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_ext.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
;
428 bool _debug_guardpage_enabled __read_mostly
;
430 static bool need_debug_guardpage(void)
435 static void init_debug_guardpage(void)
437 _debug_guardpage_enabled
= true;
440 struct page_ext_operations debug_guardpage_ops
= {
441 .need
= need_debug_guardpage
,
442 .init
= init_debug_guardpage
,
445 static int __init
debug_guardpage_minorder_setup(char *buf
)
449 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
450 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
453 _debug_guardpage_minorder
= res
;
454 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
457 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
459 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
460 unsigned int order
, int migratetype
)
462 struct page_ext
*page_ext
;
464 if (!debug_guardpage_enabled())
467 page_ext
= lookup_page_ext(page
);
468 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
470 INIT_LIST_HEAD(&page
->lru
);
471 set_page_private(page
, order
);
472 /* Guard pages are not available for any usage */
473 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
476 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
477 unsigned int order
, int migratetype
)
479 struct page_ext
*page_ext
;
481 if (!debug_guardpage_enabled())
484 page_ext
= lookup_page_ext(page
);
485 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
487 set_page_private(page
, 0);
488 if (!is_migrate_isolate(migratetype
))
489 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
492 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
493 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
494 unsigned int order
, int migratetype
) {}
495 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
496 unsigned int order
, int migratetype
) {}
499 static inline void set_page_order(struct page
*page
, unsigned int order
)
501 set_page_private(page
, order
);
502 __SetPageBuddy(page
);
505 static inline void rmv_page_order(struct page
*page
)
507 __ClearPageBuddy(page
);
508 set_page_private(page
, 0);
512 * This function checks whether a page is free && is the buddy
513 * we can do coalesce a page and its buddy if
514 * (a) the buddy is not in a hole &&
515 * (b) the buddy is in the buddy system &&
516 * (c) a page and its buddy have the same order &&
517 * (d) a page and its buddy are in the same zone.
519 * For recording whether a page is in the buddy system, we set ->_mapcount
520 * PAGE_BUDDY_MAPCOUNT_VALUE.
521 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
522 * serialized by zone->lock.
524 * For recording page's order, we use page_private(page).
526 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
529 if (!pfn_valid_within(page_to_pfn(buddy
)))
532 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
533 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
535 if (page_zone_id(page
) != page_zone_id(buddy
))
541 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
542 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
545 * zone check is done late to avoid uselessly
546 * calculating zone/node ids for pages that could
549 if (page_zone_id(page
) != page_zone_id(buddy
))
558 * Freeing function for a buddy system allocator.
560 * The concept of a buddy system is to maintain direct-mapped table
561 * (containing bit values) for memory blocks of various "orders".
562 * The bottom level table contains the map for the smallest allocatable
563 * units of memory (here, pages), and each level above it describes
564 * pairs of units from the levels below, hence, "buddies".
565 * At a high level, all that happens here is marking the table entry
566 * at the bottom level available, and propagating the changes upward
567 * as necessary, plus some accounting needed to play nicely with other
568 * parts of the VM system.
569 * At each level, we keep a list of pages, which are heads of continuous
570 * free pages of length of (1 << order) and marked with _mapcount
571 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
573 * So when we are allocating or freeing one, we can derive the state of the
574 * other. That is, if we allocate a small block, and both were
575 * free, the remainder of the region must be split into blocks.
576 * If a block is freed, and its buddy is also free, then this
577 * triggers coalescing into a block of larger size.
582 static inline void __free_one_page(struct page
*page
,
584 struct zone
*zone
, unsigned int order
,
587 unsigned long page_idx
;
588 unsigned long combined_idx
;
589 unsigned long uninitialized_var(buddy_idx
);
591 int max_order
= MAX_ORDER
;
593 VM_BUG_ON(!zone_is_initialized(zone
));
595 if (unlikely(PageCompound(page
)))
596 if (unlikely(destroy_compound_page(page
, order
)))
599 VM_BUG_ON(migratetype
== -1);
600 if (is_migrate_isolate(migratetype
)) {
602 * We restrict max order of merging to prevent merge
603 * between freepages on isolate pageblock and normal
604 * pageblock. Without this, pageblock isolation
605 * could cause incorrect freepage accounting.
607 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
609 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
612 page_idx
= pfn
& ((1 << max_order
) - 1);
614 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
615 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
617 while (order
< max_order
- 1) {
618 buddy_idx
= __find_buddy_index(page_idx
, order
);
619 buddy
= page
+ (buddy_idx
- page_idx
);
620 if (!page_is_buddy(page
, buddy
, order
))
623 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
624 * merge with it and move up one order.
626 if (page_is_guard(buddy
)) {
627 clear_page_guard(zone
, buddy
, order
, migratetype
);
629 list_del(&buddy
->lru
);
630 zone
->free_area
[order
].nr_free
--;
631 rmv_page_order(buddy
);
633 combined_idx
= buddy_idx
& page_idx
;
634 page
= page
+ (combined_idx
- page_idx
);
635 page_idx
= combined_idx
;
638 set_page_order(page
, order
);
641 * If this is not the largest possible page, check if the buddy
642 * of the next-highest order is free. If it is, it's possible
643 * that pages are being freed that will coalesce soon. In case,
644 * that is happening, add the free page to the tail of the list
645 * so it's less likely to be used soon and more likely to be merged
646 * as a higher order page
648 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
649 struct page
*higher_page
, *higher_buddy
;
650 combined_idx
= buddy_idx
& page_idx
;
651 higher_page
= page
+ (combined_idx
- page_idx
);
652 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
653 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
654 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
655 list_add_tail(&page
->lru
,
656 &zone
->free_area
[order
].free_list
[migratetype
]);
661 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
663 zone
->free_area
[order
].nr_free
++;
666 static inline int free_pages_check(struct page
*page
)
668 const char *bad_reason
= NULL
;
669 unsigned long bad_flags
= 0;
671 if (unlikely(page_mapcount(page
)))
672 bad_reason
= "nonzero mapcount";
673 if (unlikely(page
->mapping
!= NULL
))
674 bad_reason
= "non-NULL mapping";
675 if (unlikely(atomic_read(&page
->_count
) != 0))
676 bad_reason
= "nonzero _count";
677 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
678 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
679 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
682 if (unlikely(page
->mem_cgroup
))
683 bad_reason
= "page still charged to cgroup";
685 if (unlikely(bad_reason
)) {
686 bad_page(page
, bad_reason
, bad_flags
);
689 page_cpupid_reset_last(page
);
690 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
691 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
696 * Frees a number of pages from the PCP lists
697 * Assumes all pages on list are in same zone, and of same order.
698 * count is the number of pages to free.
700 * If the zone was previously in an "all pages pinned" state then look to
701 * see if this freeing clears that state.
703 * And clear the zone's pages_scanned counter, to hold off the "all pages are
704 * pinned" detection logic.
706 static void free_pcppages_bulk(struct zone
*zone
, int count
,
707 struct per_cpu_pages
*pcp
)
712 unsigned long nr_scanned
;
714 spin_lock(&zone
->lock
);
715 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
717 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
721 struct list_head
*list
;
724 * Remove pages from lists in a round-robin fashion. A
725 * batch_free count is maintained that is incremented when an
726 * empty list is encountered. This is so more pages are freed
727 * off fuller lists instead of spinning excessively around empty
732 if (++migratetype
== MIGRATE_PCPTYPES
)
734 list
= &pcp
->lists
[migratetype
];
735 } while (list_empty(list
));
737 /* This is the only non-empty list. Free them all. */
738 if (batch_free
== MIGRATE_PCPTYPES
)
739 batch_free
= to_free
;
742 int mt
; /* migratetype of the to-be-freed page */
744 page
= list_entry(list
->prev
, struct page
, lru
);
745 /* must delete as __free_one_page list manipulates */
746 list_del(&page
->lru
);
747 mt
= get_freepage_migratetype(page
);
748 if (unlikely(has_isolate_pageblock(zone
)))
749 mt
= get_pageblock_migratetype(page
);
751 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
752 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
753 trace_mm_page_pcpu_drain(page
, 0, mt
);
754 } while (--to_free
&& --batch_free
&& !list_empty(list
));
756 spin_unlock(&zone
->lock
);
759 static void free_one_page(struct zone
*zone
,
760 struct page
*page
, unsigned long pfn
,
764 unsigned long nr_scanned
;
765 spin_lock(&zone
->lock
);
766 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
768 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
770 if (unlikely(has_isolate_pageblock(zone
) ||
771 is_migrate_isolate(migratetype
))) {
772 migratetype
= get_pfnblock_migratetype(page
, pfn
);
774 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
775 spin_unlock(&zone
->lock
);
778 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
783 VM_BUG_ON_PAGE(PageTail(page
), page
);
784 VM_BUG_ON_PAGE(PageHead(page
) && compound_order(page
) != order
, page
);
786 trace_mm_page_free(page
, order
);
787 kmemcheck_free_shadow(page
, order
);
790 page
->mapping
= NULL
;
791 for (i
= 0; i
< (1 << order
); i
++)
792 bad
+= free_pages_check(page
+ i
);
796 if (!PageHighMem(page
)) {
797 debug_check_no_locks_freed(page_address(page
),
799 debug_check_no_obj_freed(page_address(page
),
802 arch_free_page(page
, order
);
803 kernel_map_pages(page
, 1 << order
, 0);
808 static void __free_pages_ok(struct page
*page
, unsigned int order
)
812 unsigned long pfn
= page_to_pfn(page
);
814 if (!free_pages_prepare(page
, order
))
817 migratetype
= get_pfnblock_migratetype(page
, pfn
);
818 local_irq_save(flags
);
819 __count_vm_events(PGFREE
, 1 << order
);
820 set_freepage_migratetype(page
, migratetype
);
821 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
822 local_irq_restore(flags
);
825 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
827 unsigned int nr_pages
= 1 << order
;
828 struct page
*p
= page
;
832 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
834 __ClearPageReserved(p
);
835 set_page_count(p
, 0);
837 __ClearPageReserved(p
);
838 set_page_count(p
, 0);
840 page_zone(page
)->managed_pages
+= nr_pages
;
841 set_page_refcounted(page
);
842 __free_pages(page
, order
);
846 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
847 void __init
init_cma_reserved_pageblock(struct page
*page
)
849 unsigned i
= pageblock_nr_pages
;
850 struct page
*p
= page
;
853 __ClearPageReserved(p
);
854 set_page_count(p
, 0);
857 set_pageblock_migratetype(page
, MIGRATE_CMA
);
859 if (pageblock_order
>= MAX_ORDER
) {
860 i
= pageblock_nr_pages
;
863 set_page_refcounted(p
);
864 __free_pages(p
, MAX_ORDER
- 1);
865 p
+= MAX_ORDER_NR_PAGES
;
866 } while (i
-= MAX_ORDER_NR_PAGES
);
868 set_page_refcounted(page
);
869 __free_pages(page
, pageblock_order
);
872 adjust_managed_page_count(page
, pageblock_nr_pages
);
877 * The order of subdivision here is critical for the IO subsystem.
878 * Please do not alter this order without good reasons and regression
879 * testing. Specifically, as large blocks of memory are subdivided,
880 * the order in which smaller blocks are delivered depends on the order
881 * they're subdivided in this function. This is the primary factor
882 * influencing the order in which pages are delivered to the IO
883 * subsystem according to empirical testing, and this is also justified
884 * by considering the behavior of a buddy system containing a single
885 * large block of memory acted on by a series of small allocations.
886 * This behavior is a critical factor in sglist merging's success.
890 static inline void expand(struct zone
*zone
, struct page
*page
,
891 int low
, int high
, struct free_area
*area
,
894 unsigned long size
= 1 << high
;
900 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
902 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
903 debug_guardpage_enabled() &&
904 high
< debug_guardpage_minorder()) {
906 * Mark as guard pages (or page), that will allow to
907 * merge back to allocator when buddy will be freed.
908 * Corresponding page table entries will not be touched,
909 * pages will stay not present in virtual address space
911 set_page_guard(zone
, &page
[size
], high
, migratetype
);
914 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
916 set_page_order(&page
[size
], high
);
921 * This page is about to be returned from the page allocator
923 static inline int check_new_page(struct page
*page
)
925 const char *bad_reason
= NULL
;
926 unsigned long bad_flags
= 0;
928 if (unlikely(page_mapcount(page
)))
929 bad_reason
= "nonzero mapcount";
930 if (unlikely(page
->mapping
!= NULL
))
931 bad_reason
= "non-NULL mapping";
932 if (unlikely(atomic_read(&page
->_count
) != 0))
933 bad_reason
= "nonzero _count";
934 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
935 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
936 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
939 if (unlikely(page
->mem_cgroup
))
940 bad_reason
= "page still charged to cgroup";
942 if (unlikely(bad_reason
)) {
943 bad_page(page
, bad_reason
, bad_flags
);
949 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
)
953 for (i
= 0; i
< (1 << order
); i
++) {
954 struct page
*p
= page
+ i
;
955 if (unlikely(check_new_page(p
)))
959 set_page_private(page
, 0);
960 set_page_refcounted(page
);
962 arch_alloc_page(page
, order
);
963 kernel_map_pages(page
, 1 << order
, 1);
965 if (gfp_flags
& __GFP_ZERO
)
966 prep_zero_page(page
, order
, gfp_flags
);
968 if (order
&& (gfp_flags
& __GFP_COMP
))
969 prep_compound_page(page
, order
);
975 * Go through the free lists for the given migratetype and remove
976 * the smallest available page from the freelists
979 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
982 unsigned int current_order
;
983 struct free_area
*area
;
986 /* Find a page of the appropriate size in the preferred list */
987 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
988 area
= &(zone
->free_area
[current_order
]);
989 if (list_empty(&area
->free_list
[migratetype
]))
992 page
= list_entry(area
->free_list
[migratetype
].next
,
994 list_del(&page
->lru
);
995 rmv_page_order(page
);
997 expand(zone
, page
, order
, current_order
, area
, migratetype
);
998 set_freepage_migratetype(page
, migratetype
);
1007 * This array describes the order lists are fallen back to when
1008 * the free lists for the desirable migrate type are depleted
1010 static int fallbacks
[MIGRATE_TYPES
][4] = {
1011 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1012 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1014 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
1015 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
1017 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
1019 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
1020 #ifdef CONFIG_MEMORY_ISOLATION
1021 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
1026 * Move the free pages in a range to the free lists of the requested type.
1027 * Note that start_page and end_pages are not aligned on a pageblock
1028 * boundary. If alignment is required, use move_freepages_block()
1030 int move_freepages(struct zone
*zone
,
1031 struct page
*start_page
, struct page
*end_page
,
1035 unsigned long order
;
1036 int pages_moved
= 0;
1038 #ifndef CONFIG_HOLES_IN_ZONE
1040 * page_zone is not safe to call in this context when
1041 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1042 * anyway as we check zone boundaries in move_freepages_block().
1043 * Remove at a later date when no bug reports exist related to
1044 * grouping pages by mobility
1046 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1049 for (page
= start_page
; page
<= end_page
;) {
1050 /* Make sure we are not inadvertently changing nodes */
1051 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1053 if (!pfn_valid_within(page_to_pfn(page
))) {
1058 if (!PageBuddy(page
)) {
1063 order
= page_order(page
);
1064 list_move(&page
->lru
,
1065 &zone
->free_area
[order
].free_list
[migratetype
]);
1066 set_freepage_migratetype(page
, migratetype
);
1068 pages_moved
+= 1 << order
;
1074 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1077 unsigned long start_pfn
, end_pfn
;
1078 struct page
*start_page
, *end_page
;
1080 start_pfn
= page_to_pfn(page
);
1081 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1082 start_page
= pfn_to_page(start_pfn
);
1083 end_page
= start_page
+ pageblock_nr_pages
- 1;
1084 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1086 /* Do not cross zone boundaries */
1087 if (!zone_spans_pfn(zone
, start_pfn
))
1089 if (!zone_spans_pfn(zone
, end_pfn
))
1092 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1095 static void change_pageblock_range(struct page
*pageblock_page
,
1096 int start_order
, int migratetype
)
1098 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1100 while (nr_pageblocks
--) {
1101 set_pageblock_migratetype(pageblock_page
, migratetype
);
1102 pageblock_page
+= pageblock_nr_pages
;
1107 * If breaking a large block of pages, move all free pages to the preferred
1108 * allocation list. If falling back for a reclaimable kernel allocation, be
1109 * more aggressive about taking ownership of free pages.
1111 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1112 * nor move CMA pages to different free lists. We don't want unmovable pages
1113 * to be allocated from MIGRATE_CMA areas.
1115 * Returns the new migratetype of the pageblock (or the same old migratetype
1116 * if it was unchanged).
1118 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1119 int start_type
, int fallback_type
)
1121 int current_order
= page_order(page
);
1124 * When borrowing from MIGRATE_CMA, we need to release the excess
1125 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1126 * is set to CMA so it is returned to the correct freelist in case
1127 * the page ends up being not actually allocated from the pcp lists.
