying to wait on pages are rare.
3936 * But in fact, the number of active page waitqueues on typical
3937 * systems is ridiculously low, less than 200. So this is even
3938 * conservative, even though it seems large.
3940 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3941 * waitqueues, i.e. the size of the waitq table given the number of pages.
3943 #define PAGES_PER_WAITQUEUE 256
3945 #ifndef CONFIG_MEMORY_HOTPLUG
3946 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3948 unsigned long size
= 1;
3950 pages
/= PAGES_PER_WAITQUEUE
;
3952 while (size
< pages
)
3956 * Once we have dozens or even hundreds of threads sleeping
3957 * on IO we've got bigger problems than wait queue collision.
3958 * Limit the size of the wait table to a reasonable size.
3960 size
= min(size
, 4096UL);
3962 return max(size
, 4UL);
3966 * A zone's size might be changed by hot-add, so it is not possible to determine
3967 * a suitable size for its wait_table. So we use the maximum size now.
3969 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3971 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3972 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3973 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3975 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3976 * or more by the traditional way. (See above). It equals:
3978 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3979 * ia64(16K page size) : = ( 8G + 4M)byte.
3980 * powerpc (64K page size) : = (32G +16M)byte.
3982 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3989 * This is an integer logarithm so that shifts can be used later
3990 * to extract the more random high bits from the multiplicative
3991 * hash function before the remainder is taken.
3993 static inline unsigned long wait_table_bits(unsigned long size
)
3999 * Check if a pageblock contains reserved pages
4001 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
4005 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4006 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
4013 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4014 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4015 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4016 * higher will lead to a bigger reserve which will get freed as contiguous
4017 * blocks as reclaim kicks in
4019 static void setup_zone_migrate_reserve(struct zone
*zone
)
4021 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4023 unsigned long block_migratetype
;
4028 * Get the start pfn, end pfn and the number of blocks to reserve
4029 * We have to be careful to be aligned to pageblock_nr_pages to
4030 * make sure that we always check pfn_valid for the first page in
4033 start_pfn
= zone
->zone_start_pfn
;
4034 end_pfn
= zone_end_pfn(zone
);
4035 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4036 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4040 * Reserve blocks are generally in place to help high-order atomic
4041 * allocations that are short-lived. A min_free_kbytes value that
4042 * would result in more than 2 reserve blocks for atomic allocations
4043 * is assumed to be in place to help anti-fragmentation for the
4044 * future allocation of hugepages at runtime.
4046 reserve
= min(2, reserve
);
4047 old_reserve
= zone
->nr_migrate_reserve_block
;
4049 /* When memory hot-add, we almost always need to do nothing */
4050 if (reserve
== old_reserve
)
4052 zone
->nr_migrate_reserve_block
= reserve
;
4054 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4055 if (!pfn_valid(pfn
))
4057 page
= pfn_to_page(pfn
);
4059 /* Watch out for overlapping nodes */
4060 if (page_to_nid(page
) != zone_to_nid(zone
))
4063 block_migratetype
= get_pageblock_migratetype(page
);
4065 /* Only test what is necessary when the reserves are not met */
4068 * Blocks with reserved pages will never free, skip
4071 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4072 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4075 /* If this block is reserved, account for it */
4076 if (block_migratetype
== MIGRATE_RESERVE
) {
4081 /* Suitable for reserving if this block is movable */
4082 if (block_migratetype
== MIGRATE_MOVABLE
) {
4083 set_pageblock_migratetype(page
,
4085 move_freepages_block(zone
, page
,
4090 } else if (!old_reserve
) {
4092 * At boot time we don't need to scan the whole zone
4093 * for turning off MIGRATE_RESERVE.
4099 * If the reserve is met and this is a previous reserved block,
4102 if (block_migratetype
== MIGRATE_RESERVE
) {
4103 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4104 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4110 * Initially all pages are reserved - free ones are freed
4111 * up by free_all_bootmem() once the early boot process is
4112 * done. Non-atomic initialization, single-pass.
4114 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4115 unsigned long start_pfn
, enum memmap_context context
)
4118 unsigned long end_pfn
= start_pfn
+ size
;
4122 if (highest_memmap_pfn
< end_pfn
- 1)
4123 highest_memmap_pfn
= end_pfn
- 1;
4125 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4126 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4128 * There can be holes in boot-time mem_map[]s
4129 * handed to this function. They do not
4130 * exist on hotplugged memory.
4132 if (context
== MEMMAP_EARLY
) {
4133 if (!early_pfn_valid(pfn
))
4135 if (!early_pfn_in_nid(pfn
, nid
))
4138 page
= pfn_to_page(pfn
);
4139 set_page_links(page
, zone
, nid
, pfn
);
4140 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4141 init_page_count(page
);
4142 page_mapcount_reset(page
);
4143 page_cpupid_reset_last(page
);
4144 SetPageReserved(page
);
4146 * Mark the block movable so that blocks are reserved for
4147 * movable at startup. This will force kernel allocations
4148 * to reserve their blocks rather than leaking throughout
4149 * the address space during boot when many long-lived
4150 * kernel allocations are made. Later some blocks near
4151 * the start are marked MIGRATE_RESERVE by
4152 * setup_zone_migrate_reserve()
4154 * bitmap is created for zone's valid pfn range. but memmap
4155 * can be created for invalid pages (for alignment)
4156 * check here not to call set_pageblock_migratetype() against
4159 if ((z
->zone_start_pfn
<= pfn
)
4160 && (pfn
< zone_end_pfn(z
))
4161 && !(pfn
& (pageblock_nr_pages
- 1)))
4162 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4164 INIT_LIST_HEAD(&page
->lru
);
4165 #ifdef WANT_PAGE_VIRTUAL
4166 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4167 if (!is_highmem_idx(zone
))
4168 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4173 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4175 unsigned int order
, t
;
4176 for_each_migratetype_order(order
, t
) {
4177 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4178 zone
->free_area
[order
].nr_free
= 0;
4182 #ifndef __HAVE_ARCH_MEMMAP_INIT
4183 #define memmap_init(size, nid, zone, start_pfn) \
4184 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4187 static int zone_batchsize(struct zone
*zone
)
4193 * The per-cpu-pages pools are set to around 1000th of the
4194 * size of the zone. But no more than 1/2 of a meg.
4196 * OK, so we don't know how big the cache is. So guess.
4198 batch
= zone
->managed_pages
/ 1024;
4199 if (batch
* PAGE_SIZE
> 512 * 1024)
4200 batch
= (512 * 1024) / PAGE_SIZE
;
4201 batch
/= 4; /* We effectively *= 4 below */
4206 * Clamp the batch to a 2^n - 1 value. Having a power
4207 * of 2 value was found to be more likely to have
4208 * suboptimal cache aliasing properties in some cases.
4210 * For example if 2 tasks are alternately allocating
4211 * batches of pages, one task can end up with a lot
4212 * of pages of one half of the possible page colors
4213 * and the other with pages of the other colors.
4215 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4220 /* The deferral and batching of frees should be suppressed under NOMMU
4223 * The problem is that NOMMU needs to be able to allocate large chunks
4224 * of contiguous memory as there's no hardware page translation to
4225 * assemble apparent contiguous memory from discontiguous pages.
4227 * Queueing large contiguous runs of pages for batching, however,
4228 * causes the pages to actually be freed in smaller chunks. As there
4229 * can be a significant delay between the individual batches being
4230 * recycled, this leads to the once large chunks of space being
4231 * fragmented and becoming unavailable for high-order allocations.
4238 * pcp->high and pcp->batch values are related and dependent on one another:
4239 * ->batch must never be higher then ->high.
4240 * The following function updates them in a safe manner without read side
4243 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4244 * those fields changing asynchronously (acording the the above rule).
