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 int percpu_pagelist_fraction
;
119 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
122 * A cached value of the page's pageblock's migratetype, used when the page is
123 * put on a pcplist. Used to avoid the pageblock migratetype lookup when
124 * freeing from pcplists in most cases, at the cost of possibly becoming stale.
125 * Also the migratetype set in the page does not necessarily match the pcplist
126 * index, e.g. page might have MIGRATE_CMA set but be on a pcplist with any
127 * other index - this ensures that it will be put on the correct CMA freelist.
129 static inline int get_pcppage_migratetype(struct page
*page
)
134 static inline void set_pcppage_migratetype(struct page
*page
, int migratetype
)
136 page
->index
= migratetype
;
139 #ifdef CONFIG_PM_SLEEP
141 * The following functions are used by the suspend/hibernate code to temporarily
142 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
143 * while devices are suspended. To avoid races with the suspend/hibernate code,
144 * they should always be called with pm_mutex held (gfp_allowed_mask also should
145 * only be modified with pm_mutex held, unless the suspend/hibernate code is
146 * guaranteed not to run in parallel with that modification).
149 static gfp_t saved_gfp_mask
;
151 void pm_restore_gfp_mask(void)
153 WARN_ON(!mutex_is_locked(&pm_mutex
));
154 if (saved_gfp_mask
) {
155 gfp_allowed_mask
= saved_gfp_mask
;
160 void pm_restrict_gfp_mask(void)
162 WARN_ON(!mutex_is_locked(&pm_mutex
));
163 WARN_ON(saved_gfp_mask
);
164 saved_gfp_mask
= gfp_allowed_mask
;
165 gfp_allowed_mask
&= ~(__GFP_IO
| __GFP_FS
);
168 bool pm_suspended_storage(void)
170 if ((gfp_allowed_mask
& (__GFP_IO
| __GFP_FS
)) == (__GFP_IO
| __GFP_FS
))
174 #endif /* CONFIG_PM_SLEEP */
176 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
177 unsigned int pageblock_order __read_mostly
;
180 static void __free_pages_ok(struct page
*page
, unsigned int order
);
183 * results with 256, 32 in the lowmem_reserve sysctl:
184 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
185 * 1G machine -> (16M dma, 784M normal, 224M high)
186 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
187 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
188 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
190 * TBD: should special case ZONE_DMA32 machines here - in those we normally
191 * don't need any ZONE_NORMAL reservation
193 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
200 #ifdef CONFIG_HIGHMEM
206 EXPORT_SYMBOL(totalram_pages
);
208 static char * const zone_names
[MAX_NR_ZONES
] = {
209 #ifdef CONFIG_ZONE_DMA
212 #ifdef CONFIG_ZONE_DMA32
216 #ifdef CONFIG_HIGHMEM
220 #ifdef CONFIG_ZONE_DEVICE
225 static void free_compound_page(struct page
*page
);
226 compound_page_dtor
* const compound_page_dtors
[] = {
229 #ifdef CONFIG_HUGETLB_PAGE
234 int min_free_kbytes
= 1024;
235 int user_min_free_kbytes
= -1;
237 static unsigned long __meminitdata nr_kernel_pages
;
238 static unsigned long __meminitdata nr_all_pages
;
239 static unsigned long __meminitdata dma_reserve
;
241 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
242 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
243 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
244 static unsigned long __initdata required_kernelcore
;
245 static unsigned long __initdata required_movablecore
;
246 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
248 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
250 EXPORT_SYMBOL(movable_zone
);
251 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
254 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
255 int nr_online_nodes __read_mostly
= 1;
256 EXPORT_SYMBOL(nr_node_ids
);
257 EXPORT_SYMBOL(nr_online_nodes
);
260 int page_group_by_mobility_disabled __read_mostly
;
262 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
263 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
265 pgdat
->first_deferred_pfn
= ULONG_MAX
;
268 /* Returns true if the struct page for the pfn is uninitialised */
269 static inline bool __meminit
early_page_uninitialised(unsigned long pfn
)
271 if (pfn
>= NODE_DATA(early_pfn_to_nid(pfn
))->first_deferred_pfn
)
277 static inline bool early_page_nid_uninitialised(unsigned long pfn
, int nid
)
279 if (pfn
>= NODE_DATA(nid
)->first_deferred_pfn
)
286 * Returns false when the remaining initialisation should be deferred until
287 * later in the boot cycle when it can be parallelised.
289 static inline bool update_defer_init(pg_data_t
*pgdat
,
290 unsigned long pfn
, unsigned long zone_end
,
291 unsigned long *nr_initialised
)
293 /* Always populate low zones for address-contrained allocations */
294 if (zone_end
< pgdat_end_pfn(pgdat
))
297 /* Initialise at least 2G of the highest zone */
299 if (*nr_initialised
> (2UL << (30 - PAGE_SHIFT
)) &&
300 (pfn
& (PAGES_PER_SECTION
- 1)) == 0) {
301 pgdat
->first_deferred_pfn
= pfn
;
308 static inline void reset_deferred_meminit(pg_data_t
*pgdat
)
312 static inline bool early_page_uninitialised(unsigned long pfn
)
317 static inline bool early_page_nid_uninitialised(unsigned long pfn
, int nid
)
322 static inline bool update_defer_init(pg_data_t
*pgdat
,
323 unsigned long pfn
, unsigned long zone_end
,
324 unsigned long *nr_initialised
)
331 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
333 if (unlikely(page_group_by_mobility_disabled
&&
334 migratetype
< MIGRATE_PCPTYPES
))
335 migratetype
= MIGRATE_UNMOVABLE
;
337 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
338 PB_migrate
, PB_migrate_end
);
341 #ifdef CONFIG_DEBUG_VM
342 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
346 unsigned long pfn
= page_to_pfn(page
);
347 unsigned long sp
, start_pfn
;
350 seq
= zone_span_seqbegin(zone
);
351 start_pfn
= zone
->zone_start_pfn
;
352 sp
= zone
->spanned_pages
;
353 if (!zone_spans_pfn(zone
, pfn
))
355 } while (zone_span_seqretry(zone
, seq
));
358 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
359 pfn
, zone_to_nid(zone
), zone
->name
,
360 start_pfn
, start_pfn
+ sp
);
365 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
367 if (!pfn_valid_within(page_to_pfn(page
)))
369 if (zone
!= page_zone(page
))
375 * Temporary debugging check for pages not lying within a given zone.
377 static int bad_range(struct zone
*zone
, struct page
*page
)
379 if (page_outside_zone_boundaries(zone
, page
))
381 if (!page_is_consistent(zone
, page
))
387 static inline int bad_range(struct zone
*zone
, struct page
*page
)
393 static void bad_page(struct page
*page
, const char *reason
,
394 unsigned long bad_flags
)
396 static unsigned long resume
;
397 static unsigned long nr_shown
;
398 static unsigned long nr_unshown
;
400 /* Don't complain about poisoned pages */
401 if (PageHWPoison(page
)) {
402 page_mapcount_reset(page
); /* remove PageBuddy */
407 * Allow a burst of 60 reports, then keep quiet for that minute;
408 * or allow a steady drip of one report per second.
410 if (nr_shown
== 60) {
411 if (time_before(jiffies
, resume
)) {
417 "BUG: Bad page state: %lu messages suppressed\n",
424 resume
= jiffies
+ 60 * HZ
;
426 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
427 current
->comm
, page_to_pfn(page
));
428 dump_page_badflags(page
, reason
, bad_flags
);
433 /* Leave bad fields for debug, except PageBuddy could make trouble */
434 page_mapcount_reset(page
); /* remove PageBuddy */
435 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
439 * Higher-order pages are called "compound pages". They are structured thusly:
441 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
443 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
444 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
446 * The first tail page's ->compound_dtor holds the offset in array of compound
447 * page destructors. See compound_page_dtors.
449 * The first tail page's ->compound_order holds the order of allocation.
450 * This usage means that zero-order pages may not be compound.
453 static void free_compound_page(struct page
*page
)
455 __free_pages_ok(page
, compound_order(page
));
458 void prep_compound_page(struct page
*page
, unsigned int order
)
461 int nr_pages
= 1 << order
;
463 set_compound_page_dtor(page
, COMPOUND_PAGE_DTOR
);
464 set_compound_order(page
, order
);
466 for (i
= 1; i
< nr_pages
; i
++) {
467 struct page
*p
= page
+ i
;
468 set_page_count(p
, 0);
469 set_compound_head(p
, page
);
473 #ifdef CONFIG_DEBUG_PAGEALLOC
474 unsigned int _debug_guardpage_minorder
;
475 bool _debug_pagealloc_enabled __read_mostly
;
476 bool _debug_guardpage_enabled __read_mostly
;
478 static int __init
early_debug_pagealloc(char *buf
)
483 if (strcmp(buf
, "on") == 0)
484 _debug_pagealloc_enabled
= true;
488 early_param("debug_pagealloc", early_debug_pagealloc
);
490 static bool need_debug_guardpage(void)
492 /* If we don't use debug_pagealloc, we don't need guard page */
493 if (!debug_pagealloc_enabled())
499 static void init_debug_guardpage(void)
501 if (!debug_pagealloc_enabled())
504 _debug_guardpage_enabled
= true;
507 struct page_ext_operations debug_guardpage_ops
= {
508 .need
= need_debug_guardpage
,
509 .init
= init_debug_guardpage
,
512 static int __init
debug_guardpage_minorder_setup(char *buf
)
516 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
517 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
520 _debug_guardpage_minorder
= res
;
521 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
524 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
526 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
527 unsigned int order
, int migratetype
)
529 struct page_ext
*page_ext
;
531 if (!debug_guardpage_enabled())
534 page_ext
= lookup_page_ext(page
);
535 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
537 INIT_LIST_HEAD(&page
->lru
);
538 set_page_private(page
, order
);
539 /* Guard pages are not available for any usage */
540 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
543 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
544 unsigned int order
, int migratetype
)
546 struct page_ext
*page_ext
;
548 if (!debug_guardpage_enabled())
551 page_ext
= lookup_page_ext(page
);
552 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
554 set_page_private(page
, 0);
555 if (!is_migrate_isolate(migratetype
))
556 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
559 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
560 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
561 unsigned int order
, int migratetype
) {}
562 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
563 unsigned int order
, int migratetype
) {}
566 static inline void set_page_order(struct page
*page
, unsigned int order
)
568 set_page_private(page
, order
);
569 __SetPageBuddy(page
);
572 static inline void rmv_page_order(struct page
*page
)
574 __ClearPageBuddy(page
);
575 set_page_private(page
, 0);
579 * This function checks whether a page is free && is the buddy
580 * we can do coalesce a page and its buddy if
581 * (a) the buddy is not in a hole &&
582 * (b) the buddy is in the buddy system &&
583 * (c) a page and its buddy have the same order &&
584 * (d) a page and its buddy are in the same zone.
586 * For recording whether a page is in the buddy system, we set ->_mapcount
587 * PAGE_BUDDY_MAPCOUNT_VALUE.
588 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
589 * serialized by zone->lock.
591 * For recording page's order, we use page_private(page).
593 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
596 if (!pfn_valid_within(page_to_pfn(buddy
)))
599 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
600 if (page_zone_id(page
) != page_zone_id(buddy
))
603 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
608 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
610 * zone check is done late to avoid uselessly
611 * calculating zone/node ids for pages that could
614 if (page_zone_id(page
) != page_zone_id(buddy
))
617 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
625 * Freeing function for a buddy system allocator.
627 * The concept of a buddy system is to maintain direct-mapped table
628 * (containing bit values) for memory blocks of various "orders".
629 * The bottom level table contains the map for the smallest allocatable
630 * units of memory (here, pages), and each level above it describes
631 * pairs of units from the levels below, hence, "buddies".
632 * At a high level, all that happens here is marking the table entry
633 * at the bottom level available, and propagating the changes upward
634 * as necessary, plus some accounting needed to play nicely with other
635 * parts of the VM system.
636 * At each level, we keep a list of pages, which are heads of continuous
637 * free pages of length of (1 << order) and marked with _mapcount
638 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
640 * So when we are allocating or freeing one, we can derive the state of the
641 * other. That is, if we allocate a small block, and both were
642 * free, the remainder of the region must be split into blocks.
643 * If a block is freed, and its buddy is also free, then this
644 * triggers coalescing into a block of larger size.
649 static inline void __free_one_page(struct page
*page
,
651 struct zone
*zone
, unsigned int order
,
654 unsigned long page_idx
;
655 unsigned long combined_idx
;
656 unsigned long uninitialized_var(buddy_idx
);
658 unsigned int max_order
= MAX_ORDER
;
660 VM_BUG_ON(!zone_is_initialized(zone
));
661 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
663 VM_BUG_ON(migratetype
== -1);
664 if (is_migrate_isolate(migratetype
)) {
666 * We restrict max order of merging to prevent merge
667 * between freepages on isolate pageblock and normal
668 * pageblock. Without this, pageblock isolation
669 * could cause incorrect freepage accounting.
671 max_order
= min_t(unsigned int, MAX_ORDER
, pageblock_order
+ 1);
673 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
676 page_idx
= pfn
& ((1 << max_order
) - 1);
678 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
679 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
681 while (order
< max_order
- 1) {
682 buddy_idx
= __find_buddy_index(page_idx
, order
);
683 buddy
= page
+ (buddy_idx
- page_idx
);
684 if (!page_is_buddy(page
, buddy
, order
))
687 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
688 * merge with it and move up one order.
690 if (page_is_guard(buddy
)) {
691 clear_page_guard(zone
, buddy
, order
, migratetype
);
693 list_del(&buddy
->lru
);
694 zone
->free_area
[order
].nr_free
--;
695 rmv_page_order(buddy
);
697 combined_idx
= buddy_idx
& page_idx
;
698 page
= page
+ (combined_idx
- page_idx
);
699 page_idx
= combined_idx
;
702 set_page_order(page
, order
);
705 * If this is not the largest possible page, check if the buddy
706 * of the next-highest order is free. If it is, it's possible
707 * that pages are being freed that will coalesce soon. In case,
708 * that is happening, add the free page to the tail of the list
709 * so it's less likely to be used soon and more likely to be merged
710 * as a higher order page
712 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
713 struct page
*higher_page
, *higher_buddy
;
714 combined_idx
= buddy_idx
& page_idx
;
715 higher_page
= page
+ (combined_idx
- page_idx
);
716 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
717 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
718 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
719 list_add_tail(&page
->lru
,
720 &zone
->free_area
[order
].free_list
[migratetype
]);
725 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
727 zone
->free_area
[order
].nr_free
++;
730 static inline int free_pages_check(struct page
*page
)
732 const char *bad_reason
= NULL
;
733 unsigned long bad_flags
= 0;
735 if (unlikely(page_mapcount(page
)))
736 bad_reason
= "nonzero mapcount";
737 if (unlikely(page
->mapping
!= NULL
))
738 bad_reason
= "non-NULL mapping";
739 if (unlikely(atomic_read(&page
->_count
) != 0))
740 bad_reason
= "nonzero _count";
741 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
742 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
743 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
746 if (unlikely(page
->mem_cgroup
))
747 bad_reason
= "page still charged to cgroup";
749 if (unlikely(bad_reason
)) {
750 bad_page(page
, bad_reason
, bad_flags
);
753 page_cpupid_reset_last(page
);
754 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
755 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
760 * Frees a number of pages from the PCP lists
761 * Assumes all pages on list are in same zone, and of same order.
762 * count is the number of pages to free.
764 * If the zone was previously in an "all pages pinned" state then look to
765 * see if this freeing clears that state.
767 * And clear the zone's pages_scanned counter, to hold off the "all pages are
768 * pinned" detection logic.