1129 if (is_migrate_cma(fallback_type
))
1130 return fallback_type
;
1132 /* Take ownership for orders >= pageblock_order */
1133 if (current_order
>= pageblock_order
) {
1134 change_pageblock_range(page
, current_order
, start_type
);
1138 if (current_order
>= pageblock_order
/ 2 ||
1139 start_type
== MIGRATE_RECLAIMABLE
||
1140 page_group_by_mobility_disabled
) {
1143 pages
= move_freepages_block(zone
, page
, start_type
);
1145 /* Claim the whole block if over half of it is free */
1146 if (pages
>= (1 << (pageblock_order
-1)) ||
1147 page_group_by_mobility_disabled
) {
1149 set_pageblock_migratetype(page
, start_type
);
1155 return fallback_type
;
1158 /* Remove an element from the buddy allocator from the fallback list */
1159 static inline struct page
*
1160 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1162 struct free_area
*area
;
1163 unsigned int current_order
;
1165 int migratetype
, new_type
, i
;
1167 /* Find the largest possible block of pages in the other list */
1168 for (current_order
= MAX_ORDER
-1;
1169 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1172 migratetype
= fallbacks
[start_migratetype
][i
];
1174 /* MIGRATE_RESERVE handled later if necessary */
1175 if (migratetype
== MIGRATE_RESERVE
)
1178 area
= &(zone
->free_area
[current_order
]);
1179 if (list_empty(&area
->free_list
[migratetype
]))
1182 page
= list_entry(area
->free_list
[migratetype
].next
,
1186 new_type
= try_to_steal_freepages(zone
, page
,
1190 /* Remove the page from the freelists */
1191 list_del(&page
->lru
);
1192 rmv_page_order(page
);
1194 expand(zone
, page
, order
, current_order
, area
,
1196 /* The freepage_migratetype may differ from pageblock's
1197 * migratetype depending on the decisions in
1198 * try_to_steal_freepages. This is OK as long as it does
1199 * not differ for MIGRATE_CMA type.
1201 set_freepage_migratetype(page
, new_type
);
1203 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1204 start_migratetype
, migratetype
, new_type
);
1214 * Do the hard work of removing an element from the buddy allocator.
1215 * Call me with the zone->lock already held.
1217 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1223 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1225 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1226 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1229 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1230 * is used because __rmqueue_smallest is an inline function
1231 * and we want just one call site
1234 migratetype
= MIGRATE_RESERVE
;
1239 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1244 * Obtain a specified number of elements from the buddy allocator, all under
1245 * a single hold of the lock, for efficiency. Add them to the supplied list.
1246 * Returns the number of new pages which were placed at *list.
1248 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1249 unsigned long count
, struct list_head
*list
,
1250 int migratetype
, bool cold
)
1254 spin_lock(&zone
->lock
);
1255 for (i
= 0; i
< count
; ++i
) {
1256 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1257 if (unlikely(page
== NULL
))
1261 * Split buddy pages returned by expand() are received here
1262 * in physical page order. The page is added to the callers and
1263 * list and the list head then moves forward. From the callers
1264 * perspective, the linked list is ordered by page number in
1265 * some conditions. This is useful for IO devices that can
1266 * merge IO requests if the physical pages are ordered
1270 list_add(&page
->lru
, list
);
1272 list_add_tail(&page
->lru
, list
);
1274 if (is_migrate_cma(get_freepage_migratetype(page
)))
1275 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1278 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1279 spin_unlock(&zone
->lock
);
1285 * Called from the vmstat counter updater to drain pagesets of this
1286 * currently executing processor on remote nodes after they have
1289 * Note that this function must be called with the thread pinned to
1290 * a single processor.
1292 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1294 unsigned long flags
;
1295 int to_drain
, batch
;
1297 local_irq_save(flags
);
1298 batch
= ACCESS_ONCE(pcp
->batch
);
1299 to_drain
= min(pcp
->count
, batch
);
1301 free_pcppages_bulk(zone
, to_drain
, pcp
);
1302 pcp
->count
-= to_drain
;
1304 local_irq_restore(flags
);
1309 * Drain pcplists of the indicated processor and zone.
1311 * The processor must either be the current processor and the
1312 * thread pinned to the current processor or a processor that
1315 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1317 unsigned long flags
;
1318 struct per_cpu_pageset
*pset
;
1319 struct per_cpu_pages
*pcp
;
1321 local_irq_save(flags
);
1322 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1326 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1329 local_irq_restore(flags
);
1333 * Drain pcplists of all zones on the indicated processor.
1335 * The processor must either be the current processor and the
1336 * thread pinned to the current processor or a processor that
1339 static void drain_pages(unsigned int cpu
)
1343 for_each_populated_zone(zone
) {
1344 drain_pages_zone(cpu
, zone
);
1349 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1351 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1352 * the single zone's pages.
1354 void drain_local_pages(struct zone
*zone
)
1356 int cpu
= smp_processor_id();
1359 drain_pages_zone(cpu
, zone
);
1365 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1367 * When zone parameter is non-NULL, spill just the single zone's pages.
1369 * Note that this code is protected against sending an IPI to an offline
1370 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1371 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1372 * nothing keeps CPUs from showing up after we populated the cpumask and
1373 * before the call to on_each_cpu_mask().
1375 void drain_all_pages(struct zone
*zone
)
1380 * Allocate in the BSS so we wont require allocation in
1381 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1383 static cpumask_t cpus_with_pcps
;
1386 * We don't care about racing with CPU hotplug event
1387 * as offline notification will cause the notified
1388 * cpu to drain that CPU pcps and on_each_cpu_mask
1389 * disables preemption as part of its processing
1391 for_each_online_cpu(cpu
) {
1392 struct per_cpu_pageset
*pcp
;
1394 bool has_pcps
= false;
1397 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1401 for_each_populated_zone(z
) {
1402 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1403 if (pcp
->pcp
.count
) {
1411 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1413 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1415 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1419 #ifdef CONFIG_HIBERNATION
1421 void mark_free_pages(struct zone
*zone
)
1423 unsigned long pfn
, max_zone_pfn
;
1424 unsigned long flags
;
1425 unsigned int order
, t
;
1426 struct list_head
*curr
;
1428 if (zone_is_empty(zone
))
1431 spin_lock_irqsave(&zone
->lock
, flags
);
1433 max_zone_pfn
= zone_end_pfn(zone
);
1434 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1435 if (pfn_valid(pfn
)) {
1436 struct page
*page
= pfn_to_page(pfn
);
1438 if (!swsusp_page_is_forbidden(page
))
1439 swsusp_unset_page_free(page
);
1442 for_each_migratetype_order(order
, t
) {
1443 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1446 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1447 for (i
= 0; i
< (1UL << order
); i
++)
1448 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1451 spin_unlock_irqrestore(&zone
->lock
, flags
);
1453 #endif /* CONFIG_PM */
1456 * Free a 0-order page
1457 * cold == true ? free a cold page : free a hot page
1459 void free_hot_cold_page(struct page
*page
, bool cold
)
1461 struct zone
*zone
= page_zone(page
);
1462 struct per_cpu_pages
*pcp
;
1463 unsigned long flags
;
1464 unsigned long pfn
= page_to_pfn(page
);
1467 if (!free_pages_prepare(page
, 0))
1470 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1471 set_freepage_migratetype(page
, migratetype
);
1472 local_irq_save(flags
);
1473 __count_vm_event(PGFREE
);
1476 * We only track unmovable, reclaimable and movable on pcp lists.
1477 * Free ISOLATE pages back to the allocator because they are being
1478 * offlined but treat RESERVE as movable pages so we can get those
1479 * areas back if necessary. Otherwise, we may have to free
1480 * excessively into the page allocator
1482 if (migratetype
>= MIGRATE_PCPTYPES
) {
1483 if (unlikely(is_migrate_isolate(migratetype
))) {
1484 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1487 migratetype
= MIGRATE_MOVABLE
;
1490 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1492 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1494 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1496 if (pcp
->count
>= pcp
->high
) {
1497 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1498 free_pcppages_bulk(zone
, batch
, pcp
);
1499 pcp
->count
-= batch
;
1503 local_irq_restore(flags
);
1507 * Free a list of 0-order pages
1509 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1511 struct page
*page
, *next
;
1513 list_for_each_entry_safe(page
, next
, list
, lru
) {
1514 trace_mm_page_free_batched(page
, cold
);
1515 free_hot_cold_page(page
, cold
);
1520 * split_page takes a non-compound higher-order page, and splits it into
1521 * n (1<<order) sub-pages: page[0..n]
1522 * Each sub-page must be freed individually.
1524 * Note: this is probably too low level an operation for use in drivers.
1525 * Please consult with lkml before using this in your driver.
1527 void split_page(struct page
*page
, unsigned int order
)
1531 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1532 VM_BUG_ON_PAGE(!page_count(page
), page
);
1534 #ifdef CONFIG_KMEMCHECK
1536 * Split shadow pages too, because free(page[0]) would
1537 * otherwise free the whole shadow.
1539 if (kmemcheck_page_is_tracked(page
))
1540 split_page(virt_to_page(page
[0].shadow
), order
);
1543 for (i
= 1; i
< (1 << order
); i
++)
1544 set_page_refcounted(page
+ i
);
1546 EXPORT_SYMBOL_GPL(split_page
);
1548 int __isolate_free_page(struct page
*page
, unsigned int order
)
1550 unsigned long watermark
;
1554 BUG_ON(!PageBuddy(page
));
1556 zone
= page_zone(page
);
1557 mt
= get_pageblock_migratetype(page
);
1559 if (!is_migrate_isolate(mt
)) {
1560 /* Obey watermarks as if the page was being allocated */
1561 watermark
= low_wmark_pages(zone
) + (1 << order
);
1562 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1565 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1568 /* Remove page from free list */
1569 list_del(&page
->lru
);
1570 zone
->free_area
[order
].nr_free
--;
1571 rmv_page_order(page
);
1573 /* Set the pageblock if the isolated page is at least a pageblock */
1574 if (order
>= pageblock_order
- 1) {
1575 struct page
*endpage
= page
+ (1 << order
) - 1;
1576 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1577 int mt
= get_pageblock_migratetype(page
);
1578 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1579 set_pageblock_migratetype(page
,
1584 return 1UL << order
;
1588 * Similar to split_page except the page is already free. As this is only
1589 * being used for migration, the migratetype of the block also changes.
1590 * As this is called with interrupts disabled, the caller is responsible
1591 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1594 * Note: this is probably too low level an operation for use in drivers.
1595 * Please consult with lkml before using this in your driver.
1597 int split_free_page(struct page
*page
)
1602 order
= page_order(page
);
1604 nr_pages
= __isolate_free_page(page
, order
);
1608 /* Split into individual pages */
1609 set_page_refcounted(page
);
1610 split_page(page
, order
);
1615 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1616 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1620 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1621 struct zone
*zone
, unsigned int order
,
1622 gfp_t gfp_flags
, int migratetype
)
1624 unsigned long flags
;
1626 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1629 if (likely(order
== 0)) {
1630 struct per_cpu_pages
*pcp
;
1631 struct list_head
*list
;
1633 local_irq_save(flags
);
1634 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1635 list
= &pcp
->lists
[migratetype
];
1636 if (list_empty(list
)) {
1637 pcp
->count
+= rmqueue_bulk(zone
, 0,
1640 if (unlikely(list_empty(list
)))
1645 page
= list_entry(list
->prev
, struct page
, lru
);
1647 page
= list_entry(list
->next
, struct page
, lru
);
1649 list_del(&page
->lru
);
1652 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1654 * __GFP_NOFAIL is not to be used in new code.
1656 * All __GFP_NOFAIL callers should be fixed so that they
1657 * properly detect and handle allocation failures.
1659 * We most definitely don't want callers attempting to
1660 * allocate greater than order-1 page units with
1663 WARN_ON_ONCE(order
> 1);
1665 spin_lock_irqsave(&zone
->lock
, flags
);
1666 page
= __rmqueue(zone
, order
, migratetype
);
1667 spin_unlock(&zone
->lock
);
1670 __mod_zone_freepage_state(zone
, -(1 << order
),
1671 get_freepage_migratetype(page
));
1674 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1675 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1676 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1677 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1679 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1680 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1681 local_irq_restore(flags
);
1683 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1684 if (prep_new_page(page
, order
, gfp_flags
))
1689 local_irq_restore(flags
);
1693 #ifdef CONFIG_FAIL_PAGE_ALLOC
1696 struct fault_attr attr
;
1698 u32 ignore_gfp_highmem
;
1699 u32 ignore_gfp_wait
;
1701 } fail_page_alloc
= {
1702 .attr
= FAULT_ATTR_INITIALIZER
,
1703 .ignore_gfp_wait
= 1,
1704 .ignore_gfp_highmem
= 1,
1708 static int __init
setup_fail_page_alloc(char *str
)
1710 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1712 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1714 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1716 if (order
< fail_page_alloc
.min_order
)
1718 if (gfp_mask
& __GFP_NOFAIL
)
1720 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1722 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1725 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1728 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1730 static int __init
fail_page_alloc_debugfs(void)
1732 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1735 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1736 &fail_page_alloc
.attr
);
1738 return PTR_ERR(dir
);
1740 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1741 &fail_page_alloc
.ignore_gfp_wait
))
1743 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1744 &fail_page_alloc
.ignore_gfp_highmem
))
1746 if (!debugfs_create_u32("min-order", mode
, dir
,
1747 &fail_page_alloc
.min_order
))
1752 debugfs_remove_recursive(dir
);
1757 late_initcall(fail_page_alloc_debugfs
);
1759 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1761 #else /* CONFIG_FAIL_PAGE_ALLOC */
1763 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1768 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1771 * Return true if free pages are above 'mark'. This takes into account the order
1772 * of the allocation.
1774 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1775 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1778 /* free_pages may go negative - that's OK */
1783 free_pages
-= (1 << order
) - 1;
1784 if (alloc_flags
& ALLOC_HIGH
)
1786 if (alloc_flags
& ALLOC_HARDER
)
1789 /* If allocation can't use CMA areas don't use free CMA pages */
1790 if (!(alloc_flags
& ALLOC_CMA
))
1791 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1794 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1796 for (o
= 0; o
< order
; o
++) {
1797 /* At the next order, this order's pages become unavailable */
1798 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1800 /* Require fewer higher order pages to be free */
1803 if (free_pages
<= min
)
1809 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1810 int classzone_idx
, int alloc_flags
)
1812 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1813 zone_page_state(z
, NR_FREE_PAGES
));
1816 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1817 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1819 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1821 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1822 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1824 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1830 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1831 * skip over zones that are not allowed by the cpuset, or that have
1832 * been recently (in last second) found to be nearly full. See further
1833 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1834 * that have to skip over a lot of full or unallowed zones.
1836 * If the zonelist cache is present in the passed zonelist, then
1837 * returns a pointer to the allowed node mask (either the current
1838 * tasks mems_allowed, or node_states[N_MEMORY].)
1840 * If the zonelist cache is not available for this zonelist, does
1841 * nothing and returns NULL.
1843 * If the fullzones BITMAP in the zonelist cache is stale (more than
1844 * a second since last zap'd) then we zap it out (clear its bits.)
1846 * We hold off even calling zlc_setup, until after we've checked the
1847 * first zone in the zonelist, on the theory that most allocations will
1848 * be satisfied from that first zone, so best to examine that zone as
1849 * quickly as we can.
1851 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1853 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1854 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1856 zlc
= zonelist
->zlcache_ptr
;
1860 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1861 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1862 zlc
->last_full_zap
= jiffies
;
1865 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1866 &cpuset_current_mems_allowed
:
1867 &node_states
[N_MEMORY
];
1868 return allowednodes
;
1872 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1873 * if it is worth looking at further for free memory:
1874 * 1) Check that the zone isn't thought to be full (doesn't have its
1875 * bit set in the zonelist_cache fullzones BITMAP).
1876 * 2) Check that the zones node (obtained from the zonelist_cache
1877 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1878 * Return true (non-zero) if zone is worth looking at further, or
1879 * else return false (zero) if it is not.
1881 * This check -ignores- the distinction between various watermarks,
1882 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1883 * found to be full for any variation of these watermarks, it will
1884 * be considered full for up to one second by all requests, unless
1885 * we are so low on memory on all allowed nodes that we are forced
1886 * into the second scan of the zonelist.
1888 * In the second scan we ignore this zonelist cache and exactly
1889 * apply the watermarks to all zones, even it is slower to do so.
1890 * We are low on memory in the second scan, and should leave no stone
1891 * unturned looking for a free page.
1893 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1894 nodemask_t
*allowednodes
)
1896 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1897 int i
; /* index of *z in zonelist zones */
1898 int n
; /* node that zone *z is on */
1900 zlc
= zonelist
->zlcache_ptr
;
1904 i
= z
- zonelist
->_zonerefs
;
1907 /* This zone is worth trying if it is allowed but not full */
1908 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1912 * Given 'z' scanning a zonelist, set the corresponding bit in
1913 * zlc->fullzones, so that subsequent attempts to allocate a page
1914 * from that zone don't waste time re-examining it.
1916 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1918 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1919 int i
; /* index of *z in zonelist zones */
1921 zlc
= zonelist
->zlcache_ptr
;
1925 i
= z
- zonelist
->_zonerefs
;
1927 set_bit(i
, zlc
->fullzones
);
1931 * clear all zones full, called after direct reclaim makes progress so that
1932 * a zone that was recently full is not skipped over for up to a second
1934 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1936 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1938 zlc
= zonelist
->zlcache_ptr
;
1942 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1945 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1947 return local_zone
->node
== zone
->node
;
1950 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1952 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1956 #else /* CONFIG_NUMA */
1958 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1963 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1964 nodemask_t
*allowednodes
)
1969 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1973 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1977 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1982 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1987 #endif /* CONFIG_NUMA */
1989 static void reset_alloc_batches(struct zone
*preferred_zone
)
1991 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
1994 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
1995 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
1996 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
1997 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1998 } while (zone
++ != preferred_zone
);
2002 * get_page_from_freelist goes through the zonelist trying to allocate
2005 static struct page
*
2006 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
2007 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
2008 struct zone
*preferred_zone
, int classzone_idx
, int migratetype
)
2011 struct page
*page
= NULL
;
2013 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
2014 int zlc_active
= 0; /* set if using zonelist_cache */
2015 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
2016 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
2017 (gfp_mask
& __GFP_WRITE
);
2018 int nr_fair_skipped
= 0;
2019 bool zonelist_rescan
;
2022 zonelist_rescan
= false;
2025 * Scan zonelist, looking for a zone with enough free.
