2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
64 #include <linux/kthread.h>
66 #include <asm/sections.h>
67 #include <asm/tlbflush.h>
68 #include <asm/div64.h>
71 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
72 static DEFINE_MUTEX(pcp_batch_high_lock
);
73 #define MIN_PERCPU_PAGELIST_FRACTION (8)
75 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
76 DEFINE_PER_CPU(int, numa_node
);
77 EXPORT_PER_CPU_SYMBOL(numa_node
);
80 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
82 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
83 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
84 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
85 * defined in <linux/topology.h>.
87 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
88 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
89 int _node_numa_mem_
[MAX_NUMNODES
];
93 * Array of node states.
95 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
96 [N_POSSIBLE
] = NODE_MASK_ALL
,
97 [N_ONLINE
] = { { [0] = 1UL } },
99 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
100 #ifdef CONFIG_HIGHMEM
101 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
103 #ifdef CONFIG_MOVABLE_NODE
104 [N_MEMORY
] = { { [0] = 1UL } },
106 [N_CPU
] = { { [0] = 1UL } },
109 EXPORT_SYMBOL(node_states
);
111 /* Protect totalram_pages and zone->managed_pages */
112 static DEFINE_SPINLOCK(managed_page_count_lock
);
114 unsigned long totalram_pages __read_mostly
;
115 unsigned long totalreserve_pages __read_mostly
;
116 unsigned long totalcma_pages __read_mostly
;
118 * When calculating the number of globally allowed dirty pages, there
119 * is a certain number of per-zone reserves that should not be
120 * considered dirtyable memory. This is the sum of those reserves
121 * over all existing zones that contribute dirtyable memory.
123 unsigned long dirty_balance_reserve __read_mostly
;
125 int percpu_pagelist_fraction
;
126 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
129 * A cached value of the page's pageblock's migratetype, used when the page is
130 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
131 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
132 * Also the migratetype set in the page does not necessarily match the pcplist
133 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
134 * other index - this ensures that it will be put on the correct CMA freelist.
136 static inline int get_pcppage_migratetype(struct page
*page
)
141 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
143 page
->index
= migratetype
;
146 #ifdef CONFIG_PM_SLEEP
148 * The following functions are used by the suspend/hibernate code to temporarily
149 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
150 * while devices are suspended. To avoid races with the suspend/hibernate code,
151 * they should always be called with pm_mutex held (gfp_allowed_mask also should
152 * only be modified with pm_mutex held, unless the suspend/hibernate code is
153 * guaranteed not to run in parallel with that modification).
156 static gfp_t saved_gfp_mask
;
158 void pm_restore_gfp_mask(void)
160 WARN_ON(!mutex_is_locked(&pm_mutex
));
161 if (saved_gfp_mask
) {
162 gfp_allowed_mask
= saved_gfp_mask
;
167 void pm_restrict_gfp_mask(void)
169 WARN_ON(!mutex_is_locked(&pm_mutex
));
170 WARN_ON(saved_gfp_mask
);
171 saved_gfp_mask
= gfp_allowed_mask
;
172 gfp_allowed_mask
&= ~GFP_IOFS
;
175 bool pm_suspended_storage(void)
177 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
181 #endif /* CONFIG_PM_SLEEP */
183 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
184 int pageblock_order __read_mostly
;
187 static void __free_pages_ok(struct page
*page
, unsigned int order
);
190 * results with 256, 32 in the lowmem_reserve sysctl:
191 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
192 * 1G machine -> (16M dma, 784M normal, 224M high)
193 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
194 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
195 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
197 * TBD: should special case ZONE_DMA32 machines here - in those we normally
198 * don't need any ZONE_NORMAL reservation
200 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
201 #ifdef CONFIG_ZONE_DMA
204 #ifdef CONFIG_ZONE_DMA32
207 #ifdef CONFIG_HIGHMEM
213 EXPORT_SYMBOL(totalram_pages
);
215 static char * const zone_names
[MAX_NR_ZONES
] = {
216 #ifdef CONFIG_ZONE_DMA
219 #ifdef CONFIG_ZONE_DMA32
223 #ifdef CONFIG_HIGHMEM
229 int min_free_kbytes
= 1024;
230 int user_min_free_kbytes
= -1;
232 static unsigned long __meminitdata nr_kernel_pages
;
233 static unsigned long __meminitdata nr_all_pages
;
234 static unsigned long __meminitdata dma_reserve
;
236 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
237 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
238 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
239 static unsigned long __initdata required_kernelcore
;
240 static unsigned long __initdata required_movablecore
;
241 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
243 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
245 EXPORT_SYMBOL(movable_zone
);
246 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
249 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
250 int nr_online_nodes __read_mostly
= 1;
251 EXPORT_SYMBOL(nr_node_ids
);
252 EXPORT_SYMBOL(nr_online_nodes
);
255 int page_group_by_mobility_disabled __read_mostly
;
257 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
258 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
260 pgdat
->first_deferred_pfn
= ULONG_MAX
;
263 /* Returns true if the struct page for the pfn is uninitialised */
264 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
266 if (pfn
>= NODE_DATA(early_pfn_to_nid(pfn
))->first_deferred_pfn
)
272 static inline bool early_page_nid_uninitialised(unsigned long pfn
, int nid
)
274 if (pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
281 * Returns false when the remaining initialisation should be deferred until
282 * later in the boot cycle when it can be parallelised.
284 static inline bool update_defer_init(pg_data_t
*pgdat
,
285 unsigned long pfn
, unsigned long zone_end
,
286 unsigned long *nr_initialised
)
288 /* Always populate low zones for address-contrained allocations */
289 if (zone_end
< pgdat_end_pfn(pgdat
))
292 /* Initialise at least 2G of the highest zone */
294 if (*nr_initialised
> (2UL << (30 - PAGE_SHIFT
)) &&
295 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
296 pgdat
->first_deferred_pfn
= pfn
;
303 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
307 static inline bool early_page_uninitialised(unsigned long pfn
)
312 static inline bool early_page_nid_uninitialised(unsigned long pfn
, int nid
)
317 static inline bool update_defer_init(pg_data_t
*pgdat
,
318 unsigned long pfn
, unsigned long zone_end
,
319 unsigned long *nr_initialised
)
326 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
328 if (unlikely(page_group_by_mobility_disabled
&&
329 migratetype
< MIGRATE_PCPTYPES
))
330 migratetype
= MIGRATE_UNMOVABLE
;
332 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
333 PB_migrate
, PB_migrate_end
);
336 #ifdef CONFIG_DEBUG_VM
337 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
341 unsigned long pfn
= page_to_pfn(page
);
342 unsigned long sp
, start_pfn
;
345 seq
= zone_span_seqbegin(zone
);
346 start_pfn
= zone
->zone_start_pfn
;
347 sp
= zone
->spanned_pages
;
348 if (!zone_spans_pfn(zone
, pfn
))
350 } while (zone_span_seqretry(zone
, seq
));
353 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
354 pfn
, zone_to_nid(zone
), zone
->name
,
355 start_pfn
, start_pfn
+ sp
);
360 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
362 if (!pfn_valid_within(page_to_pfn(page
)))
364 if (zone
!= page_zone(page
))
370 * Temporary debugging check for pages not lying within a given zone.
372 static int bad_range(struct zone
*zone
, struct page
*page
)
374 if (page_outside_zone_boundaries(zone
, page
))
376 if (!page_is_consistent(zone
, page
))
382 static inline int bad_range(struct zone
*zone
, struct page
*page
)
388 static void bad_page(struct page
*page
, const char *reason
,
389 unsigned long bad_flags
)
391 static unsigned long resume
;
392 static unsigned long nr_shown
;
393 static unsigned long nr_unshown
;
395 /* Don't complain about poisoned pages */
396 if (PageHWPoison(page
)) {
397 page_mapcount_reset(page
); /* remove PageBuddy */
402 * Allow a burst of 60 reports, then keep quiet for that minute;
403 * or allow a steady drip of one report per second.
405 if (nr_shown
== 60) {
406 if (time_before(jiffies
, resume
)) {
412 "BUG: Bad page state: %lu messages suppressed\n",
419 resume
= jiffies
+ 60 * HZ
;
421 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
422 current
->comm
, page_to_pfn(page
));
423 dump_page_badflags(page
, reason
, bad_flags
);
428 /* Leave bad fields for debug, except PageBuddy could make trouble */
429 page_mapcount_reset(page
); /* remove PageBuddy */
430 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
434 * Higher-order pages are called "compound pages". They are structured thusly:
436 * The first PAGE_SIZE page is called the "head page".
438 * The remaining PAGE_SIZE pages are called "tail pages".
440 * All pages have PG_compound set. All tail pages have their ->first_page
441 * pointing at the head page.
443 * The first tail page's ->lru.next holds the address of the compound page's
444 * put_page() function. Its ->lru.prev holds the order of allocation.
445 * This usage means that zero-order pages may not be compound.
448 static void free_compound_page(struct page
*page
)
450 __free_pages_ok(page
, compound_order(page
));
453 void prep_compound_page(struct page
*page
, unsigned long order
)
456 int nr_pages
= 1 << order
;
458 set_compound_page_dtor(page
, free_compound_page
);
459 set_compound_order(page
, order
);
461 for (i
= 1; i
< nr_pages
; i
++) {
462 struct page
*p
= page
+ i
;
463 set_page_count(p
, 0);
464 p
->first_page
= page
;
465 /* Make sure p->first_page is always valid for PageTail() */
471 #ifdef CONFIG_DEBUG_PAGEALLOC
472 unsigned int _debug_guardpage_minorder
;
473 bool _debug_pagealloc_enabled __read_mostly
;
474 bool _debug_guardpage_enabled __read_mostly
;
476 static int __init
early_debug_pagealloc(char *buf
)
481 if (strcmp(buf
, "on") == 0)
482 _debug_pagealloc_enabled
= true;
486 early_param("debug_pagealloc", early_debug_pagealloc
);
488 static bool need_debug_guardpage(void)
490 /* If we don't use debug_pagealloc, we don't need guard page */
491 if (!debug_pagealloc_enabled())
497 static void init_debug_guardpage(void)
499 if (!debug_pagealloc_enabled())
502 _debug_guardpage_enabled
= true;
505 struct page_ext_operations debug_guardpage_ops
= {
506 .need
= need_debug_guardpage
,
507 .init
= init_debug_guardpage
,
510 static int __init
debug_guardpage_minorder_setup(char *buf
)
514 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
515 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
518 _debug_guardpage_minorder
= res
;
519 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
522 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
524 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
525 unsigned int order
, int migratetype
)
527 struct page_ext
*page_ext
;
529 if (!debug_guardpage_enabled())
532 page_ext
= lookup_page_ext(page
);
533 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
535 INIT_LIST_HEAD(&page
->lru
);
536 set_page_private(page
, order
);
537 /* Guard pages are not available for any usage */
538 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
541 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
542 unsigned int order
, int migratetype
)
544 struct page_ext
*page_ext
;
546 if (!debug_guardpage_enabled())
549 page_ext
= lookup_page_ext(page
);
550 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
552 set_page_private(page
, 0);
553 if (!is_migrate_isolate(migratetype
))
554 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
557 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
558 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
559 unsigned int order
, int migratetype
) {}
560 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
561 unsigned int order
, int migratetype
) {}
564 static inline void set_page_order(struct page
*page
, unsigned int order
)
566 set_page_private(page
, order
);
567 __SetPageBuddy(page
);
570 static inline void rmv_page_order(struct page
*page
)
572 __ClearPageBuddy(page
);
573 set_page_private(page
, 0);
577 * This function checks whether a page is free && is the buddy
578 * we can do coalesce a page and its buddy if
579 * (a) the buddy is not in a hole &&
580 * (b) the buddy is in the buddy system &&
581 * (c) a page and its buddy have the same order &&
582 * (d) a page and its buddy are in the same zone.
584 * For recording whether a page is in the buddy system, we set ->_mapcount
585 * PAGE_BUDDY_MAPCOUNT_VALUE.
586 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
587 * serialized by zone->lock.
589 * For recording page's order, we use page_private(page).
591 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
594 if (!pfn_valid_within(page_to_pfn(buddy
)))
597 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
598 if (page_zone_id(page
) != page_zone_id(buddy
))
601 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
606 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
608 * zone check is done late to avoid uselessly
609 * calculating zone/node ids for pages that could
612 if (page_zone_id(page
) != page_zone_id(buddy
))
615 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
623 * Freeing function for a buddy system allocator.
625 * The concept of a buddy system is to maintain direct-mapped table
626 * (containing bit values) for memory blocks of various "orders".
627 * The bottom level table contains the map for the smallest allocatable
628 * units of memory (here, pages), and each level above it describes
629 * pairs of units from the levels below, hence, "buddies".
630 * At a high level, all that happens here is marking the table entry
631 * at the bottom level available, and propagating the changes upward
632 * as necessary, plus some accounting needed to play nicely with other
633 * parts of the VM system.
634 * At each level, we keep a list of pages, which are heads of continuous
635 * free pages of length of (1 << order) and marked with _mapcount
636 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
638 * So when we are allocating or freeing one, we can derive the state of the
639 * other. That is, if we allocate a small block, and both were
640 * free, the remainder of the region must be split into blocks.
641 * If a block is freed, and its buddy is also free, then this
642 * triggers coalescing into a block of larger size.
647 static inline void __free_one_page(struct page
*page
,
649 struct zone
*zone
, unsigned int order
,
652 unsigned long page_idx
;
653 unsigned long combined_idx
;
654 unsigned long uninitialized_var(buddy_idx
);
656 int max_order
= MAX_ORDER
;
658 VM_BUG_ON(!zone_is_initialized(zone
));
659 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
661 VM_BUG_ON(migratetype
== -1);
662 if (is_migrate_isolate(migratetype
)) {
664 * We restrict max order of merging to prevent merge
665 * between freepages on isolate pageblock and normal
666 * pageblock. Without this, pageblock isolation
667 * could cause incorrect freepage accounting.
669 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
671 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
674 page_idx
= pfn
& ((1 << max_order
) - 1);
676 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
677 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
679 while (order
< max_order
- 1) {
680 buddy_idx
= __find_buddy_index(page_idx
, order
);
681 buddy
= page
+ (buddy_idx
- page_idx
);
682 if (!page_is_buddy(page
, buddy
, order
))
685 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
686 * merge with it and move up one order.
688 if (page_is_guard(buddy
)) {
689 clear_page_guard(zone
, buddy
, order
, migratetype
);
691 list_del(&buddy
->lru
);
692 zone
->free_area
[order
].nr_free
--;
693 rmv_page_order(buddy
);
695 combined_idx
= buddy_idx
& page_idx
;
696 page
= page
+ (combined_idx
- page_idx
);
697 page_idx
= combined_idx
;
700 set_page_order(page
, order
);
703 * If this is not the largest possible page, check if the buddy
704 * of the next-highest order is free. If it is, it's possible
705 * that pages are being freed that will coalesce soon. In case,
706 * that is happening, add the free page to the tail of the list
707 * so it's less likely to be used soon and more likely to be merged
708 * as a higher order page
710 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
711 struct page
*higher_page
, *higher_buddy
;
712 combined_idx
= buddy_idx
& page_idx
;
713 higher_page
= page
+ (combined_idx
- page_idx
);
714 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
715 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
716 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
717 list_add_tail(&page
->lru
,
718 &zone
->free_area
[order
].free_list
[migratetype
]);
723 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
725 zone
->free_area
[order
].nr_free
++;
728 static inline int free_pages_check(struct page
*page
)
730 const char *bad_reason
= NULL
;
731 unsigned long bad_flags
= 0;
733 if (unlikely(page_mapcount(page
)))
734 bad_reason
= "nonzero mapcount";
735 if (unlikely(page
->mapping
!= NULL
))
736 bad_reason
= "non-NULL mapping";
737 if (unlikely(atomic_read(&page
->_count
) != 0))
738 bad_reason
= "nonzero _count";
739 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
740 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
741 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
744 if (unlikely(page
->mem_cgroup
))
745 bad_reason
= "page still charged to cgroup";
747 if (unlikely(bad_reason
)) {
748 bad_page(page
, bad_reason
, bad_flags
);
751 page_cpupid_reset_last(page
);
752 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
753 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
758 * Frees a number of pages from the PCP lists
759 * Assumes all pages on list are in same zone, and of same order.
760 * count is the number of pages to free.
762 * If the zone was previously in an "all pages pinned" state then look to
763 * see if this freeing clears that state.
765 * And clear the zone's pages_scanned counter, to hold off the "all pages are
766 * pinned" detection logic.
768 static void free_pcppages_bulk(struct zone
*zone
, int count
,
769 struct per_cpu_pages
*pcp
)
774 unsigned long nr_scanned
;
776 spin_lock(&zone
->lock
);
777 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
779 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
783 struct list_head
*list
;
786 * Remove pages from lists in a round-robin fashion. A
787 * batch_free count is maintained that is incremented when an
788 * empty list is encountered. This is so more pages are freed
789 * off fuller lists instead of spinning excessively around empty
794 if (++migratetype
== MIGRATE_PCPTYPES
)
796 list
= &pcp
->lists
[migratetype
];
797 } while (list_empty(list
));
799 /* This is the only non-empty list. Free them all. */
800 if (batch_free
== MIGRATE_PCPTYPES
)
801 batch_free
= to_free
;
804 int mt
; /* migratetype of the to-be-freed page */
806 page
= list_entry(list
->prev
, struct page
, lru
);
807 /* must delete as __free_one_page list manipulates */
808 list_del(&page
->lru
);
810 mt
= get_pcppage_migratetype(page
);
811 /* MIGRATE_ISOLATE page should not go to pcplists */
812 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
813 /* Pageblock could have been isolated meanwhile */
814 if (unlikely(has_isolate_pageblock(zone
)))
815 mt
= get_pageblock_migratetype(page
);
817 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
818 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
819 trace_mm_page_pcpu_drain(page
, 0, mt
);
820 } while (--to_free
&& --batch_free
&& !list_empty(list
));
822 spin_unlock(&zone
->lock
);
825 static void free_one_page(struct zone
*zone
,
826 struct page
*page
, unsigned long pfn
,
830 unsigned long nr_scanned
;
831 spin_lock(&zone
->lock
);
832 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
834 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
836 if (unlikely(has_isolate_pageblock(zone
) ||
837 is_migrate_isolate(migratetype
))) {
838 migratetype
= get_pfnblock_migratetype(page
, pfn
);
840 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
841 spin_unlock(&zone
->lock
);
844 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
846 if (!IS_ENABLED(CONFIG_DEBUG_VM
))
848 if (unlikely(!PageTail(page
))) {
849 bad_page(page
, "PageTail not set", 0);
852 if (unlikely(page
->first_page
!= head_page
)) {
853 bad_page(page
, "first_page not consistent", 0);
859 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
860 unsigned long zone
, int nid
)
862 set_page_links(page
, zone
, nid
, pfn
);
863 init_page_count(page
);
864 page_mapcount_reset(page
);
865 page_cpupid_reset_last(page
);
867 INIT_LIST_HEAD(&page
->lru
);
868 #ifdef WANT_PAGE_VIRTUAL
869 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
870 if (!is_highmem_idx(zone
))
871 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
875 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
878 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
881 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
882 static void init_reserved_page(unsigned long pfn
)
887 if (!early_page_uninitialised(pfn
))
890 nid
= early_pfn_to_nid(pfn
);
891 pgdat
= NODE_DATA(nid
);
893 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
894 struct zone
*zone
= &pgdat
->node_zones
[zid
];
896 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
899 __init_single_pfn(pfn
, zid
, nid
);
902 static inline void init_reserved_page(unsigned long pfn
)
905 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
908 * Initialised pages do not have PageReserved set. This function is
909 * called for each range allocated by the bootmem allocator and
910 * marks the pages PageReserved. The remaining valid pages are later
911 * sent to the buddy page allocator.