770 static void free_pcppages_bulk(struct zone
*zone
, int count
,
771 struct per_cpu_pages
*pcp
)
776 unsigned long nr_scanned
;
778 spin_lock(&zone
->lock
);
779 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
781 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
785 struct list_head
*list
;
788 * Remove pages from lists in a round-robin fashion. A
789 * batch_free count is maintained that is incremented when an
790 * empty list is encountered. This is so more pages are freed
791 * off fuller lists instead of spinning excessively around empty
796 if (++migratetype
== MIGRATE_PCPTYPES
)
798 list
= &pcp
->lists
[migratetype
];
799 } while (list_empty(list
));
801 /* This is the only non-empty list. Free them all. */
802 if (batch_free
== MIGRATE_PCPTYPES
)
803 batch_free
= to_free
;
806 int mt
; /* migratetype of the to-be-freed page */
808 page
= list_last_entry(list
, struct page
, lru
);
809 /* must delete as __free_one_page list manipulates */
810 list_del(&page
->lru
);
812 mt
= get_pcppage_migratetype(page
);
813 /* MIGRATE_ISOLATE page should not go to pcplists */
814 VM_BUG_ON_PAGE(is_migrate_isolate(mt
), page
);
815 /* Pageblock could have been isolated meanwhile */
816 if (unlikely(has_isolate_pageblock(zone
)))
817 mt
= get_pageblock_migratetype(page
);
819 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
820 trace_mm_page_pcpu_drain(page
, 0, mt
);
821 } while (--to_free
&& --batch_free
&& !list_empty(list
));
823 spin_unlock(&zone
->lock
);
826 static void free_one_page(struct zone
*zone
,
827 struct page
*page
, unsigned long pfn
,
831 unsigned long nr_scanned
;
832 spin_lock(&zone
->lock
);
833 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
835 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
837 if (unlikely(has_isolate_pageblock(zone
) ||
838 is_migrate_isolate(migratetype
))) {
839 migratetype
= get_pfnblock_migratetype(page
, pfn
);
841 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
842 spin_unlock(&zone
->lock
);
845 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
850 * We rely page->lru.next never has bit 0 set, unless the page
851 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
853 BUILD_BUG_ON((unsigned long)LIST_POISON1
& 1);
855 if (!IS_ENABLED(CONFIG_DEBUG_VM
)) {
859 if (unlikely(!PageTail(page
))) {
860 bad_page(page
, "PageTail not set", 0);
863 if (unlikely(compound_head(page
) != head_page
)) {
864 bad_page(page
, "compound_head not consistent", 0);
869 clear_compound_head(page
);
873 static void __meminit
__init_single_page(struct page
*page
, unsigned long pfn
,
874 unsigned long zone
, int nid
)
876 set_page_links(page
, zone
, nid
, pfn
);
877 init_page_count(page
);
878 page_mapcount_reset(page
);
879 page_cpupid_reset_last(page
);
881 INIT_LIST_HEAD(&page
->lru
);
882 #ifdef WANT_PAGE_VIRTUAL
883 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
884 if (!is_highmem_idx(zone
))
885 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
889 static void __meminit
__init_single_pfn(unsigned long pfn
, unsigned long zone
,
892 return __init_single_page(pfn_to_page(pfn
), pfn
, zone
, nid
);
895 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
896 static void init_reserved_page(unsigned long pfn
)
901 if (!early_page_uninitialised(pfn
))
904 nid
= early_pfn_to_nid(pfn
);
905 pgdat
= NODE_DATA(nid
);
907 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
908 struct zone
*zone
= &pgdat
->node_zones
[zid
];
910 if (pfn
>= zone
->zone_start_pfn
&& pfn
< zone_end_pfn(zone
))
913 __init_single_pfn(pfn
, zid
, nid
);
916 static inline void init_reserved_page(unsigned long pfn
)
919 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
922 * Initialised pages do not have PageReserved set. This function is
923 * called for each range allocated by the bootmem allocator and
924 * marks the pages PageReserved. The remaining valid pages are later
925 * sent to the buddy page allocator.
927 void __meminit
reserve_bootmem_region(unsigned long start
, unsigned long end
)
929 unsigned long start_pfn
= PFN_DOWN(start
);
930 unsigned long end_pfn
= PFN_UP(end
);
932 for (; start_pfn
< end_pfn
; start_pfn
++) {
933 if (pfn_valid(start_pfn
)) {
934 struct page
*page
= pfn_to_page(start_pfn
);
936 init_reserved_page(start_pfn
);
938 /* Avoid false-positive PageTail() */
939 INIT_LIST_HEAD(&page
->lru
);
941 SetPageReserved(page
);
946 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
948 bool compound
= PageCompound(page
);
951 VM_BUG_ON_PAGE(PageTail(page
), page
);
952 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
954 trace_mm_page_free(page
, order
);
955 kmemcheck_free_shadow(page
, order
);
956 kasan_free_pages(page
, order
);
959 page
->mapping
= NULL
;
960 bad
+= free_pages_check(page
);
961 for (i
= 1; i
< (1 << order
); i
++) {
963 bad
+= free_tail_pages_check(page
, page
+ i
);
964 bad
+= free_pages_check(page
+ i
);
969 reset_page_owner(page
, order
);
971 if (!PageHighMem(page
)) {
972 debug_check_no_locks_freed(page_address(page
),
974 debug_check_no_obj_freed(page_address(page
),
977 arch_free_page(page
, order
);
978 kernel_map_pages(page
, 1 << order
, 0);
983 static void __free_pages_ok(struct page
*page
, unsigned int order
)
987 unsigned long pfn
= page_to_pfn(page
);
989 if (!free_pages_prepare(page
, order
))
992 migratetype
= get_pfnblock_migratetype(page
, pfn
);
993 local_irq_save(flags
);
994 __count_vm_events(PGFREE
, 1 << order
);
995 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
996 local_irq_restore(flags
);
999 static void __init
__free_pages_boot_core(struct page
*page
,
1000 unsigned long pfn
, unsigned int order
)
1002 unsigned int nr_pages
= 1 << order
;
1003 struct page
*p
= page
;
1007 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
1009 __ClearPageReserved(p
);
1010 set_page_count(p
, 0);
1012 __ClearPageReserved(p
);
1013 set_page_count(p
, 0);
1015 page_zone(page
)->managed_pages
+= nr_pages
;
1016 set_page_refcounted(page
);
1017 __free_pages(page
, order
);
1020 #if defined(CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID) || \
1021 defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP)
1023 static struct mminit_pfnnid_cache early_pfnnid_cache __meminitdata
;
1025 int __meminit
early_pfn_to_nid(unsigned long pfn
)
1027 static DEFINE_SPINLOCK(early_pfn_lock
);
1030 spin_lock(&early_pfn_lock
);
1031 nid
= __early_pfn_to_nid(pfn
, &early_pfnnid_cache
);
1034 spin_unlock(&early_pfn_lock
);
1040 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
1041 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1042 struct mminit_pfnnid_cache
*state
)
1046 nid
= __early_pfn_to_nid(pfn
, state
);
1047 if (nid
>= 0 && nid
!= node
)
1052 /* Only safe to use early in boot when initialisation is single-threaded */
1053 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1055 return meminit_pfn_in_nid(pfn
, node
, &early_pfnnid_cache
);
1060 static inline bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
1064 static inline bool __meminit
meminit_pfn_in_nid(unsigned long pfn
, int node
,
1065 struct mminit_pfnnid_cache
*state
)
1072 void __init
__free_pages_bootmem(struct page
*page
, unsigned long pfn
,
1075 if (early_page_uninitialised(pfn
))
1077 return __free_pages_boot_core(page
, pfn
, order
);
1080 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
1081 static void __init
deferred_free_range(struct page
*page
,
1082 unsigned long pfn
, int nr_pages
)
1089 /* Free a large naturally-aligned chunk if possible */
1090 if (nr_pages
== MAX_ORDER_NR_PAGES
&&
1091 (pfn
& (MAX_ORDER_NR_PAGES
-1)) == 0) {
1092 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1093 __free_pages_boot_core(page
, pfn
, MAX_ORDER
-1);
1097 for (i
= 0; i
< nr_pages
; i
++, page
++, pfn
++)
1098 __free_pages_boot_core(page
, pfn
, 0);
1101 /* Completion tracking for deferred_init_memmap() threads */
1102 static atomic_t pgdat_init_n_undone __initdata
;
1103 static __initdata
DECLARE_COMPLETION(pgdat_init_all_done_comp
);
1105 static inline void __init
pgdat_init_report_one_done(void)
1107 if (atomic_dec_and_test(&pgdat_init_n_undone
))
1108 complete(&pgdat_init_all_done_comp
);
1111 /* Initialise remaining memory on a node */
1112 static int __init
deferred_init_memmap(void *data
)
1114 pg_data_t
*pgdat
= data
;
1115 int nid
= pgdat
->node_id
;
1116 struct mminit_pfnnid_cache nid_init_state
= { };
1117 unsigned long start
= jiffies
;
1118 unsigned long nr_pages
= 0;
1119 unsigned long walk_start
, walk_end
;
1122 unsigned long first_init_pfn
= pgdat
->first_deferred_pfn
;
1123 const struct cpumask
*cpumask
= cpumask_of_node(pgdat
->node_id
);
1125 if (first_init_pfn
== ULONG_MAX
) {
1126 pgdat_init_report_one_done();
1130 /* Bind memory initialisation thread to a local node if possible */
1131 if (!cpumask_empty(cpumask
))
1132 set_cpus_allowed_ptr(current
, cpumask
);
1134 /* Sanity check boundaries */
1135 BUG_ON(pgdat
->first_deferred_pfn
< pgdat
->node_start_pfn
);
1136 BUG_ON(pgdat
->first_deferred_pfn
> pgdat_end_pfn(pgdat
));
1137 pgdat
->first_deferred_pfn
= ULONG_MAX
;
1139 /* Only the highest zone is deferred so find it */
1140 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
1141 zone
= pgdat
->node_zones
+ zid
;
1142 if (first_init_pfn
< zone_end_pfn(zone
))
1146 for_each_mem_pfn_range(i
, nid
, &walk_start
, &walk_end
, NULL
) {
1147 unsigned long pfn
, end_pfn
;
1148 struct page
*page
= NULL
;
1149 struct page
*free_base_page
= NULL
;
1150 unsigned long free_base_pfn
= 0;
1153 end_pfn
= min(walk_end
, zone_end_pfn(zone
));
1154 pfn
= first_init_pfn
;
1155 if (pfn
< walk_start
)
1157 if (pfn
< zone
->zone_start_pfn
)
1158 pfn
= zone
->zone_start_pfn
;
1160 for (; pfn
< end_pfn
; pfn
++) {
1161 if (!pfn_valid_within(pfn
))
1165 * Ensure pfn_valid is checked every
1166 * MAX_ORDER_NR_PAGES for memory holes
1168 if ((pfn
& (MAX_ORDER_NR_PAGES
- 1)) == 0) {
1169 if (!pfn_valid(pfn
)) {
1175 if (!meminit_pfn_in_nid(pfn
, nid
, &nid_init_state
)) {
1180 /* Minimise pfn page lookups and scheduler checks */
1181 if (page
&& (pfn
& (MAX_ORDER_NR_PAGES
- 1)) != 0) {
1184 nr_pages
+= nr_to_free
;
1185 deferred_free_range(free_base_page
,
1186 free_base_pfn
, nr_to_free
);
1187 free_base_page
= NULL
;
1188 free_base_pfn
= nr_to_free
= 0;
1190 page
= pfn_to_page(pfn
);
1195 VM_BUG_ON(page_zone(page
) != zone
);
1199 __init_single_page(page
, pfn
, zid
, nid
);
1200 if (!free_base_page
) {
1201 free_base_page
= page
;
1202 free_base_pfn
= pfn
;
1207 /* Where possible, batch up pages for a single free */
1210 /* Free the current block of pages to allocator */
1211 nr_pages
+= nr_to_free
;
1212 deferred_free_range(free_base_page
, free_base_pfn
,
1214 free_base_page
= NULL
;
1215 free_base_pfn
= nr_to_free
= 0;
1218 first_init_pfn
= max(end_pfn
, first_init_pfn
);
1221 /* Sanity check that the next zone really is unpopulated */
1222 WARN_ON(++zid
< MAX_NR_ZONES
&& populated_zone(++zone
));
1224 pr_info("node %d initialised, %lu pages in %ums\n", nid
, nr_pages
,
1225 jiffies_to_msecs(jiffies
- start
));
1227 pgdat_init_report_one_done();
1231 void __init
page_alloc_init_late(void)
1235 /* There will be num_node_state(N_MEMORY) threads */
1236 atomic_set(&pgdat_init_n_undone
, num_node_state(N_MEMORY
));
1237 for_each_node_state(nid
, N_MEMORY
) {
1238 kthread_run(deferred_init_memmap
, NODE_DATA(nid
), "pgdatinit%d", nid
);
1241 /* Block until all are initialised */
1242 wait_for_completion(&pgdat_init_all_done_comp
);
1244 /* Reinit limits that are based on free pages after the kernel is up */
1245 files_maxfiles_init();
1247 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
1250 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
1251 void __init
init_cma_reserved_pageblock(struct page
*page
)
1253 unsigned i
= pageblock_nr_pages
;
1254 struct page
*p
= page
;
1257 __ClearPageReserved(p
);
1258 set_page_count(p
, 0);
1261 set_pageblock_migratetype(page
, MIGRATE_CMA
);
1263 if (pageblock_order
>= MAX_ORDER
) {
1264 i
= pageblock_nr_pages
;
1267 set_page_refcounted(p
);
1268 __free_pages(p
, MAX_ORDER
- 1);
1269 p
+= MAX_ORDER_NR_PAGES
;
1270 } while (i
-= MAX_ORDER_NR_PAGES
);
1272 set_page_refcounted(page
);
1273 __free_pages(page
, pageblock_order
);
1276 adjust_managed_page_count(page
, pageblock_nr_pages
);
1281 * The order of subdivision here is critical for the IO subsystem.
1282 * Please do not alter this order without good reasons and regression
1283 * testing. Specifically, as large blocks of memory are subdivided,
1284 * the order in which smaller blocks are delivered depends on the order
1285 * they're subdivided in this function. This is the primary factor
1286 * influencing the order in which pages are delivered to the IO
1287 * subsystem according to empirical testing, and this is also justified
1288 * by considering the behavior of a buddy system containing a single
1289 * large block of memory acted on by a series of small allocations.
1290 * This behavior is a critical factor in sglist merging's success.
1294 static inline void expand(struct zone
*zone
, struct page
*page
,
1295 int low
, int high
, struct free_area
*area
,
1298 unsigned long size
= 1 << high
;
1300 while (high
> low
) {
1304 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
1306 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
1307 debug_guardpage_enabled() &&
1308 high
< debug_guardpage_minorder()) {
1310 * Mark as guard pages (or page), that will allow to
1311 * merge back to allocator when buddy will be freed.
1312 * Corresponding page table entries will not be touched,
1313 * pages will stay not present in virtual address space
1315 set_page_guard(zone
, &page
[size
], high
, migratetype
);
1318 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
1320 set_page_order(&page
[size
], high
);
1325 * This page is about to be returned from the page allocator
1327 static inline int check_new_page(struct page
*page
)
1329 const char *bad_reason
= NULL
;
1330 unsigned long bad_flags
= 0;
1332 if (unlikely(page_mapcount(page
)))
1333 bad_reason
= "nonzero mapcount";
1334 if (unlikely(page
->mapping
!= NULL
))
1335 bad_reason
= "non-NULL mapping";
1336 if (unlikely(atomic_read(&page
->_count
) != 0))
1337 bad_reason
= "nonzero _count";
1338 if (unlikely(page
->flags
& __PG_HWPOISON
)) {
1339 bad_reason
= "HWPoisoned (hardware-corrupted)";
1340 bad_flags
= __PG_HWPOISON
;
1342 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
1343 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
1344 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
1347 if (unlikely(page
->mem_cgroup
))
1348 bad_reason
= "page still charged to cgroup";
1350 if (unlikely(bad_reason
)) {
1351 bad_page(page
, bad_reason
, bad_flags
);
1357 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
1362 for (i
= 0; i
< (1 << order
); i
++) {
1363 struct page
*p
= page
+ i
;
1364 if (unlikely(check_new_page(p
)))
1368 set_page_private(page
, 0);
1369 set_page_refcounted(page
);
1371 arch_alloc_page(page
, order
);
1372 kernel_map_pages(page
, 1 << order
, 1);
1373 kasan_alloc_pages(page
, order
);
1375 if (gfp_flags
& __GFP_ZERO
)
1376 for (i
= 0; i
< (1 << order
); i
++)
1377 clear_highpage(page
+ i
);
1379 if (order
&& (gfp_flags
& __GFP_COMP
))
1380 prep_compound_page(page
, order
);
1382 set_page_owner(page
, order
, gfp_flags
);
1385 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1386 * allocate the page. The expectation is that the caller is taking
1387 * steps that will free more memory. The caller should avoid the page
1388 * being used for !PFMEMALLOC purposes.
1390 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
1391 set_page_pfmemalloc(page
);
1393 clear_page_pfmemalloc(page
);
1399 * Go through the free lists for the given migratetype and remove
1400 * the smallest available page from the freelists
1403 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1406 unsigned int current_order
;
1407 struct free_area
*area
;
1410 /* Find a page of the appropriate size in the preferred list */
1411 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1412 area
= &(zone
->free_area
[current_order
]);
1413 page
= list_first_entry_or_null(&area
->free_list
[migratetype
],
1417 list_del(&page
->lru
);
1418 rmv_page_order(page
);
1420 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1421 set_pcppage_migratetype(page
, migratetype
);
1430 * This array describes the order lists are fallen back to when
1431 * the free lists for the desirable migrate type are depleted
1433 static int fallbacks
[MIGRATE_TYPES
][4] = {
1434 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1435 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_TYPES
},
1436 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_TYPES
},
1438 [MIGRATE_CMA
] = { MIGRATE_TYPES
}, /* Never used */
1440 #ifdef CONFIG_MEMORY_ISOLATION
1441 [MIGRATE_ISOLATE
] = { MIGRATE_TYPES
}, /* Never used */
1446 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1449 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1452 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1453 unsigned int order
) { return NULL
; }
1457 * Move the free pages in a range to the free lists of the requested type.