2026 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2028 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2029 high_zoneidx
, nodemask
) {
2032 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2033 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2035 if (cpusets_enabled() &&
2036 (alloc_flags
& ALLOC_CPUSET
) &&
2037 !cpuset_zone_allowed(zone
, gfp_mask
))
2040 * Distribute pages in proportion to the individual
2041 * zone size to ensure fair page aging. The zone a
2042 * page was allocated in should have no effect on the
2043 * time the page has in memory before being reclaimed.
2045 if (alloc_flags
& ALLOC_FAIR
) {
2046 if (!zone_local(preferred_zone
, zone
))
2048 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2054 * When allocating a page cache page for writing, we
2055 * want to get it from a zone that is within its dirty
2056 * limit, such that no single zone holds more than its
2057 * proportional share of globally allowed dirty pages.
2058 * The dirty limits take into account the zone's
2059 * lowmem reserves and high watermark so that kswapd
2060 * should be able to balance it without having to
2061 * write pages from its LRU list.
2063 * This may look like it could increase pressure on
2064 * lower zones by failing allocations in higher zones
2065 * before they are full. But the pages that do spill
2066 * over are limited as the lower zones are protected
2067 * by this very same mechanism. It should not become
2068 * a practical burden to them.
2070 * XXX: For now, allow allocations to potentially
2071 * exceed the per-zone dirty limit in the slowpath
2072 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2073 * which is important when on a NUMA setup the allowed
2074 * zones are together not big enough to reach the
2075 * global limit. The proper fix for these situations
2076 * will require awareness of zones in the
2077 * dirty-throttling and the flusher threads.
2079 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2082 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2083 if (!zone_watermark_ok(zone
, order
, mark
,
2084 classzone_idx
, alloc_flags
)) {
2087 /* Checked here to keep the fast path fast */
2088 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2089 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2092 if (IS_ENABLED(CONFIG_NUMA
) &&
2093 !did_zlc_setup
&& nr_online_nodes
> 1) {
2095 * we do zlc_setup if there are multiple nodes
2096 * and before considering the first zone allowed
2099 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2104 if (zone_reclaim_mode
== 0 ||
2105 !zone_allows_reclaim(preferred_zone
, zone
))
2106 goto this_zone_full
;
2109 * As we may have just activated ZLC, check if the first
2110 * eligible zone has failed zone_reclaim recently.
2112 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2113 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2116 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2118 case ZONE_RECLAIM_NOSCAN
:
2121 case ZONE_RECLAIM_FULL
:
2122 /* scanned but unreclaimable */
2125 /* did we reclaim enough */
2126 if (zone_watermark_ok(zone
, order
, mark
,
2127 classzone_idx
, alloc_flags
))
2131 * Failed to reclaim enough to meet watermark.
2132 * Only mark the zone full if checking the min
2133 * watermark or if we failed to reclaim just
2134 * 1<<order pages or else the page allocator
2135 * fastpath will prematurely mark zones full
2136 * when the watermark is between the low and
2139 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2140 ret
== ZONE_RECLAIM_SOME
)
2141 goto this_zone_full
;
2148 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2149 gfp_mask
, migratetype
);
2153 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2154 zlc_mark_zone_full(zonelist
, z
);
2159 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2160 * necessary to allocate the page. The expectation is
2161 * that the caller is taking steps that will free more
2162 * memory. The caller should avoid the page being used
2163 * for !PFMEMALLOC purposes.
2165 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2170 * The first pass makes sure allocations are spread fairly within the
2171 * local node. However, the local node might have free pages left
2172 * after the fairness batches are exhausted, and remote zones haven't
2173 * even been considered yet. Try once more without fairness, and
2174 * include remote zones now, before entering the slowpath and waking
2175 * kswapd: prefer spilling to a remote zone over swapping locally.
2177 if (alloc_flags
& ALLOC_FAIR
) {
2178 alloc_flags
&= ~ALLOC_FAIR
;
2179 if (nr_fair_skipped
) {
2180 zonelist_rescan
= true;
2181 reset_alloc_batches(preferred_zone
);
2183 if (nr_online_nodes
> 1)
2184 zonelist_rescan
= true;
2187 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2188 /* Disable zlc cache for second zonelist scan */
2190 zonelist_rescan
= true;
2193 if (zonelist_rescan
)
2200 * Large machines with many possible nodes should not always dump per-node
2201 * meminfo in irq context.
2203 static inline bool should_suppress_show_mem(void)
2208 ret
= in_interrupt();
2213 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2214 DEFAULT_RATELIMIT_INTERVAL
,
2215 DEFAULT_RATELIMIT_BURST
);
2217 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2219 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2221 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2222 debug_guardpage_minorder() > 0)
2226 * This documents exceptions given to allocations in certain
2227 * contexts that are allowed to allocate outside current's set
2230 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2231 if (test_thread_flag(TIF_MEMDIE
) ||
2232 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2233 filter
&= ~SHOW_MEM_FILTER_NODES
;
2234 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2235 filter
&= ~SHOW_MEM_FILTER_NODES
;
2238 struct va_format vaf
;
2241 va_start(args
, fmt
);
2246 pr_warn("%pV", &vaf
);
2251 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2252 current
->comm
, order
, gfp_mask
);
2255 if (!should_suppress_show_mem())
2260 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2261 unsigned long did_some_progress
,
2262 unsigned long pages_reclaimed
)
2264 /* Do not loop if specifically requested */
2265 if (gfp_mask
& __GFP_NORETRY
)
2268 /* Always retry if specifically requested */
2269 if (gfp_mask
& __GFP_NOFAIL
)
2273 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2274 * making forward progress without invoking OOM. Suspend also disables
2275 * storage devices so kswapd will not help. Bail if we are suspending.
2277 if (!did_some_progress
&& pm_suspended_storage())
2281 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2282 * means __GFP_NOFAIL, but that may not be true in other
2285 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2289 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2290 * specified, then we retry until we no longer reclaim any pages
2291 * (above), or we've reclaimed an order of pages at least as
2292 * large as the allocation's order. In both cases, if the
2293 * allocation still fails, we stop retrying.
2295 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2301 static inline struct page
*
2302 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2303 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2304 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2305 int classzone_idx
, int migratetype
)
2309 /* Acquire the per-zone oom lock for each zone */
2310 if (!oom_zonelist_trylock(zonelist
, gfp_mask
)) {
2311 schedule_timeout_uninterruptible(1);
2316 * PM-freezer should be notified that there might be an OOM killer on
2317 * its way to kill and wake somebody up. This is too early and we might
2318 * end up not killing anything but false positives are acceptable.
2319 * See freeze_processes.
2324 * Go through the zonelist yet one more time, keep very high watermark
2325 * here, this is only to catch a parallel oom killing, we must fail if
2326 * we're still under heavy pressure.
2328 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2329 order
, zonelist
, high_zoneidx
,
2330 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2331 preferred_zone
, classzone_idx
, migratetype
);
2335 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2336 /* The OOM killer will not help higher order allocs */
2337 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2339 /* The OOM killer does not needlessly kill tasks for lowmem */
2340 if (high_zoneidx
< ZONE_NORMAL
)
2343 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2344 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2345 * The caller should handle page allocation failure by itself if
2346 * it specifies __GFP_THISNODE.
2347 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2349 if (gfp_mask
& __GFP_THISNODE
)
2352 /* Exhausted what can be done so it's blamo time */
2353 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2356 oom_zonelist_unlock(zonelist
, gfp_mask
);
2360 #ifdef CONFIG_COMPACTION
2361 /* Try memory compaction for high-order allocations before reclaim */
2362 static struct page
*
2363 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2364 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2365 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2366 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2367 int *contended_compaction
, bool *deferred_compaction
)
2369 unsigned long compact_result
;
2375 current
->flags
|= PF_MEMALLOC
;
2376 compact_result
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2378 contended_compaction
,
2379 alloc_flags
, classzone_idx
);
2380 current
->flags
&= ~PF_MEMALLOC
;
2382 switch (compact_result
) {
2383 case COMPACT_DEFERRED
:
2384 *deferred_compaction
= true;
2386 case COMPACT_SKIPPED
:
2393 * At least in one zone compaction wasn't deferred or skipped, so let's
2394 * count a compaction stall
2396 count_vm_event(COMPACTSTALL
);
2398 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2399 order
, zonelist
, high_zoneidx
,
2400 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2401 preferred_zone
, classzone_idx
, migratetype
);
2404 struct zone
*zone
= page_zone(page
);
2406 zone
->compact_blockskip_flush
= false;
2407 compaction_defer_reset(zone
, order
, true);
2408 count_vm_event(COMPACTSUCCESS
);
2413 * It's bad if compaction run occurs and fails. The most likely reason
2414 * is that pages exist, but not enough to satisfy watermarks.
2416 count_vm_event(COMPACTFAIL
);
2423 static inline struct page
*
2424 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2425 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2426 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2427 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2428 int *contended_compaction
, bool *deferred_compaction
)
2432 #endif /* CONFIG_COMPACTION */
2434 /* Perform direct synchronous page reclaim */
2436 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2437 nodemask_t
*nodemask
)
2439 struct reclaim_state reclaim_state
;
2444 /* We now go into synchronous reclaim */
2445 cpuset_memory_pressure_bump();
2446 current
->flags
|= PF_MEMALLOC
;
2447 lockdep_set_current_reclaim_state(gfp_mask
);
2448 reclaim_state
.reclaimed_slab
= 0;
2449 current
->reclaim_state
= &reclaim_state
;
2451 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2453 current
->reclaim_state
= NULL
;
2454 lockdep_clear_current_reclaim_state();
2455 current
->flags
&= ~PF_MEMALLOC
;
2462 /* The really slow allocator path where we enter direct reclaim */
2463 static inline struct page
*
2464 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2465 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2466 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2467 int classzone_idx
, int migratetype
, unsigned long *did_some_progress
)
2469 struct page
*page
= NULL
;
2470 bool drained
= false;
2472 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2474 if (unlikely(!(*did_some_progress
)))
2477 /* After successful reclaim, reconsider all zones for allocation */
2478 if (IS_ENABLED(CONFIG_NUMA
))
2479 zlc_clear_zones_full(zonelist
);
2482 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2483 zonelist
, high_zoneidx
,
2484 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2485 preferred_zone
, classzone_idx
,
2489 * If an allocation failed after direct reclaim, it could be because
2490 * pages are pinned on the per-cpu lists. Drain them and try again
2492 if (!page
&& !drained
) {
2493 drain_all_pages(NULL
);
2502 * This is called in the allocator slow-path if the allocation request is of
2503 * sufficient urgency to ignore watermarks and take other desperate measures
2505 static inline struct page
*
2506 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2507 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2508 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2509 int classzone_idx
, int migratetype
)
2514 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2515 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2516 preferred_zone
, classzone_idx
, migratetype
);
2518 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2519 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2520 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2525 static void wake_all_kswapds(unsigned int order
,
2526 struct zonelist
*zonelist
,
2527 enum zone_type high_zoneidx
,
2528 struct zone
*preferred_zone
,
2529 nodemask_t
*nodemask
)
2534 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2535 high_zoneidx
, nodemask
)
2536 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2540 gfp_to_alloc_flags(gfp_t gfp_mask
)
2542 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2543 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2545 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2546 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2549 * The caller may dip into page reserves a bit more if the caller
2550 * cannot run direct reclaim, or if the caller has realtime scheduling
2551 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2552 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2554 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2558 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2559 * if it can't schedule.
2561 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2562 alloc_flags
|= ALLOC_HARDER
;
2564 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2565 * comment for __cpuset_node_allowed().
2567 alloc_flags
&= ~ALLOC_CPUSET
;
2568 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2569 alloc_flags
|= ALLOC_HARDER
;
2571 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2572 if (gfp_mask
& __GFP_MEMALLOC
)
2573 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2574 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2575 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2576 else if (!in_interrupt() &&
2577 ((current
->flags
& PF_MEMALLOC
) ||
2578 unlikely(test_thread_flag(TIF_MEMDIE
))))
2579 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2582 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2583 alloc_flags
|= ALLOC_CMA
;
2588 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2590 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2593 static inline struct page
*
2594 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2595 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2596 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2597 int classzone_idx
, int migratetype
)
2599 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2600 struct page
*page
= NULL
;
2602 unsigned long pages_reclaimed
= 0;
2603 unsigned long did_some_progress
;
2604 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2605 bool deferred_compaction
= false;
2606 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2609 * In the slowpath, we sanity check order to avoid ever trying to
2610 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2611 * be using allocators in order of preference for an area that is
2614 if (order
>= MAX_ORDER
) {
2615 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2620 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2621 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2622 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2623 * using a larger set of nodes after it has established that the
2624 * allowed per node queues are empty and that nodes are
2627 if (IS_ENABLED(CONFIG_NUMA
) &&
2628 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2632 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2633 wake_all_kswapds(order
, zonelist
, high_zoneidx
,
2634 preferred_zone
, nodemask
);
2637 * OK, we're below the kswapd watermark and have kicked background
2638 * reclaim. Now things get more complex, so set up alloc_flags according
2639 * to how we want to proceed.
2641 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2644 * Find the true preferred zone if the allocation is unconstrained by
2647 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
) {
2648 struct zoneref
*preferred_zoneref
;
2649 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2650 NULL
, &preferred_zone
);
2651 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2655 /* This is the last chance, in general, before the goto nopage. */
2656 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2657 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2658 preferred_zone
, classzone_idx
, migratetype
);
2662 /* Allocate without watermarks if the context allows */
2663 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2665 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2666 * the allocation is high priority and these type of
2667 * allocations are system rather than user orientated
2669 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2671 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2672 zonelist
, high_zoneidx
, nodemask
,
2673 preferred_zone
, classzone_idx
, migratetype
);
2679 /* Atomic allocations - we can't balance anything */
2682 * All existing users of the deprecated __GFP_NOFAIL are
2683 * blockable, so warn of any new users that actually allow this
2684 * type of allocation to fail.
2686 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2690 /* Avoid recursion of direct reclaim */
2691 if (current
->flags
& PF_MEMALLOC
)
2694 /* Avoid allocations with no watermarks from looping endlessly */
2695 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2699 * Try direct compaction. The first pass is asynchronous. Subsequent
2700 * attempts after direct reclaim are synchronous
2702 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2703 high_zoneidx
, nodemask
, alloc_flags
,
2705 classzone_idx
, migratetype
,
2706 migration_mode
, &contended_compaction
,
2707 &deferred_compaction
);
2711 /* Checks for THP-specific high-order allocations */
2712 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2714 * If compaction is deferred for high-order allocations, it is
2715 * because sync compaction recently failed. If this is the case
2716 * and the caller requested a THP allocation, we do not want
2717 * to heavily disrupt the system, so we fail the allocation
2718 * instead of entering direct reclaim.
2720 if (deferred_compaction
)
2724 * In all zones where compaction was attempted (and not
2725 * deferred or skipped), lock contention has been detected.
2726 * For THP allocation we do not want to disrupt the others
2727 * so we fallback to base pages instead.
2729 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2733 * If compaction was aborted due to need_resched(), we do not
2734 * want to further increase allocation latency, unless it is
2735 * khugepaged trying to collapse.
2737 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2738 && !(current
->flags
& PF_KTHREAD
))
2743 * It can become very expensive to allocate transparent hugepages at
2744 * fault, so use asynchronous memory compaction for THP unless it is
2745 * khugepaged trying to collapse.
2747 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2748 (current
->flags
& PF_KTHREAD
))
2749 migration_mode
= MIGRATE_SYNC_LIGHT
;
2751 /* Try direct reclaim and then allocating */
2752 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2753 zonelist
, high_zoneidx
,
2755 alloc_flags
, preferred_zone
,
2756 classzone_idx
, migratetype
,
2757 &did_some_progress
);
2762 * If we failed to make any progress reclaiming, then we are
2763 * running out of options and have to consider going OOM
2765 if (!did_some_progress
) {
2766 if (oom_gfp_allowed(gfp_mask
)) {
2767 if (oom_killer_disabled
)
2769 /* Coredumps can quickly deplete all memory reserves */
2770 if ((current
->flags
& PF_DUMPCORE
) &&
2771 !(gfp_mask
& __GFP_NOFAIL
))
2773 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2774 zonelist
, high_zoneidx
,
2775 nodemask
, preferred_zone
,
2776 classzone_idx
, migratetype
);
2780 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2782 * The oom killer is not called for high-order
2783 * allocations that may fail, so if no progress
2784 * is being made, there are no other options and
2785 * retrying is unlikely to help.
2787 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2790 * The oom killer is not called for lowmem
2791 * allocations to prevent needlessly killing
2794 if (high_zoneidx
< ZONE_NORMAL
)
2802 /* Check if we should retry the allocation */
2803 pages_reclaimed
+= did_some_progress
;
2804 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2806 /* Wait for some write requests to complete then retry */
2807 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2811 * High-order allocations do not necessarily loop after
2812 * direct reclaim and reclaim/compaction depends on compaction
2813 * being called after reclaim so call directly if necessary
2815 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2816 high_zoneidx
, nodemask
, alloc_flags
,
2818 classzone_idx
, migratetype
,
2819 migration_mode
, &contended_compaction
,
2820 &deferred_compaction
);
2826 warn_alloc_failed(gfp_mask
, order
, NULL
);
2829 if (kmemcheck_enabled
)
2830 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2836 * This is the 'heart' of the zoned buddy allocator.