913 void __meminit
reserve_bootmem_region(unsigned long start
, unsigned long end
)
915 unsigned long start_pfn
= PFN_DOWN(start
);
916 unsigned long end_pfn
= PFN_UP(end
);
918 for (; start_pfn
< end_pfn
; start_pfn
++) {
919 if (pfn_valid(start_pfn
)) {
920 struct page
*page
= pfn_to_page(start_pfn
);
922 init_reserved_page(start_pfn
);
923 SetPageReserved(page
);
928 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
930 bool compound
= PageCompound(page
);
933 VM_BUG_ON_PAGE(PageTail(page
), page
);
934 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
936 trace_mm_page_free(page
, order
);
937 kmemcheck_free_shadow(page
, order
);
938 kasan_free_pages(page
, order
);
941 page
->mapping
= NULL
;
942 bad
+= free_pages_check(page
);
943 for (i
= 1; i
< (1 << order
); i
++) {
945 bad
+= free_tail_pages_check(page
, page
+ i
);
946 bad
+= free_pages_check(page
+ i
);
951 reset_page_owner(page
, order
);
953 if (!PageHighMem(page
)) {
954 debug_check_no_locks_freed(page_address(page
),
956 debug_check_no_obj_freed(page_address(page
),
959 arch_free_page(page
, order
);
960 kernel_map_pages(page
, 1 << order
, 0);
965 static void __free_pages_ok(struct page
*page
, unsigned int order
)
969 unsigned long pfn
= page_to_pfn(page
);
971 if (!free_pages_prepare(page
, order
))
974 migratetype
= get_pfnblock_migratetype(page
, pfn
);
975 local_irq_save(flags
);
976 __count_vm_events(PGFREE
, 1 << order
);
977 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
978 local_irq_restore(flags
);
981 static void __init
__free_pages_boot_core(struct page
*page
,
982 unsigned long pfn
, unsigned int order
)
984 unsigned int nr_pages
= 1 << order
;
985 struct page
*p
= page
;
989 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
991 __ClearPageReserved(p
);
992 set_page_count(p
, 0);
994 __ClearPageReserved(p
);
995 set_page_count(p
, 0);
997 page_zone(page
)->managed_pages
+= nr_pages
;
998 set_page_refcounted(page
);
999 __free_pages(page
, order
);
1002 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1003 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1005 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1007 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1009 static DEFINE_SPINLOCK(early_pfn_lock
);
1012 spin_lock(&early_pfn_lock
);
1013 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1016 spin_unlock(&early_pfn_lock
);
1022 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1023 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1024 struct mminit_pfnnid_cache
*state
)
1028 nid
= __early_pfn_to_nid(pfn
, state
);
1029 if (nid
>= 0 && nid
!= node
)
1034 /* Only safe to use early in boot when initialisation is single-threaded */
1035 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1037 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1042 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1046 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1047 struct mminit_pfnnid_cache
*state
)
1054 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1057 if (early_page_uninitialised(pfn
))
1059 return __free_pages_boot_core(page
, pfn
, order
);
1062 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1063 static void __init
deferred_free_range(struct page
*page
,
1064 unsigned long pfn
, int nr_pages
)
1071 /* Free a large naturally-aligned chunk if possible */
1072 if (nr_pages
== MAX_ORDER_NR_PAGES
&&
1073 (pfn
& (MAX_ORDER_NR_PAGES
-1)) == 0) {
1074 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1075 __free_pages_boot_core(page
, pfn
, MAX_ORDER
-1);
1079 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++)
1080 __free_pages_boot_core(page
, pfn
, 0);
1083 /* Completion tracking for deferred_init_memmap() threads */
1084 static atomic_t pgdat_init_n_undone __initdata
;
1085 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1087 static inline void __init
pgdat_init_report_one_done(void)
1089 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1090 complete(&pgdat_init_all_done_comp
);
1093 /* Initialise remaining memory on a node */
1094 static int __init
deferred_init_memmap(void *data
)
1096 pg_data_t
*pgdat
= data
;
1097 int nid
= pgdat
->node_id
;
1098 struct mminit_pfnnid_cache nid_init_state
= { };
1099 unsigned long start
= jiffies
;
1100 unsigned long nr_pages
= 0;
1101 unsigned long walk_start
, walk_end
;
1104 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1105 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1107 if (first_init_pfn
== ULONG_MAX
) {
1108 pgdat_init_report_one_done();
1112 /* Bind memory initialisation thread to a local node if possible */
1113 if (!cpumask_empty(cpumask
))
1114 set_cpus_allowed_ptr(current
, cpumask
);
1116 /* Sanity check boundaries */
1117 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1118 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1119 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1121 /* Only the highest zone is deferred so find it */
1122 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1123 zone
= pgdat
->node_zones
+ zid
;
1124 if (first_init_pfn
< zone_end_pfn(zone
))
1128 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1129 unsigned long pfn
, end_pfn
;
1130 struct page
*page
= NULL
;
1131 struct page
*free_base_page
= NULL
;
1132 unsigned long free_base_pfn
= 0;
1135 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1136 pfn
= first_init_pfn
;
1137 if (pfn
< walk_start
)
1139 if (pfn
< zone
->zone_start_pfn
)
1140 pfn
= zone
->zone_start_pfn
;
1142 for (; pfn
< end_pfn
; pfn
++) {
1143 if (!pfn_valid_within(pfn
))
1147 * Ensure pfn_valid is checked every
1148 * MAX_ORDER_NR_PAGES for memory holes
1150 if ((pfn
& (MAX_ORDER_NR_PAGES
- 1)) == 0) {
1151 if (!pfn_valid(pfn
)) {
1157 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1162 /* Minimise pfn page lookups and scheduler checks */
1163 if (page
&& (pfn
& (MAX_ORDER_NR_PAGES
- 1)) != 0) {
1166 nr_pages
+= nr_to_free
;
1167 deferred_free_range(free_base_page
,
1168 free_base_pfn
, nr_to_free
);
1169 free_base_page
= NULL
;
1170 free_base_pfn
= nr_to_free
= 0;
1172 page
= pfn_to_page(pfn
);
1177 VM_BUG_ON(page_zone(page
) != zone
);
1181 __init_single_page(page
, pfn
, zid
, nid
);
1182 if (!free_base_page
) {
1183 free_base_page
= page
;
1184 free_base_pfn
= pfn
;
1189 /* Where possible, batch up pages for a single free */
1192 /* Free the current block of pages to allocator */
1193 nr_pages
+= nr_to_free
;
1194 deferred_free_range(free_base_page
, free_base_pfn
,
1196 free_base_page
= NULL
;
1197 free_base_pfn
= nr_to_free
= 0;
1200 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1203 /* Sanity check that the next zone really is unpopulated */
1204 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1206 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1207 jiffies_to_msecs(jiffies
- start
));
1209 pgdat_init_report_one_done();
1213 void __init
page_alloc_init_late(void)
1217 /* There will be num_node_state(N_MEMORY) threads */
1218 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1219 for_each_node_state(nid
, N_MEMORY
) {
1220 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1223 /* Block until all are initialised */
1224 wait_for_completion(&pgdat_init_all_done_comp
);
1226 /* Reinit limits that are based on free pages after the kernel is up */
1227 files_maxfiles_init();
1229 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1232 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1233 void __init
init_cma_reserved_pageblock(struct page
*page
)
1235 unsigned i
= pageblock_nr_pages
;
1236 struct page
*p
= page
;
1239 __ClearPageReserved(p
);
1240 set_page_count(p
, 0);
1243 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1245 if (pageblock_order
>= MAX_ORDER
) {
1246 i
= pageblock_nr_pages
;
1249 set_page_refcounted(p
);
1250 __free_pages(p
, MAX_ORDER
- 1);
1251 p
+= MAX_ORDER_NR_PAGES
;
1252 } while (i
-= MAX_ORDER_NR_PAGES
);
1254 set_page_refcounted(page
);
1255 __free_pages(page
, pageblock_order
);
1258 adjust_managed_page_count(page
, pageblock_nr_pages
);
1263 * The order of subdivision here is critical for the IO subsystem.
1264 * Please do not alter this order without good reasons and regression
1265 * testing. Specifically, as large blocks of memory are subdivided,
1266 * the order in which smaller blocks are delivered depends on the order
1267 * they're subdivided in this function. This is the primary factor
1268 * influencing the order in which pages are delivered to the IO
1269 * subsystem according to empirical testing, and this is also justified
1270 * by considering the behavior of a buddy system containing a single
1271 * large block of memory acted on by a series of small allocations.
1272 * This behavior is a critical factor in sglist merging's success.
1276 static inline void expand(struct zone
*zone
, struct page
*page
,
1277 int low
, int high
, struct free_area
*area
,
1280 unsigned long size
= 1 << high
;
1282 while (high
> low
) {
1286 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1288 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
1289 debug_guardpage_enabled() &&
1290 high
< debug_guardpage_minorder()) {
1292 * Mark as guard pages (or page), that will allow to
1293 * merge back to allocator when buddy will be freed.
1294 * Corresponding page table entries will not be touched,
1295 * pages will stay not present in virtual address space
1297 set_page_guard(zone
, &page
[size
], high
, migratetype
);
1300 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1302 set_page_order(&page
[size
], high
);
1307 * This page is about to be returned from the page allocator
1309 static inline int check_new_page(struct page
*page
)
1311 const char *bad_reason
= NULL
;
1312 unsigned long bad_flags
= 0;
1314 if (unlikely(page_mapcount(page
)))
1315 bad_reason
= "nonzero mapcount";
1316 if (unlikely(page
->mapping
!= NULL
))
1317 bad_reason
= "non-NULL mapping";
1318 if (unlikely(atomic_read(&page
->_count
) != 0))
1319 bad_reason
= "nonzero _count";
1320 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1321 bad_reason
= "HWPoisoned (hardware-corrupted)";
1322 bad_flags
= __PG_HWPOISON
;
1324 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1325 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1326 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1329 if (unlikely(page
->mem_cgroup
))
1330 bad_reason
= "page still charged to cgroup";
1332 if (unlikely(bad_reason
)) {
1333 bad_page(page
, bad_reason
, bad_flags
);
1339 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1344 for (i
= 0; i
< (1 << order
); i
++) {
1345 struct page
*p
= page
+ i
;
1346 if (unlikely(check_new_page(p
)))
1350 set_page_private(page
, 0);
1351 set_page_refcounted(page
);
1353 arch_alloc_page(page
, order
);
1354 kernel_map_pages(page
, 1 << order
, 1);
1355 kasan_alloc_pages(page
, order
);
1357 if (gfp_flags
& __GFP_ZERO
)
1358 for (i
= 0; i
< (1 << order
); i
++)
1359 clear_highpage(page
+ i
);
1361 if (order
&& (gfp_flags
& __GFP_COMP
))
1362 prep_compound_page(page
, order
);
1364 set_page_owner(page
, order
, gfp_flags
);
1367 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1368 * allocate the page. The expectation is that the caller is taking
1369 * steps that will free more memory. The caller should avoid the page
1370 * being used for !PFMEMALLOC purposes.
1372 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1373 set_page_pfmemalloc(page
);
1375 clear_page_pfmemalloc(page
);
1381 * Go through the free lists for the given migratetype and remove
1382 * the smallest available page from the freelists
1385 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1388 unsigned int current_order
;
1389 struct free_area
*area
;
1392 /* Find a page of the appropriate size in the preferred list */
1393 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1394 area
= &(zone
->free_area
[current_order
]);
1395 if (list_empty(&area
->free_list
[migratetype
]))
1398 page
= list_entry(area
->free_list
[migratetype
].next
,
1400 list_del(&page
->lru
);
1401 rmv_page_order(page
);
1403 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1404 set_pcppage_migratetype(page
, migratetype
);
1413 * This array describes the order lists are fallen back to when
1414 * the free lists for the desirable migrate type are depleted
1416 static int fallbacks
[MIGRATE_TYPES
][4] = {
1417 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1418 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1419 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
1421 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
1423 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
1424 #ifdef CONFIG_MEMORY_ISOLATION
1425 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
1430 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1433 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1436 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1437 unsigned int order
) { return NULL
; }
1441 * Move the free pages in a range to the free lists of the requested type.
1442 * Note that start_page and end_pages are not aligned on a pageblock
1443 * boundary. If alignment is required, use move_freepages_block()
1445 int move_freepages(struct zone
*zone
,
1446 struct page
*start_page
, struct page
*end_page
,
1450 unsigned long order
;
1451 int pages_moved
= 0;
1453 #ifndef CONFIG_HOLES_IN_ZONE
1455 * page_zone is not safe to call in this context when
1456 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1457 * anyway as we check zone boundaries in move_freepages_block().
1458 * Remove at a later date when no bug reports exist related to
1459 * grouping pages by mobility
1461 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1464 for (page
= start_page
; page
<= end_page
;) {
1465 /* Make sure we are not inadvertently changing nodes */
1466 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1468 if (!pfn_valid_within(page_to_pfn(page
))) {
1473 if (!PageBuddy(page
)) {
1478 order
= page_order(page
);
1479 list_move(&page
->lru
,
1480 &zone
->free_area
[order
].free_list
[migratetype
]);
1482 pages_moved
+= 1 << order
;
1488 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1491 unsigned long start_pfn
, end_pfn
;
1492 struct page
*start_page
, *end_page
;
1494 start_pfn
= page_to_pfn(page
);
1495 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1496 start_page
= pfn_to_page(start_pfn
);
1497 end_page
= start_page
+ pageblock_nr_pages
- 1;
1498 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1500 /* Do not cross zone boundaries */
1501 if (!zone_spans_pfn(zone
, start_pfn
))
1503 if (!zone_spans_pfn(zone
, end_pfn
))
1506 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1509 static void change_pageblock_range(struct page
*pageblock_page
,
1510 int start_order
, int migratetype
)
1512 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1514 while (nr_pageblocks
--) {
1515 set_pageblock_migratetype(pageblock_page
, migratetype
);
1516 pageblock_page
+= pageblock_nr_pages
;
1521 * When we are falling back to another migratetype during allocation, try to
1522 * steal extra free pages from the same pageblocks to satisfy further
1523 * allocations, instead of polluting multiple pageblocks.
1525 * If we are stealing a relatively large buddy page, it is likely there will
1526 * be more free pages in the pageblock, so try to steal them all. For
1527 * reclaimable and unmovable allocations, we steal regardless of page size,
1528 * as fragmentation caused by those allocations polluting movable pageblocks
1529 * is worse than movable allocations stealing from unmovable and reclaimable
1532 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1535 * Leaving this order check is intended, although there is
1536 * relaxed order check in next check. The reason is that
1537 * we can actually steal whole pageblock if this condition met,
1538 * but, below check doesn't guarantee it and that is just heuristic
1539 * so could be changed anytime.
1541 if (order
>= pageblock_order
)
1544 if (order
>= pageblock_order
/ 2 ||
1545 start_mt
== MIGRATE_RECLAIMABLE
||
1546 start_mt
== MIGRATE_UNMOVABLE
||
1547 page_group_by_mobility_disabled
)
1554 * This function implements actual steal behaviour. If order is large enough,
1555 * we can steal whole pageblock. If not, we first move freepages in this
1556 * pageblock and check whether half of pages are moved or not. If half of
1557 * pages are moved, we can change migratetype of pageblock and permanently
1558 * use it's pages as requested migratetype in the future.
1560 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1563 int current_order
= page_order(page
);
1566 /* Take ownership for orders >= pageblock_order */
1567 if (current_order
>= pageblock_order
) {
1568 change_pageblock_range(page
, current_order
, start_type
);
1572 pages
= move_freepages_block(zone
, page
, start_type
);
1574 /* Claim the whole block if over half of it is free */
1575 if (pages
>= (1 << (pageblock_order
-1)) ||
1576 page_group_by_mobility_disabled
)
1577 set_pageblock_migratetype(page
, start_type
);
1581 * Check whether there is a suitable fallback freepage with requested order.
1582 * If only_stealable is true, this function returns fallback_mt only if
1583 * we can steal other freepages all together. This would help to reduce
1584 * fragmentation due to mixed migratetype pages in one pageblock.
1586 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1587 int migratetype
, bool only_stealable
, bool *can_steal
)
1592 if (area
->nr_free
== 0)
1597 fallback_mt
= fallbacks
[migratetype
][i
];
1598 if (fallback_mt
== MIGRATE_RESERVE
)
1601 if (list_empty(&area
->free_list
[fallback_mt
]))
1604 if (can_steal_fallback(order
, migratetype
))
1607 if (!only_stealable
)
1617 /* Remove an element from the buddy allocator from the fallback list */
1618 static inline struct page
*
1619 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1621 struct free_area
*area
;
1622 unsigned int current_order
;
1627 /* Find the largest possible block of pages in the other list */
1628 for (current_order
= MAX_ORDER
-1;
1629 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1631 area
= &(zone
->free_area
[current_order
]);
1632 fallback_mt
= find_suitable_fallback(area
, current_order
,
1633 start_migratetype
, false, &can_steal
);
1634 if (fallback_mt
== -1)
1637 page
= list_entry(area
->free_list
[fallback_mt
].next
,
1640 steal_suitable_fallback(zone
, page
, start_migratetype
);
1642 /* Remove the page from the freelists */
1644 list_del(&page
->lru
);
1645 rmv_page_order(page
);
1647 expand(zone
, page
, order
, current_order
, area
,
1650 * The pcppage_migratetype may differ from pageblock's
1651 * migratetype depending on the decisions in
1652 * find_suitable_fallback(). This is OK as long as it does not
1653 * differ for MIGRATE_CMA pageblocks. Those can be used as
1654 * fallback only via special __rmqueue_cma_fallback() function
1656 set_pcppage_migratetype(page
, start_migratetype
);
1658 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1659 start_migratetype
, fallback_mt
);
1668 * Do the hard work of removing an element from the buddy allocator.
1669 * Call me with the zone->lock already held.
1671 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1677 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1679 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1680 if (migratetype
== MIGRATE_MOVABLE
)
1681 page
= __rmqueue_cma_fallback(zone
, order
);
1684 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1687 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1688 * is used because __rmqueue_smallest is an inline function
1689 * and we want just one call site
1692 migratetype
= MIGRATE_RESERVE
;
1697 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1702 * Obtain a specified number of elements from the buddy allocator, all under
1703 * a single hold of the lock, for efficiency. Add them to the supplied list.
1704 * Returns the number of new pages which were placed at *list.
1706 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1707 unsigned long count
, struct list_head
*list
,
1708 int migratetype
, bool cold
)
1712 spin_lock(&zone
->lock
);
1713 for (i
= 0; i
< count
; ++i
) {
1714 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1715 if (unlikely(page
== NULL
))
1719 * Split buddy pages returned by expand() are received here
1720 * in physical page order. The page is added to the callers and
1721 * list and the list head then moves forward. From the callers
1722 * perspective, the linked list is ordered by page number in
1723 * some conditions. This is useful for IO devices that can
1724 * merge IO requests if the physical pages are ordered
1728 list_add(&page
->lru
, list
);
1730 list_add_tail(&page
->lru
, list
);
1732 if (is_migrate_cma(get_pcppage_migratetype(page
)))
1733 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1736 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1737 spin_unlock(&zone
->lock
);
1743 * Called from the vmstat counter updater to drain pagesets of this
1744 * currently executing processor on remote nodes after they have
1747 * Note that this function must be called with the thread pinned to
1748 * a single processor.
1750 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1752 unsigned long flags
;
1753 int to_drain
, batch
;
1755 local_irq_save(flags
);
1756 batch
= READ_ONCE(pcp
->batch
);
1757 to_drain
= min(pcp
->count
, batch
);
1759 free_pcppages_bulk(zone
, to_drain
, pcp
);
1760 pcp
->count
-= to_drain
;
1762 local_irq_restore(flags
);
1767 * Drain pcplists of the indicated processor and zone.
1769 * The processor must either be the current processor and the
1770 * thread pinned to the current processor or a processor that
1773 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1775 unsigned long flags
;
1776 struct per_cpu_pageset
*pset
;
1777 struct per_cpu_pages
*pcp
;
1779 local_irq_save(flags
);
1780 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1784 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1787 local_irq_restore(flags
);
1791 * Drain pcplists of all zones on the indicated processor.
1793 * The processor must either be the current processor and the
1794 * thread pinned to the current processor or a processor that
1797 static void drain_pages(unsigned int cpu
)
1801 for_each_populated_zone(zone
) {
1802 drain_pages_zone(cpu
, zone
);
1807 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1809 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1810 * the single zone's pages.
1812 void drain_local_pages(struct zone
*zone
)
1814 int cpu
= smp_processor_id();
1817 drain_pages_zone(cpu
, zone
);
1823 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1825 * When zone parameter is non-NULL, spill just the single zone's pages.