1458 * Note that start_page and end_pages are not aligned on a pageblock
1459 * boundary. If alignment is required, use move_freepages_block()
1461 int move_freepages(struct zone
*zone
,
1462 struct page
*start_page
, struct page
*end_page
,
1467 int pages_moved
= 0;
1469 #ifndef CONFIG_HOLES_IN_ZONE
1471 * page_zone is not safe to call in this context when
1472 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1473 * anyway as we check zone boundaries in move_freepages_block().
1474 * Remove at a later date when no bug reports exist related to
1475 * grouping pages by mobility
1477 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1480 for (page
= start_page
; page
<= end_page
;) {
1481 /* Make sure we are not inadvertently changing nodes */
1482 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1484 if (!pfn_valid_within(page_to_pfn(page
))) {
1489 if (!PageBuddy(page
)) {
1494 order
= page_order(page
);
1495 list_move(&page
->lru
,
1496 &zone
->free_area
[order
].free_list
[migratetype
]);
1498 pages_moved
+= 1 << order
;
1504 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1507 unsigned long start_pfn
, end_pfn
;
1508 struct page
*start_page
, *end_page
;
1510 start_pfn
= page_to_pfn(page
);
1511 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1512 start_page
= pfn_to_page(start_pfn
);
1513 end_page
= start_page
+ pageblock_nr_pages
- 1;
1514 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1516 /* Do not cross zone boundaries */
1517 if (!zone_spans_pfn(zone
, start_pfn
))
1519 if (!zone_spans_pfn(zone
, end_pfn
))
1522 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1525 static void change_pageblock_range(struct page
*pageblock_page
,
1526 int start_order
, int migratetype
)
1528 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1530 while (nr_pageblocks
--) {
1531 set_pageblock_migratetype(pageblock_page
, migratetype
);
1532 pageblock_page
+= pageblock_nr_pages
;
1537 * When we are falling back to another migratetype during allocation, try to
1538 * steal extra free pages from the same pageblocks to satisfy further
1539 * allocations, instead of polluting multiple pageblocks.
1541 * If we are stealing a relatively large buddy page, it is likely there will
1542 * be more free pages in the pageblock, so try to steal them all. For
1543 * reclaimable and unmovable allocations, we steal regardless of page size,
1544 * as fragmentation caused by those allocations polluting movable pageblocks
1545 * is worse than movable allocations stealing from unmovable and reclaimable
1548 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1551 * Leaving this order check is intended, although there is
1552 * relaxed order check in next check. The reason is that
1553 * we can actually steal whole pageblock if this condition met,
1554 * but, below check doesn't guarantee it and that is just heuristic
1555 * so could be changed anytime.
1557 if (order
>= pageblock_order
)
1560 if (order
>= pageblock_order
/ 2 ||
1561 start_mt
== MIGRATE_RECLAIMABLE
||
1562 start_mt
== MIGRATE_UNMOVABLE
||
1563 page_group_by_mobility_disabled
)
1570 * This function implements actual steal behaviour. If order is large enough,
1571 * we can steal whole pageblock. If not, we first move freepages in this
1572 * pageblock and check whether half of pages are moved or not. If half of
1573 * pages are moved, we can change migratetype of pageblock and permanently
1574 * use it's pages as requested migratetype in the future.
1576 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1579 unsigned int current_order
= page_order(page
);
1582 /* Take ownership for orders >= pageblock_order */
1583 if (current_order
>= pageblock_order
) {
1584 change_pageblock_range(page
, current_order
, start_type
);
1588 pages
= move_freepages_block(zone
, page
, start_type
);
1590 /* Claim the whole block if over half of it is free */
1591 if (pages
>= (1 << (pageblock_order
-1)) ||
1592 page_group_by_mobility_disabled
)
1593 set_pageblock_migratetype(page
, start_type
);
1597 * Check whether there is a suitable fallback freepage with requested order.
1598 * If only_stealable is true, this function returns fallback_mt only if
1599 * we can steal other freepages all together. This would help to reduce
1600 * fragmentation due to mixed migratetype pages in one pageblock.
1602 int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1603 int migratetype
, bool only_stealable
, bool *can_steal
)
1608 if (area
->nr_free
== 0)
1613 fallback_mt
= fallbacks
[migratetype
][i
];
1614 if (fallback_mt
== MIGRATE_TYPES
)
1617 if (list_empty(&area
->free_list
[fallback_mt
]))
1620 if (can_steal_fallback(order
, migratetype
))
1623 if (!only_stealable
)
1634 * Reserve a pageblock for exclusive use of high-order atomic allocations if
1635 * there are no empty page blocks that contain a page with a suitable order
1637 static void reserve_highatomic_pageblock(struct page
*page
, struct zone
*zone
,
1638 unsigned int alloc_order
)
1641 unsigned long max_managed
, flags
;
1644 * Limit the number reserved to 1 pageblock or roughly 1% of a zone.
1645 * Check is race-prone but harmless.
1647 max_managed
= (zone
->managed_pages
/ 100) + pageblock_nr_pages
;
1648 if (zone
->nr_reserved_highatomic
>= max_managed
)
1651 spin_lock_irqsave(&zone
->lock
, flags
);
1653 /* Recheck the nr_reserved_highatomic limit under the lock */
1654 if (zone
->nr_reserved_highatomic
>= max_managed
)
1658 mt
= get_pageblock_migratetype(page
);
1659 if (mt
!= MIGRATE_HIGHATOMIC
&&
1660 !is_migrate_isolate(mt
) && !is_migrate_cma(mt
)) {
1661 zone
->nr_reserved_highatomic
+= pageblock_nr_pages
;
1662 set_pageblock_migratetype(page
, MIGRATE_HIGHATOMIC
);
1663 move_freepages_block(zone
, page
, MIGRATE_HIGHATOMIC
);
1667 spin_unlock_irqrestore(&zone
->lock
, flags
);
1671 * Used when an allocation is about to fail under memory pressure. This
1672 * potentially hurts the reliability of high-order allocations when under
1673 * intense memory pressure but failed atomic allocations should be easier
1674 * to recover from than an OOM.
1676 static void unreserve_highatomic_pageblock(const struct alloc_context
*ac
)
1678 struct zonelist
*zonelist
= ac
->zonelist
;
1679 unsigned long flags
;
1685 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
1687 /* Preserve at least one pageblock */
1688 if (zone
->nr_reserved_highatomic
<= pageblock_nr_pages
)
1691 spin_lock_irqsave(&zone
->lock
, flags
);
1692 for (order
= 0; order
< MAX_ORDER
; order
++) {
1693 struct free_area
*area
= &(zone
->free_area
[order
]);
1695 page
= list_first_entry_or_null(
1696 &area
->free_list
[MIGRATE_HIGHATOMIC
],
1702 * It should never happen but changes to locking could
1703 * inadvertently allow a per-cpu drain to add pages
1704 * to MIGRATE_HIGHATOMIC while unreserving so be safe
1705 * and watch for underflows.
1707 zone
->nr_reserved_highatomic
-= min(pageblock_nr_pages
,
1708 zone
->nr_reserved_highatomic
);
1711 * Convert to ac->migratetype and avoid the normal
1712 * pageblock stealing heuristics. Minimally, the caller
1713 * is doing the work and needs the pages. More
1714 * importantly, if the block was always converted to
1715 * MIGRATE_UNMOVABLE or another type then the number
1716 * of pageblocks that cannot be completely freed
1719 set_pageblock_migratetype(page
, ac
->migratetype
);
1720 move_freepages_block(zone
, page
, ac
->migratetype
);
1721 spin_unlock_irqrestore(&zone
->lock
, flags
);
1724 spin_unlock_irqrestore(&zone
->lock
, flags
);
1728 /* Remove an element from the buddy allocator from the fallback list */
1729 static inline struct page
*
1730 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1732 struct free_area
*area
;
1733 unsigned int current_order
;
1738 /* Find the largest possible block of pages in the other list */
1739 for (current_order
= MAX_ORDER
-1;
1740 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1742 area
= &(zone
->free_area
[current_order
]);
1743 fallback_mt
= find_suitable_fallback(area
, current_order
,
1744 start_migratetype
, false, &can_steal
);
1745 if (fallback_mt
== -1)
1748 page
= list_first_entry(&area
->free_list
[fallback_mt
],
1751 steal_suitable_fallback(zone
, page
, start_migratetype
);
1753 /* Remove the page from the freelists */
1755 list_del(&page
->lru
);
1756 rmv_page_order(page
);
1758 expand(zone
, page
, order
, current_order
, area
,
1761 * The pcppage_migratetype may differ from pageblock's
1762 * migratetype depending on the decisions in
1763 * find_suitable_fallback(). This is OK as long as it does not
1764 * differ for MIGRATE_CMA pageblocks. Those can be used as
1765 * fallback only via special __rmqueue_cma_fallback() function
1767 set_pcppage_migratetype(page
, start_migratetype
);
1769 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1770 start_migratetype
, fallback_mt
);
1779 * Do the hard work of removing an element from the buddy allocator.
1780 * Call me with the zone->lock already held.
1782 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1787 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1788 if (unlikely(!page
)) {
1789 if (migratetype
== MIGRATE_MOVABLE
)
1790 page
= __rmqueue_cma_fallback(zone
, order
);
1793 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1796 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1801 * Obtain a specified number of elements from the buddy allocator, all under
1802 * a single hold of the lock, for efficiency. Add them to the supplied list.
1803 * Returns the number of new pages which were placed at *list.
1805 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1806 unsigned long count
, struct list_head
*list
,
1807 int migratetype
, bool cold
)
1811 spin_lock(&zone
->lock
);
1812 for (i
= 0; i
< count
; ++i
) {
1813 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1814 if (unlikely(page
== NULL
))
1818 * Split buddy pages returned by expand() are received here
1819 * in physical page order. The page is added to the callers and
1820 * list and the list head then moves forward. From the callers
1821 * perspective, the linked list is ordered by page number in
1822 * some conditions. This is useful for IO devices that can
1823 * merge IO requests if the physical pages are ordered
1827 list_add(&page
->lru
, list
);
1829 list_add_tail(&page
->lru
, list
);
1831 if (is_migrate_cma(get_pcppage_migratetype(page
)))
1832 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1835 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1836 spin_unlock(&zone
->lock
);
1842 * Called from the vmstat counter updater to drain pagesets of this
1843 * currently executing processor on remote nodes after they have
1846 * Note that this function must be called with the thread pinned to
1847 * a single processor.
1849 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1851 unsigned long flags
;
1852 int to_drain
, batch
;
1854 local_irq_save(flags
);
1855 batch
= READ_ONCE(pcp
->batch
);
1856 to_drain
= min(pcp
->count
, batch
);
1858 free_pcppages_bulk(zone
, to_drain
, pcp
);
1859 pcp
->count
-= to_drain
;
1861 local_irq_restore(flags
);
1866 * Drain pcplists of the indicated processor and zone.
1868 * The processor must either be the current processor and the
1869 * thread pinned to the current processor or a processor that
1872 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1874 unsigned long flags
;
1875 struct per_cpu_pageset
*pset
;
1876 struct per_cpu_pages
*pcp
;
1878 local_irq_save(flags
);
1879 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1883 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1886 local_irq_restore(flags
);
1890 * Drain pcplists of all zones on the indicated processor.
1892 * The processor must either be the current processor and the
1893 * thread pinned to the current processor or a processor that
1896 static void drain_pages(unsigned int cpu
)
1900 for_each_populated_zone(zone
) {
1901 drain_pages_zone(cpu
, zone
);
1906 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1908 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1909 * the single zone's pages.
1911 void drain_local_pages(struct zone
*zone
)
1913 int cpu
= smp_processor_id();
1916 drain_pages_zone(cpu
, zone
);
1922 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1924 * When zone parameter is non-NULL, spill just the single zone's pages.
1926 * Note that this code is protected against sending an IPI to an offline
1927 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1928 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1929 * nothing keeps CPUs from showing up after we populated the cpumask and
1930 * before the call to on_each_cpu_mask().
1932 void drain_all_pages(struct zone
*zone
)
1937 * Allocate in the BSS so we wont require allocation in
1938 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1940 static cpumask_t cpus_with_pcps
;
1943 * We don't care about racing with CPU hotplug event
1944 * as offline notification will cause the notified
1945 * cpu to drain that CPU pcps and on_each_cpu_mask
1946 * disables preemption as part of its processing
1948 for_each_online_cpu(cpu
) {
1949 struct per_cpu_pageset
*pcp
;
1951 bool has_pcps
= false;
1954 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1958 for_each_populated_zone(z
) {
1959 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1960 if (pcp
->pcp
.count
) {
1968 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1970 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1972 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1976 #ifdef CONFIG_HIBERNATION
1978 void mark_free_pages(struct zone
*zone
)
1980 unsigned long pfn
, max_zone_pfn
;
1981 unsigned long flags
;
1982 unsigned int order
, t
;
1983 struct list_head
*curr
;
1985 if (zone_is_empty(zone
))
1988 spin_lock_irqsave(&zone
->lock
, flags
);
1990 max_zone_pfn
= zone_end_pfn(zone
);
1991 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1992 if (pfn_valid(pfn
)) {
1993 struct page
*page
= pfn_to_page(pfn
);
1995 if (!swsusp_page_is_forbidden(page
))
1996 swsusp_unset_page_free(page
);
1999 for_each_migratetype_order(order
, t
) {
2000 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
2003 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
2004 for (i
= 0; i
< (1UL << order
); i
++)
2005 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
2008 spin_unlock_irqrestore(&zone
->lock
, flags
);
2010 #endif /* CONFIG_PM */
2013 * Free a 0-order page
2014 * cold == true ? free a cold page : free a hot page
2016 void free_hot_cold_page(struct page
*page
, bool cold
)
2018 struct zone
*zone
= page_zone(page
);
2019 struct per_cpu_pages
*pcp
;
2020 unsigned long flags
;
2021 unsigned long pfn
= page_to_pfn(page
);
2024 if (!free_pages_prepare(page
, 0))
2027 migratetype
= get_pfnblock_migratetype(page
, pfn
);
2028 set_pcppage_migratetype(page
, migratetype
);
2029 local_irq_save(flags
);
2030 __count_vm_event(PGFREE
);
2033 * We only track unmovable, reclaimable and movable on pcp lists.
2034 * Free ISOLATE pages back to the allocator because they are being
2035 * offlined but treat RESERVE as movable pages so we can get those
2036 * areas back if necessary. Otherwise, we may have to free
2037 * excessively into the page allocator
2039 if (migratetype
>= MIGRATE_PCPTYPES
) {
2040 if (unlikely(is_migrate_isolate(migratetype
))) {
2041 free_one_page(zone
, page
, pfn
, 0, migratetype
);
2044 migratetype
= MIGRATE_MOVABLE
;
2047 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2049 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
2051 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
2053 if (pcp
->count
>= pcp
->high
) {
2054 unsigned long batch
= READ_ONCE(pcp
->batch
);
2055 free_pcppages_bulk(zone
, batch
, pcp
);
2056 pcp
->count
-= batch
;
2060 local_irq_restore(flags
);
2064 * Free a list of 0-order pages
2066 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
2068 struct page
*page
, *next
;
2070 list_for_each_entry_safe(page
, next
, list
, lru
) {
2071 trace_mm_page_free_batched(page
, cold
);
2072 free_hot_cold_page(page
, cold
);
2077 * split_page takes a non-compound higher-order page, and splits it into
2078 * n (1<<order) sub-pages: page[0..n]
2079 * Each sub-page must be freed individually.
2081 * Note: this is probably too low level an operation for use in drivers.
2082 * Please consult with lkml before using this in your driver.
2084 void split_page(struct page
*page
, unsigned int order
)
2089 VM_BUG_ON_PAGE(PageCompound(page
), page
);
2090 VM_BUG_ON_PAGE(!page_count(page
), page
);
2092 #ifdef CONFIG_KMEMCHECK
2094 * Split shadow pages too, because free(page[0]) would
2095 * otherwise free the whole shadow.
2097 if (kmemcheck_page_is_tracked(page
))
2098 split_page(virt_to_page(page
[0].shadow
), order
);
2101 gfp_mask
= get_page_owner_gfp(page
);
2102 set_page_owner(page
, 0, gfp_mask
);
2103 for (i
= 1; i
< (1 << order
); i
++) {
2104 set_page_refcounted(page
+ i
);
2105 set_page_owner(page
+ i
, 0, gfp_mask
);
2108 EXPORT_SYMBOL_GPL(split_page
);
2110 int __isolate_free_page(struct page
*page
, unsigned int order
)
2112 unsigned long watermark
;
2116 BUG_ON(!PageBuddy(page
));
2118 zone
= page_zone(page
);
2119 mt
= get_pageblock_migratetype(page
);
2121 if (!is_migrate_isolate(mt
)) {
2122 /* Obey watermarks as if the page was being allocated */
2123 watermark
= low_wmark_pages(zone
) + (1 << order
);
2124 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
2127 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
2130 /* Remove page from free list */
2131 list_del(&page
->lru
);
2132 zone
->free_area
[order
].nr_free
--;
2133 rmv_page_order(page
);
2135 set_page_owner(page
, order
, __GFP_MOVABLE
);
2137 /* Set the pageblock if the isolated page is at least a pageblock */
2138 if (order
>= pageblock_order
- 1) {
2139 struct page
*endpage
= page
+ (1 << order
) - 1;
2140 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
2141 int mt
= get_pageblock_migratetype(page
);
2142 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
2143 set_pageblock_migratetype(page
,
2149 return 1UL << order
;
2153 * Similar to split_page except the page is already free. As this is only
2154 * being used for migration, the migratetype of the block also changes.