2839 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2840 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2842 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2843 struct zone
*preferred_zone
;
2844 struct zoneref
*preferred_zoneref
;
2845 struct page
*page
= NULL
;
2846 int migratetype
= gfpflags_to_migratetype(gfp_mask
);
2847 unsigned int cpuset_mems_cookie
;
2848 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2851 gfp_mask
&= gfp_allowed_mask
;
2853 lockdep_trace_alloc(gfp_mask
);
2855 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2857 if (should_fail_alloc_page(gfp_mask
, order
))
2861 * Check the zones suitable for the gfp_mask contain at least one
2862 * valid zone. It's possible to have an empty zonelist as a result
2863 * of GFP_THISNODE and a memoryless node
2865 if (unlikely(!zonelist
->_zonerefs
->zone
))
2868 if (IS_ENABLED(CONFIG_CMA
) && migratetype
== MIGRATE_MOVABLE
)
2869 alloc_flags
|= ALLOC_CMA
;
2872 cpuset_mems_cookie
= read_mems_allowed_begin();
2874 /* The preferred zone is used for statistics later */
2875 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2876 nodemask
? : &cpuset_current_mems_allowed
,
2878 if (!preferred_zone
)
2880 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2882 /* First allocation attempt */
2883 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2884 zonelist
, high_zoneidx
, alloc_flags
,
2885 preferred_zone
, classzone_idx
, migratetype
);
2886 if (unlikely(!page
)) {
2888 * Runtime PM, block IO and its error handling path
2889 * can deadlock because I/O on the device might not
2892 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2893 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2894 zonelist
, high_zoneidx
, nodemask
,
2895 preferred_zone
, classzone_idx
, migratetype
);
2898 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2902 * When updating a task's mems_allowed, it is possible to race with
2903 * parallel threads in such a way that an allocation can fail while
2904 * the mask is being updated. If a page allocation is about to fail,
2905 * check if the cpuset changed during allocation and if so, retry.
2907 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2912 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2915 * Common helper functions.
2917 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2922 * __get_free_pages() returns a 32-bit address, which cannot represent
2925 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2927 page
= alloc_pages(gfp_mask
, order
);
2930 return (unsigned long) page_address(page
);
2932 EXPORT_SYMBOL(__get_free_pages
);
2934 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2936 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2938 EXPORT_SYMBOL(get_zeroed_page
);
2940 void __free_pages(struct page
*page
, unsigned int order
)
2942 if (put_page_testzero(page
)) {
2944 free_hot_cold_page(page
, false);
2946 __free_pages_ok(page
, order
);
2950 EXPORT_SYMBOL(__free_pages
);
2952 void free_pages(unsigned long addr
, unsigned int order
)
2955 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2956 __free_pages(virt_to_page((void *)addr
), order
);
2960 EXPORT_SYMBOL(free_pages
);
2963 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2964 * of the current memory cgroup.
2966 * It should be used when the caller would like to use kmalloc, but since the
2967 * allocation is large, it has to fall back to the page allocator.
2969 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2972 struct mem_cgroup
*memcg
= NULL
;
2974 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2976 page
= alloc_pages(gfp_mask
, order
);
2977 memcg_kmem_commit_charge(page
, memcg
, order
);
2981 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2984 struct mem_cgroup
*memcg
= NULL
;
2986 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2988 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2989 memcg_kmem_commit_charge(page
, memcg
, order
);
2994 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2997 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2999 memcg_kmem_uncharge_pages(page
, order
);
3000 __free_pages(page
, order
);
3003 void free_kmem_pages(unsigned long addr
, unsigned int order
)
3006 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3007 __free_kmem_pages(virt_to_page((void *)addr
), order
);
3011 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
3014 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3015 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3017 split_page(virt_to_page((void *)addr
), order
);
3018 while (used
< alloc_end
) {
3023 return (void *)addr
;
3027 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3028 * @size: the number of bytes to allocate
3029 * @gfp_mask: GFP flags for the allocation
3031 * This function is similar to alloc_pages(), except that it allocates the
3032 * minimum number of pages to satisfy the request. alloc_pages() can only
3033 * allocate memory in power-of-two pages.
3035 * This function is also limited by MAX_ORDER.
3037 * Memory allocated by this function must be released by free_pages_exact().
3039 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3041 unsigned int order
= get_order(size
);
3044 addr
= __get_free_pages(gfp_mask
, order
);
3045 return make_alloc_exact(addr
, order
, size
);
3047 EXPORT_SYMBOL(alloc_pages_exact
);
3050 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3052 * @nid: the preferred node ID where memory should be allocated
3053 * @size: the number of bytes to allocate
3054 * @gfp_mask: GFP flags for the allocation
3056 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3058 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3061 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3063 unsigned order
= get_order(size
);
3064 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3067 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3071 * free_pages_exact - release memory allocated via alloc_pages_exact()
3072 * @virt: the value returned by alloc_pages_exact.
3073 * @size: size of allocation, same value as passed to alloc_pages_exact().
3075 * Release the memory allocated by a previous call to alloc_pages_exact.
3077 void free_pages_exact(void *virt
, size_t size
)
3079 unsigned long addr
= (unsigned long)virt
;
3080 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3082 while (addr
< end
) {
3087 EXPORT_SYMBOL(free_pages_exact
);
3090 * nr_free_zone_pages - count number of pages beyond high watermark
3091 * @offset: The zone index of the highest zone
3093 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3094 * high watermark within all zones at or below a given zone index. For each
3095 * zone, the number of pages is calculated as:
3096 * managed_pages - high_pages
3098 static unsigned long nr_free_zone_pages(int offset
)
3103 /* Just pick one node, since fallback list is circular */
3104 unsigned long sum
= 0;
3106 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3108 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3109 unsigned long size
= zone
->managed_pages
;
3110 unsigned long high
= high_wmark_pages(zone
);
3119 * nr_free_buffer_pages - count number of pages beyond high watermark
3121 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3122 * watermark within ZONE_DMA and ZONE_NORMAL.
3124 unsigned long nr_free_buffer_pages(void)
3126 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3128 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3131 * nr_free_pagecache_pages - count number of pages beyond high watermark
3133 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3134 * high watermark within all zones.
3136 unsigned long nr_free_pagecache_pages(void)
3138 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3141 static inline void show_node(struct zone
*zone
)
3143 if (IS_ENABLED(CONFIG_NUMA
))
3144 printk("Node %d ", zone_to_nid(zone
));
3147 void si_meminfo(struct sysinfo
*val
)
3149 val
->totalram
= totalram_pages
;
3150 val
->sharedram
= global_page_state(NR_SHMEM
);
3151 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3152 val
->bufferram
= nr_blockdev_pages();
3153 val
->totalhigh
= totalhigh_pages
;
3154 val
->freehigh
= nr_free_highpages();
3155 val
->mem_unit
= PAGE_SIZE
;
3158 EXPORT_SYMBOL(si_meminfo
);
3161 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3163 int zone_type
; /* needs to be signed */
3164 unsigned long managed_pages
= 0;
3165 pg_data_t
*pgdat
= NODE_DATA(nid
);
3167 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3168 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3169 val
->totalram
= managed_pages
;
3170 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3171 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3172 #ifdef CONFIG_HIGHMEM
3173 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3174 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3180 val
->mem_unit
= PAGE_SIZE
;
3185 * Determine whether the node should be displayed or not, depending on whether
3186 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3188 bool skip_free_areas_node(unsigned int flags
, int nid
)
3191 unsigned int cpuset_mems_cookie
;
3193 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3197 cpuset_mems_cookie
= read_mems_allowed_begin();
3198 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3199 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3204 #define K(x) ((x) << (PAGE_SHIFT-10))
3206 static void show_migration_types(unsigned char type
)
3208 static const char types
[MIGRATE_TYPES
] = {
3209 [MIGRATE_UNMOVABLE
] = 'U',
3210 [MIGRATE_RECLAIMABLE
] = 'E',
3211 [MIGRATE_MOVABLE
] = 'M',
3212 [MIGRATE_RESERVE
] = 'R',
3214 [MIGRATE_CMA
] = 'C',
3216 #ifdef CONFIG_MEMORY_ISOLATION
3217 [MIGRATE_ISOLATE
] = 'I',
3220 char tmp
[MIGRATE_TYPES
+ 1];
3224 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3225 if (type
& (1 << i
))
3230 printk("(%s) ", tmp
);
3234 * Show free area list (used inside shift_scroll-lock stuff)
3235 * We also calculate the percentage fragmentation. We do this by counting the
3236 * memory on each free list with the exception of the first item on the list.
3237 * Suppresses nodes that are not allowed by current's cpuset if
3238 * SHOW_MEM_FILTER_NODES is passed.
3240 void show_free_areas(unsigned int filter
)
3245 for_each_populated_zone(zone
) {
3246 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3249 printk("%s per-cpu:\n", zone
->name
);
3251 for_each_online_cpu(cpu
) {
3252 struct per_cpu_pageset
*pageset
;
3254 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3256 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3257 cpu
, pageset
->pcp
.high
,
3258 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3262 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3263 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3265 " dirty:%lu writeback:%lu unstable:%lu\n"
3266 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3267 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3269 global_page_state(NR_ACTIVE_ANON
),
3270 global_page_state(NR_INACTIVE_ANON
),
3271 global_page_state(NR_ISOLATED_ANON
),
3272 global_page_state(NR_ACTIVE_FILE
),
3273 global_page_state(NR_INACTIVE_FILE
),
3274 global_page_state(NR_ISOLATED_FILE
),
3275 global_page_state(NR_UNEVICTABLE
),
3276 global_page_state(NR_FILE_DIRTY
),
3277 global_page_state(NR_WRITEBACK
),
3278 global_page_state(NR_UNSTABLE_NFS
),
3279 global_page_state(NR_FREE_PAGES
),
3280 global_page_state(NR_SLAB_RECLAIMABLE
),
3281 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3282 global_page_state(NR_FILE_MAPPED
),
3283 global_page_state(NR_SHMEM
),
3284 global_page_state(NR_PAGETABLE
),
3285 global_page_state(NR_BOUNCE
),
3286 global_page_state(NR_FREE_CMA_PAGES
));
3288 for_each_populated_zone(zone
) {
3291 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3299 " active_anon:%lukB"
3300 " inactive_anon:%lukB"
3301 " active_file:%lukB"
3302 " inactive_file:%lukB"
3303 " unevictable:%lukB"
3304 " isolated(anon):%lukB"
3305 " isolated(file):%lukB"
3313 " slab_reclaimable:%lukB"
3314 " slab_unreclaimable:%lukB"
3315 " kernel_stack:%lukB"
3320 " writeback_tmp:%lukB"
3321 " pages_scanned:%lu"
3322 " all_unreclaimable? %s"
3325 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3326 K(min_wmark_pages(zone
)),
3327 K(low_wmark_pages(zone
)),
3328 K(high_wmark_pages(zone
)),
3329 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3330 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3331 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3332 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3333 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3334 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3335 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3336 K(zone
->present_pages
),
3337 K(zone
->managed_pages
),
3338 K(zone_page_state(zone
, NR_MLOCK
)),
3339 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3340 K(zone_page_state(zone
, NR_WRITEBACK
)),
3341 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3342 K(zone_page_state(zone
, NR_SHMEM
)),
3343 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3344 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3345 zone_page_state(zone
, NR_KERNEL_STACK
) *
3347 K(zone_page_state(zone
, NR_PAGETABLE
)),
3348 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3349 K(zone_page_state(zone
, NR_BOUNCE
)),
3350 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3351 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3352 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3353 (!zone_reclaimable(zone
) ? "yes" : "no")
3355 printk("lowmem_reserve[]:");
3356 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3357 printk(" %ld", zone
->lowmem_reserve
[i
]);
3361 for_each_populated_zone(zone
) {
3362 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3363 unsigned char types
[MAX_ORDER
];
3365 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3368 printk("%s: ", zone
->name
);
3370 spin_lock_irqsave(&zone
->lock
, flags
);
3371 for (order
= 0; order
< MAX_ORDER
; order
++) {
3372 struct free_area
*area
= &zone
->free_area
[order
];
3375 nr
[order
] = area
->nr_free
;
3376 total
+= nr
[order
] << order
;
3379 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3380 if (!list_empty(&area
->free_list
[type
]))
3381 types
[order
] |= 1 << type
;
3384 spin_unlock_irqrestore(&zone
->lock
, flags
);
3385 for (order
= 0; order
< MAX_ORDER
; order
++) {
3386 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3388 show_migration_types(types
[order
]);
3390 printk("= %lukB\n", K(total
));
3393 hugetlb_show_meminfo();
3395 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3397 show_swap_cache_info();
3400 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3402 zoneref
->zone
= zone
;
3403 zoneref
->zone_idx
= zone_idx(zone
);
3407 * Builds allocation fallback zone lists.
3409 * Add all populated zones of a node to the zonelist.
3411 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3415 enum zone_type zone_type
= MAX_NR_ZONES
;
3419 zone
= pgdat
->node_zones
+ zone_type
;
3420 if (populated_zone(zone
)) {
3421 zoneref_set_zone(zone
,
3422 &zonelist
->_zonerefs
[nr_zones
++]);
3423 check_highest_zone(zone_type
);
3425 } while (zone_type
);
3433 * 0 = automatic detection of better ordering.
3434 * 1 = order by ([node] distance, -zonetype)
3435 * 2 = order by (-zonetype, [node] distance)
3437 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3438 * the same zonelist. So only NUMA can configure this param.
3440 #define ZONELIST_ORDER_DEFAULT 0
3441 #define ZONELIST_ORDER_NODE 1
3442 #define ZONELIST_ORDER_ZONE 2
3444 /* zonelist order in the kernel.
3445 * set_zonelist_order() will set this to NODE or ZONE.
3447 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3448 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3452 /* The value user specified ....changed by config */
3453 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3454 /* string for sysctl */
3455 #define NUMA_ZONELIST_ORDER_LEN 16
3456 char numa_zonelist_order
[16] = "default";
3459 * interface for configure zonelist ordering.
3460 * command line option "numa_zonelist_order"
3461 * = "[dD]efault - default, automatic configuration.
3462 * = "[nN]ode - order by node locality, then by zone within node
3463 * = "[zZ]one - order by zone, then by locality within zone
3466 static int __parse_numa_zonelist_order(char *s
)
3468 if (*s
== 'd' || *s
== 'D') {
3469 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3470 } else if (*s
== 'n' || *s
== 'N') {
3471 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3472 } else if (*s
== 'z' || *s
== 'Z') {
3473 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3476 "Ignoring invalid numa_zonelist_order value: "
3483 static __init
int setup_numa_zonelist_order(char *s
)
3490 ret
= __parse_numa_zonelist_order(s
);
3492 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3496 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3499 * sysctl handler for numa_zonelist_order
3501 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3502 void __user
*buffer
, size_t *length
,
3505 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3507 static DEFINE_MUTEX(zl_order_mutex
);
3509 mutex_lock(&zl_order_mutex
);
3511 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3515 strcpy(saved_string
, (char *)table
->data
);
3517 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3521 int oldval
= user_zonelist_order
;
3523 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3526 * bogus value. restore saved string
3528 strncpy((char *)table
->data
, saved_string
,
3529 NUMA_ZONELIST_ORDER_LEN
);
3530 user_zonelist_order
= oldval
;
3531 } else if (oldval
!= user_zonelist_order
) {
3532 mutex_lock(&zonelists_mutex
);
3533 build_all_zonelists(NULL
, NULL
);
3534 mutex_unlock(&zonelists_mutex
);
3538 mutex_unlock(&zl_order_mutex
);
3543 #define MAX_NODE_LOAD (nr_online_nodes)
3544 static int node_load
[MAX_NUMNODES
];
3547 * find_next_best_node - find the next node that should appear in a given node's fallback list
3548 * @node: node whose fallback list we're appending
3549 * @used_node_mask: nodemask_t of already used nodes
3551 * We use a number of factors to determine which is the next node that should
3552 * appear on a given node's fallback list. The node should not have appeared
3553 * already in @node's fallback list, and it should be the next closest node
3554 * according to the distance array (which contains arbitrary distance values
3555 * from each node to each node in the system), and should also prefer nodes
3556 * with no CPUs, since presumably they'll have very little allocation pressure
3557 * on them otherwise.
3558 * It returns -1 if no node is found.
3560 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3563 int min_val
= INT_MAX
;
3564 int best_node
= NUMA_NO_NODE
;
3565 const struct cpumask
*tmp
= cpumask_of_node(0);
3567 /* Use the local node if we haven't already */
3568 if (!node_isset(node
, *used_node_mask
)) {
3569 node_set(node
, *used_node_mask
);
3573 for_each_node_state(n
, N_MEMORY
) {
3575 /* Don't want a node to appear more than once */
3576 if (node_isset(n
, *used_node_mask
))
3579 /* Use the distance array to find the distance */
3580 val
= node_distance(node
, n
);
3582 /* Penalize nodes under us ("prefer the next node") */
3585 /* Give preference to headless and unused nodes */
3586 tmp
= cpumask_of_node(n
);
3587 if (!cpumask_empty(tmp
))
3588 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3590 /* Slight preference for less loaded node */
3591 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3592 val
+= node_load
[n
];
3594 if (val
< min_val
) {
3601 node_set(best_node
, *used_node_mask
);
3608 * Build zonelists ordered by node and zones within node.
3609 * This results in maximum locality--normal zone overflows into local
3610 * DMA zone, if any--but risks exhausting DMA zone.