4246 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4247 * outside of boot time (or some other assurance that no concurrent updaters
4250 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4251 unsigned long batch
)
4253 /* start with a fail safe value for batch */
4257 /* Update high, then batch, in order */
4264 /* a companion to pageset_set_high() */
4265 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4267 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4270 static void pageset_init(struct per_cpu_pageset
*p
)
4272 struct per_cpu_pages
*pcp
;
4275 memset(p
, 0, sizeof(*p
));
4279 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4280 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4283 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4286 pageset_set_batch(p
, batch
);
4290 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4291 * to the value high for the pageset p.
4293 static void pageset_set_high(struct per_cpu_pageset
*p
,
4296 unsigned long batch
= max(1UL, high
/ 4);
4297 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4298 batch
= PAGE_SHIFT
* 8;
4300 pageset_update(&p
->pcp
, high
, batch
);
4303 static void pageset_set_high_and_batch(struct zone
*zone
,
4304 struct per_cpu_pageset
*pcp
)
4306 if (percpu_pagelist_fraction
)
4307 pageset_set_high(pcp
,
4308 (zone
->managed_pages
/
4309 percpu_pagelist_fraction
));
4311 pageset_set_batch(pcp
, zone_batchsize(zone
));
4314 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4316 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4319 pageset_set_high_and_batch(zone
, pcp
);
4322 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4325 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4326 for_each_possible_cpu(cpu
)
4327 zone_pageset_init(zone
, cpu
);
4331 * Allocate per cpu pagesets and initialize them.
4332 * Before this call only boot pagesets were available.
4334 void __init
setup_per_cpu_pageset(void)
4338 for_each_populated_zone(zone
)
4339 setup_zone_pageset(zone
);
4342 static noinline __init_refok
4343 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4349 * The per-page waitqueue mechanism uses hashed waitqueues
4352 zone
->wait_table_hash_nr_entries
=
4353 wait_table_hash_nr_entries(zone_size_pages
);
4354 zone
->wait_table_bits
=
4355 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4356 alloc_size
= zone
->wait_table_hash_nr_entries
4357 * sizeof(wait_queue_head_t
);
4359 if (!slab_is_available()) {
4360 zone
->wait_table
= (wait_queue_head_t
*)
4361 memblock_virt_alloc_node_nopanic(
4362 alloc_size
, zone
->zone_pgdat
->node_id
);
4365 * This case means that a zone whose size was 0 gets new memory
4366 * via memory hot-add.
4367 * But it may be the case that a new node was hot-added. In
4368 * this case vmalloc() will not be able to use this new node's
4369 * memory - this wait_table must be initialized to use this new
4370 * node itself as well.
4371 * To use this new node's memory, further consideration will be
4374 zone
->wait_table
= vmalloc(alloc_size
);
4376 if (!zone
->wait_table
)
4379 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4380 init_waitqueue_head(zone
->wait_table
+ i
);
4385 static __meminit
void zone_pcp_init(struct zone
*zone
)
4388 * per cpu subsystem is not up at this point. The following code
4389 * relies on the ability of the linker to provide the
4390 * offset of a (static) per cpu variable into the per cpu area.
4392 zone
->pageset
= &boot_pageset
;
4394 if (populated_zone(zone
))
4395 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4396 zone
->name
, zone
->present_pages
,
4397 zone_batchsize(zone
));
4400 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4401 unsigned long zone_start_pfn
,
4403 enum memmap_context context
)
4405 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4407 ret
= zone_wait_table_init(zone
, size
);
4410 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4412 zone
->zone_start_pfn
= zone_start_pfn
;
4414 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4415 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4417 (unsigned long)zone_idx(zone
),
4418 zone_start_pfn
, (zone_start_pfn
+ size
));
4420 zone_init_free_lists(zone
);
4425 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4426 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4428 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4430 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4432 unsigned long start_pfn
, end_pfn
;
4435 * NOTE: The following SMP-unsafe globals are only used early in boot
4436 * when the kernel is running single-threaded.
4438 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4439 static int __meminitdata last_nid
;
4441 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4444 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4446 last_start_pfn
= start_pfn
;
4447 last_end_pfn
= end_pfn
;
4453 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4455 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4459 nid
= __early_pfn_to_nid(pfn
);
4462 /* just returns 0 */
4466 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4467 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4471 nid
= __early_pfn_to_nid(pfn
);
4472 if (nid
>= 0 && nid
!= node
)
4479 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4480 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4481 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4483 * If an architecture guarantees that all ranges registered contain no holes
4484 * and may be freed, this this function may be used instead of calling
4485 * memblock_free_early_nid() manually.
4487 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4489 unsigned long start_pfn
, end_pfn
;
4492 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4493 start_pfn
= min(start_pfn
, max_low_pfn
);
4494 end_pfn
= min(end_pfn
, max_low_pfn
);
4496 if (start_pfn
< end_pfn
)
4497 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4498 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4504 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4505 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4507 * If an architecture guarantees that all ranges registered contain no holes and may
4508 * be freed, this function may be used instead of calling memory_present() manually.
4510 void __init
sparse_memory_present_with_active_regions(int nid
)
4512 unsigned long start_pfn
, end_pfn
;
4515 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4516 memory_present(this_nid
, start_pfn
, end_pfn
);
4520 * get_pfn_range_for_nid - Return the start and end page frames for a node
4521 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4522 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4523 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4525 * It returns the start and end page frame of a node based on information
4526 * provided by memblock_set_node(). If called for a node
4527 * with no available memory, a warning is printed and the start and end
4530 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4531 unsigned long *start_pfn
, unsigned long *end_pfn
)
4533 unsigned long this_start_pfn
, this_end_pfn
;
4539 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4540 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4541 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4544 if (*start_pfn
== -1UL)
4549 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4550 * assumption is made that zones within a node are ordered in monotonic
4551 * increasing memory addresses so that the "highest" populated zone is used
4553 static void __init
find_usable_zone_for_movable(void)
4556 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4557 if (zone_index
== ZONE_MOVABLE
)
4560 if (arch_zone_highest_possible_pfn
[zone_index
] >
4561 arch_zone_lowest_possible_pfn
[zone_index
])
4565 VM_BUG_ON(zone_index
== -1);
4566 movable_zone
= zone_index
;
4570 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4571 * because it is sized independent of architecture. Unlike the other zones,
4572 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4573 * in each node depending on the size of each node and how evenly kernelcore
4574 * is distributed. This helper function adjusts the zone ranges
4575 * provided by the architecture for a given node by using the end of the
4576 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4577 * zones within a node are in order of monotonic increases memory addresses
4579 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4580 unsigned long zone_type
,
4581 unsigned long node_start_pfn
,
4582 unsigned long node_end_pfn
,
4583 unsigned long *zone_start_pfn
,
4584 unsigned long *zone_end_pfn
)
4586 /* Only adjust if ZONE_MOVABLE is on this node */
4587 if (zone_movable_pfn
[nid
]) {
4588 /* Size ZONE_MOVABLE */
4589 if (zone_type
== ZONE_MOVABLE
) {
4590 *zone_start_pfn
= zone_movable_pfn
[nid
];
4591 *zone_end_pfn
= min(node_end_pfn
,
4592 arch_zone_highest_possible_pfn
[movable_zone
]);
4594 /* Adjust for ZONE_MOVABLE starting within this range */
4595 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4596 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4597 *zone_end_pfn
= zone_movable_pfn
[nid
];
4599 /* Check if this whole range is within ZONE_MOVABLE */
4600 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4601 *zone_start_pfn
= *zone_end_pfn
;
4606 * Return the number of pages a zone spans in a node, including holes
4607 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4609 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4610 unsigned long zone_type
,
4611 unsigned long node_start_pfn
,
4612 unsigned long node_end_pfn
,
4613 unsigned long *ignored
)
4615 unsigned long zone_start_pfn
, zone_end_pfn
;
4617 /* Get the start and end of the zone */
4618 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4619 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4620 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4621 node_start_pfn
, node_end_pfn
,
4622 &zone_start_pfn
, &zone_end_pfn
);
4624 /* Check that this node has pages within the zone's required range */
4625 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4628 /* Move the zone boundaries inside the node if necessary */
4629 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4630 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4632 /* Return the spanned pages */
4633 return zone_end_pfn
- zone_start_pfn
;
4637 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4638 * then all holes in the requested range will be accounted for.