2155 * As this is called with interrupts disabled, the caller is responsible
2156 * for calling arch_alloc_page() and kernel_map_page() after interrupts
2159 * Note: this is probably too low level an operation for use in drivers.
2160 * Please consult with lkml before using this in your driver.
2162 int split_free_page(struct page
*page
)
2167 order
= page_order(page
);
2169 nr_pages
= __isolate_free_page(page
, order
);
2173 /* Split into individual pages */
2174 set_page_refcounted(page
);
2175 split_page(page
, order
);
2180 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
2183 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
2184 struct zone
*zone
, unsigned int order
,
2185 gfp_t gfp_flags
, int alloc_flags
, int migratetype
)
2187 unsigned long flags
;
2189 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
2191 if (likely(order
== 0)) {
2192 struct per_cpu_pages
*pcp
;
2193 struct list_head
*list
;
2195 local_irq_save(flags
);
2196 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
2197 list
= &pcp
->lists
[migratetype
];
2198 if (list_empty(list
)) {
2199 pcp
->count
+= rmqueue_bulk(zone
, 0,
2202 if (unlikely(list_empty(list
)))
2207 page
= list_last_entry(list
, struct page
, lru
);
2209 page
= list_first_entry(list
, struct page
, lru
);
2211 list_del(&page
->lru
);
2214 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
2216 * __GFP_NOFAIL is not to be used in new code.
2218 * All __GFP_NOFAIL callers should be fixed so that they
2219 * properly detect and handle allocation failures.
2221 * We most definitely don't want callers attempting to
2222 * allocate greater than order-1 page units with
2225 WARN_ON_ONCE(order
> 1);
2227 spin_lock_irqsave(&zone
->lock
, flags
);
2230 if (alloc_flags
& ALLOC_HARDER
) {
2231 page
= __rmqueue_smallest(zone
, order
, MIGRATE_HIGHATOMIC
);
2233 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
2236 page
= __rmqueue(zone
, order
, migratetype
);
2237 spin_unlock(&zone
->lock
);
2240 __mod_zone_freepage_state(zone
, -(1 << order
),
2241 get_pcppage_migratetype(page
));
2244 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
2245 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
2246 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
2247 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2249 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
2250 zone_statistics(preferred_zone
, zone
, gfp_flags
);
2251 local_irq_restore(flags
);
2253 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
2257 local_irq_restore(flags
);
2261 #ifdef CONFIG_FAIL_PAGE_ALLOC
2264 struct fault_attr attr
;
2266 bool ignore_gfp_highmem
;
2267 bool ignore_gfp_reclaim
;
2269 } fail_page_alloc
= {
2270 .attr
= FAULT_ATTR_INITIALIZER
,
2271 .ignore_gfp_reclaim
= true,
2272 .ignore_gfp_highmem
= true,
2276 static int __init
setup_fail_page_alloc(char *str
)
2278 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
2280 __setup("fail_page_alloc=", setup_fail_page_alloc
);
2282 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2284 if (order
< fail_page_alloc
.min_order
)
2286 if (gfp_mask
& __GFP_NOFAIL
)
2288 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
2290 if (fail_page_alloc
.ignore_gfp_reclaim
&&
2291 (gfp_mask
& __GFP_DIRECT_RECLAIM
))
2294 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
2297 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
2299 static int __init
fail_page_alloc_debugfs(void)
2301 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
2304 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
2305 &fail_page_alloc
.attr
);
2307 return PTR_ERR(dir
);
2309 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
2310 &fail_page_alloc
.ignore_gfp_reclaim
))
2312 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
2313 &fail_page_alloc
.ignore_gfp_highmem
))
2315 if (!debugfs_create_u32("min-order", mode
, dir
,
2316 &fail_page_alloc
.min_order
))
2321 debugfs_remove_recursive(dir
);
2326 late_initcall(fail_page_alloc_debugfs
);
2328 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
2330 #else /* CONFIG_FAIL_PAGE_ALLOC */
2332 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
2337 #endif /* CONFIG_FAIL_PAGE_ALLOC */
2340 * Return true if free base pages are above 'mark'. For high-order checks it
2341 * will return true of the order-0 watermark is reached and there is at least
2342 * one free page of a suitable size. Checking now avoids taking the zone lock
2343 * to check in the allocation paths if no pages are free.
2345 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
2346 unsigned long mark
, int classzone_idx
, int alloc_flags
,
2351 const int alloc_harder
= (alloc_flags
& ALLOC_HARDER
);
2353 /* free_pages may go negative - that's OK */
2354 free_pages
-= (1 << order
) - 1;
2356 if (alloc_flags
& ALLOC_HIGH
)
2360 * If the caller does not have rights to ALLOC_HARDER then subtract
2361 * the high-atomic reserves. This will over-estimate the size of the
2362 * atomic reserve but it avoids a search.
2364 if (likely(!alloc_harder
))
2365 free_pages
-= z
->nr_reserved_highatomic
;
2370 /* If allocation can't use CMA areas don't use free CMA pages */
2371 if (!(alloc_flags
& ALLOC_CMA
))
2372 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
2376 * Check watermarks for an order-0 allocation request. If these
2377 * are not met, then a high-order request also cannot go ahead
2378 * even if a suitable page happened to be free.
2380 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
2383 /* If this is an order-0 request then the watermark is fine */
2387 /* For a high-order request, check at least one suitable page is free */
2388 for (o
= order
; o
< MAX_ORDER
; o
++) {
2389 struct free_area
*area
= &z
->free_area
[o
];
2398 for (mt
= 0; mt
< MIGRATE_PCPTYPES
; mt
++) {
2399 if (!list_empty(&area
->free_list
[mt
]))
2404 if ((alloc_flags
& ALLOC_CMA
) &&
2405 !list_empty(&area
->free_list
[MIGRATE_CMA
])) {
2413 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
2414 int classzone_idx
, int alloc_flags
)
2416 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
2417 zone_page_state(z
, NR_FREE_PAGES
));
2420 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
2421 unsigned long mark
, int classzone_idx
)
2423 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
2425 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
2426 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
2428 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, 0,
2433 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2435 return local_zone
->node
== zone
->node
;
2438 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2440 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2443 #else /* CONFIG_NUMA */
2444 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2449 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2453 #endif /* CONFIG_NUMA */
2455 static void reset_alloc_batches(struct zone
*preferred_zone
)
2457 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
2460 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2461 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
2462 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
2463 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2464 } while (zone
++ != preferred_zone
);
2468 * get_page_from_freelist goes through the zonelist trying to allocate
2471 static struct page
*
2472 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2473 const struct alloc_context
*ac
)
2475 struct zonelist
*zonelist
= ac
->zonelist
;
2477 struct page
*page
= NULL
;
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 (cpusets_enabled() &&
2494 (alloc_flags
& ALLOC_CPUSET
) &&
2495 !cpuset_zone_allowed(zone
, gfp_mask
))
2498 * Distribute pages in proportion to the individual
2499 * zone size to ensure fair page aging. The zone a
2500 * page was allocated in should have no effect on the
2501 * time the page has in memory before being reclaimed.
2503 if (alloc_flags
& ALLOC_FAIR
) {
2504 if (!zone_local(ac
->preferred_zone
, zone
))
2506 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2512 * When allocating a page cache page for writing, we
2513 * want to get it from a zone that is within its dirty
2514 * limit, such that no single zone holds more than its
2515 * proportional share of globally allowed dirty pages.
2516 * The dirty limits take into account the zone's
2517 * lowmem reserves and high watermark so that kswapd
2518 * should be able to balance it without having to
2519 * write pages from its LRU list.
2521 * This may look like it could increase pressure on
2522 * lower zones by failing allocations in higher zones
2523 * before they are full. But the pages that do spill
2524 * over are limited as the lower zones are protected
2525 * by this very same mechanism. It should not become
2526 * a practical burden to them.
2528 * XXX: For now, allow allocations to potentially
2529 * exceed the per-zone dirty limit in the slowpath
2530 * (spread_dirty_pages unset) before going into reclaim,
2531 * which is important when on a NUMA setup the allowed
2532 * zones are together not big enough to reach the
2533 * global limit. The proper fix for these situations
2534 * will require awareness of zones in the
2535 * dirty-throttling and the flusher threads.
2537 if (ac
->spread_dirty_pages
&& !zone_dirty_ok(zone
))
2540 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2541 if (!zone_watermark_ok(zone
, order
, mark
,
2542 ac
->classzone_idx
, alloc_flags
)) {
2545 /* Checked here to keep the fast path fast */
2546 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2547 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2550 if (zone_reclaim_mode
== 0 ||
2551 !zone_allows_reclaim(ac
->preferred_zone
, zone
))
2554 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2556 case ZONE_RECLAIM_NOSCAN
:
2559 case ZONE_RECLAIM_FULL
:
2560 /* scanned but unreclaimable */
2563 /* did we reclaim enough */
2564 if (zone_watermark_ok(zone
, order
, mark
,
2565 ac
->classzone_idx
, alloc_flags
))
2573 page
= buffered_rmqueue(ac
->preferred_zone
, zone
, order
,
2574 gfp_mask
, alloc_flags
, ac
->migratetype
);
2576 if (prep_new_page(page
, order
, gfp_mask
, alloc_flags
))
2580 * If this is a high-order atomic allocation then check
2581 * if the pageblock should be reserved for the future
2583 if (unlikely(order
&& (alloc_flags
& ALLOC_HARDER
)))
2584 reserve_highatomic_pageblock(page
, zone
, order
);
2591 * The first pass makes sure allocations are spread fairly within the
2592 * local node. However, the local node might have free pages left
2593 * after the fairness batches are exhausted, and remote zones haven't
2594 * even been considered yet. Try once more without fairness, and
2595 * include remote zones now, before entering the slowpath and waking
2596 * kswapd: prefer spilling to a remote zone over swapping locally.
2598 if (alloc_flags
& ALLOC_FAIR
) {
2599 alloc_flags
&= ~ALLOC_FAIR
;
2600 if (nr_fair_skipped
) {
2601 zonelist_rescan
= true;
2602 reset_alloc_batches(ac
->preferred_zone
);
2604 if (nr_online_nodes
> 1)
2605 zonelist_rescan
= true;
2608 if (zonelist_rescan
)
2615 * Large machines with many possible nodes should not always dump per-node
2616 * meminfo in irq context.
2618 static inline bool should_suppress_show_mem(void)
2623 ret
= in_interrupt();
2628 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2629 DEFAULT_RATELIMIT_INTERVAL
,
2630 DEFAULT_RATELIMIT_BURST
);
2632 void warn_alloc_failed(gfp_t gfp_mask
, unsigned int order
, const char *fmt
, ...)
2634 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2636 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2637 debug_guardpage_minorder() > 0)
2641 * This documents exceptions given to allocations in certain
2642 * contexts that are allowed to allocate outside current's set
2645 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2646 if (test_thread_flag(TIF_MEMDIE
) ||
2647 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2648 filter
&= ~SHOW_MEM_FILTER_NODES
;
2649 if (in_interrupt() || !(gfp_mask
& __GFP_DIRECT_RECLAIM
))
2650 filter
&= ~SHOW_MEM_FILTER_NODES
;
2653 struct va_format vaf
;
2656 va_start(args
, fmt
);
2661 pr_warn("%pV", &vaf
);
2666 pr_warn("%s: page allocation failure: order:%u, mode:0x%x\n",
2667 current
->comm
, order
, gfp_mask
);
2670 if (!should_suppress_show_mem())
2674 static inline struct page
*
2675 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2676 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
2678 struct oom_control oc
= {
2679 .zonelist
= ac
->zonelist
,
2680 .nodemask
= ac
->nodemask
,
2681 .gfp_mask
= gfp_mask
,
2686 *did_some_progress
= 0;
2689 * Acquire the oom lock. If that fails, somebody else is
2690 * making progress for us.
2692 if (!mutex_trylock(&oom_lock
)) {
2693 *did_some_progress
= 1;
2694 schedule_timeout_uninterruptible(1);
2699 * Go through the zonelist yet one more time, keep very high watermark
2700 * here, this is only to catch a parallel oom killing, we must fail if
2701 * we're still under heavy pressure.
2703 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
2704 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
2708 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2709 /* Coredumps can quickly deplete all memory reserves */
2710 if (current
->flags
& PF_DUMPCORE
)
2712 /* The OOM killer will not help higher order allocs */
2713 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2715 /* The OOM killer does not needlessly kill tasks for lowmem */
2716 if (ac
->high_zoneidx
< ZONE_NORMAL
)
2718 /* The OOM killer does not compensate for IO-less reclaim */
2719 if (!(gfp_mask
& __GFP_FS
)) {
2721 * XXX: Page reclaim didn't yield anything,
2722 * and the OOM killer can't be invoked, but
2723 * keep looping as per tradition.
2725 *did_some_progress
= 1;
2728 if (pm_suspended_storage())
2730 /* The OOM killer may not free memory on a specific node */
2731 if (gfp_mask
& __GFP_THISNODE
)
2734 /* Exhausted what can be done so it's blamo time */
2735 if (out_of_memory(&oc
) || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
))
2736 *did_some_progress
= 1;
2738 mutex_unlock(&oom_lock
);
2742 #ifdef CONFIG_COMPACTION
2743 /* Try memory compaction for high-order allocations before reclaim */
2744 static struct page
*
2745 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2746 int alloc_flags
, const struct alloc_context
*ac
,
2747 enum migrate_mode mode
, int *contended_compaction
,
2748 bool *deferred_compaction
)
2750 unsigned long compact_result
;
2756 current
->flags
|= PF_MEMALLOC
;
2757 compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
2758 mode
, contended_compaction
);
2759 current
->flags
&= ~PF_MEMALLOC
;
2761 switch (compact_result
) {
2762 case COMPACT_DEFERRED
:
2763 *deferred_compaction
= true;
2765 case COMPACT_SKIPPED
:
2772 * At least in one zone compaction wasn't deferred or skipped, so let's
2773 * count a compaction stall
2775 count_vm_event(COMPACTSTALL
);
2777 page
= get_page_from_freelist(gfp_mask
, order
,
2778 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2781 struct zone
*zone
= page_zone(page
);
2783 zone
->compact_blockskip_flush
= false;
2784 compaction_defer_reset(zone
, order
, true);
2785 count_vm_event(COMPACTSUCCESS
);
2790 * It's bad if compaction run occurs and fails. The most likely reason
2791 * is that pages exist, but not enough to satisfy watermarks.
2793 count_vm_event(COMPACTFAIL
);
2800 static inline struct page
*
2801 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2802 int alloc_flags
, const struct alloc_context
*ac
,
2803 enum migrate_mode mode
, int *contended_compaction
,
2804 bool *deferred_compaction
)
2808 #endif /* CONFIG_COMPACTION */
2810 /* Perform direct synchronous page reclaim */
2812 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
2813 const struct alloc_context
*ac
)
2815 struct reclaim_state reclaim_state
;
2820 /* We now go into synchronous reclaim */
2821 cpuset_memory_pressure_bump();
2822 current
->flags
|= PF_MEMALLOC
;
2823 lockdep_set_current_reclaim_state(gfp_mask
);
2824 reclaim_state
.reclaimed_slab
= 0;
2825 current
->reclaim_state
= &reclaim_state
;
2827 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
2830 current
->reclaim_state
= NULL
;
2831 lockdep_clear_current_reclaim_state();
2832 current
->flags
&= ~PF_MEMALLOC
;
2839 /* The really slow allocator path where we enter direct reclaim */
2840 static inline struct page
*
2841 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2842 int alloc_flags
, const struct alloc_context
*ac
,
2843 unsigned long *did_some_progress
)
2845 struct page
*page
= NULL
;
2846 bool drained
= false;
2848 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
2849 if (unlikely(!(*did_some_progress
)))
2853 page
= get_page_from_freelist(gfp_mask
, order
,
2854 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2857 * If an allocation failed after direct reclaim, it could be because
2858 * pages are pinned on the per-cpu lists or in high alloc reserves.
2859 * Shrink them them and try again
2861 if (!page
&& !drained
) {
2862 unreserve_highatomic_pageblock(ac
);
2863 drain_all_pages(NULL
);
2871 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
2876 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2877 ac
->high_zoneidx
, ac
->nodemask
)
2878 wakeup_kswapd(zone
, order
, zone_idx(ac
->preferred_zone
));
2882 gfp_to_alloc_flags(gfp_t gfp_mask
)
2884 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2886 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2887 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2890 * The caller may dip into page reserves a bit more if the caller
2891 * cannot run direct reclaim, or if the caller has realtime scheduling
2892 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2893 * set both ALLOC_HARDER (__GFP_ATOMIC) and ALLOC_HIGH (__GFP_HIGH).
2895 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2897 if (gfp_mask
& __GFP_ATOMIC
) {
2899 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2900 * if it can't schedule.