3612 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3615 struct zonelist
*zonelist
;
3617 zonelist
= &pgdat
->node_zonelists
[0];
3618 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3620 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3621 zonelist
->_zonerefs
[j
].zone
= NULL
;
3622 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3626 * Build gfp_thisnode zonelists
3628 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3631 struct zonelist
*zonelist
;
3633 zonelist
= &pgdat
->node_zonelists
[1];
3634 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3635 zonelist
->_zonerefs
[j
].zone
= NULL
;
3636 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3640 * Build zonelists ordered by zone and nodes within zones.
3641 * This results in conserving DMA zone[s] until all Normal memory is
3642 * exhausted, but results in overflowing to remote node while memory
3643 * may still exist in local DMA zone.
3645 static int node_order
[MAX_NUMNODES
];
3647 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3650 int zone_type
; /* needs to be signed */
3652 struct zonelist
*zonelist
;
3654 zonelist
= &pgdat
->node_zonelists
[0];
3656 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3657 for (j
= 0; j
< nr_nodes
; j
++) {
3658 node
= node_order
[j
];
3659 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3660 if (populated_zone(z
)) {
3662 &zonelist
->_zonerefs
[pos
++]);
3663 check_highest_zone(zone_type
);
3667 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3668 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3671 #if defined(CONFIG_64BIT)
3673 * Devices that require DMA32/DMA are relatively rare and do not justify a
3674 * penalty to every machine in case the specialised case applies. Default
3675 * to Node-ordering on 64-bit NUMA machines
3677 static int default_zonelist_order(void)
3679 return ZONELIST_ORDER_NODE
;
3683 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3684 * by the kernel. If processes running on node 0 deplete the low memory zone
3685 * then reclaim will occur more frequency increasing stalls and potentially
3686 * be easier to OOM if a large percentage of the zone is under writeback or
3687 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3688 * Hence, default to zone ordering on 32-bit.
3690 static int default_zonelist_order(void)
3692 return ZONELIST_ORDER_ZONE
;
3694 #endif /* CONFIG_64BIT */
3696 static void set_zonelist_order(void)
3698 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3699 current_zonelist_order
= default_zonelist_order();
3701 current_zonelist_order
= user_zonelist_order
;
3704 static void build_zonelists(pg_data_t
*pgdat
)
3708 nodemask_t used_mask
;
3709 int local_node
, prev_node
;
3710 struct zonelist
*zonelist
;
3711 int order
= current_zonelist_order
;
3713 /* initialize zonelists */
3714 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3715 zonelist
= pgdat
->node_zonelists
+ i
;
3716 zonelist
->_zonerefs
[0].zone
= NULL
;
3717 zonelist
->_zonerefs
[0].zone_idx
= 0;
3720 /* NUMA-aware ordering of nodes */
3721 local_node
= pgdat
->node_id
;
3722 load
= nr_online_nodes
;
3723 prev_node
= local_node
;
3724 nodes_clear(used_mask
);
3726 memset(node_order
, 0, sizeof(node_order
));
3729 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3731 * We don't want to pressure a particular node.
3732 * So adding penalty to the first node in same
3733 * distance group to make it round-robin.
3735 if (node_distance(local_node
, node
) !=
3736 node_distance(local_node
, prev_node
))
3737 node_load
[node
] = load
;
3741 if (order
== ZONELIST_ORDER_NODE
)
3742 build_zonelists_in_node_order(pgdat
, node
);
3744 node_order
[j
++] = node
; /* remember order */
3747 if (order
== ZONELIST_ORDER_ZONE
) {
3748 /* calculate node order -- i.e., DMA last! */
3749 build_zonelists_in_zone_order(pgdat
, j
);
3752 build_thisnode_zonelists(pgdat
);
3755 /* Construct the zonelist performance cache - see further mmzone.h */
3756 static void build_zonelist_cache(pg_data_t
*pgdat
)
3758 struct zonelist
*zonelist
;
3759 struct zonelist_cache
*zlc
;
3762 zonelist
= &pgdat
->node_zonelists
[0];
3763 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3764 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3765 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3766 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3769 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3771 * Return node id of node used for "local" allocations.
3772 * I.e., first node id of first zone in arg node's generic zonelist.
3773 * Used for initializing percpu 'numa_mem', which is used primarily
3774 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3776 int local_memory_node(int node
)
3780 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3781 gfp_zone(GFP_KERNEL
),
3788 #else /* CONFIG_NUMA */
3790 static void set_zonelist_order(void)
3792 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3795 static void build_zonelists(pg_data_t
*pgdat
)
3797 int node
, local_node
;
3799 struct zonelist
*zonelist
;
3801 local_node
= pgdat
->node_id
;
3803 zonelist
= &pgdat
->node_zonelists
[0];
3804 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3807 * Now we build the zonelist so that it contains the zones
3808 * of all the other nodes.
3809 * We don't want to pressure a particular node, so when
3810 * building the zones for node N, we make sure that the
3811 * zones coming right after the local ones are those from
3812 * node N+1 (modulo N)
3814 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3815 if (!node_online(node
))
3817 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3819 for (node
= 0; node
< local_node
; node
++) {
3820 if (!node_online(node
))
3822 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3825 zonelist
->_zonerefs
[j
].zone
= NULL
;
3826 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3829 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3830 static void build_zonelist_cache(pg_data_t
*pgdat
)
3832 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3835 #endif /* CONFIG_NUMA */
3838 * Boot pageset table. One per cpu which is going to be used for all
3839 * zones and all nodes. The parameters will be set in such a way
3840 * that an item put on a list will immediately be handed over to
3841 * the buddy list. This is safe since pageset manipulation is done
3842 * with interrupts disabled.
3844 * The boot_pagesets must be kept even after bootup is complete for
3845 * unused processors and/or zones. They do play a role for bootstrapping
3846 * hotplugged processors.
3848 * zoneinfo_show() and maybe other functions do
3849 * not check if the processor is online before following the pageset pointer.
3850 * Other parts of the kernel may not check if the zone is available.
3852 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3853 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3854 static void setup_zone_pageset(struct zone
*zone
);
3857 * Global mutex to protect against size modification of zonelists
3858 * as well as to serialize pageset setup for the new populated zone.
3860 DEFINE_MUTEX(zonelists_mutex
);
3862 /* return values int ....just for stop_machine() */
3863 static int __build_all_zonelists(void *data
)
3867 pg_data_t
*self
= data
;
3870 memset(node_load
, 0, sizeof(node_load
));
3873 if (self
&& !node_online(self
->node_id
)) {
3874 build_zonelists(self
);
3875 build_zonelist_cache(self
);
3878 for_each_online_node(nid
) {
3879 pg_data_t
*pgdat
= NODE_DATA(nid
);
3881 build_zonelists(pgdat
);
3882 build_zonelist_cache(pgdat
);
3886 * Initialize the boot_pagesets that are going to be used
3887 * for bootstrapping processors. The real pagesets for
3888 * each zone will be allocated later when the per cpu
3889 * allocator is available.
3891 * boot_pagesets are used also for bootstrapping offline
3892 * cpus if the system is already booted because the pagesets
3893 * are needed to initialize allocators on a specific cpu too.
3894 * F.e. the percpu allocator needs the page allocator which
3895 * needs the percpu allocator in order to allocate its pagesets
3896 * (a chicken-egg dilemma).
3898 for_each_possible_cpu(cpu
) {
3899 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3901 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3903 * We now know the "local memory node" for each node--
3904 * i.e., the node of the first zone in the generic zonelist.
3905 * Set up numa_mem percpu variable for on-line cpus. During
3906 * boot, only the boot cpu should be on-line; we'll init the
3907 * secondary cpus' numa_mem as they come on-line. During
3908 * node/memory hotplug, we'll fixup all on-line cpus.
3910 if (cpu_online(cpu
))
3911 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3919 * Called with zonelists_mutex held always
3920 * unless system_state == SYSTEM_BOOTING.
3922 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3924 set_zonelist_order();
3926 if (system_state
== SYSTEM_BOOTING
) {
3927 __build_all_zonelists(NULL
);
3928 mminit_verify_zonelist();
3929 cpuset_init_current_mems_allowed();
3931 #ifdef CONFIG_MEMORY_HOTPLUG
3933 setup_zone_pageset(zone
);
3935 /* we have to stop all cpus to guarantee there is no user
3937 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3938 /* cpuset refresh routine should be here */
3940 vm_total_pages
= nr_free_pagecache_pages();
3942 * Disable grouping by mobility if the number of pages in the
3943 * system is too low to allow the mechanism to work. It would be
3944 * more accurate, but expensive to check per-zone. This check is
3945 * made on memory-hotadd so a system can start with mobility
3946 * disabled and enable it later
3948 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3949 page_group_by_mobility_disabled
= 1;
3951 page_group_by_mobility_disabled
= 0;
3953 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3954 "Total pages: %ld\n",
3956 zonelist_order_name
[current_zonelist_order
],
3957 page_group_by_mobility_disabled
? "off" : "on",
3960 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
3965 * Helper functions to size the waitqueue hash table.
3966 * Essentially these want to choose hash table sizes sufficiently
3967 * large so that collisions trying to wait on pages are rare.
3968 * But in fact, the number of active page waitqueues on typical
3969 * systems is ridiculously low, less than 200. So this is even
3970 * conservative, even though it seems large.
3972 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3973 * waitqueues, i.e. the size of the waitq table given the number of pages.
3975 #define PAGES_PER_WAITQUEUE 256
3977 #ifndef CONFIG_MEMORY_HOTPLUG
3978 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3980 unsigned long size
= 1;
3982 pages
/= PAGES_PER_WAITQUEUE
;
3984 while (size
< pages
)
3988 * Once we have dozens or even hundreds of threads sleeping
3989 * on IO we've got bigger problems than wait queue collision.
3990 * Limit the size of the wait table to a reasonable size.
3992 size
= min(size
, 4096UL);
3994 return max(size
, 4UL);
3998 * A zone's size might be changed by hot-add, so it is not possible to determine
3999 * a suitable size for its wait_table. So we use the maximum size now.
4001 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4003 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4004 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4005 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4007 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4008 * or more by the traditional way. (See above). It equals:
4010 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4011 * ia64(16K page size) : = ( 8G + 4M)byte.
4012 * powerpc (64K page size) : = (32G +16M)byte.
4014 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4021 * This is an integer logarithm so that shifts can be used later
4022 * to extract the more random high bits from the multiplicative
4023 * hash function before the remainder is taken.
4025 static inline unsigned long wait_table_bits(unsigned long size
)
4031 * Check if a pageblock contains reserved pages
4033 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
4037 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4038 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
4045 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4046 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4047 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4048 * higher will lead to a bigger reserve which will get freed as contiguous
4049 * blocks as reclaim kicks in
4051 static void setup_zone_migrate_reserve(struct zone
*zone
)
4053 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4055 unsigned long block_migratetype
;
4060 * Get the start pfn, end pfn and the number of blocks to reserve
4061 * We have to be careful to be aligned to pageblock_nr_pages to
4062 * make sure that we always check pfn_valid for the first page in
4065 start_pfn
= zone
->zone_start_pfn
;
4066 end_pfn
= zone_end_pfn(zone
);
4067 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4068 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4072 * Reserve blocks are generally in place to help high-order atomic
4073 * allocations that are short-lived. A min_free_kbytes value that
4074 * would result in more than 2 reserve blocks for atomic allocations
4075 * is assumed to be in place to help anti-fragmentation for the
4076 * future allocation of hugepages at runtime.
4078 reserve
= min(2, reserve
);
4079 old_reserve
= zone
->nr_migrate_reserve_block
;
4081 /* When memory hot-add, we almost always need to do nothing */
4082 if (reserve
== old_reserve
)
4084 zone
->nr_migrate_reserve_block
= reserve
;
4086 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4087 if (!pfn_valid(pfn
))
4089 page
= pfn_to_page(pfn
);
4091 /* Watch out for overlapping nodes */
4092 if (page_to_nid(page
) != zone_to_nid(zone
))
4095 block_migratetype
= get_pageblock_migratetype(page
);
4097 /* Only test what is necessary when the reserves are not met */
4100 * Blocks with reserved pages will never free, skip
4103 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4104 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4107 /* If this block is reserved, account for it */
4108 if (block_migratetype
== MIGRATE_RESERVE
) {
4113 /* Suitable for reserving if this block is movable */
4114 if (block_migratetype
== MIGRATE_MOVABLE
) {
4115 set_pageblock_migratetype(page
,
4117 move_freepages_block(zone
, page
,
4122 } else if (!old_reserve
) {
4124 * At boot time we don't need to scan the whole zone
4125 * for turning off MIGRATE_RESERVE.
4131 * If the reserve is met and this is a previous reserved block,
4134 if (block_migratetype
== MIGRATE_RESERVE
) {
4135 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4136 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4142 * Initially all pages are reserved - free ones are freed
4143 * up by free_all_bootmem() once the early boot process is
4144 * done. Non-atomic initialization, single-pass.
4146 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4147 unsigned long start_pfn
, enum memmap_context context
)
4150 unsigned long end_pfn
= start_pfn
+ size
;
4154 if (highest_memmap_pfn
< end_pfn
- 1)
4155 highest_memmap_pfn
= end_pfn
- 1;
4157 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4158 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4160 * There can be holes in boot-time mem_map[]s
4161 * handed to this function. They do not
4162 * exist on hotplugged memory.
4164 if (context
== MEMMAP_EARLY
) {
4165 if (!early_pfn_valid(pfn
))
4167 if (!early_pfn_in_nid(pfn
, nid
))
4170 page
= pfn_to_page(pfn
);
4171 set_page_links(page
, zone
, nid
, pfn
);
4172 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4173 init_page_count(page
);
4174 page_mapcount_reset(page
);
4175 page_cpupid_reset_last(page
);
4176 SetPageReserved(page
);
4178 * Mark the block movable so that blocks are reserved for
4179 * movable at startup. This will force kernel allocations
4180 * to reserve their blocks rather than leaking throughout
4181 * the address space during boot when many long-lived
4182 * kernel allocations are made. Later some blocks near
4183 * the start are marked MIGRATE_RESERVE by
4184 * setup_zone_migrate_reserve()
4186 * bitmap is created for zone's valid pfn range. but memmap
4187 * can be created for invalid pages (for alignment)
4188 * check here not to call set_pageblock_migratetype() against
4191 if ((z
->zone_start_pfn
<= pfn
)
4192 && (pfn
< zone_end_pfn(z
))
4193 && !(pfn
& (pageblock_nr_pages
- 1)))
4194 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4196 INIT_LIST_HEAD(&page
->lru
);
4197 #ifdef WANT_PAGE_VIRTUAL
4198 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4199 if (!is_highmem_idx(zone
))
4200 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4205 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4207 unsigned int order
, t
;
4208 for_each_migratetype_order(order
, t
) {
4209 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4210 zone
->free_area
[order
].nr_free
= 0;
4214 #ifndef __HAVE_ARCH_MEMMAP_INIT
4215 #define memmap_init(size, nid, zone, start_pfn) \
4216 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4219 static int zone_batchsize(struct zone
*zone
)
4225 * The per-cpu-pages pools are set to around 1000th of the
4226 * size of the zone. But no more than 1/2 of a meg.
4228 * OK, so we don't know how big the cache is. So guess.
4230 batch
= zone
->managed_pages
/ 1024;
4231 if (batch
* PAGE_SIZE
> 512 * 1024)
4232 batch
= (512 * 1024) / PAGE_SIZE
;
4233 batch
/= 4; /* We effectively *= 4 below */
4238 * Clamp the batch to a 2^n - 1 value. Having a power
4239 * of 2 value was found to be more likely to have
4240 * suboptimal cache aliasing properties in some cases.
4242 * For example if 2 tasks are alternately allocating
4243 * batches of pages, one task can end up with a lot
4244 * of pages of one half of the possible page colors
4245 * and the other with pages of the other colors.
4247 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4252 /* The deferral and batching of frees should be suppressed under NOMMU
4255 * The problem is that NOMMU needs to be able to allocate large chunks
4256 * of contiguous memory as there's no hardware page translation to
4257 * assemble apparent contiguous memory from discontiguous pages.
4259 * Queueing large contiguous runs of pages for batching, however,
4260 * causes the pages to actually be freed in smaller chunks. As there
4261 * can be a significant delay between the individual batches being
4262 * recycled, this leads to the once large chunks of space being
4263 * fragmented and becoming unavailable for high-order allocations.
4270 * pcp->high and pcp->batch values are related and dependent on one another:
4271 * ->batch must never be higher then ->high.
4272 * The following function updates them in a safe manner without read side
4275 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4276 * those fields changing asynchronously (acording the the above rule).
4278 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4279 * outside of boot time (or some other assurance that no concurrent updaters
4282 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4283 unsigned long batch
)
4285 /* start with a fail safe value for batch */
4289 /* Update high, then batch, in order */
4296 /* a companion to pageset_set_high() */
4297 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4299 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4302 static void pageset_init(struct per_cpu_pageset
*p
)
4304 struct per_cpu_pages
*pcp
;
4307 memset(p
, 0, sizeof(*p
));
4311 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4312 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4315 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4318 pageset_set_batch(p
, batch
);
4322 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4323 * to the value high for the pageset p.