4640 unsigned long __meminit
__absent_pages_in_range(int nid
,
4641 unsigned long range_start_pfn
,
4642 unsigned long range_end_pfn
)
4644 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4645 unsigned long start_pfn
, end_pfn
;
4648 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4649 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4650 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4651 nr_absent
-= end_pfn
- start_pfn
;
4657 * absent_pages_in_range - Return number of page frames in holes within a range
4658 * @start_pfn: The start PFN to start searching for holes
4659 * @end_pfn: The end PFN to stop searching for holes
4661 * It returns the number of pages frames in memory holes within a range.
4663 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4664 unsigned long end_pfn
)
4666 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4669 /* Return the number of page frames in holes in a zone on a node */
4670 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4671 unsigned long zone_type
,
4672 unsigned long node_start_pfn
,
4673 unsigned long node_end_pfn
,
4674 unsigned long *ignored
)
4676 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4677 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4678 unsigned long zone_start_pfn
, zone_end_pfn
;
4680 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4681 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4683 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4684 node_start_pfn
, node_end_pfn
,
4685 &zone_start_pfn
, &zone_end_pfn
);
4686 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4689 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4690 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4691 unsigned long zone_type
,
4692 unsigned long node_start_pfn
,
4693 unsigned long node_end_pfn
,
4694 unsigned long *zones_size
)
4696 return zones_size
[zone_type
];
4699 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4700 unsigned long zone_type
,
4701 unsigned long node_start_pfn
,
4702 unsigned long node_end_pfn
,
4703 unsigned long *zholes_size
)
4708 return zholes_size
[zone_type
];
4711 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4713 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4714 unsigned long node_start_pfn
,
4715 unsigned long node_end_pfn
,
4716 unsigned long *zones_size
,
4717 unsigned long *zholes_size
)
4719 unsigned long realtotalpages
, totalpages
= 0;
4722 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4723 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4727 pgdat
->node_spanned_pages
= totalpages
;
4729 realtotalpages
= totalpages
;
4730 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4732 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4733 node_start_pfn
, node_end_pfn
,
4735 pgdat
->node_present_pages
= realtotalpages
;
4736 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4740 #ifndef CONFIG_SPARSEMEM
4742 * Calculate the size of the zone->blockflags rounded to an unsigned long
4743 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4744 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4745 * round what is now in bits to nearest long in bits, then return it in
4748 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4750 unsigned long usemapsize
;
4752 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4753 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4754 usemapsize
= usemapsize
>> pageblock_order
;
4755 usemapsize
*= NR_PAGEBLOCK_BITS
;
4756 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4758 return usemapsize
/ 8;
4761 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4763 unsigned long zone_start_pfn
,
4764 unsigned long zonesize
)
4766 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4767 zone
->pageblock_flags
= NULL
;
4769 zone
->pageblock_flags
=
4770 memblock_virt_alloc_node_nopanic(usemapsize
,
4774 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4775 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4776 #endif /* CONFIG_SPARSEMEM */
4778 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4780 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4781 void __paginginit
set_pageblock_order(void)
4785 /* Check that pageblock_nr_pages has not already been setup */
4786 if (pageblock_order
)
4789 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4790 order
= HUGETLB_PAGE_ORDER
;
4792 order
= MAX_ORDER
- 1;
4795 * Assume the largest contiguous order of interest is a huge page.
4796 * This value may be variable depending on boot parameters on IA64 and
4799 pageblock_order
= order
;
4801 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4804 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4805 * is unused as pageblock_order is set at compile-time. See
4806 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4809 void __paginginit
set_pageblock_order(void)
4813 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4815 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4816 unsigned long present_pages
)
4818 unsigned long pages
= spanned_pages
;
4821 * Provide a more accurate estimation if there are holes within
4822 * the zone and SPARSEMEM is in use. If there are holes within the
4823 * zone, each populated memory region may cost us one or two extra
4824 * memmap pages due to alignment because memmap pages for each
4825 * populated regions may not naturally algined on page boundary.
4826 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4828 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4829 IS_ENABLED(CONFIG_SPARSEMEM
))
4830 pages
= present_pages
;
4832 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4836 * Set up the zone data structures:
4837 * - mark all pages reserved
4838 * - mark all memory queues empty
4839 * - clear the memory bitmaps
4841 * NOTE: pgdat should get zeroed by caller.
4843 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4844 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4845 unsigned long *zones_size
, unsigned long *zholes_size
)
4848 int nid
= pgdat
->node_id
;
4849 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4852 pgdat_resize_init(pgdat
);
4853 #ifdef CONFIG_NUMA_BALANCING
4854 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4855 pgdat
->numabalancing_migrate_nr_pages
= 0;
4856 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4858 init_waitqueue_head(&pgdat
->kswapd_wait
);
4859 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4860 pgdat_page_ext_init(pgdat
);
4862 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4863 struct zone
*zone
= pgdat
->node_zones
+ j
;
4864 unsigned long size
, realsize
, freesize
, memmap_pages
;
4866 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4867 node_end_pfn
, zones_size
);
4868 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4874 * Adjust freesize so that it accounts for how much memory
4875 * is used by this zone for memmap. This affects the watermark
4876 * and per-cpu initialisations
4878 memmap_pages
= calc_memmap_size(size
, realsize
);
4879 if (freesize
>= memmap_pages
) {
4880 freesize
-= memmap_pages
;
4883 " %s zone: %lu pages used for memmap\n",
4884 zone_names
[j
], memmap_pages
);
4887 " %s zone: %lu pages exceeds freesize %lu\n",
4888 zone_names
[j
], memmap_pages
, freesize
);
4890 /* Account for reserved pages */
4891 if (j
== 0 && freesize
> dma_reserve
) {
4892 freesize
-= dma_reserve
;
4893 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4894 zone_names
[0], dma_reserve
);
4897 if (!is_highmem_idx(j
))
4898 nr_kernel_pages
+= freesize
;
4899 /* Charge for highmem memmap if there are enough kernel pages */
4900 else if (nr_kernel_pages
> memmap_pages
* 2)
4901 nr_kernel_pages
-= memmap_pages
;
4902 nr_all_pages
+= freesize
;
4904 zone
->spanned_pages
= size
;
4905 zone
->present_pages
= realsize
;
4907 * Set an approximate value for lowmem here, it will be adjusted
4908 * when the bootmem allocator frees pages into the buddy system.
4909 * And all highmem pages will be managed by the buddy system.
4911 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4914 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4916 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4918 zone
->name
= zone_names
[j
];
4919 spin_lock_init(&zone
->lock
);
4920 spin_lock_init(&zone
->lru_lock
);
4921 zone_seqlock_init(zone
);
4922 zone
->zone_pgdat
= pgdat
;
4923 zone_pcp_init(zone
);
4925 /* For bootup, initialized properly in watermark setup */
4926 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4928 lruvec_init(&zone
->lruvec
);
4932 set_pageblock_order();
4933 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4934 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4935 size
, MEMMAP_EARLY
);
4937 memmap_init(size
, nid
, j
, zone_start_pfn
);
4938 zone_start_pfn
+= size
;
4942 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4944 /* Skip empty nodes */
4945 if (!pgdat
->node_spanned_pages
)
4948 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4949 /* ia64 gets its own node_mem_map, before this, without bootmem */
4950 if (!pgdat
->node_mem_map
) {
4951 unsigned long size
, start
, end
;
4955 * The zone's endpoints aren't required to be MAX_ORDER
4956 * aligned but the node_mem_map endpoints must be in order
4957 * for the buddy allocator to function correctly.