1827 * Note that this code is protected against sending an IPI to an offline
1828 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1829 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1830 * nothing keeps CPUs from showing up after we populated the cpumask and
1831 * before the call to on_each_cpu_mask().
1833 void drain_all_pages(struct zone
*zone
)
1838 * Allocate in the BSS so we wont require allocation in
1839 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1841 static cpumask_t cpus_with_pcps
;
1844 * We don't care about racing with CPU hotplug event
1845 * as offline notification will cause the notified
1846 * cpu to drain that CPU pcps and on_each_cpu_mask
1847 * disables preemption as part of its processing
1849 for_each_online_cpu(cpu
) {
1850 struct per_cpu_pageset
*pcp
;
1852 bool has_pcps
= false;
1855 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1859 for_each_populated_zone(z
) {
1860 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1861 if (pcp
->pcp
.count
) {
1869 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1871 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1873 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1877 #ifdef CONFIG_HIBERNATION
1879 void mark_free_pages(struct zone
*zone
)
1881 unsigned long pfn
, max_zone_pfn
;
1882 unsigned long flags
;
1883 unsigned int order
, t
;
1884 struct list_head
*curr
;
1886 if (zone_is_empty(zone
))
1889 spin_lock_irqsave(&zone
->lock
, flags
);
1891 max_zone_pfn
= zone_end_pfn(zone
);
1892 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1893 if (pfn_valid(pfn
)) {
1894 struct page
*page
= pfn_to_page(pfn
);
1896 if (!swsusp_page_is_forbidden(page
))
1897 swsusp_unset_page_free(page
);
1900 for_each_migratetype_order(order
, t
) {
1901 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1904 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1905 for (i
= 0; i
< (1UL << order
); i
++)
1906 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1909 spin_unlock_irqrestore(&zone
->lock
, flags
);
1911 #endif /* CONFIG_PM */
1914 * Free a 0-order page
1915 * cold == true ? free a cold page : free a hot page
1917 void free_hot_cold_page(struct page
*page
, bool cold
)
1919 struct zone
*zone
= page_zone(page
);
1920 struct per_cpu_pages
*pcp
;
1921 unsigned long flags
;
1922 unsigned long pfn
= page_to_pfn(page
);
1925 if (!free_pages_prepare(page
, 0))
1928 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1929 set_pcppage_migratetype(page
, migratetype
);
1930 local_irq_save(flags
);
1931 __count_vm_event(PGFREE
);
1934 * We only track unmovable, reclaimable and movable on pcp lists.
1935 * Free ISOLATE pages back to the allocator because they are being
1936 * offlined but treat RESERVE as movable pages so we can get those
1937 * areas back if necessary. Otherwise, we may have to free
1938 * excessively into the page allocator
1940 if (migratetype
>= MIGRATE_PCPTYPES
) {
1941 if (unlikely(is_migrate_isolate(migratetype
))) {
1942 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1945 migratetype
= MIGRATE_MOVABLE
;
1948 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1950 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1952 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1954 if (pcp
->count
>= pcp
->high
) {
1955 unsigned long batch
= READ_ONCE(pcp
->batch
);
1956 free_pcppages_bulk(zone
, batch
, pcp
);
1957 pcp
->count
-= batch
;
1961 local_irq_restore(flags
);
1965 * Free a list of 0-order pages
1967 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1969 struct page
*page
, *next
;
1971 list_for_each_entry_safe(page
, next
, list
, lru
) {
1972 trace_mm_page_free_batched(page
, cold
);
1973 free_hot_cold_page(page
, cold
);
1978 * split_page takes a non-compound higher-order page, and splits it into
1979 * n (1<<order) sub-pages: page[0..n]
1980 * Each sub-page must be freed individually.
1982 * Note: this is probably too low level an operation for use in drivers.
1983 * Please consult with lkml before using this in your driver.
1985 void split_page(struct page
*page
, unsigned int order
)
1990 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1991 VM_BUG_ON_PAGE(!page_count(page
), page
);
1993 #ifdef CONFIG_KMEMCHECK
1995 * Split shadow pages too, because free(page[0]) would
1996 * otherwise free the whole shadow.
1998 if (kmemcheck_page_is_tracked(page
))
1999 split_page(virt_to_page(page
[0].shadow
), order
);
2002 gfp_mask
= get_page_owner_gfp(page
);
2003 set_page_owner(page
, 0, gfp_mask
);
2004 for (i
= 1; i
< (1 << order
); i
++) {
2005 set_page_refcounted(page
+ i
);
2006 set_page_owner(page
+ i
, 0, gfp_mask
);
2009 EXPORT_SYMBOL_GPL(split_page
);
2011 int __isolate_free_page(struct page
*page
, unsigned int order
)
2013 unsigned long watermark
;
2017 BUG_ON(!PageBuddy(page
));
2019 zone
= page_zone(page
);
2020 mt
= get_pageblock_migratetype(page
);
2022 if (!is_migrate_isolate(mt
)) {
2023 /* Obey watermarks as if the page was being allocated */
2024 watermark
= low_wmark_pages(zone
) + (1 << order
);
2025 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
2028 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2031 /* Remove page from free list */
2032 list_del(&page
->lru
);
2033 zone
->free_area
[order
].nr_free
--;
2034 rmv_page_order(page
);
2036 set_page_owner(page
, order
, __GFP_MOVABLE
);
2038 /* Set the pageblock if the isolated page is at least a pageblock */
2039 if (order
>= pageblock_order
- 1) {
2040 struct page
*endpage
= page
+ (1 << order
) - 1;
2041 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2042 int mt
= get_pageblock_migratetype(page
);
2043 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
2044 set_pageblock_migratetype(page
,
2050 return 1UL << order
;
2054 * Similar to split_page except the page is already free. As this is only
2055 * being used for migration, the migratetype of the block also changes.
2056 * As this is called with interrupts disabled, the caller is responsible
2057 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2060 * Note: this is probably too low level an operation for use in drivers.
2061 * Please consult with lkml before using this in your driver.
2063 int split_free_page(struct page
*page
)
2068 order
= page_order(page
);
2070 nr_pages
= __isolate_free_page(page
, order
);
2074 /* Split into individual pages */
2075 set_page_refcounted(page
);
2076 split_page(page
, order
);
2081 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2084 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
2085 struct zone
*zone
, unsigned int order
,
2086 gfp_t gfp_flags
, int migratetype
)
2088 unsigned long flags
;
2090 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2092 if (likely(order
== 0)) {
2093 struct per_cpu_pages
*pcp
;
2094 struct list_head
*list
;
2096 local_irq_save(flags
);
2097 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2098 list
= &pcp
->lists
[migratetype
];
2099 if (list_empty(list
)) {
2100 pcp
->count
+= rmqueue_bulk(zone
, 0,
2103 if (unlikely(list_empty(list
)))
2108 page
= list_entry(list
->prev
, struct page
, lru
);
2110 page
= list_entry(list
->next
, struct page
, lru
);
2112 list_del(&page
->lru
);
2115 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
2117 * __GFP_NOFAIL is not to be used in new code.
2119 * All __GFP_NOFAIL callers should be fixed so that they
2120 * properly detect and handle allocation failures.
2122 * We most definitely don't want callers attempting to
2123 * allocate greater than order-1 page units with
2126 WARN_ON_ONCE(order
> 1);
2128 spin_lock_irqsave(&zone
->lock
, flags
);
2129 page
= __rmqueue(zone
, order
, migratetype
);
2130 spin_unlock(&zone
->lock
);
2133 __mod_zone_freepage_state(zone
, -(1 << order
),
2134 get_pcppage_migratetype(page
));
2137 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
2138 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
2139 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
2140 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2142 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
2143 zone_statistics(preferred_zone
, zone
, gfp_flags
);
2144 local_irq_restore(flags
);
2146 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
2150 local_irq_restore(flags
);
2154 #ifdef CONFIG_FAIL_PAGE_ALLOC
2157 struct fault_attr attr
;
2159 u32 ignore_gfp_highmem
;
2160 u32 ignore_gfp_wait
;
2162 } fail_page_alloc
= {
2163 .attr
= FAULT_ATTR_INITIALIZER
,
2164 .ignore_gfp_wait
= 1,
2165 .ignore_gfp_highmem
= 1,
2169 static int __init
setup_fail_page_alloc(char *str
)
2171 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2173 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2175 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2177 if (order
< fail_page_alloc
.min_order
)
2179 if (gfp_mask
& __GFP_NOFAIL
)
2181 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2183 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
2186 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2189 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2191 static int __init
fail_page_alloc_debugfs(void)
2193 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2196 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2197 &fail_page_alloc
.attr
);
2199 return PTR_ERR(dir
);
2201 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2202 &fail_page_alloc
.ignore_gfp_wait
))
2204 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2205 &fail_page_alloc
.ignore_gfp_highmem
))
2207 if (!debugfs_create_u32("min-order", mode
, dir
,
2208 &fail_page_alloc
.min_order
))
2213 debugfs_remove_recursive(dir
);
2218 late_initcall(fail_page_alloc_debugfs
);
2220 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2222 #else /* CONFIG_FAIL_PAGE_ALLOC */
2224 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2229 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2232 * Return true if free pages are above 'mark'. This takes into account the order
2233 * of the allocation.
2235 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
2236 unsigned long mark
, int classzone_idx
, int alloc_flags
,
2239 /* free_pages may go negative - that's OK */
2244 free_pages
-= (1 << order
) - 1;
2245 if (alloc_flags
& ALLOC_HIGH
)
2247 if (alloc_flags
& ALLOC_HARDER
)
2250 /* If allocation can't use CMA areas don't use free CMA pages */
2251 if (!(alloc_flags
& ALLOC_CMA
))
2252 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2255 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2257 for (o
= 0; o
< order
; o
++) {
2258 /* At the next order, this order's pages become unavailable */
2259 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
2261 /* Require fewer higher order pages to be free */
2264 if (free_pages
<= min
)
2270 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2271 int classzone_idx
, int alloc_flags
)
2273 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2274 zone_page_state(z
, NR_FREE_PAGES
));
2277 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2278 unsigned long mark
, int classzone_idx
, int alloc_flags
)
2280 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2282 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2283 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2285 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2291 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
2292 * skip over zones that are not allowed by the cpuset, or that have
2293 * been recently (in last second) found to be nearly full. See further
2294 * comments in mmzone.h. Reduces cache footprint of zonelist scans
2295 * that have to skip over a lot of full or unallowed zones.
2297 * If the zonelist cache is present in the passed zonelist, then
2298 * returns a pointer to the allowed node mask (either the current
2299 * tasks mems_allowed, or node_states[N_MEMORY].)
2301 * If the zonelist cache is not available for this zonelist, does
2302 * nothing and returns NULL.
2304 * If the fullzones BITMAP in the zonelist cache is stale (more than
2305 * a second since last zap'd) then we zap it out (clear its bits.)
2307 * We hold off even calling zlc_setup, until after we've checked the
2308 * first zone in the zonelist, on the theory that most allocations will
2309 * be satisfied from that first zone, so best to examine that zone as
2310 * quickly as we can.
2312 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
2314 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2315 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
2317 zlc
= zonelist
->zlcache_ptr
;
2321 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
2322 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2323 zlc
->last_full_zap
= jiffies
;
2326 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
2327 &cpuset_current_mems_allowed
:
2328 &node_states
[N_MEMORY
];
2329 return allowednodes
;
2333 * Given 'z' scanning a zonelist, run a couple of quick checks to see
2334 * if it is worth looking at further for free memory:
2335 * 1) Check that the zone isn't thought to be full (doesn't have its
2336 * bit set in the zonelist_cache fullzones BITMAP).
2337 * 2) Check that the zones node (obtained from the zonelist_cache
2338 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
2339 * Return true (non-zero) if zone is worth looking at further, or
2340 * else return false (zero) if it is not.
2342 * This check -ignores- the distinction between various watermarks,
2343 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
2344 * found to be full for any variation of these watermarks, it will
2345 * be considered full for up to one second by all requests, unless
2346 * we are so low on memory on all allowed nodes that we are forced
2347 * into the second scan of the zonelist.
2349 * In the second scan we ignore this zonelist cache and exactly
2350 * apply the watermarks to all zones, even it is slower to do so.
2351 * We are low on memory in the second scan, and should leave no stone
2352 * unturned looking for a free page.
2354 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
2355 nodemask_t
*allowednodes
)
2357 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2358 int i
; /* index of *z in zonelist zones */
2359 int n
; /* node that zone *z is on */
2361 zlc
= zonelist
->zlcache_ptr
;
2365 i
= z
- zonelist
->_zonerefs
;
2368 /* This zone is worth trying if it is allowed but not full */
2369 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
2373 * Given 'z' scanning a zonelist, set the corresponding bit in
2374 * zlc->fullzones, so that subsequent attempts to allocate a page
2375 * from that zone don't waste time re-examining it.
2377 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
2379 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2380 int i
; /* index of *z in zonelist zones */
2382 zlc
= zonelist
->zlcache_ptr
;
2386 i
= z
- zonelist
->_zonerefs
;
2388 set_bit(i
, zlc
->fullzones
);
2392 * clear all zones full, called after direct reclaim makes progress so that
2393 * a zone that was recently full is not skipped over for up to a second
2395 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
2397 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2399 zlc
= zonelist
->zlcache_ptr
;
2403 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2406 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2408 return local_zone
->node
== zone
->node
;
2411 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2413 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2417 #else /* CONFIG_NUMA */
2419 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
2424 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
2425 nodemask_t
*allowednodes
)
2430 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
2434 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
2438 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2443 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2448 #endif /* CONFIG_NUMA */
2450 static void reset_alloc_batches(struct zone
*preferred_zone
)
2452 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
2455 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2456 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
2457 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
2458 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2459 } while (zone
++ != preferred_zone
);
2463 * get_page_from_freelist goes through the zonelist trying to allocate
2466 static struct page
*
2467 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2468 const struct alloc_context
*ac
)
2470 struct zonelist
*zonelist
= ac
->zonelist
;
2472 struct page
*page
= NULL
;
2474 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
2475 int zlc_active
= 0; /* set if using zonelist_cache */
2476 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
2477 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
2478 (gfp_mask
& __GFP_WRITE
);
2479 int nr_fair_skipped
= 0;
2480 bool zonelist_rescan
;
2483 zonelist_rescan
= false;
2486 * Scan zonelist, looking for a zone with enough free.
2487 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2489 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2493 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2494 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2496 if (cpusets_enabled() &&
2497 (alloc_flags
& ALLOC_CPUSET
) &&
2498 !cpuset_zone_allowed(zone
, gfp_mask
))
2501 * Distribute pages in proportion to the individual
2502 * zone size to ensure fair page aging. The zone a
2503 * page was allocated in should have no effect on the
2504 * time the page has in memory before being reclaimed.
2506 if (alloc_flags
& ALLOC_FAIR
) {
2507 if (!zone_local(ac
->preferred_zone
, zone
))
2509 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2515 * When allocating a page cache page for writing, we
2516 * want to get it from a zone that is within its dirty
2517 * limit, such that no single zone holds more than its
2518 * proportional share of globally allowed dirty pages.
2519 * The dirty limits take into account the zone's
2520 * lowmem reserves and high watermark so that kswapd
2521 * should be able to balance it without having to
2522 * write pages from its LRU list.
2524 * This may look like it could increase pressure on
2525 * lower zones by failing allocations in higher zones
2526 * before they are full. But the pages that do spill
2527 * over are limited as the lower zones are protected
2528 * by this very same mechanism. It should not become
2529 * a practical burden to them.
2531 * XXX: For now, allow allocations to potentially
2532 * exceed the per-zone dirty limit in the slowpath
2533 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2534 * which is important when on a NUMA setup the allowed
2535 * zones are together not big enough to reach the
2536 * global limit. The proper fix for these situations
2537 * will require awareness of zones in the
2538 * dirty-throttling and the flusher threads.
2540 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2543 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2544 if (!zone_watermark_ok(zone
, order
, mark
,
2545 ac
->classzone_idx
, alloc_flags
)) {
2548 /* Checked here to keep the fast path fast */
2549 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2550 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2553 if (IS_ENABLED(CONFIG_NUMA
) &&
2554 !did_zlc_setup
&& nr_online_nodes
> 1) {
2556 * we do zlc_setup if there are multiple nodes
2557 * and before considering the first zone allowed
2560 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2565 if (zone_reclaim_mode
== 0 ||
2566 !zone_allows_reclaim(ac
->preferred_zone
, zone
))
2567 goto this_zone_full
;
2570 * As we may have just activated ZLC, check if the first
2571 * eligible zone has failed zone_reclaim recently.
2573 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2574 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2577 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2579 case ZONE_RECLAIM_NOSCAN
:
2582 case ZONE_RECLAIM_FULL
:
2583 /* scanned but unreclaimable */
2586 /* did we reclaim enough */
2587 if (zone_watermark_ok(zone
, order
, mark
,
2588 ac
->classzone_idx
, alloc_flags
))
2592 * Failed to reclaim enough to meet watermark.
2593 * Only mark the zone full if checking the min
2594 * watermark or if we failed to reclaim just
2595 * 1<<order pages or else the page allocator
2596 * fastpath will prematurely mark zones full
2597 * when the watermark is between the low and
2600 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2601 ret
== ZONE_RECLAIM_SOME
)
2602 goto this_zone_full
;
2609 page
= buffered_rmqueue(ac
->preferred_zone
, zone
, order
,
2610 gfp_mask
, ac
->migratetype
);
2612 if (prep_new_page(page
, order
, gfp_mask
, alloc_flags
))
2617 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2618 zlc_mark_zone_full(zonelist
, z
);
2622 * The first pass makes sure allocations are spread fairly within the
2623 * local node. However, the local node might have free pages left
2624 * after the fairness batches are exhausted, and remote zones haven't
2625 * even been considered yet. Try once more without fairness, and
2626 * include remote zones now, before entering the slowpath and waking
2627 * kswapd: prefer spilling to a remote zone over swapping locally.
2629 if (alloc_flags
& ALLOC_FAIR
) {
2630 alloc_flags
&= ~ALLOC_FAIR
;
2631 if (nr_fair_skipped
) {
2632 zonelist_rescan
= true;
2633 reset_alloc_batches(ac
->preferred_zone
);
2635 if (nr_online_nodes
> 1)
2636 zonelist_rescan
= true;
2639 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2640 /* Disable zlc cache for second zonelist scan */
2642 zonelist_rescan
= true;
2645 if (zonelist_rescan
)
2652 * Large machines with many possible nodes should not always dump per-node
2653 * meminfo in irq context.
2655 static inline bool should_suppress_show_mem(void)
2660 ret
= in_interrupt();
2665 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2666 DEFAULT_RATELIMIT_INTERVAL
,
2667 DEFAULT_RATELIMIT_BURST
);
2669 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2671 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2673 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2674 debug_guardpage_minorder() > 0)
2678 * This documents exceptions given to allocations in certain
2679 * contexts that are allowed to allocate outside current's set
2682 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2683 if (test_thread_flag(TIF_MEMDIE
) ||
2684 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2685 filter
&= ~SHOW_MEM_FILTER_NODES
;
2686 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2687 filter
&= ~SHOW_MEM_FILTER_NODES
;
2690 struct va_format vaf
;
2693 va_start(args
, fmt
);
2698 pr_warn("%pV", &vaf
);
2703 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2704 current
->comm
, order
, gfp_mask
);
2707 if (!should_suppress_show_mem())
2711 static inline struct page
*
2712 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2713 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
2715 struct oom_control oc
= {
2716 .zonelist
= ac
->zonelist
,
2717 .nodemask
= ac
->nodemask
,
2718 .gfp_mask
= gfp_mask
,
2723 *did_some_progress
= 0;
2726 * Acquire the oom lock. If that fails, somebody else is
2727 * making progress for us.
2729 if (!mutex_trylock(&oom_lock
)) {
2730 *did_some_progress
= 1;
2731 schedule_timeout_uninterruptible(1);
2736 * Go through the zonelist yet one more time, keep very high watermark
2737 * here, this is only to catch a parallel oom killing, we must fail if
2738 * we're still under heavy pressure.