4325 static void pageset_set_high(struct per_cpu_pageset
*p
,
4328 unsigned long batch
= max(1UL, high
/ 4);
4329 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4330 batch
= PAGE_SHIFT
* 8;
4332 pageset_update(&p
->pcp
, high
, batch
);
4335 static void pageset_set_high_and_batch(struct zone
*zone
,
4336 struct per_cpu_pageset
*pcp
)
4338 if (percpu_pagelist_fraction
)
4339 pageset_set_high(pcp
,
4340 (zone
->managed_pages
/
4341 percpu_pagelist_fraction
));
4343 pageset_set_batch(pcp
, zone_batchsize(zone
));
4346 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4348 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4351 pageset_set_high_and_batch(zone
, pcp
);
4354 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4357 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4358 for_each_possible_cpu(cpu
)
4359 zone_pageset_init(zone
, cpu
);
4363 * Allocate per cpu pagesets and initialize them.
4364 * Before this call only boot pagesets were available.
4366 void __init
setup_per_cpu_pageset(void)
4370 for_each_populated_zone(zone
)
4371 setup_zone_pageset(zone
);
4374 static noinline __init_refok
4375 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4381 * The per-page waitqueue mechanism uses hashed waitqueues
4384 zone
->wait_table_hash_nr_entries
=
4385 wait_table_hash_nr_entries(zone_size_pages
);
4386 zone
->wait_table_bits
=
4387 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4388 alloc_size
= zone
->wait_table_hash_nr_entries
4389 * sizeof(wait_queue_head_t
);
4391 if (!slab_is_available()) {
4392 zone
->wait_table
= (wait_queue_head_t
*)
4393 memblock_virt_alloc_node_nopanic(
4394 alloc_size
, zone
->zone_pgdat
->node_id
);
4397 * This case means that a zone whose size was 0 gets new memory
4398 * via memory hot-add.
4399 * But it may be the case that a new node was hot-added. In
4400 * this case vmalloc() will not be able to use this new node's
4401 * memory - this wait_table must be initialized to use this new
4402 * node itself as well.
4403 * To use this new node's memory, further consideration will be
4406 zone
->wait_table
= vmalloc(alloc_size
);
4408 if (!zone
->wait_table
)
4411 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4412 init_waitqueue_head(zone
->wait_table
+ i
);
4417 static __meminit
void zone_pcp_init(struct zone
*zone
)
4420 * per cpu subsystem is not up at this point. The following code
4421 * relies on the ability of the linker to provide the
4422 * offset of a (static) per cpu variable into the per cpu area.
4424 zone
->pageset
= &boot_pageset
;
4426 if (populated_zone(zone
))
4427 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4428 zone
->name
, zone
->present_pages
,
4429 zone_batchsize(zone
));
4432 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4433 unsigned long zone_start_pfn
,
4435 enum memmap_context context
)
4437 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4439 ret
= zone_wait_table_init(zone
, size
);
4442 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4444 zone
->zone_start_pfn
= zone_start_pfn
;
4446 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4447 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4449 (unsigned long)zone_idx(zone
),
4450 zone_start_pfn
, (zone_start_pfn
+ size
));
4452 zone_init_free_lists(zone
);
4457 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4458 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4460 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4462 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4464 unsigned long start_pfn
, end_pfn
;
4467 * NOTE: The following SMP-unsafe globals are only used early in boot
4468 * when the kernel is running single-threaded.
4470 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4471 static int __meminitdata last_nid
;
4473 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4476 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4478 last_start_pfn
= start_pfn
;
4479 last_end_pfn
= end_pfn
;
4485 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4487 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4491 nid
= __early_pfn_to_nid(pfn
);
4494 /* just returns 0 */
4498 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4499 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4503 nid
= __early_pfn_to_nid(pfn
);
4504 if (nid
>= 0 && nid
!= node
)
4511 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4512 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4513 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4515 * If an architecture guarantees that all ranges registered contain no holes
4516 * and may be freed, this this function may be used instead of calling
4517 * memblock_free_early_nid() manually.
4519 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4521 unsigned long start_pfn
, end_pfn
;
4524 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4525 start_pfn
= min(start_pfn
, max_low_pfn
);
4526 end_pfn
= min(end_pfn
, max_low_pfn
);
4528 if (start_pfn
< end_pfn
)
4529 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4530 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4536 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4537 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4539 * If an architecture guarantees that all ranges registered contain no holes and may
4540 * be freed, this function may be used instead of calling memory_present() manually.
4542 void __init
sparse_memory_present_with_active_regions(int nid
)
4544 unsigned long start_pfn
, end_pfn
;
4547 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4548 memory_present(this_nid
, start_pfn
, end_pfn
);
4552 * get_pfn_range_for_nid - Return the start and end page frames for a node
4553 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4554 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4555 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4557 * It returns the start and end page frame of a node based on information
4558 * provided by memblock_set_node(). If called for a node
4559 * with no available memory, a warning is printed and the start and end
4562 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4563 unsigned long *start_pfn
, unsigned long *end_pfn
)
4565 unsigned long this_start_pfn
, this_end_pfn
;
4571 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4572 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4573 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4576 if (*start_pfn
== -1UL)
4581 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4582 * assumption is made that zones within a node are ordered in monotonic
4583 * increasing memory addresses so that the "highest" populated zone is used
4585 static void __init
find_usable_zone_for_movable(void)
4588 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4589 if (zone_index
== ZONE_MOVABLE
)
4592 if (arch_zone_highest_possible_pfn
[zone_index
] >
4593 arch_zone_lowest_possible_pfn
[zone_index
])
4597 VM_BUG_ON(zone_index
== -1);
4598 movable_zone
= zone_index
;
4602 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4603 * because it is sized independent of architecture. Unlike the other zones,
4604 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4605 * in each node depending on the size of each node and how evenly kernelcore
4606 * is distributed. This helper function adjusts the zone ranges
4607 * provided by the architecture for a given node by using the end of the
4608 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4609 * zones within a node are in order of monotonic increases memory addresses
4611 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4612 unsigned long zone_type
,
4613 unsigned long node_start_pfn
,
4614 unsigned long node_end_pfn
,
4615 unsigned long *zone_start_pfn
,
4616 unsigned long *zone_end_pfn
)
4618 /* Only adjust if ZONE_MOVABLE is on this node */
4619 if (zone_movable_pfn
[nid
]) {
4620 /* Size ZONE_MOVABLE */
4621 if (zone_type
== ZONE_MOVABLE
) {
4622 *zone_start_pfn
= zone_movable_pfn
[nid
];
4623 *zone_end_pfn
= min(node_end_pfn
,
4624 arch_zone_highest_possible_pfn
[movable_zone
]);
4626 /* Adjust for ZONE_MOVABLE starting within this range */
4627 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4628 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4629 *zone_end_pfn
= zone_movable_pfn
[nid
];
4631 /* Check if this whole range is within ZONE_MOVABLE */
4632 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4633 *zone_start_pfn
= *zone_end_pfn
;
4638 * Return the number of pages a zone spans in a node, including holes
4639 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4641 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4642 unsigned long zone_type
,
4643 unsigned long node_start_pfn
,
4644 unsigned long node_end_pfn
,
4645 unsigned long *ignored
)
4647 unsigned long zone_start_pfn
, zone_end_pfn
;
4649 /* Get the start and end of the zone */
4650 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4651 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4652 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4653 node_start_pfn
, node_end_pfn
,
4654 &zone_start_pfn
, &zone_end_pfn
);
4656 /* Check that this node has pages within the zone's required range */
4657 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4660 /* Move the zone boundaries inside the node if necessary */
4661 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4662 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4664 /* Return the spanned pages */
4665 return zone_end_pfn
- zone_start_pfn
;
4669 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4670 * then all holes in the requested range will be accounted for.
4672 unsigned long __meminit
__absent_pages_in_range(int nid
,
4673 unsigned long range_start_pfn
,
4674 unsigned long range_end_pfn
)
4676 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4677 unsigned long start_pfn
, end_pfn
;
4680 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4681 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4682 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4683 nr_absent
-= end_pfn
- start_pfn
;
4689 * absent_pages_in_range - Return number of page frames in holes within a range
4690 * @start_pfn: The start PFN to start searching for holes
4691 * @end_pfn: The end PFN to stop searching for holes
4693 * It returns the number of pages frames in memory holes within a range.
4695 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4696 unsigned long end_pfn
)
4698 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4701 /* Return the number of page frames in holes in a zone on a node */
4702 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4703 unsigned long zone_type
,
4704 unsigned long node_start_pfn
,
4705 unsigned long node_end_pfn
,
4706 unsigned long *ignored
)
4708 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4709 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4710 unsigned long zone_start_pfn
, zone_end_pfn
;
4712 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4713 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4715 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4716 node_start_pfn
, node_end_pfn
,
4717 &zone_start_pfn
, &zone_end_pfn
);
4718 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4721 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4722 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4723 unsigned long zone_type
,
4724 unsigned long node_start_pfn
,
4725 unsigned long node_end_pfn
,
4726 unsigned long *zones_size
)
4728 return zones_size
[zone_type
];
4731 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4732 unsigned long zone_type
,
4733 unsigned long node_start_pfn
,
4734 unsigned long node_end_pfn
,
4735 unsigned long *zholes_size
)
4740 return zholes_size
[zone_type
];
4743 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4745 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4746 unsigned long node_start_pfn
,
4747 unsigned long node_end_pfn
,
4748 unsigned long *zones_size
,
4749 unsigned long *zholes_size
)
4751 unsigned long realtotalpages
, totalpages
= 0;
4754 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4755 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4759 pgdat
->node_spanned_pages
= totalpages
;
4761 realtotalpages
= totalpages
;
4762 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4764 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4765 node_start_pfn
, node_end_pfn
,
4767 pgdat
->node_present_pages
= realtotalpages
;
4768 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4772 #ifndef CONFIG_SPARSEMEM
4774 * Calculate the size of the zone->blockflags rounded to an unsigned long
4775 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4776 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4777 * round what is now in bits to nearest long in bits, then return it in
4780 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4782 unsigned long usemapsize
;
4784 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4785 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4786 usemapsize
= usemapsize
>> pageblock_order
;
4787 usemapsize
*= NR_PAGEBLOCK_BITS
;
4788 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4790 return usemapsize
/ 8;
4793 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4795 unsigned long zone_start_pfn
,
4796 unsigned long zonesize
)
4798 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4799 zone
->pageblock_flags
= NULL
;
4801 zone
->pageblock_flags
=
4802 memblock_virt_alloc_node_nopanic(usemapsize
,
4806 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4807 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4808 #endif /* CONFIG_SPARSEMEM */
4810 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4812 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4813 void __paginginit
set_pageblock_order(void)
4817 /* Check that pageblock_nr_pages has not already been setup */
4818 if (pageblock_order
)
4821 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4822 order
= HUGETLB_PAGE_ORDER
;
4824 order
= MAX_ORDER
- 1;
4827 * Assume the largest contiguous order of interest is a huge page.
4828 * This value may be variable depending on boot parameters on IA64 and
4831 pageblock_order
= order
;
4833 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4836 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4837 * is unused as pageblock_order is set at compile-time. See
4838 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4841 void __paginginit
set_pageblock_order(void)
4845 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4847 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4848 unsigned long present_pages
)
4850 unsigned long pages
= spanned_pages
;
4853 * Provide a more accurate estimation if there are holes within
4854 * the zone and SPARSEMEM is in use. If there are holes within the
4855 * zone, each populated memory region may cost us one or two extra
4856 * memmap pages due to alignment because memmap pages for each
4857 * populated regions may not naturally algined on page boundary.
4858 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4860 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4861 IS_ENABLED(CONFIG_SPARSEMEM
))
4862 pages
= present_pages
;
4864 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4868 * Set up the zone data structures:
4869 * - mark all pages reserved
4870 * - mark all memory queues empty
4871 * - clear the memory bitmaps
4873 * NOTE: pgdat should get zeroed by caller.
4875 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4876 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4877 unsigned long *zones_size
, unsigned long *zholes_size
)
4880 int nid
= pgdat
->node_id
;
4881 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4884 pgdat_resize_init(pgdat
);
4885 #ifdef CONFIG_NUMA_BALANCING
4886 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4887 pgdat
->numabalancing_migrate_nr_pages
= 0;
4888 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4890 init_waitqueue_head(&pgdat
->kswapd_wait
);
4891 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4892 pgdat_page_ext_init(pgdat
);
4894 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4895 struct zone
*zone
= pgdat
->node_zones
+ j
;
4896 unsigned long size
, realsize
, freesize
, memmap_pages
;
4898 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4899 node_end_pfn
, zones_size
);
4900 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4906 * Adjust freesize so that it accounts for how much memory
4907 * is used by this zone for memmap. This affects the watermark
4908 * and per-cpu initialisations
4910 memmap_pages
= calc_memmap_size(size
, realsize
);
4911 if (freesize
>= memmap_pages
) {
4912 freesize
-= memmap_pages
;
4915 " %s zone: %lu pages used for memmap\n",
4916 zone_names
[j
], memmap_pages
);
4919 " %s zone: %lu pages exceeds freesize %lu\n",
4920 zone_names
[j
], memmap_pages
, freesize
);
4922 /* Account for reserved pages */
4923 if (j
== 0 && freesize
> dma_reserve
) {
4924 freesize
-= dma_reserve
;
4925 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4926 zone_names
[0], dma_reserve
);
4929 if (!is_highmem_idx(j
))
4930 nr_kernel_pages
+= freesize
;
4931 /* Charge for highmem memmap if there are enough kernel pages */
4932 else if (nr_kernel_pages
> memmap_pages
* 2)
4933 nr_kernel_pages
-= memmap_pages
;
4934 nr_all_pages
+= freesize
;
4936 zone
->spanned_pages
= size
;
4937 zone
->present_pages
= realsize
;
4939 * Set an approximate value for lowmem here, it will be adjusted
4940 * when the bootmem allocator frees pages into the buddy system.
4941 * And all highmem pages will be managed by the buddy system.
4943 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4946 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4948 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4950 zone
->name
= zone_names
[j
];
4951 spin_lock_init(&zone
->lock
);
4952 spin_lock_init(&zone
->lru_lock
);
4953 zone_seqlock_init(zone
);
4954 zone
->zone_pgdat
= pgdat
;
4955 zone_pcp_init(zone
);
4957 /* For bootup, initialized properly in watermark setup */
4958 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4960 lruvec_init(&zone
->lruvec
);
4964 set_pageblock_order();
4965 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4966 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4967 size
, MEMMAP_EARLY
);
4969 memmap_init(size
, nid
, j
, zone_start_pfn
);
4970 zone_start_pfn
+= size
;
4974 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4976 /* Skip empty nodes */
4977 if (!pgdat
->node_spanned_pages
)
4980 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4981 /* ia64 gets its own node_mem_map, before this, without bootmem */
4982 if (!pgdat
->node_mem_map
) {
4983 unsigned long size
, start
, end
;
4987 * The zone's endpoints aren't required to be MAX_ORDER
4988 * aligned but the node_mem_map endpoints must be in order
4989 * for the buddy allocator to function correctly.
4991 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4992 end
= pgdat_end_pfn(pgdat
);
4993 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4994 size
= (end
- start
) * sizeof(struct page
);
4995 map
= alloc_remap(pgdat
->node_id
, size
);
4997 map
= memblock_virt_alloc_node_nopanic(size
,
4999 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
5001 #ifndef CONFIG_NEED_MULTIPLE_NODES
5003 * With no DISCONTIG, the global mem_map is just set as node 0's
5005 if (pgdat
== NODE_DATA(0)) {
5006 mem_map
= NODE_DATA(0)->node_mem_map
;
5007 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5008 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5009 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
5010 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5013 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5016 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5017 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5019 pg_data_t
*pgdat
= NODE_DATA(nid
);
5020 unsigned long start_pfn
= 0;
5021 unsigned long end_pfn
= 0;
5023 /* pg_data_t should be reset to zero when it's allocated */
5024 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
5026 pgdat
->node_id
= nid
;
5027 pgdat
->node_start_pfn
= node_start_pfn
;
5028 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5029 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5030 printk(KERN_INFO
"Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid
,
5031 (u64
) start_pfn
<< PAGE_SHIFT
, (u64
) (end_pfn
<< PAGE_SHIFT
) - 1);
5033 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5034 zones_size
, zholes_size
);
5036 alloc_node_mem_map(pgdat
);
5037 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5038 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5039 nid
, (unsigned long)pgdat
,
5040 (unsigned long)pgdat
->node_mem_map
);
5043 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
5044 zones_size
, zholes_size
);
5047 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5049 #if MAX_NUMNODES > 1
5051 * Figure out the number of possible node ids.
5053 void __init
setup_nr_node_ids(void)
5056 unsigned int highest
= 0;
5058 for_each_node_mask(node
, node_possible_map
)
5060 nr_node_ids
= highest
+ 1;
5065 * node_map_pfn_alignment - determine the maximum internode alignment
5067 * This function should be called after node map is populated and sorted.
5068 * It calculates the maximum power of two alignment which can distinguish
5071 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5072 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5073 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5074 * shifted, 1GiB is enough and this function will indicate so.
5076 * This is used to test whether pfn -> nid mapping of the chosen memory
5077 * model has fine enough granularity to avoid incorrect mapping for the
5078 * populated node map.
5080 * Returns the determined alignment in pfn's. 0 if there is no alignment
5081 * requirement (single node).
5083 unsigned long __init
node_map_pfn_alignment(void)
5085 unsigned long accl_mask
= 0, last_end
= 0;
5086 unsigned long start
, end
, mask
;
5090 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5091 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5098 * Start with a mask granular enough to pin-point to the
5099 * start pfn and tick off bits one-by-one until it becomes
5100 * too coarse to separate the current node from the last.