4959 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4960 end
= pgdat_end_pfn(pgdat
);
4961 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4962 size
= (end
- start
) * sizeof(struct page
);
4963 map
= alloc_remap(pgdat
->node_id
, size
);
4965 map
= memblock_virt_alloc_node_nopanic(size
,
4967 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4969 #ifndef CONFIG_NEED_MULTIPLE_NODES
4971 * With no DISCONTIG, the global mem_map is just set as node 0's
4973 if (pgdat
== NODE_DATA(0)) {
4974 mem_map
= NODE_DATA(0)->node_mem_map
;
4975 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4976 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4977 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4978 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4981 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4984 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4985 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4987 pg_data_t
*pgdat
= NODE_DATA(nid
);
4988 unsigned long start_pfn
= 0;
4989 unsigned long end_pfn
= 0;
4991 /* pg_data_t should be reset to zero when it's allocated */
4992 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4994 pgdat
->node_id
= nid
;
4995 pgdat
->node_start_pfn
= node_start_pfn
;
4996 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4997 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4998 printk(KERN_INFO
"Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid
,
4999 (u64
) start_pfn
<< PAGE_SHIFT
, (u64
) (end_pfn
<< PAGE_SHIFT
) - 1);
5001 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5002 zones_size
, zholes_size
);
5004 alloc_node_mem_map(pgdat
);
5005 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5006 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5007 nid
, (unsigned long)pgdat
,
5008 (unsigned long)pgdat
->node_mem_map
);
5011 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
5012 zones_size
, zholes_size
);
5015 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5017 #if MAX_NUMNODES > 1
5019 * Figure out the number of possible node ids.
5021 void __init
setup_nr_node_ids(void)
5024 unsigned int highest
= 0;
5026 for_each_node_mask(node
, node_possible_map
)
5028 nr_node_ids
= highest
+ 1;
5033 * node_map_pfn_alignment - determine the maximum internode alignment
5035 * This function should be called after node map is populated and sorted.
5036 * It calculates the maximum power of two alignment which can distinguish
5039 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5040 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5041 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5042 * shifted, 1GiB is enough and this function will indicate so.
5044 * This is used to test whether pfn -> nid mapping of the chosen memory
5045 * model has fine enough granularity to avoid incorrect mapping for the
5046 * populated node map.
5048 * Returns the determined alignment in pfn's. 0 if there is no alignment
5049 * requirement (single node).
5051 unsigned long __init
node_map_pfn_alignment(void)
5053 unsigned long accl_mask
= 0, last_end
= 0;
5054 unsigned long start
, end
, mask
;
5058 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5059 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5066 * Start with a mask granular enough to pin-point to the
5067 * start pfn and tick off bits one-by-one until it becomes
5068 * too coarse to separate the current node from the last.
5070 mask
= ~((1 << __ffs(start
)) - 1);
5071 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5074 /* accumulate all internode masks */
5078 /* convert mask to number of pages */
5079 return ~accl_mask
+ 1;
5082 /* Find the lowest pfn for a node */
5083 static unsigned long __init
find_min_pfn_for_node(int nid
)
5085 unsigned long min_pfn
= ULONG_MAX
;
5086 unsigned long start_pfn
;
5089 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5090 min_pfn
= min(min_pfn
, start_pfn
);
5092 if (min_pfn
== ULONG_MAX
) {
5094 "Could not find start_pfn for node %d\n", nid
);
5102 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5104 * It returns the minimum PFN based on information provided via
5105 * memblock_set_node().
5107 unsigned long __init
find_min_pfn_with_active_regions(void)
5109 return find_min_pfn_for_node(MAX_NUMNODES
);
5113 * early_calculate_totalpages()
5114 * Sum pages in active regions for movable zone.
5115 * Populate N_MEMORY for calculating usable_nodes.
5117 static unsigned long __init
early_calculate_totalpages(void)
5119 unsigned long totalpages
= 0;
5120 unsigned long start_pfn
, end_pfn
;
5123 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5124 unsigned long pages
= end_pfn
- start_pfn
;
5126 totalpages
+= pages
;
5128 node_set_state(nid
, N_MEMORY
);
5134 * Find the PFN the Movable zone begins in each node. Kernel memory
5135 * is spread evenly between nodes as long as the nodes have enough
5136 * memory. When they don't, some nodes will have more kernelcore than
5139 static void __init
find_zone_movable_pfns_for_nodes(void)
5142 unsigned long usable_startpfn
;
5143 unsigned long kernelcore_node
, kernelcore_remaining
;
5144 /* save the state before borrow the nodemask */
5145 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5146 unsigned long totalpages
= early_calculate_totalpages();
5147 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5148 struct memblock_region
*r
;
5150 /* Need to find movable_zone earlier when movable_node is specified. */
5151 find_usable_zone_for_movable();
5154 * If movable_node is specified, ignore kernelcore and movablecore
5157 if (movable_node_is_enabled()) {
5158 for_each_memblock(memory
, r
) {
5159 if (!memblock_is_hotpluggable(r
))
5164 usable_startpfn
= PFN_DOWN(r
->base
);
5165 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5166 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5174 * If movablecore=nn[KMG] was specified, calculate what size of
5175 * kernelcore that corresponds so that memory usable for
5176 * any allocation type is evenly spread. If both kernelcore
5177 * and movablecore are specified, then the value of kernelcore
5178 * will be used for required_kernelcore if it's greater than
5179 * what movablecore would have allowed.
5181 if (required_movablecore
) {
5182 unsigned long corepages
;
5185 * Round-up so that ZONE_MOVABLE is at least as large as what
5186 * was requested by the user
5188 required_movablecore
=
5189 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5190 corepages
= totalpages
- required_movablecore
;
5192 required_kernelcore
= max(required_kernelcore
, corepages
);
5195 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5196 if (!required_kernelcore
)
5199 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5200 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5203 /* Spread kernelcore memory as evenly as possible throughout nodes */
5204 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5205 for_each_node_state(nid
, N_MEMORY
) {
5206 unsigned long start_pfn
, end_pfn
;
5209 * Recalculate kernelcore_node if the division per node
5210 * now exceeds what is necessary to satisfy the requested
5211 * amount of memory for the kernel
5213 if (required_kernelcore
< kernelcore_node
)
5214 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5217 * As the map is walked, we track how much memory is usable
5218 * by the kernel using kernelcore_remaining. When it is
5219 * 0, the rest of the node is usable by ZONE_MOVABLE
5221 kernelcore_remaining
= kernelcore_node
;
5223 /* Go through each range of PFNs within this node */
5224 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5225 unsigned long size_pages
;
5227 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5228 if (start_pfn
>= end_pfn
)
5231 /* Account for what is only usable for kernelcore */
5232 if (start_pfn
< usable_startpfn
) {
5233 unsigned long kernel_pages
;
5234 kernel_pages
= min(end_pfn
, usable_startpfn
)
5237 kernelcore_remaining
-= min(kernel_pages
,
5238 kernelcore_remaining
);
5239 required_kernelcore
-= min(kernel_pages
,
5240 required_kernelcore
);
5242 /* Continue if range is now fully accounted */
5243 if (end_pfn
<= usable_startpfn
) {
5246 * Push zone_movable_pfn to the end so
5247 * that if we have to rebalance
5248 * kernelcore across nodes, we will
5249 * not double account here
5251 zone_movable_pfn
[nid
] = end_pfn
;
5254 start_pfn
= usable_startpfn
;
5258 * The usable PFN range for ZONE_MOVABLE is from
5259 * start_pfn->end_pfn. Calculate size_pages as the
5260 * number of pages used as kernelcore
5262 size_pages
= end_pfn
- start_pfn
;
5263 if (size_pages
> kernelcore_remaining
)
5264 size_pages
= kernelcore_remaining
;
5265 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5268 * Some kernelcore has been met, update counts and
5269 * break if the kernelcore for this node has been
5272 required_kernelcore
-= min(required_kernelcore
,
5274 kernelcore_remaining
-= size_pages
;
5275 if (!kernelcore_remaining
)
5281 * If there is still required_kernelcore, we do another pass with one
5282 * less node in the count. This will push zone_movable_pfn[nid] further
5283 * along on the nodes that still have memory until kernelcore is
5287 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5291 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5292 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5293 zone_movable_pfn
[nid
] =
5294 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5297 /* restore the node_state */
5298 node_states
[N_MEMORY
] = saved_node_state
;
5301 /* Any regular or high memory on that node ? */
5302 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5304 enum zone_type zone_type
;
5306 if (N_MEMORY
== N_NORMAL_MEMORY
)
5309 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5310 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5311 if (populated_zone(zone
)) {
5312 node_set_state(nid
, N_HIGH_MEMORY
);
5313 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5314 zone_type
<= ZONE_NORMAL
)
5315 node_set_state(nid
, N_NORMAL_MEMORY
);
5322 * free_area_init_nodes - Initialise all pg_data_t and zone data
5323 * @max_zone_pfn: an array of max PFNs for each zone
5325 * This will call free_area_init_node() for each active node in the system.