2902 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2903 alloc_flags
|= ALLOC_HARDER
;
2905 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2906 * comment for __cpuset_node_allowed().
2908 alloc_flags
&= ~ALLOC_CPUSET
;
2909 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2910 alloc_flags
|= ALLOC_HARDER
;
2912 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2913 if (gfp_mask
& __GFP_MEMALLOC
)
2914 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2915 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2916 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2917 else if (!in_interrupt() &&
2918 ((current
->flags
& PF_MEMALLOC
) ||
2919 unlikely(test_thread_flag(TIF_MEMDIE
))))
2920 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2923 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2924 alloc_flags
|= ALLOC_CMA
;
2929 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2931 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2934 static inline bool is_thp_gfp_mask(gfp_t gfp_mask
)
2936 return (gfp_mask
& (GFP_TRANSHUGE
| __GFP_KSWAPD_RECLAIM
)) == GFP_TRANSHUGE
;
2939 static inline struct page
*
2940 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2941 struct alloc_context
*ac
)
2943 bool can_direct_reclaim
= gfp_mask
& __GFP_DIRECT_RECLAIM
;
2944 struct page
*page
= NULL
;
2946 unsigned long pages_reclaimed
= 0;
2947 unsigned long did_some_progress
;
2948 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2949 bool deferred_compaction
= false;
2950 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2953 * In the slowpath, we sanity check order to avoid ever trying to
2954 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2955 * be using allocators in order of preference for an area that is
2958 if (order
>= MAX_ORDER
) {
2959 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2964 * We also sanity check to catch abuse of atomic reserves being used by
2965 * callers that are not in atomic context.
2967 if (WARN_ON_ONCE((gfp_mask
& (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)) ==
2968 (__GFP_ATOMIC
|__GFP_DIRECT_RECLAIM
)))
2969 gfp_mask
&= ~__GFP_ATOMIC
;
2972 * If this allocation cannot block and it is for a specific node, then
2973 * fail early. There's no need to wakeup kswapd or retry for a
2974 * speculative node-specific allocation.
2976 if (IS_ENABLED(CONFIG_NUMA
) && (gfp_mask
& __GFP_THISNODE
) && !can_direct_reclaim
)
2980 if (gfp_mask
& __GFP_KSWAPD_RECLAIM
)
2981 wake_all_kswapds(order
, ac
);
2984 * OK, we're below the kswapd watermark and have kicked background
2985 * reclaim. Now things get more complex, so set up alloc_flags according
2986 * to how we want to proceed.
2988 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2991 * Find the true preferred zone if the allocation is unconstrained by
2994 if (!(alloc_flags
& ALLOC_CPUSET
) && !ac
->nodemask
) {
2995 struct zoneref
*preferred_zoneref
;
2996 preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
2997 ac
->high_zoneidx
, NULL
, &ac
->preferred_zone
);
2998 ac
->classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
3001 /* This is the last chance, in general, before the goto nopage. */
3002 page
= get_page_from_freelist(gfp_mask
, order
,
3003 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
3007 /* Allocate without watermarks if the context allows */
3008 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
3010 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
3011 * the allocation is high priority and these type of
3012 * allocations are system rather than user orientated
3014 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
3015 page
= get_page_from_freelist(gfp_mask
, order
,
3016 ALLOC_NO_WATERMARKS
, ac
);
3021 /* Caller is not willing to reclaim, we can't balance anything */
3022 if (!can_direct_reclaim
) {
3024 * All existing users of the __GFP_NOFAIL are blockable, so warn
3025 * of any new users that actually allow this type of allocation
3028 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
3032 /* Avoid recursion of direct reclaim */
3033 if (current
->flags
& PF_MEMALLOC
) {
3035 * __GFP_NOFAIL request from this context is rather bizarre
3036 * because we cannot reclaim anything and only can loop waiting
3037 * for somebody to do a work for us.
3039 if (WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
)) {
3046 /* Avoid allocations with no watermarks from looping endlessly */
3047 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
3051 * Try direct compaction. The first pass is asynchronous. Subsequent
3052 * attempts after direct reclaim are synchronous
3054 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
3056 &contended_compaction
,
3057 &deferred_compaction
);
3061 /* Checks for THP-specific high-order allocations */
3062 if (is_thp_gfp_mask(gfp_mask
)) {
3064 * If compaction is deferred for high-order allocations, it is
3065 * because sync compaction recently failed. If this is the case
3066 * and the caller requested a THP allocation, we do not want
3067 * to heavily disrupt the system, so we fail the allocation
3068 * instead of entering direct reclaim.
3070 if (deferred_compaction
)
3074 * In all zones where compaction was attempted (and not
3075 * deferred or skipped), lock contention has been detected.
3076 * For THP allocation we do not want to disrupt the others
3077 * so we fallback to base pages instead.
3079 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
3083 * If compaction was aborted due to need_resched(), we do not
3084 * want to further increase allocation latency, unless it is
3085 * khugepaged trying to collapse.
3087 if (contended_compaction
== COMPACT_CONTENDED_SCHED
3088 && !(current
->flags
& PF_KTHREAD
))
3093 * It can become very expensive to allocate transparent hugepages at
3094 * fault, so use asynchronous memory compaction for THP unless it is
3095 * khugepaged trying to collapse.
3097 if (!is_thp_gfp_mask(gfp_mask
) || (current
->flags
& PF_KTHREAD
))
3098 migration_mode
= MIGRATE_SYNC_LIGHT
;
3100 /* Try direct reclaim and then allocating */
3101 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
3102 &did_some_progress
);
3106 /* Do not loop if specifically requested */
3107 if (gfp_mask
& __GFP_NORETRY
)
3110 /* Keep reclaiming pages as long as there is reasonable progress */
3111 pages_reclaimed
+= did_some_progress
;
3112 if ((did_some_progress
&& order
<= PAGE_ALLOC_COSTLY_ORDER
) ||
3113 ((gfp_mask
& __GFP_REPEAT
) && pages_reclaimed
< (1 << order
))) {
3114 /* Wait for some write requests to complete then retry */
3115 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
3119 /* Reclaim has failed us, start killing things */
3120 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
, &did_some_progress
);
3124 /* Retry as long as the OOM killer is making progress */
3125 if (did_some_progress
)
3130 * High-order allocations do not necessarily loop after
3131 * direct reclaim and reclaim/compaction depends on compaction
3132 * being called after reclaim so call directly if necessary
3134 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
,
3136 &contended_compaction
,
3137 &deferred_compaction
);
3141 warn_alloc_failed(gfp_mask
, order
, NULL
);
3147 * This is the 'heart' of the zoned buddy allocator.
3150 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
3151 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
3153 struct zoneref
*preferred_zoneref
;
3154 struct page
*page
= NULL
;
3155 unsigned int cpuset_mems_cookie
;
3156 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
3157 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
3158 struct alloc_context ac
= {
3159 .high_zoneidx
= gfp_zone(gfp_mask
),
3160 .nodemask
= nodemask
,
3161 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
3164 gfp_mask
&= gfp_allowed_mask
;
3166 lockdep_trace_alloc(gfp_mask
);
3168 might_sleep_if(gfp_mask
& __GFP_DIRECT_RECLAIM
);
3170 if (should_fail_alloc_page(gfp_mask
, order
))
3174 * Check the zones suitable for the gfp_mask contain at least one
3175 * valid zone. It's possible to have an empty zonelist as a result
3176 * of __GFP_THISNODE and a memoryless node
3178 if (unlikely(!zonelist
->_zonerefs
->zone
))
3181 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
3182 alloc_flags
|= ALLOC_CMA
;
3185 cpuset_mems_cookie
= read_mems_allowed_begin();
3187 /* We set it here, as __alloc_pages_slowpath might have changed it */
3188 ac
.zonelist
= zonelist
;
3190 /* Dirty zone balancing only done in the fast path */
3191 ac
.spread_dirty_pages
= (gfp_mask
& __GFP_WRITE
);
3193 /* The preferred zone is used for statistics later */
3194 preferred_zoneref
= first_zones_zonelist(ac
.zonelist
, ac
.high_zoneidx
,
3195 ac
.nodemask
? : &cpuset_current_mems_allowed
,
3196 &ac
.preferred_zone
);
3197 if (!ac
.preferred_zone
)
3199 ac
.classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
3201 /* First allocation attempt */
3202 alloc_mask
= gfp_mask
|__GFP_HARDWALL
;
3203 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
3204 if (unlikely(!page
)) {
3206 * Runtime PM, block IO and its error handling path
3207 * can deadlock because I/O on the device might not
3210 alloc_mask
= memalloc_noio_flags(gfp_mask
);
3211 ac
.spread_dirty_pages
= false;
3213 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
3216 if (kmemcheck_enabled
&& page
)
3217 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
3219 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
3223 * When updating a task's mems_allowed, it is possible to race with
3224 * parallel threads in such a way that an allocation can fail while
3225 * the mask is being updated. If a page allocation is about to fail,
3226 * check if the cpuset changed during allocation and if so, retry.
3228 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
3233 EXPORT_SYMBOL(__alloc_pages_nodemask
);
3236 * Common helper functions.
3238 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
3243 * __get_free_pages() returns a 32-bit address, which cannot represent
3246 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
3248 page
= alloc_pages(gfp_mask
, order
);
3251 return (unsigned long) page_address(page
);
3253 EXPORT_SYMBOL(__get_free_pages
);
3255 unsigned long get_zeroed_page(gfp_t gfp_mask
)
3257 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
3259 EXPORT_SYMBOL(get_zeroed_page
);
3261 void __free_pages(struct page
*page
, unsigned int order
)
3263 if (put_page_testzero(page
)) {
3265 free_hot_cold_page(page
, false);
3267 __free_pages_ok(page
, order
);
3271 EXPORT_SYMBOL(__free_pages
);
3273 void free_pages(unsigned long addr
, unsigned int order
)
3276 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3277 __free_pages(virt_to_page((void *)addr
), order
);
3281 EXPORT_SYMBOL(free_pages
);
3285 * An arbitrary-length arbitrary-offset area of memory which resides
3286 * within a 0 or higher order page. Multiple fragments within that page
3287 * are individually refcounted, in the page's reference counter.
3289 * The page_frag functions below provide a simple allocation framework for
3290 * page fragments. This is used by the network stack and network device
3291 * drivers to provide a backing region of memory for use as either an
3292 * sk_buff->head, or to be used in the "frags" portion of skb_shared_info.
3294 static struct page
*__page_frag_refill(struct page_frag_cache
*nc
,
3297 struct page
*page
= NULL
;
3298 gfp_t gfp
= gfp_mask
;
3300 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3301 gfp_mask
|= __GFP_COMP
| __GFP_NOWARN
| __GFP_NORETRY
|
3303 page
= alloc_pages_node(NUMA_NO_NODE
, gfp_mask
,
3304 PAGE_FRAG_CACHE_MAX_ORDER
);
3305 nc
->size
= page
? PAGE_FRAG_CACHE_MAX_SIZE
: PAGE_SIZE
;
3307 if (unlikely(!page
))
3308 page
= alloc_pages_node(NUMA_NO_NODE
, gfp
, 0);
3310 nc
->va
= page
? page_address(page
) : NULL
;
3315 void *__alloc_page_frag(struct page_frag_cache
*nc
,
3316 unsigned int fragsz
, gfp_t gfp_mask
)
3318 unsigned int size
= PAGE_SIZE
;
3322 if (unlikely(!nc
->va
)) {
3324 page
= __page_frag_refill(nc
, gfp_mask
);
3328 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3329 /* if size can vary use size else just use PAGE_SIZE */
3332 /* Even if we own the page, we do not use atomic_set().
3333 * This would break get_page_unless_zero() users.
3335 atomic_add(size
- 1, &page
->_count
);
3337 /* reset page count bias and offset to start of new frag */
3338 nc
->pfmemalloc
= page_is_pfmemalloc(page
);
3339 nc
->pagecnt_bias
= size
;
3343 offset
= nc
->offset
- fragsz
;
3344 if (unlikely(offset
< 0)) {
3345 page
= virt_to_page(nc
->va
);
3347 if (!atomic_sub_and_test(nc
->pagecnt_bias
, &page
->_count
))
3350 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
3351 /* if size can vary use size else just use PAGE_SIZE */
3354 /* OK, page count is 0, we can safely set it */
3355 atomic_set(&page
->_count
, size
);
3357 /* reset page count bias and offset to start of new frag */
3358 nc
->pagecnt_bias
= size
;
3359 offset
= size
- fragsz
;
3363 nc
->offset
= offset
;
3365 return nc
->va
+ offset
;
3367 EXPORT_SYMBOL(__alloc_page_frag
);
3370 * Frees a page fragment allocated out of either a compound or order 0 page.
3372 void __free_page_frag(void *addr
)
3374 struct page
*page
= virt_to_head_page(addr
);
3376 if (unlikely(put_page_testzero(page
)))
3377 __free_pages_ok(page
, compound_order(page
));
3379 EXPORT_SYMBOL(__free_page_frag
);
3382 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
3383 * of the current memory cgroup if __GFP_ACCOUNT is set, other than that it is
3384 * equivalent to alloc_pages.
3386 * It should be used when the caller would like to use kmalloc, but since the
3387 * allocation is large, it has to fall back to the page allocator.
3389 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
3393 page
= alloc_pages(gfp_mask
, order
);
3394 if (page
&& memcg_kmem_charge(page
, gfp_mask
, order
) != 0) {
3395 __free_pages(page
, order
);
3401 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
3405 page
= alloc_pages_node(nid
, gfp_mask
, order
);
3406 if (page
&& memcg_kmem_charge(page
, gfp_mask
, order
) != 0) {
3407 __free_pages(page
, order
);
3414 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3417 void __free_kmem_pages(struct page
*page
, unsigned int order
)
3419 memcg_kmem_uncharge(page
, order
);
3420 __free_pages(page
, order
);
3423 void free_kmem_pages(unsigned long addr
, unsigned int order
)
3426 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3427 __free_kmem_pages(virt_to_page((void *)addr
), order
);
3431 static void *make_alloc_exact(unsigned long addr
, unsigned int order
,
3435 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3436 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3438 split_page(virt_to_page((void *)addr
), order
);
3439 while (used
< alloc_end
) {
3444 return (void *)addr
;
3448 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3449 * @size: the number of bytes to allocate
3450 * @gfp_mask: GFP flags for the allocation
3452 * This function is similar to alloc_pages(), except that it allocates the
3453 * minimum number of pages to satisfy the request. alloc_pages() can only
3454 * allocate memory in power-of-two pages.
3456 * This function is also limited by MAX_ORDER.
3458 * Memory allocated by this function must be released by free_pages_exact().
3460 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3462 unsigned int order
= get_order(size
);
3465 addr
= __get_free_pages(gfp_mask
, order
);
3466 return make_alloc_exact(addr
, order
, size
);
3468 EXPORT_SYMBOL(alloc_pages_exact
);
3471 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3473 * @nid: the preferred node ID where memory should be allocated
3474 * @size: the number of bytes to allocate
3475 * @gfp_mask: GFP flags for the allocation
3477 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3480 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3482 unsigned int order
= get_order(size
);
3483 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3486 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3490 * free_pages_exact - release memory allocated via alloc_pages_exact()
3491 * @virt: the value returned by alloc_pages_exact.
3492 * @size: size of allocation, same value as passed to alloc_pages_exact().
3494 * Release the memory allocated by a previous call to alloc_pages_exact.
3496 void free_pages_exact(void *virt
, size_t size
)
3498 unsigned long addr
= (unsigned long)virt
;
3499 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3501 while (addr
< end
) {
3506 EXPORT_SYMBOL(free_pages_exact
);
3509 * nr_free_zone_pages - count number of pages beyond high watermark
3510 * @offset: The zone index of the highest zone
3512 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3513 * high watermark within all zones at or below a given zone index. For each
3514 * zone, the number of pages is calculated as:
3515 * managed_pages - high_pages
3517 static unsigned long nr_free_zone_pages(int offset
)
3522 /* Just pick one node, since fallback list is circular */
3523 unsigned long sum
= 0;
3525 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3527 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3528 unsigned long size
= zone
->managed_pages
;
3529 unsigned long high
= high_wmark_pages(zone
);
3538 * nr_free_buffer_pages - count number of pages beyond high watermark
3540 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3541 * watermark within ZONE_DMA and ZONE_NORMAL.
3543 unsigned long nr_free_buffer_pages(void)
3545 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3547 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3550 * nr_free_pagecache_pages - count number of pages beyond high watermark
3552 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3553 * high watermark within all zones.