2740 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
2741 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
2745 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2746 /* Coredumps can quickly deplete all memory reserves */
2747 if (current
->flags
& PF_DUMPCORE
)
2749 /* The OOM killer will not help higher order allocs */
2750 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2752 /* The OOM killer does not needlessly kill tasks for lowmem */
2753 if (ac
->high_zoneidx
< ZONE_NORMAL
)
2755 /* The OOM killer does not compensate for IO-less reclaim */
2756 if (!(gfp_mask
& __GFP_FS
)) {
2758 * XXX: Page reclaim didn't yield anything,
2759 * and the OOM killer can't be invoked, but
2760 * keep looping as per tradition.
2762 *did_some_progress
= 1;
2765 if (pm_suspended_storage())
2767 /* The OOM killer may not free memory on a specific node */
2768 if (gfp_mask
& __GFP_THISNODE
)
2771 /* Exhausted what can be done so it's blamo time */
2772 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
))
2773 *did_some_progress
= 1;
2775 mutex_unlock(&oom_lock
);
2779 #ifdef CONFIG_COMPACTION
2780 /* Try memory compaction for high-order allocations before reclaim */
2781 static struct page
*
2782 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2783 int alloc_flags
, const struct alloc_context
*ac
,
2784 enum migrate_mode mode
, int *contended_compaction
,
2785 bool *deferred_compaction
)
2787 unsigned long compact_result
;
2793 current
->flags
|= PF_MEMALLOC
;
2794 compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
2795 mode
, contended_compaction
);
2796 current
->flags
&= ~PF_MEMALLOC
;
2798 switch (compact_result
) {
2799 case COMPACT_DEFERRED
:
2800 *deferred_compaction
= true;
2802 case COMPACT_SKIPPED
:
2809 * At least in one zone compaction wasn't deferred or skipped, so let's
2810 * count a compaction stall
2812 count_vm_event(COMPACTSTALL
);
2814 page
= get_page_from_freelist(gfp_mask
, order
,
2815 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2818 struct zone
*zone
= page_zone(page
);
2820 zone
->compact_blockskip_flush
= false;
2821 compaction_defer_reset(zone
, order
, true);
2822 count_vm_event(COMPACTSUCCESS
);
2827 * It's bad if compaction run occurs and fails. The most likely reason
2828 * is that pages exist, but not enough to satisfy watermarks.
2830 count_vm_event(COMPACTFAIL
);
2837 static inline struct page
*
2838 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2839 int alloc_flags
, const struct alloc_context
*ac
,
2840 enum migrate_mode mode
, int *contended_compaction
,
2841 bool *deferred_compaction
)
2845 #endif /* CONFIG_COMPACTION */
2847 /* Perform direct synchronous page reclaim */
2849 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
2850 const struct alloc_context
*ac
)
2852 struct reclaim_state reclaim_state
;
2857 /* We now go into synchronous reclaim */
2858 cpuset_memory_pressure_bump();
2859 current
->flags
|= PF_MEMALLOC
;
2860 lockdep_set_current_reclaim_state(gfp_mask
);
2861 reclaim_state
.reclaimed_slab
= 0;
2862 current
->reclaim_state
= &reclaim_state
;
2864 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
2867 current
->reclaim_state
= NULL
;
2868 lockdep_clear_current_reclaim_state();
2869 current
->flags
&= ~PF_MEMALLOC
;
2876 /* The really slow allocator path where we enter direct reclaim */
2877 static inline struct page
*
2878 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2879 int alloc_flags
, const struct alloc_context
*ac
,
2880 unsigned long *did_some_progress
)
2882 struct page
*page
= NULL
;
2883 bool drained
= false;
2885 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
2886 if (unlikely(!(*did_some_progress
)))
2889 /* After successful reclaim, reconsider all zones for allocation */
2890 if (IS_ENABLED(CONFIG_NUMA
))
2891 zlc_clear_zones_full(ac
->zonelist
);
2894 page
= get_page_from_freelist(gfp_mask
, order
,
2895 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2898 * If an allocation failed after direct reclaim, it could be because
2899 * pages are pinned on the per-cpu lists. Drain them and try again
2901 if (!page
&& !drained
) {
2902 drain_all_pages(NULL
);
2911 * This is called in the allocator slow-path if the allocation request is of
2912 * sufficient urgency to ignore watermarks and take other desperate measures
2914 static inline struct page
*
2915 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2916 const struct alloc_context
*ac
)
2921 page
= get_page_from_freelist(gfp_mask
, order
,
2922 ALLOC_NO_WATERMARKS
, ac
);
2924 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2925 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
,
2927 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2932 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
2937 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2938 ac
->high_zoneidx
, ac
->nodemask
)
2939 wakeup_kswapd(zone
, order
, zone_idx(ac
->preferred_zone
));
2943 gfp_to_alloc_flags(gfp_t gfp_mask
)
2945 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2946 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2948 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2949 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2952 * The caller may dip into page reserves a bit more if the caller
2953 * cannot run direct reclaim, or if the caller has realtime scheduling
2954 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2955 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2957 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2961 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2962 * if it can't schedule.
2964 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2965 alloc_flags
|= ALLOC_HARDER
;
2967 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2968 * comment for __cpuset_node_allowed().
2970 alloc_flags
&= ~ALLOC_CPUSET
;
2971 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2972 alloc_flags
|= ALLOC_HARDER
;
2974 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2975 if (gfp_mask
& __GFP_MEMALLOC
)
2976 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2977 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2978 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2979 else if (!in_interrupt() &&
2980 ((current
->flags
& PF_MEMALLOC
) ||
2981 unlikely(test_thread_flag(TIF_MEMDIE
))))
2982 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2985 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2986 alloc_flags
|= ALLOC_CMA
;
2991 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2993 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2996 static inline struct page
*
2997 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2998 struct alloc_context
*ac
)
3000 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
3001 struct page
*page
= NULL
;
3003 unsigned long pages_reclaimed
= 0;
3004 unsigned long did_some_progress
;
3005 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
3006 bool deferred_compaction
= false;
3007 int contended_compaction
= COMPACT_CONTENDED_NONE
;
3010 * In the slowpath, we sanity check order to avoid ever trying to
3011 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
3012 * be using allocators in order of preference for an area that is
3015 if (order
>= MAX_ORDER
) {
3016 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
3021 * If this allocation cannot block and it is for a specific node, then
3022 * fail early. There's no need to wakeup kswapd or retry for a
3023 * speculative node-specific allocation.
3025 if (IS_ENABLED(CONFIG_NUMA
) && (gfp_mask
& __GFP_THISNODE
) && !wait
)
3029 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
3030 wake_all_kswapds(order
, ac
);
3033 * OK, we're below the kswapd watermark and have kicked background
3034 * reclaim. Now things get more complex, so set up alloc_flags according
3035 * to how we want to proceed.
3037 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
3040 * Find the true preferred zone if the allocation is unconstrained by
3043 if (!(alloc_flags
& ALLOC_CPUSET
) && !ac
->nodemask
) {
3044 struct zoneref
*preferred_zoneref
;
3045 preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
3046 ac
->high_zoneidx
, NULL
, &ac
->preferred_zone
);
3047 ac
->classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
3050 /* This is the last chance, in general, before the goto nopage. */
3051 page
= get_page_from_freelist(gfp_mask
, order
,
3052 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3056 /* Allocate without watermarks if the context allows */
3057 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
3059 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3060 * the allocation is high priority and these type of
3061 * allocations are system rather than user orientated
3063 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3065 page
= __alloc_pages_high_priority(gfp_mask
, order
, ac
);
3072 /* Atomic allocations - we can't balance anything */
3075 * All existing users of the deprecated __GFP_NOFAIL are
3076 * blockable, so warn of any new users that actually allow this
3077 * type of allocation to fail.
3079 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3083 /* Avoid recursion of direct reclaim */
3084 if (current
->flags
& PF_MEMALLOC
)
3087 /* Avoid allocations with no watermarks from looping endlessly */
3088 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3092 * Try direct compaction. The first pass is asynchronous. Subsequent
3093 * attempts after direct reclaim are synchronous
3095 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3097 &contended_compaction
,
3098 &deferred_compaction
);
3102 /* Checks for THP-specific high-order allocations */
3103 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
3105 * If compaction is deferred for high-order allocations, it is
3106 * because sync compaction recently failed. If this is the case
3107 * and the caller requested a THP allocation, we do not want
3108 * to heavily disrupt the system, so we fail the allocation
3109 * instead of entering direct reclaim.
3111 if (deferred_compaction
)
3115 * In all zones where compaction was attempted (and not
3116 * deferred or skipped), lock contention has been detected.
3117 * For THP allocation we do not want to disrupt the others
3118 * so we fallback to base pages instead.
3120 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
3124 * If compaction was aborted due to need_resched(), we do not
3125 * want to further increase allocation latency, unless it is
3126 * khugepaged trying to collapse.
3128 if (contended_compaction
== COMPACT_CONTENDED_SCHED
3129 && !(current
->flags
& PF_KTHREAD
))
3134 * It can become very expensive to allocate transparent hugepages at
3135 * fault, so use asynchronous memory compaction for THP unless it is
3136 * khugepaged trying to collapse.
3138 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
3139 (current
->flags
& PF_KTHREAD
))
3140 migration_mode
= MIGRATE_SYNC_LIGHT
;
3142 /* Try direct reclaim and then allocating */
3143 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3144 &did_some_progress
);
3148 /* Do not loop if specifically requested */
3149 if (gfp_mask
& __GFP_NORETRY
)
3152 /* Keep reclaiming pages as long as there is reasonable progress */
3153 pages_reclaimed
+= did_some_progress
;
3154 if ((did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
) ||
3155 ((gfp_mask
& __GFP_REPEAT
) && pages_reclaimed
< (1 << order
))) {
3156 /* Wait for some write requests to complete then retry */
3157 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
3161 /* Reclaim has failed us, start killing things */
3162 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3166 /* Retry as long as the OOM killer is making progress */
3167 if (did_some_progress
)
3172 * High-order allocations do not necessarily loop after
3173 * direct reclaim and reclaim/compaction depends on compaction
3174 * being called after reclaim so call directly if necessary
3176 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
,
3178 &contended_compaction
,
3179 &deferred_compaction
);
3183 warn_alloc_failed(gfp_mask
, order
, NULL
);
3189 * This is the 'heart' of the zoned buddy allocator.
3192 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3193 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3195 struct zoneref
*preferred_zoneref
;
3196 struct page
*page
= NULL
;
3197 unsigned int cpuset_mems_cookie
;
3198 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
3199 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
3200 struct alloc_context ac
= {
3201 .high_zoneidx
= gfp_zone(gfp_mask
),
3202 .nodemask
= nodemask
,
3203 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3206 gfp_mask
&= gfp_allowed_mask
;
3208 lockdep_trace_alloc(gfp_mask
);
3210 might_sleep_if(gfp_mask
& __GFP_WAIT
);
3212 if (should_fail_alloc_page(gfp_mask
, order
))
3216 * Check the zones suitable for the gfp_mask contain at least one
3217 * valid zone. It's possible to have an empty zonelist as a result
3218 * of __GFP_THISNODE and a memoryless node
3220 if (unlikely(!zonelist
->_zonerefs
->zone
))
3223 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3224 alloc_flags
|= ALLOC_CMA
;
3227 cpuset_mems_cookie
= read_mems_allowed_begin();
3229 /* We set it here, as __alloc_pages_slowpath might have changed it */
3230 ac
.zonelist
= zonelist
;
3231 /* The preferred zone is used for statistics later */
3232 preferred_zoneref
= first_zones_zonelist(ac
.zonelist
, ac
.high_zoneidx
,
3233 ac
.nodemask
? : &cpuset_current_mems_allowed
,
3234 &ac
.preferred_zone
);
3235 if (!ac
.preferred_zone
)
3237 ac
.classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
3239 /* First allocation attempt */
3240 alloc_mask
= gfp_mask
|__GFP_HARDWALL
;
3241 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3242 if (unlikely(!page
)) {
3244 * Runtime PM, block IO and its error handling path
3245 * can deadlock because I/O on the device might not
3248 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3250 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3253 if (kmemcheck_enabled
&& page
)
3254 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3256 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3260 * When updating a task's mems_allowed, it is possible to race with
3261 * parallel threads in such a way that an allocation can fail while
3262 * the mask is being updated. If a page allocation is about to fail,
3263 * check if the cpuset changed during allocation and if so, retry.
3265 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3270 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3273 * Common helper functions.
3275 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3280 * __get_free_pages() returns a 32-bit address, which cannot represent
3283 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3285 page
= alloc_pages(gfp_mask
, order
);
3288 return (unsigned long) page_address(page
);
3290 EXPORT_SYMBOL(__get_free_pages
);
3292 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3294 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3296 EXPORT_SYMBOL(get_zeroed_page
);
3298 void __free_pages(struct page
*page
, unsigned int order
)
3300 if (put_page_testzero(page
)) {
3302 free_hot_cold_page(page
, false);
3304 __free_pages_ok(page
, order
);
3308 EXPORT_SYMBOL(__free_pages
);
3310 void free_pages(unsigned long addr
, unsigned int order
)
3313 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3314 __free_pages(virt_to_page((void *)addr
), order
);
3318 EXPORT_SYMBOL(free_pages
);
3322 * An arbitrary-length arbitrary-offset area of memory which resides
3323 * within a 0 or higher order page. Multiple fragments within that page
3324 * are individually refcounted, in the page's reference counter.
3326 * The page_frag functions below provide a simple allocation framework for
3327 * page fragments. This is used by the network stack and network device
3328 * drivers to provide a backing region of memory for use as either an
3329 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3331 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3334 struct page
*page
= NULL
;
3335 gfp_t gfp
= gfp_mask
;
3337 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3338 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3340 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3341 PAGE_FRAG_CACHE_MAX_ORDER
);
3342 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3344 if (unlikely(!page
))
3345 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3347 nc
->va
= page
? page_address(page
) : NULL
;
3352 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3353 unsigned int fragsz
, gfp_t gfp_mask
)
3355 unsigned int size
= PAGE_SIZE
;
3359 if (unlikely(!nc
->va
)) {
3361 page
= __page_frag_refill(nc
, gfp_mask
);
3365 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3366 /* if size can vary use size else just use PAGE_SIZE */
3369 /* Even if we own the page, we do not use atomic_set().
3370 * This would break get_page_unless_zero() users.
3372 atomic_add(size
- 1, &page
->_count
);
3374 /* reset page count bias and offset to start of new frag */
3375 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
3376 nc
->pagecnt_bias
= size
;
3380 offset
= nc
->offset
- fragsz
;
3381 if (unlikely(offset
< 0)) {
3382 page
= virt_to_page(nc
->va
);
3384 if (!atomic_sub_and_test(nc
->pagecnt_bias
, &page
->_count
))
3387 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3388 /* if size can vary use size else just use PAGE_SIZE */
3391 /* OK, page count is 0, we can safely set it */
3392 atomic_set(&page
->_count
, size
);
3394 /* reset page count bias and offset to start of new frag */
3395 nc
->pagecnt_bias
= size
;
3396 offset
= size
- fragsz
;
3400 nc
->offset
= offset
;
3402 return nc
->va
+ offset
;
3404 EXPORT_SYMBOL(__alloc_page_frag
);
3407 * Frees a page fragment allocated out of either a compound or order 0 page.
3409 void __free_page_frag(void *addr
)
3411 struct page
*page
= virt_to_head_page(addr
);
3413 if (unlikely(put_page_testzero(page
)))
3414 __free_pages_ok(page
, compound_order(page
));
3416 EXPORT_SYMBOL(__free_page_frag
);
3419 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3420 * of the current memory cgroup.
3422 * It should be used when the caller would like to use kmalloc, but since the
3423 * allocation is large, it has to fall back to the page allocator.
3425 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
3428 struct mem_cgroup
*memcg
= NULL
;
3430 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
3432 page
= alloc_pages(gfp_mask
, order
);
3433 memcg_kmem_commit_charge(page
, memcg
, order
);
3437 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
3440 struct mem_cgroup
*memcg
= NULL
;
3442 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
3444 page
= alloc_pages_node(nid
, gfp_mask
, order
);
3445 memcg_kmem_commit_charge(page
, memcg
, order
);
3450 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3453 void __free_kmem_pages(struct page
*page
, unsigned int order
)
3455 memcg_kmem_uncharge_pages(page
, order
);
3456 __free_pages(page
, order
);
3459 void free_kmem_pages(unsigned long addr
, unsigned int order
)
3462 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3463 __free_kmem_pages(virt_to_page((void *)addr
), order
);
3467 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
3470 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3471 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3473 split_page(virt_to_page((void *)addr
), order
);
3474 while (used
< alloc_end
) {
3479 return (void *)addr
;
3483 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3484 * @size: the number of bytes to allocate
3485 * @gfp_mask: GFP flags for the allocation
3487 * This function is similar to alloc_pages(), except that it allocates the
3488 * minimum number of pages to satisfy the request. alloc_pages() can only
3489 * allocate memory in power-of-two pages.
3491 * This function is also limited by MAX_ORDER.
3493 * Memory allocated by this function must be released by free_pages_exact().
3495 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3497 unsigned int order
= get_order(size
);
3500 addr
= __get_free_pages(gfp_mask
, order
);
3501 return make_alloc_exact(addr
, order
, size
);
3503 EXPORT_SYMBOL(alloc_pages_exact
);
3506 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3508 * @nid: the preferred node ID where memory should be allocated
3509 * @size: the number of bytes to allocate
3510 * @gfp_mask: GFP flags for the allocation
3512 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3514 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3517 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3519 unsigned order
= get_order(size
);
3520 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3523 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3527 * free_pages_exact - release memory allocated via alloc_pages_exact()
3528 * @virt: the value returned by alloc_pages_exact.
3529 * @size: size of allocation, same value as passed to alloc_pages_exact().
3531 * Release the memory allocated by a previous call to alloc_pages_exact.
3533 void free_pages_exact(void *virt
, size_t size
)
3535 unsigned long addr
= (unsigned long)virt
;
3536 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3538 while (addr
< end
) {
3543 EXPORT_SYMBOL(free_pages_exact
);
3546 * nr_free_zone_pages - count number of pages beyond high watermark
3547 * @offset: The zone index of the highest zone
3549 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3550 * high watermark within all zones at or below a given zone index. For each
3551 * zone, the number of pages is calculated as:
3552 * managed_pages - high_pages
3554 static unsigned long nr_free_zone_pages(int offset
)
3559 /* Just pick one node, since fallback list is circular */
3560 unsigned long sum
= 0;
3562 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3564 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3565 unsigned long size
= zone
->managed_pages
;
3566 unsigned long high
= high_wmark_pages(zone
);
3575 * nr_free_buffer_pages - count number of pages beyond high watermark
3577 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3578 * watermark within ZONE_DMA and ZONE_NORMAL.
3580 unsigned long nr_free_buffer_pages(void)
3582 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3584 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3587 * nr_free_pagecache_pages - count number of pages beyond high watermark
3589 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3590 * high watermark within all zones.
3592 unsigned long nr_free_pagecache_pages(void)
3594 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3597 static inline void show_node(struct zone
*zone
)
3599 if (IS_ENABLED(CONFIG_NUMA
))
3600 printk("Node %d ", zone_to_nid(zone
));
3603 void si_meminfo(struct sysinfo
*val
)
3605 val
->totalram
= totalram_pages
;
3606 val
->sharedram
= global_page_state(NR_SHMEM
);
3607 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3608 val
->bufferram
= nr_blockdev_pages();
3609 val
->totalhigh
= totalhigh_pages
;
3610 val
->freehigh
= nr_free_highpages();
3611 val
->mem_unit
= PAGE_SIZE
;
3614 EXPORT_SYMBOL(si_meminfo
);
3617 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3619 int zone_type
; /* needs to be signed */
3620 unsigned long managed_pages
= 0;
3621 pg_data_t
*pgdat
= NODE_DATA(nid
);
3623 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3624 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3625 val
->totalram
= managed_pages
;
3626 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3627 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3628 #ifdef CONFIG_HIGHMEM
3629 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3630 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3636 val
->mem_unit
= PAGE_SIZE
;
3641 * Determine whether the node should be displayed or not, depending on whether
3642 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3644 bool skip_free_areas_node(unsigned int flags
, int nid
)
3647 unsigned int cpuset_mems_cookie
;
3649 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3653 cpuset_mems_cookie
= read_mems_allowed_begin();
3654 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3655 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3660 #define K(x) ((x) << (PAGE_SHIFT-10))
3662 static void show_migration_types(unsigned char type
)
3664 static const char types
[MIGRATE_TYPES
] = {
3665 [MIGRATE_UNMOVABLE
] = 'U',
3666 [MIGRATE_RECLAIMABLE
] = 'E',
3667 [MIGRATE_MOVABLE
] = 'M',
3668 [MIGRATE_RESERVE
] = 'R',
3670 [MIGRATE_CMA
] = 'C',
3672 #ifdef CONFIG_MEMORY_ISOLATION
3673 [MIGRATE_ISOLATE
] = 'I',
3676 char tmp
[MIGRATE_TYPES
+ 1];
3680 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3681 if (type
& (1 << i
))
3686 printk("(%s) ", tmp
);
3690 * Show free area list (used inside shift_scroll-lock stuff)
3691 * We also calculate the percentage fragmentation. We do this by counting the
3692 * memory on each free list with the exception of the first item on the list.