5326 * Using the page ranges provided by memblock_set_node(), the size of each
5327 * zone in each node and their holes is calculated. If the maximum PFN
5328 * between two adjacent zones match, it is assumed that the zone is empty.
5329 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5330 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5331 * starts where the previous one ended. For example, ZONE_DMA32 starts
5332 * at arch_max_dma_pfn.
5334 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5336 unsigned long start_pfn
, end_pfn
;
5339 /* Record where the zone boundaries are */
5340 memset(arch_zone_lowest_possible_pfn
, 0,
5341 sizeof(arch_zone_lowest_possible_pfn
));
5342 memset(arch_zone_highest_possible_pfn
, 0,
5343 sizeof(arch_zone_highest_possible_pfn
));
5344 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5345 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5346 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5347 if (i
== ZONE_MOVABLE
)
5349 arch_zone_lowest_possible_pfn
[i
] =
5350 arch_zone_highest_possible_pfn
[i
-1];
5351 arch_zone_highest_possible_pfn
[i
] =
5352 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5354 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5355 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5357 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5358 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5359 find_zone_movable_pfns_for_nodes();
5361 /* Print out the zone ranges */
5362 pr_info("Zone ranges:\n");
5363 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5364 if (i
== ZONE_MOVABLE
)
5366 pr_info(" %-8s ", zone_names
[i
]);
5367 if (arch_zone_lowest_possible_pfn
[i
] ==
5368 arch_zone_highest_possible_pfn
[i
])
5371 pr_cont("[mem %0#10lx-%0#10lx]\n",
5372 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5373 (arch_zone_highest_possible_pfn
[i
]
5374 << PAGE_SHIFT
) - 1);
5377 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5378 pr_info("Movable zone start for each node\n");
5379 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5380 if (zone_movable_pfn
[i
])
5381 pr_info(" Node %d: %#010lx\n", i
,
5382 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5385 /* Print out the early node map */
5386 pr_info("Early memory node ranges\n");
5387 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5388 pr_info(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5389 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5391 /* Initialise every node */
5392 mminit_verify_pageflags_layout();
5393 setup_nr_node_ids();
5394 for_each_online_node(nid
) {
5395 pg_data_t
*pgdat
= NODE_DATA(nid
);
5396 free_area_init_node(nid
, NULL
,
5397 find_min_pfn_for_node(nid
), NULL
);
5399 /* Any memory on that node */
5400 if (pgdat
->node_present_pages
)
5401 node_set_state(nid
, N_MEMORY
);
5402 check_for_memory(pgdat
, nid
);
5406 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5408 unsigned long long coremem
;
5412 coremem
= memparse(p
, &p
);
5413 *core
= coremem
>> PAGE_SHIFT
;
5415 /* Paranoid check that UL is enough for the coremem value */
5416 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5422 * kernelcore=size sets the amount of memory for use for allocations that
5423 * cannot be reclaimed or migrated.
5425 static int __init
cmdline_parse_kernelcore(char *p
)
5427 return cmdline_parse_core(p
, &required_kernelcore
);
5431 * movablecore=size sets the amount of memory for use for allocations that
5432 * can be reclaimed or migrated.
5434 static int __init
cmdline_parse_movablecore(char *p
)
5436 return cmdline_parse_core(p
, &required_movablecore
);
5439 early_param("kernelcore", cmdline_parse_kernelcore
);
5440 early_param("movablecore", cmdline_parse_movablecore
);
5442 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5444 void adjust_managed_page_count(struct page
*page
, long count
)
5446 spin_lock(&managed_page_count_lock
);
5447 page_zone(page
)->managed_pages
+= count
;
5448 totalram_pages
+= count
;
5449 #ifdef CONFIG_HIGHMEM
5450 if (PageHighMem(page
))
5451 totalhigh_pages
+= count
;
5453 spin_unlock(&managed_page_count_lock
);
5455 EXPORT_SYMBOL(adjust_managed_page_count
);
5457 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5460 unsigned long pages
= 0;
5462 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5463 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5464 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5465 if ((unsigned int)poison
<= 0xFF)
5466 memset(pos
, poison
, PAGE_SIZE
);
5467 free_reserved_page(virt_to_page(pos
));
5471 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5472 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5476 EXPORT_SYMBOL(free_reserved_area
);
5478 #ifdef CONFIG_HIGHMEM
5479 void free_highmem_page(struct page
*page
)
5481 __free_reserved_page(page
);
5483 page_zone(page
)->managed_pages
++;
5489 void __init
mem_init_print_info(const char *str
)
5491 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5492 unsigned long init_code_size
, init_data_size
;
5494 physpages
= get_num_physpages();
5495 codesize
= _etext
- _stext
;
5496 datasize
= _edata
- _sdata
;
5497 rosize
= __end_rodata
- __start_rodata
;
5498 bss_size
= __bss_stop
- __bss_start
;
5499 init_data_size
= __init_end
- __init_begin
;
5500 init_code_size
= _einittext
- _sinittext
;
5503 * Detect special cases and adjust section sizes accordingly:
5504 * 1) .init.* may be embedded into .data sections
5505 * 2) .init.text.* may be out of [__init_begin, __init_end],
5506 * please refer to arch/tile/kernel/vmlinux.lds.S.
5507 * 3) .rodata.* may be embedded into .text or .data sections.
5509 #define adj_init_size(start, end, size, pos, adj) \
5511 if (start <= pos && pos < end && size > adj) \
5515 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5516 _sinittext
, init_code_size
);
5517 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5518 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5519 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5520 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5522 #undef adj_init_size
5524 pr_info("Memory: %luK/%luK available "
5525 "(%luK kernel code, %luK rwdata, %luK rodata, "
5526 "%luK init, %luK bss, %luK reserved"
5527 #ifdef CONFIG_HIGHMEM
5531 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5532 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5533 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5534 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5535 #ifdef CONFIG_HIGHMEM
5536 totalhigh_pages
<< (PAGE_SHIFT
-10),
5538 str
? ", " : "", str
? str
: "");
5542 * set_dma_reserve - set the specified number of pages reserved in the first zone
5543 * @new_dma_reserve: The number of pages to mark reserved
5545 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5546 * In the DMA zone, a significant percentage may be consumed by kernel image
5547 * and other unfreeable allocations which can skew the watermarks badly. This
5548 * function may optionally be used to account for unfreeable pages in the
5549 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5550 * smaller per-cpu batchsize.