3555 unsigned long nr_free_pagecache_pages(void)
3557 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3560 static inline void show_node(struct zone
*zone
)
3562 if (IS_ENABLED(CONFIG_NUMA
))
3563 printk("Node %d ", zone_to_nid(zone
));
3566 void si_meminfo(struct sysinfo
*val
)
3568 val
->totalram
= totalram_pages
;
3569 val
->sharedram
= global_page_state(NR_SHMEM
);
3570 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3571 val
->bufferram
= nr_blockdev_pages();
3572 val
->totalhigh
= totalhigh_pages
;
3573 val
->freehigh
= nr_free_highpages();
3574 val
->mem_unit
= PAGE_SIZE
;
3577 EXPORT_SYMBOL(si_meminfo
);
3580 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3582 int zone_type
; /* needs to be signed */
3583 unsigned long managed_pages
= 0;
3584 pg_data_t
*pgdat
= NODE_DATA(nid
);
3586 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3587 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3588 val
->totalram
= managed_pages
;
3589 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3590 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3591 #ifdef CONFIG_HIGHMEM
3592 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3593 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3599 val
->mem_unit
= PAGE_SIZE
;
3604 * Determine whether the node should be displayed or not, depending on whether
3605 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3607 bool skip_free_areas_node(unsigned int flags
, int nid
)
3610 unsigned int cpuset_mems_cookie
;
3612 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3616 cpuset_mems_cookie
= read_mems_allowed_begin();
3617 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3618 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3623 #define K(x) ((x) << (PAGE_SHIFT-10))
3625 static void show_migration_types(unsigned char type
)
3627 static const char types
[MIGRATE_TYPES
] = {
3628 [MIGRATE_UNMOVABLE
] = 'U',
3629 [MIGRATE_MOVABLE
] = 'M',
3630 [MIGRATE_RECLAIMABLE
] = 'E',
3631 [MIGRATE_HIGHATOMIC
] = 'H',
3633 [MIGRATE_CMA
] = 'C',
3635 #ifdef CONFIG_MEMORY_ISOLATION
3636 [MIGRATE_ISOLATE
] = 'I',
3639 char tmp
[MIGRATE_TYPES
+ 1];
3643 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3644 if (type
& (1 << i
))
3649 printk("(%s) ", tmp
);
3653 * Show free area list (used inside shift_scroll-lock stuff)
3654 * We also calculate the percentage fragmentation. We do this by counting the
3655 * memory on each free list with the exception of the first item on the list.
3658 * SHOW_MEM_FILTER_NODES: suppress nodes that are not allowed by current's
3661 void show_free_areas(unsigned int filter
)
3663 unsigned long free_pcp
= 0;
3667 for_each_populated_zone(zone
) {
3668 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3671 for_each_online_cpu(cpu
)
3672 free_pcp
+= per_cpu_ptr(zone
->pageset
, cpu
)->pcp
.count
;
3675 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3676 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3677 " unevictable:%lu dirty:%lu writeback:%lu unstable:%lu\n"
3678 " slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3679 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3680 " free:%lu free_pcp:%lu free_cma:%lu\n",
3681 global_page_state(NR_ACTIVE_ANON
),
3682 global_page_state(NR_INACTIVE_ANON
),
3683 global_page_state(NR_ISOLATED_ANON
),
3684 global_page_state(NR_ACTIVE_FILE
),
3685 global_page_state(NR_INACTIVE_FILE
),
3686 global_page_state(NR_ISOLATED_FILE
),
3687 global_page_state(NR_UNEVICTABLE
),
3688 global_page_state(NR_FILE_DIRTY
),
3689 global_page_state(NR_WRITEBACK
),
3690 global_page_state(NR_UNSTABLE_NFS
),
3691 global_page_state(NR_SLAB_RECLAIMABLE
),
3692 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3693 global_page_state(NR_FILE_MAPPED
),
3694 global_page_state(NR_SHMEM
),
3695 global_page_state(NR_PAGETABLE
),
3696 global_page_state(NR_BOUNCE
),
3697 global_page_state(NR_FREE_PAGES
),
3699 global_page_state(NR_FREE_CMA_PAGES
));
3701 for_each_populated_zone(zone
) {
3704 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
;
3717 " active_anon:%lukB"
3718 " inactive_anon:%lukB"
3719 " active_file:%lukB"
3720 " inactive_file:%lukB"
3721 " unevictable:%lukB"
3722 " isolated(anon):%lukB"
3723 " isolated(file):%lukB"
3731 " slab_reclaimable:%lukB"
3732 " slab_unreclaimable:%lukB"
3733 " kernel_stack:%lukB"
3740 " writeback_tmp:%lukB"
3741 " pages_scanned:%lu"
3742 " all_unreclaimable? %s"
3745 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3746 K(min_wmark_pages(zone
)),
3747 K(low_wmark_pages(zone
)),
3748 K(high_wmark_pages(zone
)),
3749 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3750 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3751 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3752 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3753 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3754 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3755 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3756 K(zone
->present_pages
),
3757 K(zone
->managed_pages
),
3758 K(zone_page_state(zone
, NR_MLOCK
)),
3759 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3760 K(zone_page_state(zone
, NR_WRITEBACK
)),
3761 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3762 K(zone_page_state(zone
, NR_SHMEM
)),
3763 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3764 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3765 zone_page_state(zone
, NR_KERNEL_STACK
) *
3767 K(zone_page_state(zone
, NR_PAGETABLE
)),
3768 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3769 K(zone_page_state(zone
, NR_BOUNCE
)),
3771 K(this_cpu_read(zone
->pageset
->pcp
.count
)),
3772 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3773 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3774 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3775 (!zone_reclaimable(zone
) ? "yes" : "no")
3777 printk("lowmem_reserve[]:");
3778 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3779 printk(" %ld", zone
->lowmem_reserve
[i
]);
3783 for_each_populated_zone(zone
) {
3785 unsigned long nr
[MAX_ORDER
], flags
, total
= 0;
3786 unsigned char types
[MAX_ORDER
];
3788 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3791 printk("%s: ", zone
->name
);
3793 spin_lock_irqsave(&zone
->lock
, flags
);
3794 for (order
= 0; order
< MAX_ORDER
; order
++) {
3795 struct free_area
*area
= &zone
->free_area
[order
];
3798 nr
[order
] = area
->nr_free
;
3799 total
+= nr
[order
] << order
;
3802 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3803 if (!list_empty(&area
->free_list
[type
]))
3804 types
[order
] |= 1 << type
;
3807 spin_unlock_irqrestore(&zone
->lock
, flags
);
3808 for (order
= 0; order
< MAX_ORDER
; order
++) {
3809 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3811 show_migration_types(types
[order
]);
3813 printk("= %lukB\n", K(total
));
3816 hugetlb_show_meminfo();
3818 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3820 show_swap_cache_info();
3823 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3825 zoneref
->zone
= zone
;
3826 zoneref
->zone_idx
= zone_idx(zone
);
3830 * Builds allocation fallback zone lists.
3832 * Add all populated zones of a node to the zonelist.
3834 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3838 enum zone_type zone_type
= MAX_NR_ZONES
;
3842 zone
= pgdat
->node_zones
+ zone_type
;
3843 if (populated_zone(zone
)) {
3844 zoneref_set_zone(zone
,
3845 &zonelist
->_zonerefs
[nr_zones
++]);
3846 check_highest_zone(zone_type
);
3848 } while (zone_type
);
3856 * 0 = automatic detection of better ordering.
3857 * 1 = order by ([node] distance, -zonetype)
3858 * 2 = order by (-zonetype, [node] distance)
3860 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3861 * the same zonelist. So only NUMA can configure this param.
3863 #define ZONELIST_ORDER_DEFAULT 0
3864 #define ZONELIST_ORDER_NODE 1
3865 #define ZONELIST_ORDER_ZONE 2
3867 /* zonelist order in the kernel.
3868 * set_zonelist_order() will set this to NODE or ZONE.
3870 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3871 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3875 /* The value user specified ....changed by config */
3876 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3877 /* string for sysctl */
3878 #define NUMA_ZONELIST_ORDER_LEN 16
3879 char numa_zonelist_order
[16] = "default";
3882 * interface for configure zonelist ordering.
3883 * command line option "numa_zonelist_order"
3884 * = "[dD]efault - default, automatic configuration.
3885 * = "[nN]ode - order by node locality, then by zone within node
3886 * = "[zZ]one - order by zone, then by locality within zone
3889 static int __parse_numa_zonelist_order(char *s
)
3891 if (*s
== 'd' || *s
== 'D') {
3892 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3893 } else if (*s
== 'n' || *s
== 'N') {
3894 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3895 } else if (*s
== 'z' || *s
== 'Z') {
3896 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3899 "Ignoring invalid numa_zonelist_order value: "
3906 static __init
int setup_numa_zonelist_order(char *s
)
3913 ret
= __parse_numa_zonelist_order(s
);
3915 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3919 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3922 * sysctl handler for numa_zonelist_order
3924 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3925 void __user
*buffer
, size_t *length
,
3928 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3930 static DEFINE_MUTEX(zl_order_mutex
);
3932 mutex_lock(&zl_order_mutex
);
3934 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3938 strcpy(saved_string
, (char *)table
->data
);
3940 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3944 int oldval
= user_zonelist_order
;
3946 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3949 * bogus value. restore saved string
3951 strncpy((char *)table
->data
, saved_string
,
3952 NUMA_ZONELIST_ORDER_LEN
);
3953 user_zonelist_order
= oldval
;
3954 } else if (oldval
!= user_zonelist_order
) {
3955 mutex_lock(&zonelists_mutex
);
3956 build_all_zonelists(NULL
, NULL
);
3957 mutex_unlock(&zonelists_mutex
);
3961 mutex_unlock(&zl_order_mutex
);
3966 #define MAX_NODE_LOAD (nr_online_nodes)
3967 static int node_load
[MAX_NUMNODES
];
3970 * find_next_best_node - find the next node that should appear in a given node's fallback list
3971 * @node: node whose fallback list we're appending
3972 * @used_node_mask: nodemask_t of already used nodes
3974 * We use a number of factors to determine which is the next node that should
3975 * appear on a given node's fallback list. The node should not have appeared
3976 * already in @node's fallback list, and it should be the next closest node
3977 * according to the distance array (which contains arbitrary distance values
3978 * from each node to each node in the system), and should also prefer nodes
3979 * with no CPUs, since presumably they'll have very little allocation pressure
3980 * on them otherwise.
3981 * It returns -1 if no node is found.
3983 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3986 int min_val
= INT_MAX
;
3987 int best_node
= NUMA_NO_NODE
;
3988 const struct cpumask
*tmp
= cpumask_of_node(0);
3990 /* Use the local node if we haven't already */
3991 if (!node_isset(node
, *used_node_mask
)) {
3992 node_set(node
, *used_node_mask
);
3996 for_each_node_state(n
, N_MEMORY
) {
3998 /* Don't want a node to appear more than once */
3999 if (node_isset(n
, *used_node_mask
))
4002 /* Use the distance array to find the distance */
4003 val
= node_distance(node
, n
);
4005 /* Penalize nodes under us ("prefer the next node") */
4008 /* Give preference to headless and unused nodes */
4009 tmp
= cpumask_of_node(n
);
4010 if (!cpumask_empty(tmp
))
4011 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
4013 /* Slight preference for less loaded node */
4014 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
4015 val
+= node_load
[n
];
4017 if (val
< min_val
) {
4024 node_set(best_node
, *used_node_mask
);
4031 * Build zonelists ordered by node and zones within node.
4032 * This results in maximum locality--normal zone overflows into local
4033 * DMA zone, if any--but risks exhausting DMA zone.
4035 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
4038 struct zonelist
*zonelist
;
4040 zonelist
= &pgdat
->node_zonelists
[0];
4041 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
4043 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4044 zonelist
->_zonerefs
[j
].zone
= NULL
;
4045 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4049 * Build gfp_thisnode zonelists
4051 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
4054 struct zonelist
*zonelist
;
4056 zonelist
= &pgdat
->node_zonelists
[1];
4057 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4058 zonelist
->_zonerefs
[j
].zone
= NULL
;
4059 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4063 * Build zonelists ordered by zone and nodes within zones.
4064 * This results in conserving DMA zone[s] until all Normal memory is
4065 * exhausted, but results in overflowing to remote node while memory
4066 * may still exist in local DMA zone.
4068 static int node_order
[MAX_NUMNODES
];
4070 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
4073 int zone_type
; /* needs to be signed */
4075 struct zonelist
*zonelist
;
4077 zonelist
= &pgdat
->node_zonelists
[0];
4079 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
4080 for (j
= 0; j
< nr_nodes
; j
++) {
4081 node
= node_order
[j
];
4082 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
4083 if (populated_zone(z
)) {
4085 &zonelist
->_zonerefs
[pos
++]);
4086 check_highest_zone(zone_type
);
4090 zonelist
->_zonerefs
[pos
].zone
= NULL
;
4091 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
4094 #if defined(CONFIG_64BIT)
4096 * Devices that require DMA32/DMA are relatively rare and do not justify a
4097 * penalty to every machine in case the specialised case applies. Default
4098 * to Node-ordering on 64-bit NUMA machines
4100 static int default_zonelist_order(void)
4102 return ZONELIST_ORDER_NODE
;
4106 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
4107 * by the kernel. If processes running on node 0 deplete the low memory zone
4108 * then reclaim will occur more frequency increasing stalls and potentially
4109 * be easier to OOM if a large percentage of the zone is under writeback or
4110 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
4111 * Hence, default to zone ordering on 32-bit.
4113 static int default_zonelist_order(void)
4115 return ZONELIST_ORDER_ZONE
;
4117 #endif /* CONFIG_64BIT */
4119 static void set_zonelist_order(void)
4121 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
4122 current_zonelist_order
= default_zonelist_order();
4124 current_zonelist_order
= user_zonelist_order
;
4127 static void build_zonelists(pg_data_t
*pgdat
)
4130 nodemask_t used_mask
;
4131 int local_node
, prev_node
;
4132 struct zonelist
*zonelist
;
4133 unsigned int order
= current_zonelist_order
;
4135 /* initialize zonelists */
4136 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
4137 zonelist
= pgdat
->node_zonelists
+ i
;
4138 zonelist
->_zonerefs
[0].zone
= NULL
;
4139 zonelist
->_zonerefs
[0].zone_idx
= 0;
4142 /* NUMA-aware ordering of nodes */
4143 local_node
= pgdat
->node_id
;
4144 load
= nr_online_nodes
;
4145 prev_node
= local_node
;
4146 nodes_clear(used_mask
);
4148 memset(node_order
, 0, sizeof(node_order
));
4151 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
4153 * We don't want to pressure a particular node.
4154 * So adding penalty to the first node in same
4155 * distance group to make it round-robin.
4157 if (node_distance(local_node
, node
) !=
4158 node_distance(local_node
, prev_node
))
4159 node_load
[node
] = load
;
4163 if (order
== ZONELIST_ORDER_NODE
)
4164 build_zonelists_in_node_order(pgdat
, node
);
4166 node_order
[i
++] = node
; /* remember order */
4169 if (order
== ZONELIST_ORDER_ZONE
) {
4170 /* calculate node order -- i.e., DMA last! */
4171 build_zonelists_in_zone_order(pgdat
, i
);
4174 build_thisnode_zonelists(pgdat
);
4177 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4179 * Return node id of node used for "local" allocations.
4180 * I.e., first node id of first zone in arg node's generic zonelist.
4181 * Used for initializing percpu 'numa_mem', which is used primarily
4182 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
4184 int local_memory_node(int node
)
4188 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
4189 gfp_zone(GFP_KERNEL
),
4196 #else /* CONFIG_NUMA */
4198 static void set_zonelist_order(void)
4200 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
4203 static void build_zonelists(pg_data_t
*pgdat
)
4205 int node
, local_node
;
4207 struct zonelist
*zonelist
;
4209 local_node
= pgdat
->node_id
;
4211 zonelist
= &pgdat
->node_zonelists
[0];
4212 j
= build_zonelists_node(pgdat
, zonelist
, 0);
4215 * Now we build the zonelist so that it contains the zones
4216 * of all the other nodes.
4217 * We don't want to pressure a particular node, so when
4218 * building the zones for node N, we make sure that the
4219 * zones coming right after the local ones are those from
4220 * node N+1 (modulo N)
4222 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
4223 if (!node_online(node
))
4225 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4227 for (node
= 0; node
< local_node
; node
++) {
4228 if (!node_online(node
))
4230 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
4233 zonelist
->_zonerefs
[j
].zone
= NULL
;
4234 zonelist
->_zonerefs
[j
].zone_idx
= 0;
4237 #endif /* CONFIG_NUMA */
4240 * Boot pageset table. One per cpu which is going to be used for all
4241 * zones and all nodes. The parameters will be set in such a way
4242 * that an item put on a list will immediately be handed over to
4243 * the buddy list. This is safe since pageset manipulation is done
4244 * with interrupts disabled.
4246 * The boot_pagesets must be kept even after bootup is complete for
4247 * unused processors and/or zones. They do play a role for bootstrapping
4248 * hotplugged processors.
4250 * zoneinfo_show() and maybe other functions do
4251 * not check if the processor is online before following the pageset pointer.
4252 * Other parts of the kernel may not check if the zone is available.
4254 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
4255 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
4256 static void setup_zone_pageset(struct zone
*zone
);
4259 * Global mutex to protect against size modification of zonelists
4260 * as well as to serialize pageset setup for the new populated zone.