3695 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3698 void show_free_areas(unsigned int filter
)
3700 unsigned long free_pcp
= 0;
3704 for_each_populated_zone(zone
) {
3705 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3708 for_each_online_cpu(cpu
)
3709 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
3712 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3713 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3714 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3715 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3716 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3717 " free:%lu free_pcp:%lu free_cma:%lu\n",
3718 global_page_state(NR_ACTIVE_ANON
),
3719 global_page_state(NR_INACTIVE_ANON
),
3720 global_page_state(NR_ISOLATED_ANON
),
3721 global_page_state(NR_ACTIVE_FILE
),
3722 global_page_state(NR_INACTIVE_FILE
),
3723 global_page_state(NR_ISOLATED_FILE
),
3724 global_page_state(NR_UNEVICTABLE
),
3725 global_page_state(NR_FILE_DIRTY
),
3726 global_page_state(NR_WRITEBACK
),
3727 global_page_state(NR_UNSTABLE_NFS
),
3728 global_page_state(NR_SLAB_RECLAIMABLE
),
3729 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3730 global_page_state(NR_FILE_MAPPED
),
3731 global_page_state(NR_SHMEM
),
3732 global_page_state(NR_PAGETABLE
),
3733 global_page_state(NR_BOUNCE
),
3734 global_page_state(NR_FREE_PAGES
),
3736 global_page_state(NR_FREE_CMA_PAGES
));
3738 for_each_populated_zone(zone
) {
3741 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3745 for_each_online_cpu(cpu
)
3746 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
3754 " active_anon:%lukB"
3755 " inactive_anon:%lukB"
3756 " active_file:%lukB"
3757 " inactive_file:%lukB"
3758 " unevictable:%lukB"
3759 " isolated(anon):%lukB"
3760 " isolated(file):%lukB"
3768 " slab_reclaimable:%lukB"
3769 " slab_unreclaimable:%lukB"
3770 " kernel_stack:%lukB"
3777 " writeback_tmp:%lukB"
3778 " pages_scanned:%lu"
3779 " all_unreclaimable? %s"
3782 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3783 K(min_wmark_pages(zone
)),
3784 K(low_wmark_pages(zone
)),
3785 K(high_wmark_pages(zone
)),
3786 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3787 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3788 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3789 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3790 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3791 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3792 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3793 K(zone
->present_pages
),
3794 K(zone
->managed_pages
),
3795 K(zone_page_state(zone
, NR_MLOCK
)),
3796 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3797 K(zone_page_state(zone
, NR_WRITEBACK
)),
3798 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3799 K(zone_page_state(zone
, NR_SHMEM
)),
3800 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3801 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3802 zone_page_state(zone
, NR_KERNEL_STACK
) *
3804 K(zone_page_state(zone
, NR_PAGETABLE
)),
3805 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3806 K(zone_page_state(zone
, NR_BOUNCE
)),
3808 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
3809 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3810 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3811 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3812 (!zone_reclaimable(zone
) ? "yes" : "no")
3814 printk("lowmem_reserve[]:");
3815 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3816 printk(" %ld", zone
->lowmem_reserve
[i
]);
3820 for_each_populated_zone(zone
) {
3821 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3822 unsigned char types
[MAX_ORDER
];
3824 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3827 printk("%s: ", zone
->name
);
3829 spin_lock_irqsave(&zone
->lock
, flags
);
3830 for (order
= 0; order
< MAX_ORDER
; order
++) {
3831 struct free_area
*area
= &zone
->free_area
[order
];
3834 nr
[order
] = area
->nr_free
;
3835 total
+= nr
[order
] << order
;
3838 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3839 if (!list_empty(&area
->free_list
[type
]))
3840 types
[order
] |= 1 << type
;
3843 spin_unlock_irqrestore(&zone
->lock
, flags
);
3844 for (order
= 0; order
< MAX_ORDER
; order
++) {
3845 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3847 show_migration_types(types
[order
]);
3849 printk("= %lukB\n", K(total
));
3852 hugetlb_show_meminfo();
3854 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3856 show_swap_cache_info();
3859 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3861 zoneref
->zone
= zone
;
3862 zoneref
->zone_idx
= zone_idx(zone
);
3866 * Builds allocation fallback zone lists.
3868 * Add all populated zones of a node to the zonelist.
3870 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3874 enum zone_type zone_type
= MAX_NR_ZONES
;
3878 zone
= pgdat
->node_zones
+ zone_type
;
3879 if (populated_zone(zone
)) {
3880 zoneref_set_zone(zone
,
3881 &zonelist
->_zonerefs
[nr_zones
++]);
3882 check_highest_zone(zone_type
);
3884 } while (zone_type
);
3892 * 0 = automatic detection of better ordering.
3893 * 1 = order by ([node] distance, -zonetype)
3894 * 2 = order by (-zonetype, [node] distance)
3896 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3897 * the same zonelist. So only NUMA can configure this param.
3899 #define ZONELIST_ORDER_DEFAULT 0
3900 #define ZONELIST_ORDER_NODE 1
3901 #define ZONELIST_ORDER_ZONE 2
3903 /* zonelist order in the kernel.
3904 * set_zonelist_order() will set this to NODE or ZONE.
3906 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3907 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3911 /* The value user specified ....changed by config */
3912 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3913 /* string for sysctl */
3914 #define NUMA_ZONELIST_ORDER_LEN 16
3915 char numa_zonelist_order
[16] = "default";
3918 * interface for configure zonelist ordering.
3919 * command line option "numa_zonelist_order"
3920 * = "[dD]efault - default, automatic configuration.
3921 * = "[nN]ode - order by node locality, then by zone within node
3922 * = "[zZ]one - order by zone, then by locality within zone
3925 static int __parse_numa_zonelist_order(char *s
)
3927 if (*s
== 'd' || *s
== 'D') {
3928 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3929 } else if (*s
== 'n' || *s
== 'N') {
3930 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3931 } else if (*s
== 'z' || *s
== 'Z') {
3932 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3935 "Ignoring invalid numa_zonelist_order value: "
3942 static __init
int setup_numa_zonelist_order(char *s
)
3949 ret
= __parse_numa_zonelist_order(s
);
3951 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3955 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3958 * sysctl handler for numa_zonelist_order
3960 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3961 void __user
*buffer
, size_t *length
,
3964 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3966 static DEFINE_MUTEX(zl_order_mutex
);
3968 mutex_lock(&zl_order_mutex
);
3970 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3974 strcpy(saved_string
, (char *)table
->data
);
3976 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3980 int oldval
= user_zonelist_order
;
3982 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3985 * bogus value. restore saved string
3987 strncpy((char *)table
->data
, saved_string
,
3988 NUMA_ZONELIST_ORDER_LEN
);
3989 user_zonelist_order
= oldval
;
3990 } else if (oldval
!= user_zonelist_order
) {
3991 mutex_lock(&zonelists_mutex
);
3992 build_all_zonelists(NULL
, NULL
);
3993 mutex_unlock(&zonelists_mutex
);
3997 mutex_unlock(&zl_order_mutex
);
4002 #define MAX_NODE_LOAD (nr_online_nodes)
4003 static int node_load
[MAX_NUMNODES
];
4006 * find_next_best_node - find the next node that should appear in a given node's fallback list
4007 * @node: node whose fallback list we're appending
4008 * @used_node_mask: nodemask_t of already used nodes
4010 * We use a number of factors to determine which is the next node that should
4011 * appear on a given node's fallback list. The node should not have appeared
4012 * already in @node's fallback list, and it should be the next closest node
4013 * according to the distance array (which contains arbitrary distance values
4014 * from each node to each node in the system), and should also prefer nodes
4015 * with no CPUs, since presumably they'll have very little allocation pressure
4016 * on them otherwise.
4017 * It returns -1 if no node is found.
4019 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
4022 int min_val
= INT_MAX
;
4023 int best_node
= NUMA_NO_NODE
;
4024 const struct cpumask
*tmp
= cpumask_of_node(0);
4026 /* Use the local node if we haven't already */
4027 if (!node_isset(node
, *used_node_mask
)) {
4028 node_set(node
, *used_node_mask
);
4032 for_each_node_state(n
, N_MEMORY
) {
4034 /* Don't want a node to appear more than once */
4035 if (node_isset(n
, *used_node_mask
))
4038 /* Use the distance array to find the distance */
4039 val
= node_distance(node
, n
);
4041 /* Penalize nodes under us ("prefer the next node") */
4044 /* Give preference to headless and unused nodes */
4045 tmp
= cpumask_of_node(n
);
4046 if (!cpumask_empty(tmp
))
4047 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4049 /* Slight preference for less loaded node */
4050 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4051 val
+= node_load
[n
];
4053 if (val
< min_val
) {
4060 node_set(best_node
, *used_node_mask
);
4067 * Build zonelists ordered by node and zones within node.
4068 * This results in maximum locality--normal zone overflows into local
4069 * DMA zone, if any--but risks exhausting DMA zone.
4071 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4074 struct zonelist
*zonelist
;
4076 zonelist
= &pgdat
->node_zonelists
[0];
4077 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4079 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4080 zonelist
->_zonerefs
[j
].zone
= NULL
;
4081 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4085 * Build gfp_thisnode zonelists
4087 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4090 struct zonelist
*zonelist
;
4092 zonelist
= &pgdat
->node_zonelists
[1];
4093 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4094 zonelist
->_zonerefs
[j
].zone
= NULL
;
4095 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4099 * Build zonelists ordered by zone and nodes within zones.
4100 * This results in conserving DMA zone[s] until all Normal memory is
4101 * exhausted, but results in overflowing to remote node while memory
4102 * may still exist in local DMA zone.
4104 static int node_order
[MAX_NUMNODES
];
4106 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4109 int zone_type
; /* needs to be signed */
4111 struct zonelist
*zonelist
;
4113 zonelist
= &pgdat
->node_zonelists
[0];
4115 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4116 for (j
= 0; j
< nr_nodes
; j
++) {
4117 node
= node_order
[j
];
4118 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4119 if (populated_zone(z
)) {
4121 &zonelist
->_zonerefs
[pos
++]);
4122 check_highest_zone(zone_type
);
4126 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4127 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4130 #if defined(CONFIG_64BIT)
4132 * Devices that require DMA32/DMA are relatively rare and do not justify a
4133 * penalty to every machine in case the specialised case applies. Default
4134 * to Node-ordering on 64-bit NUMA machines
4136 static int default_zonelist_order(void)
4138 return ZONELIST_ORDER_NODE
;
4142 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4143 * by the kernel. If processes running on node 0 deplete the low memory zone
4144 * then reclaim will occur more frequency increasing stalls and potentially
4145 * be easier to OOM if a large percentage of the zone is under writeback or
4146 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4147 * Hence, default to zone ordering on 32-bit.
4149 static int default_zonelist_order(void)
4151 return ZONELIST_ORDER_ZONE
;
4153 #endif /* CONFIG_64BIT */
4155 static void set_zonelist_order(void)
4157 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4158 current_zonelist_order
= default_zonelist_order();
4160 current_zonelist_order
= user_zonelist_order
;
4163 static void build_zonelists(pg_data_t
*pgdat
)
4167 nodemask_t used_mask
;
4168 int local_node
, prev_node
;
4169 struct zonelist
*zonelist
;
4170 int order
= current_zonelist_order
;
4172 /* initialize zonelists */
4173 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4174 zonelist
= pgdat
->node_zonelists
+ i
;
4175 zonelist
->_zonerefs
[0].zone
= NULL
;
4176 zonelist
->_zonerefs
[0].zone_idx
= 0;
4179 /* NUMA-aware ordering of nodes */
4180 local_node
= pgdat
->node_id
;
4181 load
= nr_online_nodes
;
4182 prev_node
= local_node
;
4183 nodes_clear(used_mask
);
4185 memset(node_order
, 0, sizeof(node_order
));
4188 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4190 * We don't want to pressure a particular node.
4191 * So adding penalty to the first node in same
4192 * distance group to make it round-robin.
4194 if (node_distance(local_node
, node
) !=
4195 node_distance(local_node
, prev_node
))
4196 node_load
[node
] = load
;
4200 if (order
== ZONELIST_ORDER_NODE
)
4201 build_zonelists_in_node_order(pgdat
, node
);
4203 node_order
[j
++] = node
; /* remember order */
4206 if (order
== ZONELIST_ORDER_ZONE
) {
4207 /* calculate node order -- i.e., DMA last! */
4208 build_zonelists_in_zone_order(pgdat
, j
);
4211 build_thisnode_zonelists(pgdat
);
4214 /* Construct the zonelist performance cache - see further mmzone.h */
4215 static void build_zonelist_cache(pg_data_t
*pgdat
)
4217 struct zonelist
*zonelist
;
4218 struct zonelist_cache
*zlc
;
4221 zonelist
= &pgdat
->node_zonelists
[0];
4222 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
4223 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
4224 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
4225 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
4228 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4230 * Return node id of node used for "local" allocations.
4231 * I.e., first node id of first zone in arg node's generic zonelist.
4232 * Used for initializing percpu 'numa_mem', which is used primarily
4233 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4235 int local_memory_node(int node
)
4239 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4240 gfp_zone(GFP_KERNEL
),
4247 #else /* CONFIG_NUMA */
4249 static void set_zonelist_order(void)
4251 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4254 static void build_zonelists(pg_data_t
*pgdat
)
4256 int node
, local_node
;
4258 struct zonelist
*zonelist
;
4260 local_node
= pgdat
->node_id
;
4262 zonelist
= &pgdat
->node_zonelists
[0];
4263 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4266 * Now we build the zonelist so that it contains the zones
4267 * of all the other nodes.
4268 * We don't want to pressure a particular node, so when
4269 * building the zones for node N, we make sure that the
4270 * zones coming right after the local ones are those from
4271 * node N+1 (modulo N)
4273 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4274 if (!node_online(node
))
4276 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4278 for (node
= 0; node
< local_node
; node
++) {
4279 if (!node_online(node
))
4281 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4284 zonelist
->_zonerefs
[j
].zone
= NULL
;
4285 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4288 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
4289 static void build_zonelist_cache(pg_data_t
*pgdat
)
4291 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
4294 #endif /* CONFIG_NUMA */
4297 * Boot pageset table. One per cpu which is going to be used for all
4298 * zones and all nodes. The parameters will be set in such a way
4299 * that an item put on a list will immediately be handed over to
4300 * the buddy list. This is safe since pageset manipulation is done
4301 * with interrupts disabled.
4303 * The boot_pagesets must be kept even after bootup is complete for
4304 * unused processors and/or zones. They do play a role for bootstrapping
4305 * hotplugged processors.
4307 * zoneinfo_show() and maybe other functions do
4308 * not check if the processor is online before following the pageset pointer.
4309 * Other parts of the kernel may not check if the zone is available.
4311 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4312 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4313 static void setup_zone_pageset(struct zone
*zone
);
4316 * Global mutex to protect against size modification of zonelists
4317 * as well as to serialize pageset setup for the new populated zone.
4319 DEFINE_MUTEX(zonelists_mutex
);
4321 /* return values int ....just for stop_machine() */
4322 static int __build_all_zonelists(void *data
)
4326 pg_data_t
*self
= data
;
4329 memset(node_load
, 0, sizeof(node_load
));
4332 if (self
&& !node_online(self
->node_id
)) {
4333 build_zonelists(self
);
4334 build_zonelist_cache(self
);
4337 for_each_online_node(nid
) {
4338 pg_data_t
*pgdat
= NODE_DATA(nid
);
4340 build_zonelists(pgdat
);
4341 build_zonelist_cache(pgdat
);
4345 * Initialize the boot_pagesets that are going to be used
4346 * for bootstrapping processors. The real pagesets for
4347 * each zone will be allocated later when the per cpu
4348 * allocator is available.
4350 * boot_pagesets are used also for bootstrapping offline
4351 * cpus if the system is already booted because the pagesets
4352 * are needed to initialize allocators on a specific cpu too.
4353 * F.e. the percpu allocator needs the page allocator which
4354 * needs the percpu allocator in order to allocate its pagesets
4355 * (a chicken-egg dilemma).
4357 for_each_possible_cpu(cpu
) {
4358 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4360 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4362 * We now know the "local memory node" for each node--
4363 * i.e., the node of the first zone in the generic zonelist.
4364 * Set up numa_mem percpu variable for on-line cpus. During
4365 * boot, only the boot cpu should be on-line; we'll init the
4366 * secondary cpus' numa_mem as they come on-line. During
4367 * node/memory hotplug, we'll fixup all on-line cpus.
4369 if (cpu_online(cpu
))
4370 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4377 static noinline
void __init
4378 build_all_zonelists_init(void)
4380 __build_all_zonelists(NULL
);
4381 mminit_verify_zonelist();
4382 cpuset_init_current_mems_allowed();
4386 * Called with zonelists_mutex held always
4387 * unless system_state == SYSTEM_BOOTING.
4389 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4390 * [we're only called with non-NULL zone through __meminit paths] and
4391 * (2) call of __init annotated helper build_all_zonelists_init
4392 * [protected by SYSTEM_BOOTING].
4394 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4396 set_zonelist_order();
4398 if (system_state
== SYSTEM_BOOTING
) {
4399 build_all_zonelists_init();
4401 #ifdef CONFIG_MEMORY_HOTPLUG
4403 setup_zone_pageset(zone
);
4405 /* we have to stop all cpus to guarantee there is no user
4407 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4408 /* cpuset refresh routine should be here */
4410 vm_total_pages
= nr_free_pagecache_pages();
4412 * Disable grouping by mobility if the number of pages in the
4413 * system is too low to allow the mechanism to work. It would be
4414 * more accurate, but expensive to check per-zone. This check is
4415 * made on memory-hotadd so a system can start with mobility
4416 * disabled and enable it later
4418 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4419 page_group_by_mobility_disabled
= 1;
4421 page_group_by_mobility_disabled
= 0;
4423 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4424 "Total pages: %ld\n",
4426 zonelist_order_name
[current_zonelist_order
],
4427 page_group_by_mobility_disabled
? "off" : "on",
4430 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4435 * Helper functions to size the waitqueue hash table.
4436 * Essentially these want to choose hash table sizes sufficiently
4437 * large so that collisions trying to wait on pages are rare.
4438 * But in fact, the number of active page waitqueues on typical
4439 * systems is ridiculously low, less than 200. So this is even
4440 * conservative, even though it seems large.
4442 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4443 * waitqueues, i.e. the size of the waitq table given the number of pages.
4445 #define PAGES_PER_WAITQUEUE 256
4447 #ifndef CONFIG_MEMORY_HOTPLUG
4448 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4450 unsigned long size
= 1;
4452 pages
/= PAGES_PER_WAITQUEUE
;
4454 while (size
< pages
)
4458 * Once we have dozens or even hundreds of threads sleeping
4459 * on IO we've got bigger problems than wait queue collision.
4460 * Limit the size of the wait table to a reasonable size.
4462 size
= min(size
, 4096UL);
4464 return max(size
, 4UL);
4468 * A zone's size might be changed by hot-add, so it is not possible to determine
4469 * a suitable size for its wait_table. So we use the maximum size now.
4471 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4473 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4474 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4475 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4477 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4478 * or more by the traditional way. (See above). It equals:
4480 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4481 * ia64(16K page size) : = ( 8G + 4M)byte.
4482 * powerpc (64K page size) : = (32G +16M)byte.
4484 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4491 * This is an integer logarithm so that shifts can be used later
4492 * to extract the more random high bits from the multiplicative
4493 * hash function before the remainder is taken.