5552 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5554 dma_reserve
= new_dma_reserve
;
5557 void __init
free_area_init(unsigned long *zones_size
)
5559 free_area_init_node(0, zones_size
,
5560 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5563 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5564 unsigned long action
, void *hcpu
)
5566 int cpu
= (unsigned long)hcpu
;
5568 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5569 lru_add_drain_cpu(cpu
);
5573 * Spill the event counters of the dead processor
5574 * into the current processors event counters.
5575 * This artificially elevates the count of the current
5578 vm_events_fold_cpu(cpu
);
5581 * Zero the differential counters of the dead processor
5582 * so that the vm statistics are consistent.
5584 * This is only okay since the processor is dead and cannot
5585 * race with what we are doing.
5587 cpu_vm_stats_fold(cpu
);
5592 void __init
page_alloc_init(void)
5594 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5598 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5599 * or min_free_kbytes changes.
5601 static void calculate_totalreserve_pages(void)
5603 struct pglist_data
*pgdat
;
5604 unsigned long reserve_pages
= 0;
5605 enum zone_type i
, j
;
5607 for_each_online_pgdat(pgdat
) {
5608 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5609 struct zone
*zone
= pgdat
->node_zones
+ i
;
5612 /* Find valid and maximum lowmem_reserve in the zone */
5613 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5614 if (zone
->lowmem_reserve
[j
] > max
)
5615 max
= zone
->lowmem_reserve
[j
];
5618 /* we treat the high watermark as reserved pages. */
5619 max
+= high_wmark_pages(zone
);
5621 if (max
> zone
->managed_pages
)
5622 max
= zone
->managed_pages
;
5623 reserve_pages
+= max
;
5625 * Lowmem reserves are not available to
5626 * GFP_HIGHUSER page cache allocations and
5627 * kswapd tries to balance zones to their high
5628 * watermark. As a result, neither should be
5629 * regarded as dirtyable memory, to prevent a
5630 * situation where reclaim has to clean pages
5631 * in order to balance the zones.
5633 zone
->dirty_balance_reserve
= max
;
5636 dirty_balance_reserve
= reserve_pages
;
5637 totalreserve_pages
= reserve_pages
;
5641 * setup_per_zone_lowmem_reserve - called whenever
5642 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5643 * has a correct pages reserved value, so an adequate number of
5644 * pages are left in the zone after a successful __alloc_pages().
5646 static void setup_per_zone_lowmem_reserve(void)
5648 struct pglist_data
*pgdat
;
5649 enum zone_type j
, idx
;
5651 for_each_online_pgdat(pgdat
) {
5652 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5653 struct zone
*zone
= pgdat
->node_zones
+ j
;
5654 unsigned long managed_pages
= zone
->managed_pages
;
5656 zone
->lowmem_reserve
[j
] = 0;
5660 struct zone
*lower_zone
;
5664 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5665 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5667 lower_zone
= pgdat
->node_zones
+ idx
;
5668 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5669 sysctl_lowmem_reserve_ratio
[idx
];
5670 managed_pages
+= lower_zone
->managed_pages
;
5675 /* update totalreserve_pages */
5676 calculate_totalreserve_pages();
5679 static void __setup_per_zone_wmarks(void)
5681 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5682 unsigned long lowmem_pages
= 0;
5684 unsigned long flags
;
5686 /* Calculate total number of !ZONE_HIGHMEM pages */
5687 for_each_zone(zone
) {
5688 if (!is_highmem(zone
))
5689 lowmem_pages
+= zone
->managed_pages
;
5692 for_each_zone(zone
) {
5695 spin_lock_irqsave(&zone
->lock
, flags
);
5696 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5697 do_div(tmp
, lowmem_pages
);
5698 if (is_highmem(zone
)) {
5700 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5701 * need highmem pages, so cap pages_min to a small
5704 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5705 * deltas controls asynch page reclaim, and so should
5706 * not be capped for highmem.
5708 unsigned long min_pages
;
5710 min_pages
= zone
->managed_pages
/ 1024;
5711 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5712 zone
->watermark
[WMARK_MIN
] = min_pages
;
5715 * If it's a lowmem zone, reserve a number of pages
5716 * proportionate to the zone's size.
5718 zone
->watermark
[WMARK_MIN
] = tmp
;
5721 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5722 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5724 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5725 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5726 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5728 setup_zone_migrate_reserve(zone
);
5729 spin_unlock_irqrestore(&zone
->lock
, flags
);
5732 /* update totalreserve_pages */
5733 calculate_totalreserve_pages();
5737 * setup_per_zone_wmarks - called when min_free_kbytes changes
5738 * or when memory is hot-{added|removed}
5740 * Ensures that the watermark[min,low,high] values for each zone are set
5741 * correctly with respect to min_free_kbytes.
5743 void setup_per_zone_wmarks(void)
5745 mutex_lock(&zonelists_mutex
);
5746 __setup_per_zone_wmarks();
5747 mutex_unlock(&zonelists_mutex
);
5751 * The inactive anon list should be small enough that the VM never has to
5752 * do too much work, but large enough that each inactive page has a chance
5753 * to be referenced again before it is swapped out.
5755 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5756 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5757 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5758 * the anonymous pages are kept on the inactive list.
5761 * memory ratio inactive anon
5762 * -------------------------------------
5771 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5773 unsigned int gb
, ratio
;
5775 /* Zone size in gigabytes */
5776 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5778 ratio
= int_sqrt(10 * gb
);
5782 zone
->inactive_ratio
= ratio
;
5785 static void __meminit
setup_per_zone_inactive_ratio(void)
5790 calculate_zone_inactive_ratio(zone
);
5794 * Initialise min_free_kbytes.
5796 * For small machines we want it small (128k min). For large machines
5797 * we want it large (64MB max). But it is not linear, because network
5798 * bandwidth does not increase linearly with machine size. We use
5800 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5801 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5817 int __meminit
init_per_zone_wmark_min(void)
5819 unsigned long lowmem_kbytes
;
5820 int new_min_free_kbytes
;
5822 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5823 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5825 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5826 min_free_kbytes
= new_min_free_kbytes
;
5827 if (min_free_kbytes
< 128)
5828 min_free_kbytes
= 128;
5829 if (min_free_kbytes
> 65536)
5830 min_free_kbytes
= 65536;
5832 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5833 new_min_free_kbytes
, user_min_free_kbytes
);
5835 setup_per_zone_wmarks();
5836 refresh_zone_stat_thresholds();
5837 setup_per_zone_lowmem_reserve();
5838 setup_per_zone_inactive_ratio();
5841 module_init(init_per_zone_wmark_min
)
5844 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5845 * that we can call two helper functions whenever min_free_kbytes
5848 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5849 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5853 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5858 user_min_free_kbytes
= min_free_kbytes
;
5859 setup_per_zone_wmarks();
5865 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5866 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5871 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5876 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5877 sysctl_min_unmapped_ratio
) / 100;
5881 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5882 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5887 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5892 zone
->min_slab_pages
= (zone
->managed_pages
*
5893 sysctl_min_slab_ratio
) / 100;
5899 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5900 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5901 * whenever sysctl_lowmem_reserve_ratio changes.
5903 * The reserve ratio obviously has absolutely no relation with the
5904 * minimum watermarks. The lowmem reserve ratio can only make sense
5905 * if in function of the boot time zone sizes.
5907 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5908 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5910 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5911 setup_per_zone_lowmem_reserve();
5916 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5917 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5918 * pagelist can have before it gets flushed back to buddy allocator.