4262 DEFINE_MUTEX(zonelists_mutex
);
4264 /* return values int ....just for stop_machine() */
4265 static int __build_all_zonelists(void *data
)
4269 pg_data_t
*self
= data
;
4272 memset(node_load
, 0, sizeof(node_load
));
4275 if (self
&& !node_online(self
->node_id
)) {
4276 build_zonelists(self
);
4279 for_each_online_node(nid
) {
4280 pg_data_t
*pgdat
= NODE_DATA(nid
);
4282 build_zonelists(pgdat
);
4286 * Initialize the boot_pagesets that are going to be used
4287 * for bootstrapping processors. The real pagesets for
4288 * each zone will be allocated later when the per cpu
4289 * allocator is available.
4291 * boot_pagesets are used also for bootstrapping offline
4292 * cpus if the system is already booted because the pagesets
4293 * are needed to initialize allocators on a specific cpu too.
4294 * F.e. the percpu allocator needs the page allocator which
4295 * needs the percpu allocator in order to allocate its pagesets
4296 * (a chicken-egg dilemma).
4298 for_each_possible_cpu(cpu
) {
4299 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
4301 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
4303 * We now know the "local memory node" for each node--
4304 * i.e., the node of the first zone in the generic zonelist.
4305 * Set up numa_mem percpu variable for on-line cpus. During
4306 * boot, only the boot cpu should be on-line; we'll init the
4307 * secondary cpus' numa_mem as they come on-line. During
4308 * node/memory hotplug, we'll fixup all on-line cpus.
4310 if (cpu_online(cpu
))
4311 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
4318 static noinline
void __init
4319 build_all_zonelists_init(void)
4321 __build_all_zonelists(NULL
);
4322 mminit_verify_zonelist();
4323 cpuset_init_current_mems_allowed();
4327 * Called with zonelists_mutex held always
4328 * unless system_state == SYSTEM_BOOTING.
4330 * __ref due to (1) call of __meminit annotated setup_zone_pageset
4331 * [we're only called with non-NULL zone through __meminit paths] and
4332 * (2) call of __init annotated helper build_all_zonelists_init
4333 * [protected by SYSTEM_BOOTING].
4335 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
4337 set_zonelist_order();
4339 if (system_state
== SYSTEM_BOOTING
) {
4340 build_all_zonelists_init();
4342 #ifdef CONFIG_MEMORY_HOTPLUG
4344 setup_zone_pageset(zone
);
4346 /* we have to stop all cpus to guarantee there is no user
4348 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
4349 /* cpuset refresh routine should be here */
4351 vm_total_pages
= nr_free_pagecache_pages();
4353 * Disable grouping by mobility if the number of pages in the
4354 * system is too low to allow the mechanism to work. It would be
4355 * more accurate, but expensive to check per-zone. This check is
4356 * made on memory-hotadd so a system can start with mobility
4357 * disabled and enable it later
4359 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
4360 page_group_by_mobility_disabled
= 1;
4362 page_group_by_mobility_disabled
= 0;
4364 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
4365 "Total pages: %ld\n",
4367 zonelist_order_name
[current_zonelist_order
],
4368 page_group_by_mobility_disabled
? "off" : "on",
4371 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
4376 * Helper functions to size the waitqueue hash table.
4377 * Essentially these want to choose hash table sizes sufficiently
4378 * large so that collisions trying to wait on pages are rare.
4379 * But in fact, the number of active page waitqueues on typical
4380 * systems is ridiculously low, less than 200. So this is even
4381 * conservative, even though it seems large.
4383 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
4384 * waitqueues, i.e. the size of the waitq table given the number of pages.
4386 #define PAGES_PER_WAITQUEUE 256
4388 #ifndef CONFIG_MEMORY_HOTPLUG
4389 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4391 unsigned long size
= 1;
4393 pages
/= PAGES_PER_WAITQUEUE
;
4395 while (size
< pages
)
4399 * Once we have dozens or even hundreds of threads sleeping
4400 * on IO we've got bigger problems than wait queue collision.
4401 * Limit the size of the wait table to a reasonable size.
4403 size
= min(size
, 4096UL);
4405 return max(size
, 4UL);
4409 * A zone's size might be changed by hot-add, so it is not possible to determine
4410 * a suitable size for its wait_table. So we use the maximum size now.
4412 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4414 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4415 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4416 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4418 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4419 * or more by the traditional way. (See above). It equals:
4421 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4422 * ia64(16K page size) : = ( 8G + 4M)byte.
4423 * powerpc (64K page size) : = (32G +16M)byte.
4425 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4432 * This is an integer logarithm so that shifts can be used later
4433 * to extract the more random high bits from the multiplicative
4434 * hash function before the remainder is taken.
4436 static inline unsigned long wait_table_bits(unsigned long size
)
4442 * Initially all pages are reserved - free ones are freed
4443 * up by free_all_bootmem() once the early boot process is
4444 * done. Non-atomic initialization, single-pass.
4446 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4447 unsigned long start_pfn
, enum memmap_context context
)
4449 pg_data_t
*pgdat
= NODE_DATA(nid
);
4450 unsigned long end_pfn
= start_pfn
+ size
;
4453 unsigned long nr_initialised
= 0;
4455 if (highest_memmap_pfn
< end_pfn
- 1)
4456 highest_memmap_pfn
= end_pfn
- 1;
4458 z
= &pgdat
->node_zones
[zone
];
4459 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4461 * There can be holes in boot-time mem_map[]s
4462 * handed to this function. They do not
4463 * exist on hotplugged memory.
4465 if (context
== MEMMAP_EARLY
) {
4466 if (!early_pfn_valid(pfn
))
4468 if (!early_pfn_in_nid(pfn
, nid
))
4470 if (!update_defer_init(pgdat
, pfn
, end_pfn
,
4476 * Mark the block movable so that blocks are reserved for
4477 * movable at startup. This will force kernel allocations
4478 * to reserve their blocks rather than leaking throughout
4479 * the address space during boot when many long-lived
4480 * kernel allocations are made.
4482 * bitmap is created for zone's valid pfn range. but memmap
4483 * can be created for invalid pages (for alignment)
4484 * check here not to call set_pageblock_migratetype() against
4487 if (!(pfn
& (pageblock_nr_pages
- 1))) {
4488 struct page
*page
= pfn_to_page(pfn
);
4490 __init_single_page(page
, pfn
, zone
, nid
);
4491 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4493 __init_single_pfn(pfn
, zone
, nid
);
4498 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4500 unsigned int order
, t
;
4501 for_each_migratetype_order(order
, t
) {
4502 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4503 zone
->free_area
[order
].nr_free
= 0;
4507 #ifndef __HAVE_ARCH_MEMMAP_INIT
4508 #define memmap_init(size, nid, zone, start_pfn) \
4509 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4512 static int zone_batchsize(struct zone
*zone
)
4518 * The per-cpu-pages pools are set to around 1000th of the
4519 * size of the zone. But no more than 1/2 of a meg.
4521 * OK, so we don't know how big the cache is. So guess.
4523 batch
= zone
->managed_pages
/ 1024;
4524 if (batch
* PAGE_SIZE
> 512 * 1024)
4525 batch
= (512 * 1024) / PAGE_SIZE
;
4526 batch
/= 4; /* We effectively *= 4 below */
4531 * Clamp the batch to a 2^n - 1 value. Having a power
4532 * of 2 value was found to be more likely to have
4533 * suboptimal cache aliasing properties in some cases.
4535 * For example if 2 tasks are alternately allocating
4536 * batches of pages, one task can end up with a lot
4537 * of pages of one half of the possible page colors
4538 * and the other with pages of the other colors.
4540 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4545 /* The deferral and batching of frees should be suppressed under NOMMU
4548 * The problem is that NOMMU needs to be able to allocate large chunks
4549 * of contiguous memory as there's no hardware page translation to
4550 * assemble apparent contiguous memory from discontiguous pages.
4552 * Queueing large contiguous runs of pages for batching, however,
4553 * causes the pages to actually be freed in smaller chunks. As there
4554 * can be a significant delay between the individual batches being
4555 * recycled, this leads to the once large chunks of space being
4556 * fragmented and becoming unavailable for high-order allocations.
4563 * pcp->high and pcp->batch values are related and dependent on one another:
4564 * ->batch must never be higher then ->high.
4565 * The following function updates them in a safe manner without read side
4568 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4569 * those fields changing asynchronously (acording the the above rule).
4571 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4572 * outside of boot time (or some other assurance that no concurrent updaters
4575 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4576 unsigned long batch
)
4578 /* start with a fail safe value for batch */
4582 /* Update high, then batch, in order */
4589 /* a companion to pageset_set_high() */
4590 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4592 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4595 static void pageset_init(struct per_cpu_pageset
*p
)
4597 struct per_cpu_pages
*pcp
;
4600 memset(p
, 0, sizeof(*p
));
4604 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4605 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4608 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4611 pageset_set_batch(p
, batch
);
4615 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4616 * to the value high for the pageset p.
4618 static void pageset_set_high(struct per_cpu_pageset
*p
,
4621 unsigned long batch
= max(1UL, high
/ 4);
4622 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4623 batch
= PAGE_SHIFT
* 8;
4625 pageset_update(&p
->pcp
, high
, batch
);
4628 static void pageset_set_high_and_batch(struct zone
*zone
,
4629 struct per_cpu_pageset
*pcp
)
4631 if (percpu_pagelist_fraction
)
4632 pageset_set_high(pcp
,
4633 (zone
->managed_pages
/
4634 percpu_pagelist_fraction
));
4636 pageset_set_batch(pcp
, zone_batchsize(zone
));
4639 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4641 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4644 pageset_set_high_and_batch(zone
, pcp
);
4647 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4650 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4651 for_each_possible_cpu(cpu
)
4652 zone_pageset_init(zone
, cpu
);
4656 * Allocate per cpu pagesets and initialize them.
4657 * Before this call only boot pagesets were available.
4659 void __init
setup_per_cpu_pageset(void)
4663 for_each_populated_zone(zone
)
4664 setup_zone_pageset(zone
);
4667 static noinline __init_refok
4668 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4674 * The per-page waitqueue mechanism uses hashed waitqueues
4677 zone
->wait_table_hash_nr_entries
=
4678 wait_table_hash_nr_entries(zone_size_pages
);
4679 zone
->wait_table_bits
=
4680 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4681 alloc_size
= zone
->wait_table_hash_nr_entries
4682 * sizeof(wait_queue_head_t
);
4684 if (!slab_is_available()) {
4685 zone
->wait_table
= (wait_queue_head_t
*)
4686 memblock_virt_alloc_node_nopanic(
4687 alloc_size
, zone
->zone_pgdat
->node_id
);
4690 * This case means that a zone whose size was 0 gets new memory
4691 * via memory hot-add.
4692 * But it may be the case that a new node was hot-added. In
4693 * this case vmalloc() will not be able to use this new node's
4694 * memory - this wait_table must be initialized to use this new
4695 * node itself as well.
4696 * To use this new node's memory, further consideration will be
4699 zone
->wait_table
= vmalloc(alloc_size
);
4701 if (!zone
->wait_table
)
4704 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4705 init_waitqueue_head(zone
->wait_table
+ i
);
4710 static __meminit
void zone_pcp_init(struct zone
*zone
)
4713 * per cpu subsystem is not up at this point. The following code
4714 * relies on the ability of the linker to provide the
4715 * offset of a (static) per cpu variable into the per cpu area.
4717 zone
->pageset
= &boot_pageset
;
4719 if (populated_zone(zone
))
4720 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4721 zone
->name
, zone
->present_pages
,
4722 zone_batchsize(zone
));
4725 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4726 unsigned long zone_start_pfn
,
4729 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4731 ret
= zone_wait_table_init(zone
, size
);
4734 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4736 zone
->zone_start_pfn
= zone_start_pfn
;
4738 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4739 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4741 (unsigned long)zone_idx(zone
),
4742 zone_start_pfn
, (zone_start_pfn
+ size
));
4744 zone_init_free_lists(zone
);
4749 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4750 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4753 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4755 int __meminit
__early_pfn_to_nid(unsigned long pfn
,
4756 struct mminit_pfnnid_cache
*state
)
4758 unsigned long start_pfn
, end_pfn
;
4761 if (state
->last_start
<= pfn
&& pfn
< state
->last_end
)
4762 return state
->last_nid
;
4764 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4766 state
->last_start
= start_pfn
;
4767 state
->last_end
= end_pfn
;
4768 state
->last_nid
= nid
;
4773 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4776 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4777 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4778 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4780 * If an architecture guarantees that all ranges registered contain no holes
4781 * and may be freed, this this function may be used instead of calling
4782 * memblock_free_early_nid() manually.
4784 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4786 unsigned long start_pfn
, end_pfn
;
4789 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4790 start_pfn
= min(start_pfn
, max_low_pfn
);
4791 end_pfn
= min(end_pfn
, max_low_pfn
);
4793 if (start_pfn
< end_pfn
)
4794 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4795 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4801 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4802 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4804 * If an architecture guarantees that all ranges registered contain no holes and may
4805 * be freed, this function may be used instead of calling memory_present() manually.
4807 void __init
sparse_memory_present_with_active_regions(int nid
)
4809 unsigned long start_pfn
, end_pfn
;
4812 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4813 memory_present(this_nid
, start_pfn
, end_pfn
);
4817 * get_pfn_range_for_nid - Return the start and end page frames for a node
4818 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4819 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4820 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4822 * It returns the start and end page frame of a node based on information
4823 * provided by memblock_set_node(). If called for a node
4824 * with no available memory, a warning is printed and the start and end
4827 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4828 unsigned long *start_pfn
, unsigned long *end_pfn
)
4830 unsigned long this_start_pfn
, this_end_pfn
;
4836 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4837 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4838 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4841 if (*start_pfn
== -1UL)
4846 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4847 * assumption is made that zones within a node are ordered in monotonic
4848 * increasing memory addresses so that the "highest" populated zone is used
4850 static void __init
find_usable_zone_for_movable(void)
4853 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4854 if (zone_index
== ZONE_MOVABLE
)
4857 if (arch_zone_highest_possible_pfn
[zone_index
] >
4858 arch_zone_lowest_possible_pfn
[zone_index
])
4862 VM_BUG_ON(zone_index
== -1);
4863 movable_zone
= zone_index
;
4867 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4868 * because it is sized independent of architecture. Unlike the other zones,
4869 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4870 * in each node depending on the size of each node and how evenly kernelcore
4871 * is distributed. This helper function adjusts the zone ranges
4872 * provided by the architecture for a given node by using the end of the
4873 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4874 * zones within a node are in order of monotonic increases memory addresses
4876 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4877 unsigned long zone_type
,
4878 unsigned long node_start_pfn
,
4879 unsigned long node_end_pfn
,
4880 unsigned long *zone_start_pfn
,
4881 unsigned long *zone_end_pfn
)
4883 /* Only adjust if ZONE_MOVABLE is on this node */
4884 if (zone_movable_pfn
[nid
]) {
4885 /* Size ZONE_MOVABLE */
4886 if (zone_type
== ZONE_MOVABLE
) {
4887 *zone_start_pfn
= zone_movable_pfn
[nid
];
4888 *zone_end_pfn
= min(node_end_pfn
,
4889 arch_zone_highest_possible_pfn
[movable_zone
]);
4891 /* Adjust for ZONE_MOVABLE starting within this range */
4892 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4893 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4894 *zone_end_pfn
= zone_movable_pfn
[nid
];
4896 /* Check if this whole range is within ZONE_MOVABLE */
4897 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4898 *zone_start_pfn
= *zone_end_pfn
;
4903 * Return the number of pages a zone spans in a node, including holes
4904 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4906 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4907 unsigned long zone_type
,
4908 unsigned long node_start_pfn
,
4909 unsigned long node_end_pfn
,
4910 unsigned long *ignored
)
4912 unsigned long zone_start_pfn
, zone_end_pfn
;
4914 /* When hotadd a new node from cpu_up(), the node should be empty */
4915 if (!node_start_pfn
&& !node_end_pfn
)
4918 /* Get the start and end of the zone */
4919 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4920 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4921 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4922 node_start_pfn
, node_end_pfn
,
4923 &zone_start_pfn
, &zone_end_pfn
);
4925 /* Check that this node has pages within the zone's required range */
4926 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4929 /* Move the zone boundaries inside the node if necessary */
4930 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4931 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4933 /* Return the spanned pages */
4934 return zone_end_pfn
- zone_start_pfn
;
4938 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4939 * then all holes in the requested range will be accounted for.
4941 unsigned long __meminit
__absent_pages_in_range(int nid
,
4942 unsigned long range_start_pfn
,
4943 unsigned long range_end_pfn
)
4945 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4946 unsigned long start_pfn
, end_pfn
;
4949 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4950 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4951 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4952 nr_absent
-= end_pfn
- start_pfn
;
4958 * absent_pages_in_range - Return number of page frames in holes within a range
4959 * @start_pfn: The start PFN to start searching for holes
4960 * @end_pfn: The end PFN to stop searching for holes
4962 * It returns the number of pages frames in memory holes within a range.