5102 mask
= ~((1 << __ffs(start
)) - 1);
5103 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5106 /* accumulate all internode masks */
5110 /* convert mask to number of pages */
5111 return ~accl_mask
+ 1;
5114 /* Find the lowest pfn for a node */
5115 static unsigned long __init
find_min_pfn_for_node(int nid
)
5117 unsigned long min_pfn
= ULONG_MAX
;
5118 unsigned long start_pfn
;
5121 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5122 min_pfn
= min(min_pfn
, start_pfn
);
5124 if (min_pfn
== ULONG_MAX
) {
5126 "Could not find start_pfn for node %d\n", nid
);
5134 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5136 * It returns the minimum PFN based on information provided via
5137 * memblock_set_node().
5139 unsigned long __init
find_min_pfn_with_active_regions(void)
5141 return find_min_pfn_for_node(MAX_NUMNODES
);
5145 * early_calculate_totalpages()
5146 * Sum pages in active regions for movable zone.
5147 * Populate N_MEMORY for calculating usable_nodes.
5149 static unsigned long __init
early_calculate_totalpages(void)
5151 unsigned long totalpages
= 0;
5152 unsigned long start_pfn
, end_pfn
;
5155 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5156 unsigned long pages
= end_pfn
- start_pfn
;
5158 totalpages
+= pages
;
5160 node_set_state(nid
, N_MEMORY
);
5166 * Find the PFN the Movable zone begins in each node. Kernel memory
5167 * is spread evenly between nodes as long as the nodes have enough
5168 * memory. When they don't, some nodes will have more kernelcore than
5171 static void __init
find_zone_movable_pfns_for_nodes(void)
5174 unsigned long usable_startpfn
;
5175 unsigned long kernelcore_node
, kernelcore_remaining
;
5176 /* save the state before borrow the nodemask */
5177 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5178 unsigned long totalpages
= early_calculate_totalpages();
5179 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5180 struct memblock_region
*r
;
5182 /* Need to find movable_zone earlier when movable_node is specified. */
5183 find_usable_zone_for_movable();
5186 * If movable_node is specified, ignore kernelcore and movablecore
5189 if (movable_node_is_enabled()) {
5190 for_each_memblock(memory
, r
) {
5191 if (!memblock_is_hotpluggable(r
))
5196 usable_startpfn
= PFN_DOWN(r
->base
);
5197 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5198 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5206 * If movablecore=nn[KMG] was specified, calculate what size of
5207 * kernelcore that corresponds so that memory usable for
5208 * any allocation type is evenly spread. If both kernelcore
5209 * and movablecore are specified, then the value of kernelcore
5210 * will be used for required_kernelcore if it's greater than
5211 * what movablecore would have allowed.
5213 if (required_movablecore
) {
5214 unsigned long corepages
;
5217 * Round-up so that ZONE_MOVABLE is at least as large as what
5218 * was requested by the user
5220 required_movablecore
=
5221 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5222 corepages
= totalpages
- required_movablecore
;
5224 required_kernelcore
= max(required_kernelcore
, corepages
);
5227 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5228 if (!required_kernelcore
)
5231 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5232 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5235 /* Spread kernelcore memory as evenly as possible throughout nodes */
5236 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5237 for_each_node_state(nid
, N_MEMORY
) {
5238 unsigned long start_pfn
, end_pfn
;
5241 * Recalculate kernelcore_node if the division per node
5242 * now exceeds what is necessary to satisfy the requested
5243 * amount of memory for the kernel
5245 if (required_kernelcore
< kernelcore_node
)
5246 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5249 * As the map is walked, we track how much memory is usable
5250 * by the kernel using kernelcore_remaining. When it is
5251 * 0, the rest of the node is usable by ZONE_MOVABLE
5253 kernelcore_remaining
= kernelcore_node
;
5255 /* Go through each range of PFNs within this node */
5256 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5257 unsigned long size_pages
;
5259 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5260 if (start_pfn
>= end_pfn
)
5263 /* Account for what is only usable for kernelcore */
5264 if (start_pfn
< usable_startpfn
) {
5265 unsigned long kernel_pages
;
5266 kernel_pages
= min(end_pfn
, usable_startpfn
)
5269 kernelcore_remaining
-= min(kernel_pages
,
5270 kernelcore_remaining
);
5271 required_kernelcore
-= min(kernel_pages
,
5272 required_kernelcore
);
5274 /* Continue if range is now fully accounted */
5275 if (end_pfn
<= usable_startpfn
) {
5278 * Push zone_movable_pfn to the end so
5279 * that if we have to rebalance
5280 * kernelcore across nodes, we will
5281 * not double account here
5283 zone_movable_pfn
[nid
] = end_pfn
;
5286 start_pfn
= usable_startpfn
;
5290 * The usable PFN range for ZONE_MOVABLE is from
5291 * start_pfn->end_pfn. Calculate size_pages as the
5292 * number of pages used as kernelcore
5294 size_pages
= end_pfn
- start_pfn
;
5295 if (size_pages
> kernelcore_remaining
)
5296 size_pages
= kernelcore_remaining
;
5297 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5300 * Some kernelcore has been met, update counts and
5301 * break if the kernelcore for this node has been
5304 required_kernelcore
-= min(required_kernelcore
,
5306 kernelcore_remaining
-= size_pages
;
5307 if (!kernelcore_remaining
)
5313 * If there is still required_kernelcore, we do another pass with one
5314 * less node in the count. This will push zone_movable_pfn[nid] further
5315 * along on the nodes that still have memory until kernelcore is
5319 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5323 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5324 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5325 zone_movable_pfn
[nid
] =
5326 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5329 /* restore the node_state */
5330 node_states
[N_MEMORY
] = saved_node_state
;
5333 /* Any regular or high memory on that node ? */
5334 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5336 enum zone_type zone_type
;
5338 if (N_MEMORY
== N_NORMAL_MEMORY
)
5341 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5342 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5343 if (populated_zone(zone
)) {
5344 node_set_state(nid
, N_HIGH_MEMORY
);
5345 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5346 zone_type
<= ZONE_NORMAL
)
5347 node_set_state(nid
, N_NORMAL_MEMORY
);
5354 * free_area_init_nodes - Initialise all pg_data_t and zone data
5355 * @max_zone_pfn: an array of max PFNs for each zone
5357 * This will call free_area_init_node() for each active node in the system.
5358 * Using the page ranges provided by memblock_set_node(), the size of each
5359 * zone in each node and their holes is calculated. If the maximum PFN
5360 * between two adjacent zones match, it is assumed that the zone is empty.
5361 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5362 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5363 * starts where the previous one ended. For example, ZONE_DMA32 starts
5364 * at arch_max_dma_pfn.
5366 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5368 unsigned long start_pfn
, end_pfn
;
5371 /* Record where the zone boundaries are */
5372 memset(arch_zone_lowest_possible_pfn
, 0,
5373 sizeof(arch_zone_lowest_possible_pfn
));
5374 memset(arch_zone_highest_possible_pfn
, 0,
5375 sizeof(arch_zone_highest_possible_pfn
));
5376 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5377 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5378 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5379 if (i
== ZONE_MOVABLE
)
5381 arch_zone_lowest_possible_pfn
[i
] =
5382 arch_zone_highest_possible_pfn
[i
-1];
5383 arch_zone_highest_possible_pfn
[i
] =
5384 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5386 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5387 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5389 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5390 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5391 find_zone_movable_pfns_for_nodes();
5393 /* Print out the zone ranges */
5394 pr_info("Zone ranges:\n");
5395 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5396 if (i
== ZONE_MOVABLE
)
5398 pr_info(" %-8s ", zone_names
[i
]);
5399 if (arch_zone_lowest_possible_pfn
[i
] ==
5400 arch_zone_highest_possible_pfn
[i
])
5403 pr_cont("[mem %0#10lx-%0#10lx]\n",
5404 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5405 (arch_zone_highest_possible_pfn
[i
]
5406 << PAGE_SHIFT
) - 1);
5409 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5410 pr_info("Movable zone start for each node\n");
5411 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5412 if (zone_movable_pfn
[i
])
5413 pr_info(" Node %d: %#010lx\n", i
,
5414 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5417 /* Print out the early node map */
5418 pr_info("Early memory node ranges\n");
5419 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5420 pr_info(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5421 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5423 /* Initialise every node */
5424 mminit_verify_pageflags_layout();
5425 setup_nr_node_ids();
5426 for_each_online_node(nid
) {
5427 pg_data_t
*pgdat
= NODE_DATA(nid
);
5428 free_area_init_node(nid
, NULL
,
5429 find_min_pfn_for_node(nid
), NULL
);
5431 /* Any memory on that node */
5432 if (pgdat
->node_present_pages
)
5433 node_set_state(nid
, N_MEMORY
);
5434 check_for_memory(pgdat
, nid
);
5438 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5440 unsigned long long coremem
;
5444 coremem
= memparse(p
, &p
);
5445 *core
= coremem
>> PAGE_SHIFT
;
5447 /* Paranoid check that UL is enough for the coremem value */
5448 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5454 * kernelcore=size sets the amount of memory for use for allocations that
5455 * cannot be reclaimed or migrated.
5457 static int __init
cmdline_parse_kernelcore(char *p
)
5459 return cmdline_parse_core(p
, &required_kernelcore
);
5463 * movablecore=size sets the amount of memory for use for allocations that
5464 * can be reclaimed or migrated.
5466 static int __init
cmdline_parse_movablecore(char *p
)
5468 return cmdline_parse_core(p
, &required_movablecore
);
5471 early_param("kernelcore", cmdline_parse_kernelcore
);
5472 early_param("movablecore", cmdline_parse_movablecore
);
5474 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5476 void adjust_managed_page_count(struct page
*page
, long count
)
5478 spin_lock(&managed_page_count_lock
);
5479 page_zone(page
)->managed_pages
+= count
;
5480 totalram_pages
+= count
;
5481 #ifdef CONFIG_HIGHMEM
5482 if (PageHighMem(page
))
5483 totalhigh_pages
+= count
;
5485 spin_unlock(&managed_page_count_lock
);
5487 EXPORT_SYMBOL(adjust_managed_page_count
);
5489 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5492 unsigned long pages
= 0;
5494 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5495 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5496 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5497 if ((unsigned int)poison
<= 0xFF)
5498 memset(pos
, poison
, PAGE_SIZE
);
5499 free_reserved_page(virt_to_page(pos
));
5503 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5504 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5508 EXPORT_SYMBOL(free_reserved_area
);
5510 #ifdef CONFIG_HIGHMEM
5511 void free_highmem_page(struct page
*page
)
5513 __free_reserved_page(page
);
5515 page_zone(page
)->managed_pages
++;
5521 void __init
mem_init_print_info(const char *str
)
5523 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5524 unsigned long init_code_size
, init_data_size
;
5526 physpages
= get_num_physpages();
5527 codesize
= _etext
- _stext
;
5528 datasize
= _edata
- _sdata
;
5529 rosize
= __end_rodata
- __start_rodata
;
5530 bss_size
= __bss_stop
- __bss_start
;
5531 init_data_size
= __init_end
- __init_begin
;
5532 init_code_size
= _einittext
- _sinittext
;
5535 * Detect special cases and adjust section sizes accordingly:
5536 * 1) .init.* may be embedded into .data sections
5537 * 2) .init.text.* may be out of [__init_begin, __init_end],
5538 * please refer to arch/tile/kernel/vmlinux.lds.S.
5539 * 3) .rodata.* may be embedded into .text or .data sections.
5541 #define adj_init_size(start, end, size, pos, adj) \
5543 if (start <= pos && pos < end && size > adj) \
5547 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5548 _sinittext
, init_code_size
);
5549 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5550 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5551 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5552 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5554 #undef adj_init_size
5556 pr_info("Memory: %luK/%luK available "
5557 "(%luK kernel code, %luK rwdata, %luK rodata, "
5558 "%luK init, %luK bss, %luK reserved"
5559 #ifdef CONFIG_HIGHMEM
5563 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5564 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5565 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5566 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5567 #ifdef CONFIG_HIGHMEM
5568 totalhigh_pages
<< (PAGE_SHIFT
-10),
5570 str
? ", " : "", str
? str
: "");
5574 * set_dma_reserve - set the specified number of pages reserved in the first zone
5575 * @new_dma_reserve: The number of pages to mark reserved
5577 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5578 * In the DMA zone, a significant percentage may be consumed by kernel image
5579 * and other unfreeable allocations which can skew the watermarks badly. This
5580 * function may optionally be used to account for unfreeable pages in the
5581 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5582 * smaller per-cpu batchsize.
5584 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5586 dma_reserve
= new_dma_reserve
;
5589 void __init
free_area_init(unsigned long *zones_size
)
5591 free_area_init_node(0, zones_size
,
5592 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5595 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5596 unsigned long action
, void *hcpu
)
5598 int cpu
= (unsigned long)hcpu
;
5600 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5601 lru_add_drain_cpu(cpu
);
5605 * Spill the event counters of the dead processor
5606 * into the current processors event counters.
5607 * This artificially elevates the count of the current
5610 vm_events_fold_cpu(cpu
);
5613 * Zero the differential counters of the dead processor
5614 * so that the vm statistics are consistent.
5616 * This is only okay since the processor is dead and cannot
5617 * race with what we are doing.
5619 cpu_vm_stats_fold(cpu
);
5624 void __init
page_alloc_init(void)
5626 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5630 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5631 * or min_free_kbytes changes.
5633 static void calculate_totalreserve_pages(void)
5635 struct pglist_data
*pgdat
;
5636 unsigned long reserve_pages
= 0;
5637 enum zone_type i
, j
;
5639 for_each_online_pgdat(pgdat
) {
5640 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5641 struct zone
*zone
= pgdat
->node_zones
+ i
;
5644 /* Find valid and maximum lowmem_reserve in the zone */
5645 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5646 if (zone
->lowmem_reserve
[j
] > max
)
5647 max
= zone
->lowmem_reserve
[j
];
5650 /* we treat the high watermark as reserved pages. */
5651 max
+= high_wmark_pages(zone
);
5653 if (max
> zone
->managed_pages
)
5654 max
= zone
->managed_pages
;
5655 reserve_pages
+= max
;
5657 * Lowmem reserves are not available to
5658 * GFP_HIGHUSER page cache allocations and
5659 * kswapd tries to balance zones to their high
5660 * watermark. As a result, neither should be
5661 * regarded as dirtyable memory, to prevent a
5662 * situation where reclaim has to clean pages
5663 * in order to balance the zones.
5665 zone
->dirty_balance_reserve
= max
;
5668 dirty_balance_reserve
= reserve_pages
;
5669 totalreserve_pages
= reserve_pages
;
5673 * setup_per_zone_lowmem_reserve - called whenever
5674 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5675 * has a correct pages reserved value, so an adequate number of
5676 * pages are left in the zone after a successful __alloc_pages().
5678 static void setup_per_zone_lowmem_reserve(void)
5680 struct pglist_data
*pgdat
;
5681 enum zone_type j
, idx
;
5683 for_each_online_pgdat(pgdat
) {
5684 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5685 struct zone
*zone
= pgdat
->node_zones
+ j
;
5686 unsigned long managed_pages
= zone
->managed_pages
;
5688 zone
->lowmem_reserve
[j
] = 0;
5692 struct zone
*lower_zone
;
5696 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5697 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5699 lower_zone
= pgdat
->node_zones
+ idx
;
5700 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5701 sysctl_lowmem_reserve_ratio
[idx
];
5702 managed_pages
+= lower_zone
->managed_pages
;
5707 /* update totalreserve_pages */
5708 calculate_totalreserve_pages();
5711 static void __setup_per_zone_wmarks(void)
5713 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5714 unsigned long lowmem_pages
= 0;
5716 unsigned long flags
;
5718 /* Calculate total number of !ZONE_HIGHMEM pages */
5719 for_each_zone(zone
) {
5720 if (!is_highmem(zone
))
5721 lowmem_pages
+= zone
->managed_pages
;
5724 for_each_zone(zone
) {
5727 spin_lock_irqsave(&zone
->lock
, flags
);
5728 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5729 do_div(tmp
, lowmem_pages
);
5730 if (is_highmem(zone
)) {
5732 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5733 * need highmem pages, so cap pages_min to a small
5736 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5737 * deltas controls asynch page reclaim, and so should
5738 * not be capped for highmem.
5740 unsigned long min_pages
;
5742 min_pages
= zone
->managed_pages
/ 1024;
5743 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5744 zone
->watermark
[WMARK_MIN
] = min_pages
;
5747 * If it's a lowmem zone, reserve a number of pages
5748 * proportionate to the zone's size.
5750 zone
->watermark
[WMARK_MIN
] = tmp
;
5753 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5754 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5756 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5757 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5758 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5760 setup_zone_migrate_reserve(zone
);
5761 spin_unlock_irqrestore(&zone
->lock
, flags
);
5764 /* update totalreserve_pages */
5765 calculate_totalreserve_pages();
5769 * setup_per_zone_wmarks - called when min_free_kbytes changes
5770 * or when memory is hot-{added|removed}
5772 * Ensures that the watermark[min,low,high] values for each zone are set
5773 * correctly with respect to min_free_kbytes.
5775 void setup_per_zone_wmarks(void)
5777 mutex_lock(&zonelists_mutex
);
5778 __setup_per_zone_wmarks();
5779 mutex_unlock(&zonelists_mutex
);
5783 * The inactive anon list should be small enough that the VM never has to
5784 * do too much work, but large enough that each inactive page has a chance
5785 * to be referenced again before it is swapped out.
5787 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5788 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5789 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5790 * the anonymous pages are kept on the inactive list.