5920 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5921 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5924 int old_percpu_pagelist_fraction
;
5927 mutex_lock(&pcp_batch_high_lock
);
5928 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5930 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5931 if (!write
|| ret
< 0)
5934 /* Sanity checking to avoid pcp imbalance */
5935 if (percpu_pagelist_fraction
&&
5936 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5937 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5943 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
5946 for_each_populated_zone(zone
) {
5949 for_each_possible_cpu(cpu
)
5950 pageset_set_high_and_batch(zone
,
5951 per_cpu_ptr(zone
->pageset
, cpu
));
5954 mutex_unlock(&pcp_batch_high_lock
);
5958 int hashdist
= HASHDIST_DEFAULT
;
5961 static int __init
set_hashdist(char *str
)
5965 hashdist
= simple_strtoul(str
, &str
, 0);
5968 __setup("hashdist=", set_hashdist
);
5972 * allocate a large system hash table from bootmem
5973 * - it is assumed that the hash table must contain an exact power-of-2
5974 * quantity of entries
5975 * - limit is the number of hash buckets, not the total allocation size
5977 void *__init
alloc_large_system_hash(const char *tablename
,
5978 unsigned long bucketsize
,
5979 unsigned long numentries
,
5982 unsigned int *_hash_shift
,
5983 unsigned int *_hash_mask
,
5984 unsigned long low_limit
,
5985 unsigned long high_limit
)
5987 unsigned long long max
= high_limit
;
5988 unsigned long log2qty
, size
;
5991 /* allow the kernel cmdline to have a say */
5993 /* round applicable memory size up to nearest megabyte */
5994 numentries
= nr_kernel_pages
;
5996 /* It isn't necessary when PAGE_SIZE >= 1MB */
5997 if (PAGE_SHIFT
< 20)
5998 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6000 /* limit to 1 bucket per 2^scale bytes of low memory */
6001 if (scale
> PAGE_SHIFT
)
6002 numentries
>>= (scale
- PAGE_SHIFT
);
6004 numentries
<<= (PAGE_SHIFT
- scale
);
6006 /* Make sure we've got at least a 0-order allocation.. */
6007 if (unlikely(flags
& HASH_SMALL
)) {
6008 /* Makes no sense without HASH_EARLY */
6009 WARN_ON(!(flags
& HASH_EARLY
));
6010 if (!(numentries
>> *_hash_shift
)) {
6011 numentries
= 1UL << *_hash_shift
;
6012 BUG_ON(!numentries
);
6014 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6015 numentries
= PAGE_SIZE
/ bucketsize
;
6017 numentries
= roundup_pow_of_two(numentries
);
6019 /* limit allocation size to 1/16 total memory by default */
6021 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6022 do_div(max
, bucketsize
);
6024 max
= min(max
, 0x80000000ULL
);
6026 if (numentries
< low_limit
)
6027 numentries
= low_limit
;
6028 if (numentries
> max
)
6031 log2qty
= ilog2(numentries
);
6034 size
= bucketsize
<< log2qty
;
6035 if (flags
& HASH_EARLY
)
6036 table
= memblock_virt_alloc_nopanic(size
, 0);
6038 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6041 * If bucketsize is not a power-of-two, we may free
6042 * some pages at the end of hash table which
6043 * alloc_pages_exact() automatically does
6045 if (get_order(size
) < MAX_ORDER
) {
6046 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6047 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6050 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6053 panic("Failed to allocate %s hash table\n", tablename
);
6055 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6058 ilog2(size
) - PAGE_SHIFT
,
6062 *_hash_shift
= log2qty
;
6064 *_hash_mask
= (1 << log2qty
) - 1;
6069 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6070 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6073 #ifdef CONFIG_SPARSEMEM
6074 return __pfn_to_section(pfn
)->pageblock_flags
;
6076 return zone
->pageblock_flags
;
6077 #endif /* CONFIG_SPARSEMEM */
6080 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6082 #ifdef CONFIG_SPARSEMEM
6083 pfn
&= (PAGES_PER_SECTION
-1);
6084 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6086 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6087 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6088 #endif /* CONFIG_SPARSEMEM */
6092 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6093 * @page: The page within the block of interest
6094 * @pfn: The target page frame number
6095 * @end_bitidx: The last bit of interest to retrieve
6096 * @mask: mask of bits that the caller is interested in
6098 * Return: pageblock_bits flags
6100 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6101 unsigned long end_bitidx
,
6105 unsigned long *bitmap
;
6106 unsigned long bitidx
, word_bitidx
;
6109 zone
= page_zone(page
);
6110 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6111 bitidx
= pfn_to_bitidx(zone
, pfn
);
6112 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6113 bitidx
&= (BITS_PER_LONG
-1);
6115 word
= bitmap
[word_bitidx
];
6116 bitidx
+= end_bitidx
;
6117 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6121 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6122 * @page: The page within the block of interest
6123 * @flags: The flags to set
6124 * @pfn: The target page frame number
6125 * @end_bitidx: The last bit of interest
6126 * @mask: mask of bits that the caller is interested in
6128 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6130 unsigned long end_bitidx
,
6134 unsigned long *bitmap
;
6135 unsigned long bitidx
, word_bitidx
;
6136 unsigned long old_word
, word
;
6138 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6140 zone
= page_zone(page
);
6141 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6142 bitidx
= pfn_to_bitidx(zone
, pfn
);
6143 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6144 bitidx
&= (BITS_PER_LONG
-1);
6146 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6148 bitidx
+= end_bitidx
;
6149 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6150 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6152 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6154 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6155 if (word
== old_word
)
6162 * This function checks whether pageblock includes unmovable pages or not.
6163 * If @count is not zero, it is okay to include less @count unmovable pages
6165 * PageLRU check without isolation or lru_lock could race so that
6166 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6167 * expect this function should be exact.
6169 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6170 bool skip_hwpoisoned_pages
)
6172 unsigned long pfn
, iter
, found
;
6176 * For avoiding noise data, lru_add_drain_all() should be called
6177 * If ZONE_MOVABLE, the zone never contains unmovable pages
6179 if (zone_idx(zone
) == ZONE_MOVABLE
)
6181 mt
= get_pageblock_migratetype(page
);
6182 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6185 pfn
= page_to_pfn(page
);
6186 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6187 unsigned long check
= pfn
+ iter
;
6189 if (!pfn_valid_within(check
))
6192 page
= pfn_to_page(check
);
6195 * Hugepages are not in LRU lists, but they're movable.
6196 * We need not scan over tail pages bacause we don't
6197 * handle each tail page individually in migration.
6199 if (PageHuge(page
)) {
6200 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6205 * We can't use page_count without pin a page
6206 * because another CPU can free compound page.
6207 * This check already skips compound tails of THP
6208 * because their page->_count is zero at all time.
6210 if (!atomic_read(&page
->_count
)) {
6211 if (PageBuddy(page
))
6212 iter
+= (1 << page_order(page
)) - 1;
6217 * The HWPoisoned page may be not in buddy system, and
6218 * page_count() is not 0.
6220 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6226 * If there are RECLAIMABLE pages, we need to check it.
6227 * But now, memory offline itself doesn't call shrink_slab()
6228 * and it still to be fixed.
6231 * If the page is not RAM, page_count()should be 0.
6232 * we don't need more check. This is an _used_ not-movable page.
6234 * The problematic thing here is PG_reserved pages. PG_reserved
6235 * is set to both of a memory hole page and a _used_ kernel
6244 bool is_pageblock_removable_nolock(struct page
*page
)
6250 * We have to be careful here because we are iterating over memory
6251 * sections which are not zone aware so we might end up outside of
6252 * the zone but still within the section.
6253 * We have to take care about the node as well. If the node is offline
6254 * its NODE_DATA will be NULL - see page_zone.