4495 static inline unsigned long wait_table_bits(unsigned long size
)
4501 * Check if a pageblock contains reserved pages
4503 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
4507 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4508 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
4515 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4516 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4517 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4518 * higher will lead to a bigger reserve which will get freed as contiguous
4519 * blocks as reclaim kicks in
4521 static void setup_zone_migrate_reserve(struct zone
*zone
)
4523 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4525 unsigned long block_migratetype
;
4530 * Get the start pfn, end pfn and the number of blocks to reserve
4531 * We have to be careful to be aligned to pageblock_nr_pages to
4532 * make sure that we always check pfn_valid for the first page in
4535 start_pfn
= zone
->zone_start_pfn
;
4536 end_pfn
= zone_end_pfn(zone
);
4537 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4538 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4542 * Reserve blocks are generally in place to help high-order atomic
4543 * allocations that are short-lived. A min_free_kbytes value that
4544 * would result in more than 2 reserve blocks for atomic allocations
4545 * is assumed to be in place to help anti-fragmentation for the
4546 * future allocation of hugepages at runtime.
4548 reserve
= min(2, reserve
);
4549 old_reserve
= zone
->nr_migrate_reserve_block
;
4551 /* When memory hot-add, we almost always need to do nothing */
4552 if (reserve
== old_reserve
)
4554 zone
->nr_migrate_reserve_block
= reserve
;
4556 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4557 if (!early_page_nid_uninitialised(pfn
, zone_to_nid(zone
)))
4560 if (!pfn_valid(pfn
))
4562 page
= pfn_to_page(pfn
);
4564 /* Watch out for overlapping nodes */
4565 if (page_to_nid(page
) != zone_to_nid(zone
))
4568 block_migratetype
= get_pageblock_migratetype(page
);
4570 /* Only test what is necessary when the reserves are not met */
4573 * Blocks with reserved pages will never free, skip
4576 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4577 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4580 /* If this block is reserved, account for it */
4581 if (block_migratetype
== MIGRATE_RESERVE
) {
4586 /* Suitable for reserving if this block is movable */
4587 if (block_migratetype
== MIGRATE_MOVABLE
) {
4588 set_pageblock_migratetype(page
,
4590 move_freepages_block(zone
, page
,
4595 } else if (!old_reserve
) {
4597 * At boot time we don't need to scan the whole zone
4598 * for turning off MIGRATE_RESERVE.
4604 * If the reserve is met and this is a previous reserved block,
4607 if (block_migratetype
== MIGRATE_RESERVE
) {
4608 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4609 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4615 * Initially all pages are reserved - free ones are freed
4616 * up by free_all_bootmem() once the early boot process is
4617 * done. Non-atomic initialization, single-pass.
4619 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4620 unsigned long start_pfn
, enum memmap_context context
)
4622 pg_data_t
*pgdat
= NODE_DATA(nid
);
4623 unsigned long end_pfn
= start_pfn
+ size
;
4626 unsigned long nr_initialised
= 0;
4628 if (highest_memmap_pfn
< end_pfn
- 1)
4629 highest_memmap_pfn
= end_pfn
- 1;
4631 z
= &pgdat
->node_zones
[zone
];
4632 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4634 * There can be holes in boot-time mem_map[]s
4635 * handed to this function. They do not
4636 * exist on hotplugged memory.
4638 if (context
== MEMMAP_EARLY
) {
4639 if (!early_pfn_valid(pfn
))
4641 if (!early_pfn_in_nid(pfn
, nid
))
4643 if (!update_defer_init(pgdat
, pfn
, end_pfn
,
4649 * Mark the block movable so that blocks are reserved for
4650 * movable at startup. This will force kernel allocations
4651 * to reserve their blocks rather than leaking throughout
4652 * the address space during boot when many long-lived
4653 * kernel allocations are made. Later some blocks near
4654 * the start are marked MIGRATE_RESERVE by
4655 * setup_zone_migrate_reserve()
4657 * bitmap is created for zone's valid pfn range. but memmap
4658 * can be created for invalid pages (for alignment)
4659 * check here not to call set_pageblock_migratetype() against
4662 if (!(pfn
& (pageblock_nr_pages
- 1))) {
4663 struct page
*page
= pfn_to_page(pfn
);
4665 __init_single_page(page
, pfn
, zone
, nid
);
4666 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4668 __init_single_pfn(pfn
, zone
, nid
);
4673 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4675 unsigned int order
, t
;
4676 for_each_migratetype_order(order
, t
) {
4677 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4678 zone
->free_area
[order
].nr_free
= 0;
4682 #ifndef __HAVE_ARCH_MEMMAP_INIT
4683 #define memmap_init(size, nid, zone, start_pfn) \
4684 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4687 static int zone_batchsize(struct zone
*zone
)
4693 * The per-cpu-pages pools are set to around 1000th of the
4694 * size of the zone. But no more than 1/2 of a meg.
4696 * OK, so we don't know how big the cache is. So guess.
4698 batch
= zone
->managed_pages
/ 1024;
4699 if (batch
* PAGE_SIZE
> 512 * 1024)
4700 batch
= (512 * 1024) / PAGE_SIZE
;
4701 batch
/= 4; /* We effectively *= 4 below */
4706 * Clamp the batch to a 2^n - 1 value. Having a power
4707 * of 2 value was found to be more likely to have
4708 * suboptimal cache aliasing properties in some cases.
4710 * For example if 2 tasks are alternately allocating
4711 * batches of pages, one task can end up with a lot
4712 * of pages of one half of the possible page colors
4713 * and the other with pages of the other colors.
4715 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4720 /* The deferral and batching of frees should be suppressed under NOMMU
4723 * The problem is that NOMMU needs to be able to allocate large chunks
4724 * of contiguous memory as there's no hardware page translation to
4725 * assemble apparent contiguous memory from discontiguous pages.
4727 * Queueing large contiguous runs of pages for batching, however,
4728 * causes the pages to actually be freed in smaller chunks. As there
4729 * can be a significant delay between the individual batches being
4730 * recycled, this leads to the once large chunks of space being
4731 * fragmented and becoming unavailable for high-order allocations.
4738 * pcp->high and pcp->batch values are related and dependent on one another:
4739 * ->batch must never be higher then ->high.
4740 * The following function updates them in a safe manner without read side
4743 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4744 * those fields changing asynchronously (acording the the above rule).
4746 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4747 * outside of boot time (or some other assurance that no concurrent updaters
4750 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4751 unsigned long batch
)
4753 /* start with a fail safe value for batch */
4757 /* Update high, then batch, in order */
4764 /* a companion to pageset_set_high() */
4765 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4767 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4770 static void pageset_init(struct per_cpu_pageset
*p
)
4772 struct per_cpu_pages
*pcp
;
4775 memset(p
, 0, sizeof(*p
));
4779 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4780 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4783 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4786 pageset_set_batch(p
, batch
);
4790 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4791 * to the value high for the pageset p.
4793 static void pageset_set_high(struct per_cpu_pageset
*p
,
4796 unsigned long batch
= max(1UL, high
/ 4);
4797 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4798 batch
= PAGE_SHIFT
* 8;
4800 pageset_update(&p
->pcp
, high
, batch
);
4803 static void pageset_set_high_and_batch(struct zone
*zone
,
4804 struct per_cpu_pageset
*pcp
)
4806 if (percpu_pagelist_fraction
)
4807 pageset_set_high(pcp
,
4808 (zone
->managed_pages
/
4809 percpu_pagelist_fraction
));
4811 pageset_set_batch(pcp
, zone_batchsize(zone
));
4814 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4816 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4819 pageset_set_high_and_batch(zone
, pcp
);
4822 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4825 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4826 for_each_possible_cpu(cpu
)
4827 zone_pageset_init(zone
, cpu
);
4831 * Allocate per cpu pagesets and initialize them.
4832 * Before this call only boot pagesets were available.
4834 void __init
setup_per_cpu_pageset(void)
4838 for_each_populated_zone(zone
)
4839 setup_zone_pageset(zone
);
4842 static noinline __init_refok
4843 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4849 * The per-page waitqueue mechanism uses hashed waitqueues
4852 zone
->wait_table_hash_nr_entries
=
4853 wait_table_hash_nr_entries(zone_size_pages
);
4854 zone
->wait_table_bits
=
4855 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4856 alloc_size
= zone
->wait_table_hash_nr_entries
4857 * sizeof(wait_queue_head_t
);
4859 if (!slab_is_available()) {
4860 zone
->wait_table
= (wait_queue_head_t
*)
4861 memblock_virt_alloc_node_nopanic(
4862 alloc_size
, zone
->zone_pgdat
->node_id
);
4865 * This case means that a zone whose size was 0 gets new memory
4866 * via memory hot-add.
4867 * But it may be the case that a new node was hot-added. In
4868 * this case vmalloc() will not be able to use this new node's
4869 * memory - this wait_table must be initialized to use this new
4870 * node itself as well.
4871 * To use this new node's memory, further consideration will be
4874 zone
->wait_table
= vmalloc(alloc_size
);
4876 if (!zone
->wait_table
)
4879 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4880 init_waitqueue_head(zone
->wait_table
+ i
);
4885 static __meminit
void zone_pcp_init(struct zone
*zone
)
4888 * per cpu subsystem is not up at this point. The following code
4889 * relies on the ability of the linker to provide the
4890 * offset of a (static) per cpu variable into the per cpu area.
4892 zone
->pageset
= &boot_pageset
;
4894 if (populated_zone(zone
))
4895 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4896 zone
->name
, zone
->present_pages
,
4897 zone_batchsize(zone
));
4900 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4901 unsigned long zone_start_pfn
,
4903 enum memmap_context context
)
4905 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4907 ret
= zone_wait_table_init(zone
, size
);
4910 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4912 zone
->zone_start_pfn
= zone_start_pfn
;
4914 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4915 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4917 (unsigned long)zone_idx(zone
),
4918 zone_start_pfn
, (zone_start_pfn
+ size
));
4920 zone_init_free_lists(zone
);
4925 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4926 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4929 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4931 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
4932 struct mminit_pfnnid_cache
*state
)
4934 unsigned long start_pfn
, end_pfn
;
4937 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
4938 return state
->last_nid
;
4940 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4942 state
->last_start
= start_pfn
;
4943 state
->last_end
= end_pfn
;
4944 state
->last_nid
= nid
;
4949 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4952 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4953 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4954 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4956 * If an architecture guarantees that all ranges registered contain no holes
4957 * and may be freed, this this function may be used instead of calling
4958 * memblock_free_early_nid() manually.
4960 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4962 unsigned long start_pfn
, end_pfn
;
4965 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4966 start_pfn
= min(start_pfn
, max_low_pfn
);
4967 end_pfn
= min(end_pfn
, max_low_pfn
);
4969 if (start_pfn
< end_pfn
)
4970 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4971 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4977 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4978 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4980 * If an architecture guarantees that all ranges registered contain no holes and may
4981 * be freed, this function may be used instead of calling memory_present() manually.
4983 void __init
sparse_memory_present_with_active_regions(int nid
)
4985 unsigned long start_pfn
, end_pfn
;
4988 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4989 memory_present(this_nid
, start_pfn
, end_pfn
);
4993 * get_pfn_range_for_nid - Return the start and end page frames for a node
4994 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4995 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4996 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4998 * It returns the start and end page frame of a node based on information
4999 * provided by memblock_set_node(). If called for a node
5000 * with no available memory, a warning is printed and the start and end
5003 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
5004 unsigned long *start_pfn
, unsigned long *end_pfn
)
5006 unsigned long this_start_pfn
, this_end_pfn
;
5012 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
5013 *start_pfn
= min(*start_pfn
, this_start_pfn
);
5014 *end_pfn
= max(*end_pfn
, this_end_pfn
);
5017 if (*start_pfn
== -1UL)
5022 * This finds a zone that can be used for ZONE_MOVABLE pages. The
5023 * assumption is made that zones within a node are ordered in monotonic
5024 * increasing memory addresses so that the "highest" populated zone is used
5026 static void __init
find_usable_zone_for_movable(void)
5029 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
5030 if (zone_index
== ZONE_MOVABLE
)
5033 if (arch_zone_highest_possible_pfn
[zone_index
] >
5034 arch_zone_lowest_possible_pfn
[zone_index
])
5038 VM_BUG_ON(zone_index
== -1);
5039 movable_zone
= zone_index
;
5043 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
5044 * because it is sized independent of architecture. Unlike the other zones,
5045 * the starting point for ZONE_MOVABLE is not fixed. It may be different
5046 * in each node depending on the size of each node and how evenly kernelcore
5047 * is distributed. This helper function adjusts the zone ranges
5048 * provided by the architecture for a given node by using the end of the
5049 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
5050 * zones within a node are in order of monotonic increases memory addresses
5052 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
5053 unsigned long zone_type
,
5054 unsigned long node_start_pfn
,
5055 unsigned long node_end_pfn
,
5056 unsigned long *zone_start_pfn
,
5057 unsigned long *zone_end_pfn
)
5059 /* Only adjust if ZONE_MOVABLE is on this node */
5060 if (zone_movable_pfn
[nid
]) {
5061 /* Size ZONE_MOVABLE */
5062 if (zone_type
== ZONE_MOVABLE
) {
5063 *zone_start_pfn
= zone_movable_pfn
[nid
];
5064 *zone_end_pfn
= min(node_end_pfn
,
5065 arch_zone_highest_possible_pfn
[movable_zone
]);
5067 /* Adjust for ZONE_MOVABLE starting within this range */
5068 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
5069 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
5070 *zone_end_pfn
= zone_movable_pfn
[nid
];
5072 /* Check if this whole range is within ZONE_MOVABLE */
5073 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
5074 *zone_start_pfn
= *zone_end_pfn
;
5079 * Return the number of pages a zone spans in a node, including holes
5080 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
5082 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5083 unsigned long zone_type
,
5084 unsigned long node_start_pfn
,
5085 unsigned long node_end_pfn
,
5086 unsigned long *ignored
)
5088 unsigned long zone_start_pfn
, zone_end_pfn
;
5090 /* When hotadd a new node, the node should be empty */
5091 if (!node_start_pfn
&& !node_end_pfn
)
5094 /* Get the start and end of the zone */
5095 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
5096 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
5097 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5098 node_start_pfn
, node_end_pfn
,
5099 &zone_start_pfn
, &zone_end_pfn
);
5101 /* Check that this node has pages within the zone's required range */
5102 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
5105 /* Move the zone boundaries inside the node if necessary */
5106 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
5107 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
5109 /* Return the spanned pages */
5110 return zone_end_pfn
- zone_start_pfn
;
5114 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
5115 * then all holes in the requested range will be accounted for.
5117 unsigned long __meminit
__absent_pages_in_range(int nid
,
5118 unsigned long range_start_pfn
,
5119 unsigned long range_end_pfn
)
5121 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
5122 unsigned long start_pfn
, end_pfn
;
5125 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5126 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
5127 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
5128 nr_absent
-= end_pfn
- start_pfn
;
5134 * absent_pages_in_range - Return number of page frames in holes within a range
5135 * @start_pfn: The start PFN to start searching for holes
5136 * @end_pfn: The end PFN to stop searching for holes
5138 * It returns the number of pages frames in memory holes within a range.
5140 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
5141 unsigned long end_pfn
)
5143 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
5146 /* Return the number of page frames in holes in a zone on a node */
5147 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5148 unsigned long zone_type
,
5149 unsigned long node_start_pfn
,
5150 unsigned long node_end_pfn
,
5151 unsigned long *ignored
)
5153 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
5154 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
5155 unsigned long zone_start_pfn
, zone_end_pfn
;
5157 /* When hotadd a new node, the node should be empty */
5158 if (!node_start_pfn
&& !node_end_pfn
)
5161 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
5162 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
5164 adjust_zone_range_for_zone_movable(nid
, zone_type
,
5165 node_start_pfn
, node_end_pfn
,
5166 &zone_start_pfn
, &zone_end_pfn
);
5167 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
5170 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5171 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
5172 unsigned long zone_type
,
5173 unsigned long node_start_pfn
,
5174 unsigned long node_end_pfn
,
5175 unsigned long *zones_size
)
5177 return zones_size
[zone_type
];
5180 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5181 unsigned long zone_type
,
5182 unsigned long node_start_pfn
,
5183 unsigned long node_end_pfn
,
5184 unsigned long *zholes_size
)
5189 return zholes_size
[zone_type
];
5192 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5194 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5195 unsigned long node_start_pfn
,
5196 unsigned long node_end_pfn
,
5197 unsigned long *zones_size
,
5198 unsigned long *zholes_size
)
5200 unsigned long realtotalpages
= 0, totalpages
= 0;
5203 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5204 struct zone
*zone
= pgdat
->node_zones
+ i
;
5205 unsigned long size
, real_size
;
5207 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5211 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5212 node_start_pfn
, node_end_pfn
,
5214 zone
->spanned_pages
= size
;
5215 zone
->present_pages
= real_size
;
5218 realtotalpages
+= real_size
;
5221 pgdat
->node_spanned_pages
= totalpages
;
5222 pgdat
->node_present_pages
= realtotalpages
;
5223 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5227 #ifndef CONFIG_SPARSEMEM
5229 * Calculate the size of the zone->blockflags rounded to an unsigned long
5230 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5231 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5232 * round what is now in bits to nearest long in bits, then return it in
5235 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5237 unsigned long usemapsize
;
5239 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5240 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5241 usemapsize
= usemapsize
>> pageblock_order
;
5242 usemapsize
*= NR_PAGEBLOCK_BITS
;
5243 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5245 return usemapsize
/ 8;
5248 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5250 unsigned long zone_start_pfn
,
5251 unsigned long zonesize
)
5253 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5254 zone
->pageblock_flags
= NULL
;
5256 zone
->pageblock_flags
=
5257 memblock_virt_alloc_node_nopanic(usemapsize
,
5261 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5262 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5263 #endif /* CONFIG_SPARSEMEM */
5265 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5267 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5268 void __paginginit
set_pageblock_order(void)
5272 /* Check that pageblock_nr_pages has not already been setup */
5273 if (pageblock_order
)
5276 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5277 order
= HUGETLB_PAGE_ORDER
;
5279 order
= MAX_ORDER
- 1;
5282 * Assume the largest contiguous order of interest is a huge page.
5283 * This value may be variable depending on boot parameters on IA64 and
5286 pageblock_order
= order
;
5288 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5291 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5292 * is unused as pageblock_order is set at compile-time. See
5293 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5296 void __paginginit
set_pageblock_order(void)
5300 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5302 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5303 unsigned long present_pages
)
5305 unsigned long pages
= spanned_pages
;
5308 * Provide a more accurate estimation if there are holes within
5309 * the zone and SPARSEMEM is in use. If there are holes within the
5310 * zone, each populated memory region may cost us one or two extra
5311 * memmap pages due to alignment because memmap pages for each
5312 * populated regions may not naturally algined on page boundary.
5313 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5315 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5316 IS_ENABLED(CONFIG_SPARSEMEM
))
5317 pages
= present_pages
;
5319 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5323 * Set up the zone data structures:
5324 * - mark all pages reserved
5325 * - mark all memory queues empty
5326 * - clear the memory bitmaps
5328 * NOTE: pgdat should get zeroed by caller.
5330 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5333 int nid
= pgdat
->node_id
;
5334 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
5337 pgdat_resize_init(pgdat
);
5338 #ifdef CONFIG_NUMA_BALANCING
5339 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5340 pgdat
->numabalancing_migrate_nr_pages
= 0;
5341 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5343 init_waitqueue_head(&pgdat
->kswapd_wait
);
5344 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5345 pgdat_page_ext_init(pgdat
);
5347 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5348 struct zone
*zone
= pgdat
->node_zones
+ j
;
5349 unsigned long size
, realsize
, freesize
, memmap_pages
;
5351 size
= zone
->spanned_pages
;
5352 realsize
= freesize
= zone
->present_pages
;
5355 * Adjust freesize so that it accounts for how much memory
5356 * is used by this zone for memmap. This affects the watermark
5357 * and per-cpu initialisations
5359 memmap_pages
= calc_memmap_size(size
, realsize
);
5360 if (!is_highmem_idx(j
)) {
5361 if (freesize
>= memmap_pages
) {
5362 freesize
-= memmap_pages
;
5365 " %s zone: %lu pages used for memmap\n",
5366 zone_names
[j
], memmap_pages
);
5369 " %s zone: %lu pages exceeds freesize %lu\n",
5370 zone_names
[j
], memmap_pages
, freesize
);
5373 /* Account for reserved pages */
5374 if (j
== 0 && freesize
> dma_reserve
) {
5375 freesize
-= dma_reserve
;
5376 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5377 zone_names
[0], dma_reserve
);
5380 if (!is_highmem_idx(j
))
5381 nr_kernel_pages
+= freesize
;
5382 /* Charge for highmem memmap if there are enough kernel pages */
5383 else if (nr_kernel_pages
> memmap_pages
* 2)
5384 nr_kernel_pages
-= memmap_pages
;
5385 nr_all_pages
+= freesize
;
5388 * Set an approximate value for lowmem here, it will be adjusted
5389 * when the bootmem allocator frees pages into the buddy system.