5793 * memory ratio inactive anon
5794 * -------------------------------------
5803 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5805 unsigned int gb
, ratio
;
5807 /* Zone size in gigabytes */
5808 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5810 ratio
= int_sqrt(10 * gb
);
5814 zone
->inactive_ratio
= ratio
;
5817 static void __meminit
setup_per_zone_inactive_ratio(void)
5822 calculate_zone_inactive_ratio(zone
);
5826 * Initialise min_free_kbytes.
5828 * For small machines we want it small (128k min). For large machines
5829 * we want it large (64MB max). But it is not linear, because network
5830 * bandwidth does not increase linearly with machine size. We use
5832 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5833 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5849 int __meminit
init_per_zone_wmark_min(void)
5851 unsigned long lowmem_kbytes
;
5852 int new_min_free_kbytes
;
5854 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5855 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5857 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5858 min_free_kbytes
= new_min_free_kbytes
;
5859 if (min_free_kbytes
< 128)
5860 min_free_kbytes
= 128;
5861 if (min_free_kbytes
> 65536)
5862 min_free_kbytes
= 65536;
5864 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5865 new_min_free_kbytes
, user_min_free_kbytes
);
5867 setup_per_zone_wmarks();
5868 refresh_zone_stat_thresholds();
5869 setup_per_zone_lowmem_reserve();
5870 setup_per_zone_inactive_ratio();
5873 module_init(init_per_zone_wmark_min
)
5876 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5877 * that we can call two helper functions whenever min_free_kbytes
5880 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5881 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5885 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5890 user_min_free_kbytes
= min_free_kbytes
;
5891 setup_per_zone_wmarks();
5897 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5898 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5903 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5908 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5909 sysctl_min_unmapped_ratio
) / 100;
5913 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5914 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5919 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5924 zone
->min_slab_pages
= (zone
->managed_pages
*
5925 sysctl_min_slab_ratio
) / 100;
5931 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5932 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5933 * whenever sysctl_lowmem_reserve_ratio changes.
5935 * The reserve ratio obviously has absolutely no relation with the
5936 * minimum watermarks. The lowmem reserve ratio can only make sense
5937 * if in function of the boot time zone sizes.
5939 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5940 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5942 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5943 setup_per_zone_lowmem_reserve();
5948 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5949 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5950 * pagelist can have before it gets flushed back to buddy allocator.
5952 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5953 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5956 int old_percpu_pagelist_fraction
;
5959 mutex_lock(&pcp_batch_high_lock
);
5960 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5962 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5963 if (!write
|| ret
< 0)
5966 /* Sanity checking to avoid pcp imbalance */
5967 if (percpu_pagelist_fraction
&&
5968 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5969 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5975 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
5978 for_each_populated_zone(zone
) {
5981 for_each_possible_cpu(cpu
)
5982 pageset_set_high_and_batch(zone
,
5983 per_cpu_ptr(zone
->pageset
, cpu
));
5986 mutex_unlock(&pcp_batch_high_lock
);
5990 int hashdist
= HASHDIST_DEFAULT
;
5993 static int __init
set_hashdist(char *str
)
5997 hashdist
= simple_strtoul(str
, &str
, 0);
6000 __setup("hashdist=", set_hashdist
);
6004 * allocate a large system hash table from bootmem
6005 * - it is assumed that the hash table must contain an exact power-of-2
6006 * quantity of entries
6007 * - limit is the number of hash buckets, not the total allocation size
6009 void *__init
alloc_large_system_hash(const char *tablename
,
6010 unsigned long bucketsize
,
6011 unsigned long numentries
,
6014 unsigned int *_hash_shift
,
6015 unsigned int *_hash_mask
,
6016 unsigned long low_limit
,
6017 unsigned long high_limit
)
6019 unsigned long long max
= high_limit
;
6020 unsigned long log2qty
, size
;
6023 /* allow the kernel cmdline to have a say */
6025 /* round applicable memory size up to nearest megabyte */
6026 numentries
= nr_kernel_pages
;
6028 /* It isn't necessary when PAGE_SIZE >= 1MB */
6029 if (PAGE_SHIFT
< 20)
6030 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6032 /* limit to 1 bucket per 2^scale bytes of low memory */
6033 if (scale
> PAGE_SHIFT
)
6034 numentries
>>= (scale
- PAGE_SHIFT
);
6036 numentries
<<= (PAGE_SHIFT
- scale
);
6038 /* Make sure we've got at least a 0-order allocation.. */
6039 if (unlikely(flags
& HASH_SMALL
)) {
6040 /* Makes no sense without HASH_EARLY */
6041 WARN_ON(!(flags
& HASH_EARLY
));
6042 if (!(numentries
>> *_hash_shift
)) {
6043 numentries
= 1UL << *_hash_shift
;
6044 BUG_ON(!numentries
);
6046 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6047 numentries
= PAGE_SIZE
/ bucketsize
;
6049 numentries
= roundup_pow_of_two(numentries
);
6051 /* limit allocation size to 1/16 total memory by default */
6053 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6054 do_div(max
, bucketsize
);
6056 max
= min(max
, 0x80000000ULL
);
6058 if (numentries
< low_limit
)
6059 numentries
= low_limit
;
6060 if (numentries
> max
)
6063 log2qty
= ilog2(numentries
);
6066 size
= bucketsize
<< log2qty
;
6067 if (flags
& HASH_EARLY
)
6068 table
= memblock_virt_alloc_nopanic(size
, 0);
6070 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6073 * If bucketsize is not a power-of-two, we may free
6074 * some pages at the end of hash table which
6075 * alloc_pages_exact() automatically does
6077 if (get_order(size
) < MAX_ORDER
) {
6078 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6079 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6082 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6085 panic("Failed to allocate %s hash table\n", tablename
);
6087 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6090 ilog2(size
) - PAGE_SHIFT
,
6094 *_hash_shift
= log2qty
;
6096 *_hash_mask
= (1 << log2qty
) - 1;
6101 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6102 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6105 #ifdef CONFIG_SPARSEMEM
6106 return __pfn_to_section(pfn
)->pageblock_flags
;
6108 return zone
->pageblock_flags
;
6109 #endif /* CONFIG_SPARSEMEM */
6112 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6114 #ifdef CONFIG_SPARSEMEM
6115 pfn
&= (PAGES_PER_SECTION
-1);
6116 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6118 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6119 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6120 #endif /* CONFIG_SPARSEMEM */
6124 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6125 * @page: The page within the block of interest
6126 * @pfn: The target page frame number
6127 * @end_bitidx: The last bit of interest to retrieve
6128 * @mask: mask of bits that the caller is interested in
6130 * Return: pageblock_bits flags
6132 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6133 unsigned long end_bitidx
,
6137 unsigned long *bitmap
;
6138 unsigned long bitidx
, word_bitidx
;
6141 zone
= page_zone(page
);
6142 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6143 bitidx
= pfn_to_bitidx(zone
, pfn
);
6144 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6145 bitidx
&= (BITS_PER_LONG
-1);
6147 word
= bitmap
[word_bitidx
];
6148 bitidx
+= end_bitidx
;
6149 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6153 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6154 * @page: The page within the block of interest
6155 * @flags: The flags to set
6156 * @pfn: The target page frame number
6157 * @end_bitidx: The last bit of interest
6158 * @mask: mask of bits that the caller is interested in
6160 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6162 unsigned long end_bitidx
,
6166 unsigned long *bitmap
;
6167 unsigned long bitidx
, word_bitidx
;
6168 unsigned long old_word
, word
;
6170 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6172 zone
= page_zone(page
);
6173 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6174 bitidx
= pfn_to_bitidx(zone
, pfn
);
6175 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6176 bitidx
&= (BITS_PER_LONG
-1);
6178 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6180 bitidx
+= end_bitidx
;
6181 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6182 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6184 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6186 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6187 if (word
== old_word
)
6194 * This function checks whether pageblock includes unmovable pages or not.
6195 * If @count is not zero, it is okay to include less @count unmovable pages
6197 * PageLRU check without isolation or lru_lock could race so that
6198 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6199 * expect this function should be exact.
6201 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6202 bool skip_hwpoisoned_pages
)
6204 unsigned long pfn
, iter
, found
;
6208 * For avoiding noise data, lru_add_drain_all() should be called
6209 * If ZONE_MOVABLE, the zone never contains unmovable pages
6211 if (zone_idx(zone
) == ZONE_MOVABLE
)
6213 mt
= get_pageblock_migratetype(page
);
6214 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6217 pfn
= page_to_pfn(page
);
6218 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6219 unsigned long check
= pfn
+ iter
;
6221 if (!pfn_valid_within(check
))
6224 page
= pfn_to_page(check
);
6227 * Hugepages are not in LRU lists, but they're movable.
6228 * We need not scan over tail pages bacause we don't
6229 * handle each tail page individually in migration.
6231 if (PageHuge(page
)) {
6232 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6237 * We can't use page_count without pin a page
6238 * because another CPU can free compound page.
6239 * This check already skips compound tails of THP
6240 * because their page->_count is zero at all time.
6242 if (!atomic_read(&page
->_count
)) {
6243 if (PageBuddy(page
))
6244 iter
+= (1 << page_order(page
)) - 1;
6249 * The HWPoisoned page may be not in buddy system, and
6250 * page_count() is not 0.
6252 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6258 * If there are RECLAIMABLE pages, we need to check it.
6259 * But now, memory offline itself doesn't call shrink_slab()
6260 * and it still to be fixed.
6263 * If the page is not RAM, page_count()should be 0.
6264 * we don't need more check. This is an _used_ not-movable page.
6266 * The problematic thing here is PG_reserved pages. PG_reserved
6267 * is set to both of a memory hole page and a _used_ kernel
6276 bool is_pageblock_removable_nolock(struct page
*page
)
6282 * We have to be careful here because we are iterating over memory
6283 * sections which are not zone aware so we might end up outside of
6284 * the zone but still within the section.
6285 * We have to take care about the node as well. If the node is offline
6286 * its NODE_DATA will be NULL - see page_zone.
6288 if (!node_online(page_to_nid(page
)))
6291 zone
= page_zone(page
);
6292 pfn
= page_to_pfn(page
);
6293 if (!zone_spans_pfn(zone
, pfn
))
6296 return !has_unmovable_pages(zone
, page
, 0, true);
6301 static unsigned long pfn_max_align_down(unsigned long pfn
)
6303 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6304 pageblock_nr_pages
) - 1);
6307 static unsigned long pfn_max_align_up(unsigned long pfn
)
6309 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6310 pageblock_nr_pages
));
6313 /* [start, end) must belong to a single zone. */
6314 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6315 unsigned long start
, unsigned long end
)
6317 /* This function is based on compact_zone() from compaction.c. */
6318 unsigned long nr_reclaimed
;
6319 unsigned long pfn
= start
;
6320 unsigned int tries
= 0;
6325 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6326 if (fatal_signal_pending(current
)) {
6331 if (list_empty(&cc
->migratepages
)) {
6332 cc
->nr_migratepages
= 0;
6333 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6339 } else if (++tries
== 5) {
6340 ret
= ret
< 0 ? ret
: -EBUSY
;
6344 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6346 cc
->nr_migratepages
-= nr_reclaimed
;
6348 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6349 NULL
, 0, cc
->mode
, MR_CMA
);
6352 putback_movable_pages(&cc
->migratepages
);
6359 * alloc_contig_range() -- tries to allocate given range of pages
6360 * @start: start PFN to allocate
6361 * @end: one-past-the-last PFN to allocate
6362 * @migratetype: migratetype of the underlaying pageblocks (either
6363 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6364 * in range must have the same migratetype and it must
6365 * be either of the two.
6367 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6368 * aligned, however it's the caller's responsibility to guarantee that
6369 * we are the only thread that changes migrate type of pageblocks the
6372 * The PFN range must belong to a single zone.
6374 * Returns zero on success or negative error code. On success all
6375 * pages which PFN is in [start, end) are allocated for the caller and
6376 * need to be freed with free_contig_range().
6378 int alloc_contig_range(unsigned long start
, unsigned long end
,
6379 unsigned migratetype
)
6381 unsigned long outer_start
, outer_end
;
6384 struct compact_control cc
= {
6385 .nr_migratepages
= 0,
6387 .zone
= page_zone(pfn_to_page(start
)),
6388 .mode
= MIGRATE_SYNC
,
6389 .ignore_skip_hint
= true,
6391 INIT_LIST_HEAD(&cc
.migratepages
);
6394 * What we do here is we mark all pageblocks in range as
6395 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6396 * have different sizes, and due to the way page allocator
6397 * work, we align the range to biggest of the two pages so
6398 * that page allocator won't try to merge buddies from
6399 * different pageblocks and change MIGRATE_ISOLATE to some
6400 * other migration type.
6402 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6403 * migrate the pages from an unaligned range (ie. pages that
6404 * we are interested in). This will put all the pages in
6405 * range back to page allocator as MIGRATE_ISOLATE.
6407 * When this is done, we take the pages in range from page
6408 * allocator removing them from the buddy system. This way
6409 * page allocator will never consider using them.
6411 * This lets us mark the pageblocks back as
6412 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6413 * aligned range but not in the unaligned, original range are
6414 * put back to page allocator so that buddy can use them.
6417 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6418 pfn_max_align_up(end
), migratetype
,
6423 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6428 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6429 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6430 * more, all pages in [start, end) are free in page allocator.
6431 * What we are going to do is to allocate all pages from
6432 * [start, end) (that is remove them from page allocator).
6434 * The only problem is that pages at the beginning and at the
6435 * end of interesting range may be not aligned with pages that
6436 * page allocator holds, ie. they can be part of higher order
6437 * pages. Because of this, we reserve the bigger range and
6438 * once this is done free the pages we are not interested in.
6440 * We don't have to hold zone->lock here because the pages are
6441 * isolated thus they won't get removed from buddy.
6444 lru_add_drain_all();
6445 drain_all_pages(cc
.zone
);
6448 outer_start
= start
;
6449 while (!PageBuddy(pfn_to_page(outer_start
))) {
6450 if (++order
>= MAX_ORDER
) {
6454 outer_start
&= ~0UL << order
;
6457 /* Make sure the range is really isolated. */
6458 if (test_pages_isolated(outer_start
, end
, false)) {
6459 pr_info("%s: [%lx, %lx) PFNs busy\n",
6460 __func__
, outer_start
, end
);
6465 /* Grab isolated pages from freelists. */
6466 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6472 /* Free head and tail (if any) */
6473 if (start
!= outer_start
)
6474 free_contig_range(outer_start
, start
- outer_start
);
6475 if (end
!= outer_end
)
6476 free_contig_range(end
, outer_end
- end
);
6479 undo_isolate_page_range(pfn_max_align_down(start
),
6480 pfn_max_align_up(end
), migratetype
);
6484 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6486 unsigned int count
= 0;
6488 for (; nr_pages
--; pfn
++) {
6489 struct page
*page
= pfn_to_page(pfn
);
6491 count
+= page_count(page
) != 1;
6494 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6498 #ifdef CONFIG_MEMORY_HOTPLUG
6500 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6501 * page high values need to be recalulated.
6503 void __meminit
zone_pcp_update(struct zone
*zone
)
6506 mutex_lock(&pcp_batch_high_lock
);
6507 for_each_possible_cpu(cpu
)
6508 pageset_set_high_and_batch(zone
,
6509 per_cpu_ptr(zone
->pageset
, cpu
));
6510 mutex_unlock(&pcp_batch_high_lock
);
6514 void zone_pcp_reset(struct zone
*zone
)
6516 unsigned long flags
;
6518 struct per_cpu_pageset
*pset
;
6520 /* avoid races with drain_pages() */
6521 local_irq_save(flags
);
6522 if (zone
->pageset
!= &boot_pageset
) {
6523 for_each_online_cpu(cpu
) {
6524 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6525 drain_zonestat(zone
, pset
);
6527 free_percpu(zone
->pageset
);
6528 zone
->pageset
= &boot_pageset
;
6530 local_irq_restore(flags
);
6533 #ifdef CONFIG_MEMORY_HOTREMOVE
6535 * All pages in the range must be isolated before calling this.
6538 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6542 unsigned int order
, i
;
6544 unsigned long flags
;
6545 /* find the first valid pfn */
6546 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6551 zone
= page_zone(pfn_to_page(pfn
));
6552 spin_lock_irqsave(&zone
->lock
, flags
);
6554 while (pfn
< end_pfn
) {
6555 if (!pfn_valid(pfn
)) {
6559 page
= pfn_to_page(pfn
);
6561 * The HWPoisoned page may be not in buddy system, and
6562 * page_count() is not 0.
6564 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6566 SetPageReserved(page
);
6570 BUG_ON(page_count(page
));
6571 BUG_ON(!PageBuddy(page
));
6572 order
= page_order(page
);
6573 #ifdef CONFIG_DEBUG_VM
6574 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6575 pfn
, 1 << order
, end_pfn
);
6577 list_del(&page
->lru
);
6578 rmv_page_order(page
);
6579 zone
->free_area
[order
].nr_free
--;
6580 for (i
= 0; i
< (1 << order
); i
++)
6581 SetPageReserved((page
+i
));
6582 pfn
+= (1 << order
);
6584 spin_unlock_irqrestore(&zone
->lock
, flags
);
6588 #ifdef CONFIG_MEMORY_FAILURE
6589 bool is_free_buddy_page(struct page
*page
)
6591 struct zone
*zone
= page_zone(page
);
6592 unsigned long pfn
= page_to_pfn(page
);
6593 unsigned long flags
;
6596 spin_lock_irqsave(&zone
->lock
, flags
);
6597 for (order
= 0; order
< MAX_ORDER
; order
++) {
6598 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6600 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6603 spin_unlock_irqrestore(&zone
->lock
, flags
);
6605 return order
< MAX_ORDER
;