6256 if (!node_online(page_to_nid(page
)))
6259 zone
= page_zone(page
);
6260 pfn
= page_to_pfn(page
);
6261 if (!zone_spans_pfn(zone
, pfn
))
6264 return !has_unmovable_pages(zone
, page
, 0, true);
6269 static unsigned long pfn_max_align_down(unsigned long pfn
)
6271 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6272 pageblock_nr_pages
) - 1);
6275 static unsigned long pfn_max_align_up(unsigned long pfn
)
6277 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6278 pageblock_nr_pages
));
6281 /* [start, end) must belong to a single zone. */
6282 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6283 unsigned long start
, unsigned long end
)
6285 /* This function is based on compact_zone() from compaction.c. */
6286 unsigned long nr_reclaimed
;
6287 unsigned long pfn
= start
;
6288 unsigned int tries
= 0;
6293 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6294 if (fatal_signal_pending(current
)) {
6299 if (list_empty(&cc
->migratepages
)) {
6300 cc
->nr_migratepages
= 0;
6301 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6307 } else if (++tries
== 5) {
6308 ret
= ret
< 0 ? ret
: -EBUSY
;
6312 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6314 cc
->nr_migratepages
-= nr_reclaimed
;
6316 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6317 NULL
, 0, cc
->mode
, MR_CMA
);
6320 putback_movable_pages(&cc
->migratepages
);
6327 * alloc_contig_range() -- tries to allocate given range of pages
6328 * @start: start PFN to allocate
6329 * @end: one-past-the-last PFN to allocate
6330 * @migratetype: migratetype of the underlaying pageblocks (either
6331 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6332 * in range must have the same migratetype and it must
6333 * be either of the two.
6335 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6336 * aligned, however it's the caller's responsibility to guarantee that
6337 * we are the only thread that changes migrate type of pageblocks the
6340 * The PFN range must belong to a single zone.
6342 * Returns zero on success or negative error code. On success all
6343 * pages which PFN is in [start, end) are allocated for the caller and
6344 * need to be freed with free_contig_range().
6346 int alloc_contig_range(unsigned long start
, unsigned long end
,
6347 unsigned migratetype
)
6349 unsigned long outer_start
, outer_end
;
6352 struct compact_control cc
= {
6353 .nr_migratepages
= 0,
6355 .zone
= page_zone(pfn_to_page(start
)),
6356 .mode
= MIGRATE_SYNC
,
6357 .ignore_skip_hint
= true,
6359 INIT_LIST_HEAD(&cc
.migratepages
);
6362 * What we do here is we mark all pageblocks in range as
6363 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6364 * have different sizes, and due to the way page allocator
6365 * work, we align the range to biggest of the two pages so
6366 * that page allocator won't try to merge buddies from
6367 * different pageblocks and change MIGRATE_ISOLATE to some
6368 * other migration type.
6370 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6371 * migrate the pages from an unaligned range (ie. pages that
6372 * we are interested in). This will put all the pages in
6373 * range back to page allocator as MIGRATE_ISOLATE.
6375 * When this is done, we take the pages in range from page
6376 * allocator removing them from the buddy system. This way
6377 * page allocator will never consider using them.
6379 * This lets us mark the pageblocks back as
6380 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6381 * aligned range but not in the unaligned, original range are
6382 * put back to page allocator so that buddy can use them.
6385 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6386 pfn_max_align_up(end
), migratetype
,
6391 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6396 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6397 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6398 * more, all pages in [start, end) are free in page allocator.
6399 * What we are going to do is to allocate all pages from
6400 * [start, end) (that is remove them from page allocator).
6402 * The only problem is that pages at the beginning and at the
6403 * end of interesting range may be not aligned with pages that
6404 * page allocator holds, ie. they can be part of higher order
6405 * pages. Because of this, we reserve the bigger range and
6406 * once this is done free the pages we are not interested in.
6408 * We don't have to hold zone->lock here because the pages are
6409 * isolated thus they won't get removed from buddy.
6412 lru_add_drain_all();
6413 drain_all_pages(cc
.zone
);
6416 outer_start
= start
;
6417 while (!PageBuddy(pfn_to_page(outer_start
))) {
6418 if (++order
>= MAX_ORDER
) {
6422 outer_start
&= ~0UL << order
;
6425 /* Make sure the range is really isolated. */
6426 if (test_pages_isolated(outer_start
, end
, false)) {
6427 pr_info("%s: [%lx, %lx) PFNs busy\n",
6428 __func__
, outer_start
, end
);
6433 /* Grab isolated pages from freelists. */
6434 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6440 /* Free head and tail (if any) */
6441 if (start
!= outer_start
)
6442 free_contig_range(outer_start
, start
- outer_start
);
6443 if (end
!= outer_end
)
6444 free_contig_range(end
, outer_end
- end
);
6447 undo_isolate_page_range(pfn_max_align_down(start
),
6448 pfn_max_align_up(end
), migratetype
);
6452 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6454 unsigned int count
= 0;
6456 for (; nr_pages
--; pfn
++) {
6457 struct page
*page
= pfn_to_page(pfn
);
6459 count
+= page_count(page
) != 1;
6462 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6466 #ifdef CONFIG_MEMORY_HOTPLUG
6468 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6469 * page high values need to be recalulated.
6471 void __meminit
zone_pcp_update(struct zone
*zone
)
6474 mutex_lock(&pcp_batch_high_lock
);
6475 for_each_possible_cpu(cpu
)
6476 pageset_set_high_and_batch(zone
,
6477 per_cpu_ptr(zone
->pageset
, cpu
));
6478 mutex_unlock(&pcp_batch_high_lock
);
6482 void zone_pcp_reset(struct zone
*zone
)
6484 unsigned long flags
;
6486 struct per_cpu_pageset
*pset
;
6488 /* avoid races with drain_pages() */
6489 local_irq_save(flags
);
6490 if (zone
->pageset
!= &boot_pageset
) {
6491 for_each_online_cpu(cpu
) {
6492 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6493 drain_zonestat(zone
, pset
);
6495 free_percpu(zone
->pageset
);
6496 zone
->pageset
= &boot_pageset
;
6498 local_irq_restore(flags
);
6501 #ifdef CONFIG_MEMORY_HOTREMOVE
6503 * All pages in the range must be isolated before calling this.
6506 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6510 unsigned int order
, i
;
6512 unsigned long flags
;
6513 /* find the first valid pfn */
6514 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6519 zone
= page_zone(pfn_to_page(pfn
));
6520 spin_lock_irqsave(&zone
->lock
, flags
);
6522 while (pfn
< end_pfn
) {
6523 if (!pfn_valid(pfn
)) {
6527 page
= pfn_to_page(pfn
);
6529 * The HWPoisoned page may be not in buddy system, and
6530 * page_count() is not 0.
6532 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6534 SetPageReserved(page
);
6538 BUG_ON(page_count(page
));
6539 BUG_ON(!PageBuddy(page
));
6540 order
= page_order(page
);
6541 #ifdef CONFIG_DEBUG_VM
6542 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6543 pfn
, 1 << order
, end_pfn
);
6545 list_del(&page
->lru
);
6546 rmv_page_order(page
);
6547 zone
->free_area
[order
].nr_free
--;
6548 for (i
= 0; i
< (1 << order
); i
++)
6549 SetPageReserved((page
+i
));
6550 pfn
+= (1 << order
);
6552 spin_unlock_irqrestore(&zone
->lock
, flags
);
6556 #ifdef CONFIG_MEMORY_FAILURE
6557 bool is_free_buddy_page(struct page
*page
)
6559 struct zone
*zone
= page_zone(page
);
6560 unsigned long pfn
= page_to_pfn(page
);
6561 unsigned long flags
;
6564 spin_lock_irqsave(&zone
->lock
, flags
);
6565 for (order
= 0; order
< MAX_ORDER
; order
++) {
6566 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6568 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6571 spin_unlock_irqrestore(&zone
->lock
, flags
);
6573 return order
< MAX_ORDER
;