5390 * And all highmem pages will be managed by the buddy system.
5392 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5395 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
5397 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
5399 zone
->name
= zone_names
[j
];
5400 spin_lock_init(&zone
->lock
);
5401 spin_lock_init(&zone
->lru_lock
);
5402 zone_seqlock_init(zone
);
5403 zone
->zone_pgdat
= pgdat
;
5404 zone_pcp_init(zone
);
5406 /* For bootup, initialized properly in watermark setup */
5407 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
5409 lruvec_init(&zone
->lruvec
);
5413 set_pageblock_order();
5414 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5415 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
5416 size
, MEMMAP_EARLY
);
5418 memmap_init(size
, nid
, j
, zone_start_pfn
);
5419 zone_start_pfn
+= size
;
5423 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
5425 /* Skip empty nodes */
5426 if (!pgdat
->node_spanned_pages
)
5429 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5430 /* ia64 gets its own node_mem_map, before this, without bootmem */
5431 if (!pgdat
->node_mem_map
) {
5432 unsigned long size
, start
, end
;
5436 * The zone's endpoints aren't required to be MAX_ORDER
5437 * aligned but the node_mem_map endpoints must be in order
5438 * for the buddy allocator to function correctly.
5440 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5441 end
= pgdat_end_pfn(pgdat
);
5442 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5443 size
= (end
- start
) * sizeof(struct page
);
5444 map
= alloc_remap(pgdat
->node_id
, size
);
5446 map
= memblock_virt_alloc_node_nopanic(size
,
5448 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
5450 #ifndef CONFIG_NEED_MULTIPLE_NODES
5452 * With no DISCONTIG, the global mem_map is just set as node 0's
5454 if (pgdat
== NODE_DATA(0)) {
5455 mem_map
= NODE_DATA(0)->node_mem_map
;
5456 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5457 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5458 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
5459 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5462 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5465 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5466 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5468 pg_data_t
*pgdat
= NODE_DATA(nid
);
5469 unsigned long start_pfn
= 0;
5470 unsigned long end_pfn
= 0;
5472 /* pg_data_t should be reset to zero when it's allocated */
5473 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
5475 reset_deferred_meminit(pgdat
);
5476 pgdat
->node_id
= nid
;
5477 pgdat
->node_start_pfn
= node_start_pfn
;
5478 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5479 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5480 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5481 (u64
)start_pfn
<< PAGE_SHIFT
, ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5483 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5484 zones_size
, zholes_size
);
5486 alloc_node_mem_map(pgdat
);
5487 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5488 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5489 nid
, (unsigned long)pgdat
,
5490 (unsigned long)pgdat
->node_mem_map
);
5493 free_area_init_core(pgdat
);
5496 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5498 #if MAX_NUMNODES > 1
5500 * Figure out the number of possible node ids.
5502 void __init
setup_nr_node_ids(void)
5504 unsigned int highest
;
5506 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
5507 nr_node_ids
= highest
+ 1;
5512 * node_map_pfn_alignment - determine the maximum internode alignment
5514 * This function should be called after node map is populated and sorted.
5515 * It calculates the maximum power of two alignment which can distinguish
5518 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5519 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5520 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5521 * shifted, 1GiB is enough and this function will indicate so.
5523 * This is used to test whether pfn -> nid mapping of the chosen memory
5524 * model has fine enough granularity to avoid incorrect mapping for the
5525 * populated node map.
5527 * Returns the determined alignment in pfn's. 0 if there is no alignment
5528 * requirement (single node).
5530 unsigned long __init
node_map_pfn_alignment(void)
5532 unsigned long accl_mask
= 0, last_end
= 0;
5533 unsigned long start
, end
, mask
;
5537 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5538 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5545 * Start with a mask granular enough to pin-point to the
5546 * start pfn and tick off bits one-by-one until it becomes
5547 * too coarse to separate the current node from the last.
5549 mask
= ~((1 << __ffs(start
)) - 1);
5550 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5553 /* accumulate all internode masks */
5557 /* convert mask to number of pages */
5558 return ~accl_mask
+ 1;
5561 /* Find the lowest pfn for a node */
5562 static unsigned long __init
find_min_pfn_for_node(int nid
)
5564 unsigned long min_pfn
= ULONG_MAX
;
5565 unsigned long start_pfn
;
5568 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5569 min_pfn
= min(min_pfn
, start_pfn
);
5571 if (min_pfn
== ULONG_MAX
) {
5573 "Could not find start_pfn for node %d\n", nid
);
5581 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5583 * It returns the minimum PFN based on information provided via
5584 * memblock_set_node().
5586 unsigned long __init
find_min_pfn_with_active_regions(void)
5588 return find_min_pfn_for_node(MAX_NUMNODES
);
5592 * early_calculate_totalpages()
5593 * Sum pages in active regions for movable zone.
5594 * Populate N_MEMORY for calculating usable_nodes.
5596 static unsigned long __init
early_calculate_totalpages(void)
5598 unsigned long totalpages
= 0;
5599 unsigned long start_pfn
, end_pfn
;
5602 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5603 unsigned long pages
= end_pfn
- start_pfn
;
5605 totalpages
+= pages
;
5607 node_set_state(nid
, N_MEMORY
);
5613 * Find the PFN the Movable zone begins in each node. Kernel memory
5614 * is spread evenly between nodes as long as the nodes have enough
5615 * memory. When they don't, some nodes will have more kernelcore than
5618 static void __init
find_zone_movable_pfns_for_nodes(void)
5621 unsigned long usable_startpfn
;
5622 unsigned long kernelcore_node
, kernelcore_remaining
;
5623 /* save the state before borrow the nodemask */
5624 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5625 unsigned long totalpages
= early_calculate_totalpages();
5626 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5627 struct memblock_region
*r
;
5629 /* Need to find movable_zone earlier when movable_node is specified. */
5630 find_usable_zone_for_movable();
5633 * If movable_node is specified, ignore kernelcore and movablecore
5636 if (movable_node_is_enabled()) {
5637 for_each_memblock(memory
, r
) {
5638 if (!memblock_is_hotpluggable(r
))
5643 usable_startpfn
= PFN_DOWN(r
->base
);
5644 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5645 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5653 * If movablecore=nn[KMG] was specified, calculate what size of
5654 * kernelcore that corresponds so that memory usable for
5655 * any allocation type is evenly spread. If both kernelcore
5656 * and movablecore are specified, then the value of kernelcore
5657 * will be used for required_kernelcore if it's greater than
5658 * what movablecore would have allowed.
5660 if (required_movablecore
) {
5661 unsigned long corepages
;
5664 * Round-up so that ZONE_MOVABLE is at least as large as what
5665 * was requested by the user
5667 required_movablecore
=
5668 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5669 corepages
= totalpages
- required_movablecore
;
5671 required_kernelcore
= max(required_kernelcore
, corepages
);
5674 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5675 if (!required_kernelcore
)
5678 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5679 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5682 /* Spread kernelcore memory as evenly as possible throughout nodes */
5683 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5684 for_each_node_state(nid
, N_MEMORY
) {
5685 unsigned long start_pfn
, end_pfn
;
5688 * Recalculate kernelcore_node if the division per node
5689 * now exceeds what is necessary to satisfy the requested
5690 * amount of memory for the kernel
5692 if (required_kernelcore
< kernelcore_node
)
5693 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5696 * As the map is walked, we track how much memory is usable
5697 * by the kernel using kernelcore_remaining. When it is
5698 * 0, the rest of the node is usable by ZONE_MOVABLE
5700 kernelcore_remaining
= kernelcore_node
;
5702 /* Go through each range of PFNs within this node */
5703 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5704 unsigned long size_pages
;
5706 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5707 if (start_pfn
>= end_pfn
)
5710 /* Account for what is only usable for kernelcore */
5711 if (start_pfn
< usable_startpfn
) {
5712 unsigned long kernel_pages
;
5713 kernel_pages
= min(end_pfn
, usable_startpfn
)
5716 kernelcore_remaining
-= min(kernel_pages
,
5717 kernelcore_remaining
);
5718 required_kernelcore
-= min(kernel_pages
,
5719 required_kernelcore
);
5721 /* Continue if range is now fully accounted */
5722 if (end_pfn
<= usable_startpfn
) {
5725 * Push zone_movable_pfn to the end so
5726 * that if we have to rebalance
5727 * kernelcore across nodes, we will
5728 * not double account here
5730 zone_movable_pfn
[nid
] = end_pfn
;
5733 start_pfn
= usable_startpfn
;
5737 * The usable PFN range for ZONE_MOVABLE is from
5738 * start_pfn->end_pfn. Calculate size_pages as the
5739 * number of pages used as kernelcore
5741 size_pages
= end_pfn
- start_pfn
;
5742 if (size_pages
> kernelcore_remaining
)
5743 size_pages
= kernelcore_remaining
;
5744 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5747 * Some kernelcore has been met, update counts and
5748 * break if the kernelcore for this node has been
5751 required_kernelcore
-= min(required_kernelcore
,
5753 kernelcore_remaining
-= size_pages
;
5754 if (!kernelcore_remaining
)
5760 * If there is still required_kernelcore, we do another pass with one
5761 * less node in the count. This will push zone_movable_pfn[nid] further
5762 * along on the nodes that still have memory until kernelcore is
5766 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5770 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5771 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5772 zone_movable_pfn
[nid
] =
5773 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5776 /* restore the node_state */
5777 node_states
[N_MEMORY
] = saved_node_state
;
5780 /* Any regular or high memory on that node ? */
5781 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5783 enum zone_type zone_type
;
5785 if (N_MEMORY
== N_NORMAL_MEMORY
)
5788 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5789 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5790 if (populated_zone(zone
)) {
5791 node_set_state(nid
, N_HIGH_MEMORY
);
5792 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5793 zone_type
<= ZONE_NORMAL
)
5794 node_set_state(nid
, N_NORMAL_MEMORY
);
5801 * free_area_init_nodes - Initialise all pg_data_t and zone data
5802 * @max_zone_pfn: an array of max PFNs for each zone
5804 * This will call free_area_init_node() for each active node in the system.
5805 * Using the page ranges provided by memblock_set_node(), the size of each
5806 * zone in each node and their holes is calculated. If the maximum PFN
5807 * between two adjacent zones match, it is assumed that the zone is empty.
5808 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5809 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5810 * starts where the previous one ended. For example, ZONE_DMA32 starts
5811 * at arch_max_dma_pfn.
5813 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5815 unsigned long start_pfn
, end_pfn
;
5818 /* Record where the zone boundaries are */
5819 memset(arch_zone_lowest_possible_pfn
, 0,
5820 sizeof(arch_zone_lowest_possible_pfn
));
5821 memset(arch_zone_highest_possible_pfn
, 0,
5822 sizeof(arch_zone_highest_possible_pfn
));
5823 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5824 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5825 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5826 if (i
== ZONE_MOVABLE
)
5828 arch_zone_lowest_possible_pfn
[i
] =
5829 arch_zone_highest_possible_pfn
[i
-1];
5830 arch_zone_highest_possible_pfn
[i
] =
5831 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5833 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5834 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5836 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5837 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5838 find_zone_movable_pfns_for_nodes();
5840 /* Print out the zone ranges */
5841 pr_info("Zone ranges:\n");
5842 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5843 if (i
== ZONE_MOVABLE
)
5845 pr_info(" %-8s ", zone_names
[i
]);
5846 if (arch_zone_lowest_possible_pfn
[i
] ==
5847 arch_zone_highest_possible_pfn
[i
])
5850 pr_cont("[mem %#018Lx-%#018Lx]\n",
5851 (u64
)arch_zone_lowest_possible_pfn
[i
]
5853 ((u64
)arch_zone_highest_possible_pfn
[i
]
5854 << PAGE_SHIFT
) - 1);
5857 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5858 pr_info("Movable zone start for each node\n");
5859 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5860 if (zone_movable_pfn
[i
])
5861 pr_info(" Node %d: %#018Lx\n", i
,
5862 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
5865 /* Print out the early node map */
5866 pr_info("Early memory node ranges\n");
5867 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5868 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
5869 (u64
)start_pfn
<< PAGE_SHIFT
,
5870 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5872 /* Initialise every node */
5873 mminit_verify_pageflags_layout();
5874 setup_nr_node_ids();
5875 for_each_online_node(nid
) {
5876 pg_data_t
*pgdat
= NODE_DATA(nid
);
5877 free_area_init_node(nid
, NULL
,
5878 find_min_pfn_for_node(nid
), NULL
);
5880 /* Any memory on that node */
5881 if (pgdat
->node_present_pages
)
5882 node_set_state(nid
, N_MEMORY
);
5883 check_for_memory(pgdat
, nid
);
5887 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5889 unsigned long long coremem
;
5893 coremem
= memparse(p
, &p
);
5894 *core
= coremem
>> PAGE_SHIFT
;
5896 /* Paranoid check that UL is enough for the coremem value */
5897 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5903 * kernelcore=size sets the amount of memory for use for allocations that
5904 * cannot be reclaimed or migrated.
5906 static int __init
cmdline_parse_kernelcore(char *p
)
5908 return cmdline_parse_core(p
, &required_kernelcore
);
5912 * movablecore=size sets the amount of memory for use for allocations that
5913 * can be reclaimed or migrated.
5915 static int __init
cmdline_parse_movablecore(char *p
)
5917 return cmdline_parse_core(p
, &required_movablecore
);
5920 early_param("kernelcore", cmdline_parse_kernelcore
);
5921 early_param("movablecore", cmdline_parse_movablecore
);
5923 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5925 void adjust_managed_page_count(struct page
*page
, long count
)
5927 spin_lock(&managed_page_count_lock
);
5928 page_zone(page
)->managed_pages
+= count
;
5929 totalram_pages
+= count
;
5930 #ifdef CONFIG_HIGHMEM
5931 if (PageHighMem(page
))
5932 totalhigh_pages
+= count
;
5934 spin_unlock(&managed_page_count_lock
);
5936 EXPORT_SYMBOL(adjust_managed_page_count
);
5938 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5941 unsigned long pages
= 0;
5943 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5944 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5945 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5946 if ((unsigned int)poison
<= 0xFF)
5947 memset(pos
, poison
, PAGE_SIZE
);
5948 free_reserved_page(virt_to_page(pos
));
5952 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5953 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5957 EXPORT_SYMBOL(free_reserved_area
);
5959 #ifdef CONFIG_HIGHMEM
5960 void free_highmem_page(struct page
*page
)
5962 __free_reserved_page(page
);
5964 page_zone(page
)->managed_pages
++;
5970 void __init
mem_init_print_info(const char *str
)
5972 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5973 unsigned long init_code_size
, init_data_size
;
5975 physpages
= get_num_physpages();
5976 codesize
= _etext
- _stext
;
5977 datasize
= _edata
- _sdata
;
5978 rosize
= __end_rodata
- __start_rodata
;
5979 bss_size
= __bss_stop
- __bss_start
;
5980 init_data_size
= __init_end
- __init_begin
;
5981 init_code_size
= _einittext
- _sinittext
;
5984 * Detect special cases and adjust section sizes accordingly:
5985 * 1) .init.* may be embedded into .data sections
5986 * 2) .init.text.* may be out of [__init_begin, __init_end],
5987 * please refer to arch/tile/kernel/vmlinux.lds.S.
5988 * 3) .rodata.* may be embedded into .text or .data sections.
5990 #define adj_init_size(start, end, size, pos, adj) \
5992 if (start <= pos && pos < end && size > adj) \
5996 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5997 _sinittext
, init_code_size
);
5998 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5999 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
6000 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
6001 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
6003 #undef adj_init_size
6005 pr_info("Memory: %luK/%luK available "
6006 "(%luK kernel code, %luK rwdata, %luK rodata, "
6007 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
6008 #ifdef CONFIG_HIGHMEM
6012 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
6013 codesize
>> 10, datasize
>> 10, rosize
>> 10,
6014 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
6015 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
-10),
6016 totalcma_pages
<< (PAGE_SHIFT
-10),
6017 #ifdef CONFIG_HIGHMEM
6018 totalhigh_pages
<< (PAGE_SHIFT
-10),
6020 str
? ", " : "", str
? str
: "");
6024 * set_dma_reserve - set the specified number of pages reserved in the first zone
6025 * @new_dma_reserve: The number of pages to mark reserved
6027 * The per-cpu batchsize and zone watermarks are determined by present_pages.
6028 * In the DMA zone, a significant percentage may be consumed by kernel image
6029 * and other unfreeable allocations which can skew the watermarks badly. This
6030 * function may optionally be used to account for unfreeable pages in the
6031 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
6032 * smaller per-cpu batchsize.
6034 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
6036 dma_reserve
= new_dma_reserve
;
6039 void __init
free_area_init(unsigned long *zones_size
)
6041 free_area_init_node(0, zones_size
,
6042 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
6045 static int page_alloc_cpu_notify(struct notifier_block
*self
,
6046 unsigned long action
, void *hcpu
)
6048 int cpu
= (unsigned long)hcpu
;
6050 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
6051 lru_add_drain_cpu(cpu
);
6055 * Spill the event counters of the dead processor
6056 * into the current processors event counters.
6057 * This artificially elevates the count of the current
6060 vm_events_fold_cpu(cpu
);
6063 * Zero the differential counters of the dead processor
6064 * so that the vm statistics are consistent.
6066 * This is only okay since the processor is dead and cannot
6067 * race with what we are doing.
6069 cpu_vm_stats_fold(cpu
);
6074 void __init
page_alloc_init(void)
6076 hotcpu_notifier(page_alloc_cpu_notify
, 0);
6080 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
6081 * or min_free_kbytes changes.
6083 static void calculate_totalreserve_pages(void)
6085 struct pglist_data
*pgdat
;
6086 unsigned long reserve_pages
= 0;
6087 enum zone_type i
, j
;
6089 for_each_online_pgdat(pgdat
) {
6090 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
6091 struct zone
*zone
= pgdat
->node_zones
+ i
;
6094 /* Find valid and maximum lowmem_reserve in the zone */
6095 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
6096 if (zone
->lowmem_reserve
[j
] > max
)
6097 max
= zone
->lowmem_reserve
[j
];
6100 /* we treat the high watermark as reserved pages. */
6101 max
+= high_wmark_pages(zone
);
6103 if (max
> zone
->managed_pages
)
6104 max
= zone
->managed_pages
;
6105 reserve_pages
+= max
;
6107 * Lowmem reserves are not available to
6108 * GFP_HIGHUSER page cache allocations and
6109 * kswapd tries to balance zones to their high
6110 * watermark. As a result, neither should be
6111 * regarded as dirtyable memory, to prevent a
6112 * situation where reclaim has to clean pages
6113 * in order to balance the zones.
6115 zone
->dirty_balance_reserve
= max
;
6118 dirty_balance_reserve
= reserve_pages
;
6119 totalreserve_pages
= reserve_pages
;
6123 * setup_per_zone_lowmem_reserve - called whenever
6124 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
6125 * has a correct pages reserved value, so an adequate number of
6126 * pages are left in the zone after a successful __alloc_pages().
6128 static void setup_per_zone_lowmem_reserve(void)
6130 struct pglist_data
*pgdat
;
6131 enum zone_type j
, idx
;
6133 for_each_online_pgdat(pgdat
) {
6134 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
6135 struct zone
*zone
= pgdat
->node_zones
+ j
;
6136 unsigned long managed_pages
= zone
->managed_pages
;
6138 zone
->lowmem_reserve
[j
] = 0;
6142 struct zone
*lower_zone
;
6146 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
6147 sysctl_lowmem_reserve_ratio
[idx
] = 1;
6149 lower_zone
= pgdat
->node_zones
+ idx
;
6150 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
6151 sysctl_lowmem_reserve_ratio
[idx
];
6152 managed_pages
+= lower_zone
->managed_pages
;
6157 /* update totalreserve_pages */
6158 calculate_totalreserve_pages();
6161 static void __setup_per_zone_wmarks(void)
6163 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
6164 unsigned long lowmem_pages
= 0;
6166 unsigned long flags
;
6168 /* Calculate total number of !ZONE_HIGHMEM pages */
6169 for_each_zone(zone
) {
6170 if (!is_highmem(zone
))
6171 lowmem_pages
+= zone
->managed_pages
;
6174 for_each_zone(zone
) {
6177 spin_lock_irqsave(&zone
->lock
, flags
);
6178 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6179 do_div(tmp
, lowmem_pages
);
6180 if (is_highmem(zone
)) {
6182 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6183 * need highmem pages, so cap pages_min to a small
6186 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6187 * deltas control asynch page reclaim, and so should
6188 * not be capped for highmem.
6190 unsigned long min_pages
;
6192 min_pages
= zone
->managed_pages
/ 1024;
6193 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6194 zone
->watermark
[WMARK_MIN
] = min_pages
;
6197 * If it's a lowmem zone, reserve a number of pages
6198 * proportionate to the zone's size.
6200 zone
->watermark
[WMARK_MIN
] = tmp
;
6203 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
6204 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
6206 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
6207 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
6208 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
6210 setup_zone_migrate_reserve(zone
);
6211 spin_unlock_irqrestore(&zone
->lock
, flags
);
6214 /* update totalreserve_pages */
6215 calculate_totalreserve_pages();
6219 * setup_per_zone_wmarks - called when min_free_kbytes changes
6220 * or when memory is hot-{added|removed}
6222 * Ensures that the watermark[min,low,high] values for each zone are set
6223 * correctly with respect to min_free_kbytes.
6225 void setup_per_zone_wmarks(void)
6227 mutex_lock(&zonelists_mutex
);
6228 __setup_per_zone_wmarks();
6229 mutex_unlock(&zonelists_mutex
);
6233 * The inactive anon list should be small enough that the VM never has to
6234 * do too much work, but large enough that each inactive page has a chance
6235 * to be referenced again before it is swapped out.
6237 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6238 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6239 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6240 * the anonymous pages are kept on the inactive list.
6243 * memory ratio inactive anon
6244 * -------------------------------------
6253 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
6255 unsigned int gb
, ratio
;
6257 /* Zone size in gigabytes */
6258 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
6260 ratio
= int_sqrt(10 * gb
);
6264 zone
->inactive_ratio
= ratio
;
6267 static void __meminit
setup_per_zone_inactive_ratio(void)
6272 calculate_zone_inactive_ratio(zone
);
6276 * Initialise min_free_kbytes.
6278 * For small machines we want it small (128k min). For large machines
6279 * we want it large (64MB max). But it is not linear, because network
6280 * bandwidth does not increase linearly with machine size. We use
6282 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6283 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6299 int __meminit
init_per_zone_wmark_min(void)
6301 unsigned long lowmem_kbytes
;
6302 int new_min_free_kbytes
;
6304 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6305 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6307 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6308 min_free_kbytes
= new_min_free_kbytes
;
6309 if (min_free_kbytes
< 128)
6310 min_free_kbytes
= 128;
6311 if (min_free_kbytes
> 65536)
6312 min_free_kbytes
= 65536;
6314 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6315 new_min_free_kbytes
, user_min_free_kbytes
);
6317 setup_per_zone_wmarks();
6318 refresh_zone_stat_thresholds();
6319 setup_per_zone_lowmem_reserve();
6320 setup_per_zone_inactive_ratio();
6323 module_init(init_per_zone_wmark_min
)
6326 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6327 * that we can call two helper functions whenever min_free_kbytes
6330 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6331 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6335 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6340 user_min_free_kbytes
= min_free_kbytes
;
6341 setup_per_zone_wmarks();
6347 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6348 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6353 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6358 zone
->min_unmapped_pages
= (zone
->managed_pages
*
6359 sysctl_min_unmapped_ratio
) / 100;
6363 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6364 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6369 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6374 zone
->min_slab_pages
= (zone
->managed_pages
*
6375 sysctl_min_slab_ratio
) / 100;
6381 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6382 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6383 * whenever sysctl_lowmem_reserve_ratio changes.
6385 * The reserve ratio obviously has absolutely no relation with the
6386 * minimum watermarks. The lowmem reserve ratio can only make sense
6387 * if in function of the boot time zone sizes.
6389 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6390 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6392 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6393 setup_per_zone_lowmem_reserve();
6398 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6399 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6400 * pagelist can have before it gets flushed back to buddy allocator.
6402 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6403 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6406 int old_percpu_pagelist_fraction
;
6409 mutex_lock(&pcp_batch_high_lock
);
6410 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6412 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6413 if (!write
|| ret
< 0)
6416 /* Sanity checking to avoid pcp imbalance */
6417 if (percpu_pagelist_fraction
&&
6418 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6419 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6425 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6428 for_each_populated_zone(zone
) {
6431 for_each_possible_cpu(cpu
)
6432 pageset_set_high_and_batch(zone
,
6433 per_cpu_ptr(zone
->pageset
, cpu
));
6436 mutex_unlock(&pcp_batch_high_lock
);
6441 int hashdist
= HASHDIST_DEFAULT
;
6443 static int __init
set_hashdist(char *str
)
6447 hashdist
= simple_strtoul(str
, &str
, 0);
6450 __setup("hashdist=", set_hashdist
);
6454 * allocate a large system hash table from bootmem
6455 * - it is assumed that the hash table must contain an exact power-of-2
6456 * quantity of entries
6457 * - limit is the number of hash buckets, not the total allocation size
6459 void *__init
alloc_large_system_hash(const char *tablename
,
6460 unsigned long bucketsize
,
6461 unsigned long numentries
,
6464 unsigned int *_hash_shift
,
6465 unsigned int *_hash_mask
,
6466 unsigned long low_limit
,
6467 unsigned long high_limit
)
6469 unsigned long long max
= high_limit
;
6470 unsigned long log2qty
, size
;
6473 /* allow the kernel cmdline to have a say */
6475 /* round applicable memory size up to nearest megabyte */
6476 numentries
= nr_kernel_pages
;
6478 /* It isn't necessary when PAGE_SIZE >= 1MB */
6479 if (PAGE_SHIFT
< 20)
6480 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6482 /* limit to 1 bucket per 2^scale bytes of low memory */
6483 if (scale
> PAGE_SHIFT
)
6484 numentries
>>= (scale
- PAGE_SHIFT
);
6486 numentries
<<= (PAGE_SHIFT
- scale
);
6488 /* Make sure we've got at least a 0-order allocation.. */
6489 if (unlikely(flags
& HASH_SMALL
)) {
6490 /* Makes no sense without HASH_EARLY */
6491 WARN_ON(!(flags
& HASH_EARLY
));
6492 if (!(numentries
>> *_hash_shift
)) {
6493 numentries
= 1UL << *_hash_shift
;
6494 BUG_ON(!numentries
);
6496 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6497 numentries
= PAGE_SIZE
/ bucketsize
;
6499 numentries
= roundup_pow_of_two(numentries
);
6501 /* limit allocation size to 1/16 total memory by default */
6503 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6504 do_div(max
, bucketsize
);
6506 max
= min(max
, 0x80000000ULL
);
6508 if (numentries
< low_limit
)
6509 numentries
= low_limit
;
6510 if (numentries
> max
)
6513 log2qty
= ilog2(numentries
);
6516 size
= bucketsize
<< log2qty
;
6517 if (flags
& HASH_EARLY
)
6518 table
= memblock_virt_alloc_nopanic(size
, 0);
6520 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6523 * If bucketsize is not a power-of-two, we may free
6524 * some pages at the end of hash table which
6525 * alloc_pages_exact() automatically does
6527 if (get_order(size
) < MAX_ORDER
) {
6528 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6529 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6532 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6535 panic("Failed to allocate %s hash table\n", tablename
);
6537 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6540 ilog2(size
) - PAGE_SHIFT
,
6544 *_hash_shift
= log2qty
;
6546 *_hash_mask
= (1 << log2qty
) - 1;
6551 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6552 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6555 #ifdef CONFIG_SPARSEMEM
6556 return __pfn_to_section(pfn
)->pageblock_flags
;
6558 return zone
->pageblock_flags
;
6559 #endif /* CONFIG_SPARSEMEM */
6562 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6564 #ifdef CONFIG_SPARSEMEM
6565 pfn
&= (PAGES_PER_SECTION
-1);
6566 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6568 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6569 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6570 #endif /* CONFIG_SPARSEMEM */
6574 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6575 * @page: The page within the block of interest
6576 * @pfn: The target page frame number
6577 * @end_bitidx: The last bit of interest to retrieve
6578 * @mask: mask of bits that the caller is interested in
6580 * Return: pageblock_bits flags
6582 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6583 unsigned long end_bitidx
,
6587 unsigned long *bitmap
;
6588 unsigned long bitidx
, word_bitidx
;
6591 zone
= page_zone(page
);
6592 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6593 bitidx
= pfn_to_bitidx(zone
, pfn
);
6594 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6595 bitidx
&= (BITS_PER_LONG
-1);
6597 word
= bitmap
[word_bitidx
];
6598 bitidx
+= end_bitidx
;
6599 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6603 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6604 * @page: The page within the block of interest
6605 * @flags: The flags to set
6606 * @pfn: The target page frame number
6607 * @end_bitidx: The last bit of interest
6608 * @mask: mask of bits that the caller is interested in
6610 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6612 unsigned long end_bitidx
,
6616 unsigned long *bitmap
;
6617 unsigned long bitidx
, word_bitidx
;
6618 unsigned long old_word
, word
;
6620 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6622 zone
= page_zone(page
);
6623 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6624 bitidx
= pfn_to_bitidx(zone
, pfn
);
6625 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6626 bitidx
&= (BITS_PER_LONG
-1);
6628 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6630 bitidx
+= end_bitidx
;
6631 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6632 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6634 word
= READ_ONCE(bitmap
[word_bitidx
]);
6636 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6637 if (word
== old_word
)
6644 * This function checks whether pageblock includes unmovable pages or not.
6645 * If @count is not zero, it is okay to include less @count unmovable pages
6647 * PageLRU check without isolation or lru_lock could race so that
6648 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6649 * expect this function should be exact.
6651 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6652 bool skip_hwpoisoned_pages
)
6654 unsigned long pfn
, iter
, found
;
6658 * For avoiding noise data, lru_add_drain_all() should be called
6659 * If ZONE_MOVABLE, the zone never contains unmovable pages
6661 if (zone_idx(zone
) == ZONE_MOVABLE
)
6663 mt
= get_pageblock_migratetype(page
);
6664 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6667 pfn
= page_to_pfn(page
);
6668 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6669 unsigned long check
= pfn
+ iter
;
6671 if (!pfn_valid_within(check
))
6674 page
= pfn_to_page(check
);
6677 * Hugepages are not in LRU lists, but they're movable.
6678 * We need not scan over tail pages bacause we don't
6679 * handle each tail page individually in migration.
6681 if (PageHuge(page
)) {
6682 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6687 * We can't use page_count without pin a page
6688 * because another CPU can free compound page.
6689 * This check already skips compound tails of THP
6690 * because their page->_count is zero at all time.
6692 if (!atomic_read(&page
->_count
)) {
6693 if (PageBuddy(page
))
6694 iter
+= (1 << page_order(page
)) - 1;
6699 * The HWPoisoned page may be not in buddy system, and
6700 * page_count() is not 0.
6702 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6708 * If there are RECLAIMABLE pages, we need to check
6709 * it. But now, memory offline itself doesn't call
6710 * shrink_node_slabs() and it still to be fixed.
6713 * If the page is not RAM, page_count()should be 0.
6714 * we don't need more check. This is an _used_ not-movable page.
6716 * The problematic thing here is PG_reserved pages. PG_reserved
6717 * is set to both of a memory hole page and a _used_ kernel
6726 bool is_pageblock_removable_nolock(struct page
*page
)
6732 * We have to be careful here because we are iterating over memory
6733 * sections which are not zone aware so we might end up outside of
6734 * the zone but still within the section.
6735 * We have to take care about the node as well. If the node is offline
6736 * its NODE_DATA will be NULL - see page_zone.
6738 if (!node_online(page_to_nid(page
)))
6741 zone
= page_zone(page
);
6742 pfn
= page_to_pfn(page
);
6743 if (!zone_spans_pfn(zone
, pfn
))
6746 return !has_unmovable_pages(zone
, page
, 0, true);
6751 static unsigned long pfn_max_align_down(unsigned long pfn
)
6753 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6754 pageblock_nr_pages
) - 1);
6757 static unsigned long pfn_max_align_up(unsigned long pfn
)
6759 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6760 pageblock_nr_pages
));
6763 /* [start, end) must belong to a single zone. */
6764 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6765 unsigned long start
, unsigned long end
)
6767 /* This function is based on compact_zone() from compaction.c. */
6768 unsigned long nr_reclaimed
;
6769 unsigned long pfn
= start
;
6770 unsigned int tries
= 0;
6775 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6776 if (fatal_signal_pending(current
)) {
6781 if (list_empty(&cc
->migratepages
)) {
6782 cc
->nr_migratepages
= 0;
6783 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6789 } else if (++tries
== 5) {
6790 ret
= ret
< 0 ? ret
: -EBUSY
;
6794 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6796 cc
->nr_migratepages
-= nr_reclaimed
;
6798 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6799 NULL
, 0, cc
->mode
, MR_CMA
);
6802 putback_movable_pages(&cc
->migratepages
);
6809 * alloc_contig_range() -- tries to allocate given range of pages
6810 * @start: start PFN to allocate
6811 * @end: one-past-the-last PFN to allocate
6812 * @migratetype: migratetype of the underlaying pageblocks (either
6813 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6814 * in range must have the same migratetype and it must
6815 * be either of the two.
6817 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6818 * aligned, however it's the caller's responsibility to guarantee that
6819 * we are the only thread that changes migrate type of pageblocks the
6822 * The PFN range must belong to a single zone.
6824 * Returns zero on success or negative error code. On success all
6825 * pages which PFN is in [start, end) are allocated for the caller and
6826 * need to be freed with free_contig_range().
6828 int alloc_contig_range(unsigned long start
, unsigned long end
,
6829 unsigned migratetype
)
6831 unsigned long outer_start
, outer_end
;
6834 struct compact_control cc
= {
6835 .nr_migratepages
= 0,
6837 .zone
= page_zone(pfn_to_page(start
)),
6838 .mode
= MIGRATE_SYNC
,
6839 .ignore_skip_hint
= true,
6841 INIT_LIST_HEAD(&cc
.migratepages
);
6844 * What we do here is we mark all pageblocks in range as
6845 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6846 * have different sizes, and due to the way page allocator
6847 * work, we align the range to biggest of the two pages so
6848 * that page allocator won't try to merge buddies from
6849 * different pageblocks and change MIGRATE_ISOLATE to some
6850 * other migration type.
6852 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6853 * migrate the pages from an unaligned range (ie. pages that
6854 * we are interested in). This will put all the pages in
6855 * range back to page allocator as MIGRATE_ISOLATE.
6857 * When this is done, we take the pages in range from page
6858 * allocator removing them from the buddy system. This way
6859 * page allocator will never consider using them.
6861 * This lets us mark the pageblocks back as
6862 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6863 * aligned range but not in the unaligned, original range are
6864 * put back to page allocator so that buddy can use them.
6867 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6868 pfn_max_align_up(end
), migratetype
,
6873 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6878 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6879 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6880 * more, all pages in [start, end) are free in page allocator.
6881 * What we are going to do is to allocate all pages from
6882 * [start, end) (that is remove them from page allocator).
6884 * The only problem is that pages at the beginning and at the
6885 * end of interesting range may be not aligned with pages that
6886 * page allocator holds, ie. they can be part of higher order
6887 * pages. Because of this, we reserve the bigger range and
6888 * once this is done free the pages we are not interested in.
6890 * We don't have to hold zone->lock here because the pages are
6891 * isolated thus they won't get removed from buddy.
6894 lru_add_drain_all();
6895 drain_all_pages(cc
.zone
);
6898 outer_start
= start
;
6899 while (!PageBuddy(pfn_to_page(outer_start
))) {
6900 if (++order
>= MAX_ORDER
) {
6904 outer_start
&= ~0UL << order
;
6907 /* Make sure the range is really isolated. */
6908 if (test_pages_isolated(outer_start
, end
, false)) {
6909 pr_info("%s: [%lx, %lx) PFNs busy\n",
6910 __func__
, outer_start
, end
);
6915 /* Grab isolated pages from freelists. */
6916 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6922 /* Free head and tail (if any) */
6923 if (start
!= outer_start
)
6924 free_contig_range(outer_start
, start
- outer_start
);
6925 if (end
!= outer_end
)
6926 free_contig_range(end
, outer_end
- end
);
6929 undo_isolate_page_range(pfn_max_align_down(start
),
6930 pfn_max_align_up(end
), migratetype
);
6934 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6936 unsigned int count
= 0;
6938 for (; nr_pages
--; pfn
++) {
6939 struct page
*page
= pfn_to_page(pfn
);
6941 count
+= page_count(page
) != 1;
6944 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6948 #ifdef CONFIG_MEMORY_HOTPLUG
6950 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6951 * page high values need to be recalulated.
6953 void __meminit
zone_pcp_update(struct zone
*zone
)
6956 mutex_lock(&pcp_batch_high_lock
);
6957 for_each_possible_cpu(cpu
)
6958 pageset_set_high_and_batch(zone
,
6959 per_cpu_ptr(zone
->pageset
, cpu
));
6960 mutex_unlock(&pcp_batch_high_lock
);
6964 void zone_pcp_reset(struct zone
*zone
)
6966 unsigned long flags
;
6968 struct per_cpu_pageset
*pset
;
6970 /* avoid races with drain_pages() */
6971 local_irq_save(flags
);
6972 if (zone
->pageset
!= &boot_pageset
) {
6973 for_each_online_cpu(cpu
) {
6974 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6975 drain_zonestat(zone
, pset
);
6977 free_percpu(zone
->pageset
);
6978 zone
->pageset
= &boot_pageset
;
6980 local_irq_restore(flags
);
6983 #ifdef CONFIG_MEMORY_HOTREMOVE
6985 * All pages in the range must be isolated before calling this.
6988 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6992 unsigned int order
, i
;
6994 unsigned long flags
;
6995 /* find the first valid pfn */
6996 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
7001 zone
= page_zone(pfn_to_page(pfn
));
7002 spin_lock_irqsave(&zone
->lock
, flags
);
7004 while (pfn
< end_pfn
) {
7005 if (!pfn_valid(pfn
)) {
7009 page
= pfn_to_page(pfn
);
7011 * The HWPoisoned page may be not in buddy system, and
7012 * page_count() is not 0.
7014 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
7016 SetPageReserved(page
);
7020 BUG_ON(page_count(page
));
7021 BUG_ON(!PageBuddy(page
));
7022 order
= page_order(page
);
7023 #ifdef CONFIG_DEBUG_VM
7024 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
7025 pfn
, 1 << order
, end_pfn
);
7027 list_del(&page
->lru
);
7028 rmv_page_order(page
);
7029 zone
->free_area
[order
].nr_free
--;
7030 for (i
= 0; i
< (1 << order
); i
++)
7031 SetPageReserved((page
+i
));
7032 pfn
+= (1 << order
);
7034 spin_unlock_irqrestore(&zone
->lock
, flags
);
7038 #ifdef CONFIG_MEMORY_FAILURE
7039 bool is_free_buddy_page(struct page
*page
)
7041 struct zone
*zone
= page_zone(page
);
7042 unsigned long pfn
= page_to_pfn(page
);
7043 unsigned long flags
;
7046 spin_lock_irqsave(&zone
->lock
, flags
);
7047 for (order
= 0; order
< MAX_ORDER
; order
++) {
7048 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
7050 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
7053 spin_unlock_irqrestore(&zone
->lock
, flags
);
7055 return order
< MAX_ORDER
;