4964 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4965 unsigned long end_pfn
)
4967 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4970 /* Return the number of page frames in holes in a zone on a node */
4971 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4972 unsigned long zone_type
,
4973 unsigned long node_start_pfn
,
4974 unsigned long node_end_pfn
,
4975 unsigned long *ignored
)
4977 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4978 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4979 unsigned long zone_start_pfn
, zone_end_pfn
;
4981 /* When hotadd a new node from cpu_up(), the node should be empty */
4982 if (!node_start_pfn
&& !node_end_pfn
)
4985 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4986 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4988 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4989 node_start_pfn
, node_end_pfn
,
4990 &zone_start_pfn
, &zone_end_pfn
);
4991 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4994 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4995 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4996 unsigned long zone_type
,
4997 unsigned long node_start_pfn
,
4998 unsigned long node_end_pfn
,
4999 unsigned long *zones_size
)
5001 return zones_size
[zone_type
];
5004 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
5005 unsigned long zone_type
,
5006 unsigned long node_start_pfn
,
5007 unsigned long node_end_pfn
,
5008 unsigned long *zholes_size
)
5013 return zholes_size
[zone_type
];
5016 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5018 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
5019 unsigned long node_start_pfn
,
5020 unsigned long node_end_pfn
,
5021 unsigned long *zones_size
,
5022 unsigned long *zholes_size
)
5024 unsigned long realtotalpages
= 0, totalpages
= 0;
5027 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5028 struct zone
*zone
= pgdat
->node_zones
+ i
;
5029 unsigned long size
, real_size
;
5031 size
= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
5035 real_size
= size
- zone_absent_pages_in_node(pgdat
->node_id
, i
,
5036 node_start_pfn
, node_end_pfn
,
5038 zone
->spanned_pages
= size
;
5039 zone
->present_pages
= real_size
;
5042 realtotalpages
+= real_size
;
5045 pgdat
->node_spanned_pages
= totalpages
;
5046 pgdat
->node_present_pages
= realtotalpages
;
5047 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
5051 #ifndef CONFIG_SPARSEMEM
5053 * Calculate the size of the zone->blockflags rounded to an unsigned long
5054 * Start by making sure zonesize is a multiple of pageblock_order by rounding
5055 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
5056 * round what is now in bits to nearest long in bits, then return it in
5059 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
5061 unsigned long usemapsize
;
5063 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
5064 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
5065 usemapsize
= usemapsize
>> pageblock_order
;
5066 usemapsize
*= NR_PAGEBLOCK_BITS
;
5067 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
5069 return usemapsize
/ 8;
5072 static void __init
setup_usemap(struct pglist_data
*pgdat
,
5074 unsigned long zone_start_pfn
,
5075 unsigned long zonesize
)
5077 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
5078 zone
->pageblock_flags
= NULL
;
5080 zone
->pageblock_flags
=
5081 memblock_virt_alloc_node_nopanic(usemapsize
,
5085 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
5086 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
5087 #endif /* CONFIG_SPARSEMEM */
5089 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
5091 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
5092 void __paginginit
set_pageblock_order(void)
5096 /* Check that pageblock_nr_pages has not already been setup */
5097 if (pageblock_order
)
5100 if (HPAGE_SHIFT
> PAGE_SHIFT
)
5101 order
= HUGETLB_PAGE_ORDER
;
5103 order
= MAX_ORDER
- 1;
5106 * Assume the largest contiguous order of interest is a huge page.
5107 * This value may be variable depending on boot parameters on IA64 and
5110 pageblock_order
= order
;
5112 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5115 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
5116 * is unused as pageblock_order is set at compile-time. See
5117 * include/linux/pageblock-flags.h for the values of pageblock_order based on
5120 void __paginginit
set_pageblock_order(void)
5124 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
5126 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
5127 unsigned long present_pages
)
5129 unsigned long pages
= spanned_pages
;
5132 * Provide a more accurate estimation if there are holes within
5133 * the zone and SPARSEMEM is in use. If there are holes within the
5134 * zone, each populated memory region may cost us one or two extra
5135 * memmap pages due to alignment because memmap pages for each
5136 * populated regions may not naturally algined on page boundary.
5137 * So the (present_pages >> 4) heuristic is a tradeoff for that.
5139 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
5140 IS_ENABLED(CONFIG_SPARSEMEM
))
5141 pages
= present_pages
;
5143 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
5147 * Set up the zone data structures:
5148 * - mark all pages reserved
5149 * - mark all memory queues empty
5150 * - clear the memory bitmaps
5152 * NOTE: pgdat should get zeroed by caller.
5154 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
)
5157 int nid
= pgdat
->node_id
;
5158 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
5161 pgdat_resize_init(pgdat
);
5162 #ifdef CONFIG_NUMA_BALANCING
5163 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
5164 pgdat
->numabalancing_migrate_nr_pages
= 0;
5165 pgdat
->numabalancing_migrate_next_window
= jiffies
;
5167 init_waitqueue_head(&pgdat
->kswapd_wait
);
5168 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
5169 pgdat_page_ext_init(pgdat
);
5171 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5172 struct zone
*zone
= pgdat
->node_zones
+ j
;
5173 unsigned long size
, realsize
, freesize
, memmap_pages
;
5175 size
= zone
->spanned_pages
;
5176 realsize
= freesize
= zone
->present_pages
;
5179 * Adjust freesize so that it accounts for how much memory
5180 * is used by this zone for memmap. This affects the watermark
5181 * and per-cpu initialisations
5183 memmap_pages
= calc_memmap_size(size
, realsize
);
5184 if (!is_highmem_idx(j
)) {
5185 if (freesize
>= memmap_pages
) {
5186 freesize
-= memmap_pages
;
5189 " %s zone: %lu pages used for memmap\n",
5190 zone_names
[j
], memmap_pages
);
5193 " %s zone: %lu pages exceeds freesize %lu\n",
5194 zone_names
[j
], memmap_pages
, freesize
);
5197 /* Account for reserved pages */
5198 if (j
== 0 && freesize
> dma_reserve
) {
5199 freesize
-= dma_reserve
;
5200 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
5201 zone_names
[0], dma_reserve
);
5204 if (!is_highmem_idx(j
))
5205 nr_kernel_pages
+= freesize
;
5206 /* Charge for highmem memmap if there are enough kernel pages */
5207 else if (nr_kernel_pages
> memmap_pages
* 2)
5208 nr_kernel_pages
-= memmap_pages
;
5209 nr_all_pages
+= freesize
;
5212 * Set an approximate value for lowmem here, it will be adjusted
5213 * when the bootmem allocator frees pages into the buddy system.
5214 * And all highmem pages will be managed by the buddy system.
5216 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
5219 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
5221 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
5223 zone
->name
= zone_names
[j
];
5224 spin_lock_init(&zone
->lock
);
5225 spin_lock_init(&zone
->lru_lock
);
5226 zone_seqlock_init(zone
);
5227 zone
->zone_pgdat
= pgdat
;
5228 zone_pcp_init(zone
);
5230 /* For bootup, initialized properly in watermark setup */
5231 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
5233 lruvec_init(&zone
->lruvec
);
5237 set_pageblock_order();
5238 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
5239 ret
= init_currently_empty_zone(zone
, zone_start_pfn
, size
);
5241 memmap_init(size
, nid
, j
, zone_start_pfn
);
5242 zone_start_pfn
+= size
;
5246 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
5248 unsigned long __maybe_unused start
= 0;
5249 unsigned long __maybe_unused offset
= 0;
5251 /* Skip empty nodes */
5252 if (!pgdat
->node_spanned_pages
)
5255 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5256 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5257 offset
= pgdat
->node_start_pfn
- start
;
5258 /* ia64 gets its own node_mem_map, before this, without bootmem */
5259 if (!pgdat
->node_mem_map
) {
5260 unsigned long size
, end
;
5264 * The zone's endpoints aren't required to be MAX_ORDER
5265 * aligned but the node_mem_map endpoints must be in order
5266 * for the buddy allocator to function correctly.
5268 end
= pgdat_end_pfn(pgdat
);
5269 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5270 size
= (end
- start
) * sizeof(struct page
);
5271 map
= alloc_remap(pgdat
->node_id
, size
);
5273 map
= memblock_virt_alloc_node_nopanic(size
,
5275 pgdat
->node_mem_map
= map
+ offset
;
5277 #ifndef CONFIG_NEED_MULTIPLE_NODES
5279 * With no DISCONTIG, the global mem_map is just set as node 0's
5281 if (pgdat
== NODE_DATA(0)) {
5282 mem_map
= NODE_DATA(0)->node_mem_map
;
5283 #if defined(CONFIG_HAVE_MEMBLOCK_NODE_MAP) || defined(CONFIG_FLATMEM)
5284 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5286 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5289 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5292 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5293 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5295 pg_data_t
*pgdat
= NODE_DATA(nid
);
5296 unsigned long start_pfn
= 0;
5297 unsigned long end_pfn
= 0;
5299 /* pg_data_t should be reset to zero when it's allocated */
5300 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
5302 reset_deferred_meminit(pgdat
);
5303 pgdat
->node_id
= nid
;
5304 pgdat
->node_start_pfn
= node_start_pfn
;
5305 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5306 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5307 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5308 (u64
)start_pfn
<< PAGE_SHIFT
,
5309 end_pfn
? ((u64
)end_pfn
<< PAGE_SHIFT
) - 1 : 0);
5311 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5312 zones_size
, zholes_size
);
5314 alloc_node_mem_map(pgdat
);
5315 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5316 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5317 nid
, (unsigned long)pgdat
,
5318 (unsigned long)pgdat
->node_mem_map
);
5321 free_area_init_core(pgdat
);
5324 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5326 #if MAX_NUMNODES > 1
5328 * Figure out the number of possible node ids.
5330 void __init
setup_nr_node_ids(void)
5332 unsigned int highest
;
5334 highest
= find_last_bit(node_possible_map
.bits
, MAX_NUMNODES
);
5335 nr_node_ids
= highest
+ 1;
5340 * node_map_pfn_alignment - determine the maximum internode alignment
5342 * This function should be called after node map is populated and sorted.
5343 * It calculates the maximum power of two alignment which can distinguish
5346 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5347 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5348 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5349 * shifted, 1GiB is enough and this function will indicate so.
5351 * This is used to test whether pfn -> nid mapping of the chosen memory
5352 * model has fine enough granularity to avoid incorrect mapping for the
5353 * populated node map.
5355 * Returns the determined alignment in pfn's. 0 if there is no alignment
5356 * requirement (single node).
5358 unsigned long __init
node_map_pfn_alignment(void)
5360 unsigned long accl_mask
= 0, last_end
= 0;
5361 unsigned long start
, end
, mask
;
5365 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5366 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5373 * Start with a mask granular enough to pin-point to the
5374 * start pfn and tick off bits one-by-one until it becomes
5375 * too coarse to separate the current node from the last.
5377 mask
= ~((1 << __ffs(start
)) - 1);
5378 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5381 /* accumulate all internode masks */
5385 /* convert mask to number of pages */
5386 return ~accl_mask
+ 1;
5389 /* Find the lowest pfn for a node */
5390 static unsigned long __init
find_min_pfn_for_node(int nid
)
5392 unsigned long min_pfn
= ULONG_MAX
;
5393 unsigned long start_pfn
;
5396 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5397 min_pfn
= min(min_pfn
, start_pfn
);
5399 if (min_pfn
== ULONG_MAX
) {
5401 "Could not find start_pfn for node %d\n", nid
);
5409 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5411 * It returns the minimum PFN based on information provided via
5412 * memblock_set_node().
5414 unsigned long __init
find_min_pfn_with_active_regions(void)
5416 return find_min_pfn_for_node(MAX_NUMNODES
);
5420 * early_calculate_totalpages()
5421 * Sum pages in active regions for movable zone.
5422 * Populate N_MEMORY for calculating usable_nodes.
5424 static unsigned long __init
early_calculate_totalpages(void)
5426 unsigned long totalpages
= 0;
5427 unsigned long start_pfn
, end_pfn
;
5430 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5431 unsigned long pages
= end_pfn
- start_pfn
;
5433 totalpages
+= pages
;
5435 node_set_state(nid
, N_MEMORY
);
5441 * Find the PFN the Movable zone begins in each node. Kernel memory
5442 * is spread evenly between nodes as long as the nodes have enough
5443 * memory. When they don't, some nodes will have more kernelcore than
5446 static void __init
find_zone_movable_pfns_for_nodes(void)
5449 unsigned long usable_startpfn
;
5450 unsigned long kernelcore_node
, kernelcore_remaining
;
5451 /* save the state before borrow the nodemask */
5452 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5453 unsigned long totalpages
= early_calculate_totalpages();
5454 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5455 struct memblock_region
*r
;
5457 /* Need to find movable_zone earlier when movable_node is specified. */
5458 find_usable_zone_for_movable();
5461 * If movable_node is specified, ignore kernelcore and movablecore
5464 if (movable_node_is_enabled()) {
5465 for_each_memblock(memory
, r
) {
5466 if (!memblock_is_hotpluggable(r
))
5471 usable_startpfn
= PFN_DOWN(r
->base
);
5472 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5473 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5481 * If movablecore=nn[KMG] was specified, calculate what size of
5482 * kernelcore that corresponds so that memory usable for
5483 * any allocation type is evenly spread. If both kernelcore
5484 * and movablecore are specified, then the value of kernelcore
5485 * will be used for required_kernelcore if it's greater than
5486 * what movablecore would have allowed.
5488 if (required_movablecore
) {
5489 unsigned long corepages
;
5492 * Round-up so that ZONE_MOVABLE is at least as large as what
5493 * was requested by the user
5495 required_movablecore
=
5496 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5497 required_movablecore
= min(totalpages
, required_movablecore
);
5498 corepages
= totalpages
- required_movablecore
;
5500 required_kernelcore
= max(required_kernelcore
, corepages
);
5504 * If kernelcore was not specified or kernelcore size is larger
5505 * than totalpages, there is no ZONE_MOVABLE.
5507 if (!required_kernelcore
|| required_kernelcore
>= totalpages
)
5510 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5511 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5514 /* Spread kernelcore memory as evenly as possible throughout nodes */
5515 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5516 for_each_node_state(nid
, N_MEMORY
) {
5517 unsigned long start_pfn
, end_pfn
;
5520 * Recalculate kernelcore_node if the division per node
5521 * now exceeds what is necessary to satisfy the requested
5522 * amount of memory for the kernel
5524 if (required_kernelcore
< kernelcore_node
)
5525 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5528 * As the map is walked, we track how much memory is usable
5529 * by the kernel using kernelcore_remaining. When it is
5530 * 0, the rest of the node is usable by ZONE_MOVABLE
5532 kernelcore_remaining
= kernelcore_node
;
5534 /* Go through each range of PFNs within this node */
5535 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5536 unsigned long size_pages
;
5538 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5539 if (start_pfn
>= end_pfn
)
5542 /* Account for what is only usable for kernelcore */
5543 if (start_pfn
< usable_startpfn
) {
5544 unsigned long kernel_pages
;
5545 kernel_pages
= min(end_pfn
, usable_startpfn
)
5548 kernelcore_remaining
-= min(kernel_pages
,
5549 kernelcore_remaining
);
5550 required_kernelcore
-= min(kernel_pages
,
5551 required_kernelcore
);
5553 /* Continue if range is now fully accounted */
5554 if (end_pfn
<= usable_startpfn
) {
5557 * Push zone_movable_pfn to the end so
5558 * that if we have to rebalance
5559 * kernelcore across nodes, we will
5560 * not double account here
5562 zone_movable_pfn
[nid
] = end_pfn
;
5565 start_pfn
= usable_startpfn
;
5569 * The usable PFN range for ZONE_MOVABLE is from
5570 * start_pfn->end_pfn. Calculate size_pages as the
5571 * number of pages used as kernelcore
5573 size_pages
= end_pfn
- start_pfn
;
5574 if (size_pages
> kernelcore_remaining
)
5575 size_pages
= kernelcore_remaining
;
5576 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5579 * Some kernelcore has been met, update counts and
5580 * break if the kernelcore for this node has been
5583 required_kernelcore
-= min(required_kernelcore
,
5585 kernelcore_remaining
-= size_pages
;
5586 if (!kernelcore_remaining
)
5592 * If there is still required_kernelcore, we do another pass with one
5593 * less node in the count. This will push zone_movable_pfn[nid] further
5594 * along on the nodes that still have memory until kernelcore is
5598 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5602 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5603 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5604 zone_movable_pfn
[nid
] =
5605 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5608 /* restore the node_state */
5609 node_states
[N_MEMORY
] = saved_node_state
;
5612 /* Any regular or high memory on that node ? */
5613 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5615 enum zone_type zone_type
;
5617 if (N_MEMORY
== N_NORMAL_MEMORY
)
5620 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5621 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5622 if (populated_zone(zone
)) {
5623 node_set_state(nid
, N_HIGH_MEMORY
);
5624 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5625 zone_type
<= ZONE_NORMAL
)
5626 node_set_state(nid
, N_NORMAL_MEMORY
);
5633 * free_area_init_nodes - Initialise all pg_data_t and zone data
5634 * @max_zone_pfn: an array of max PFNs for each zone
5636 * This will call free_area_init_node() for each active node in the system.
5637 * Using the page ranges provided by memblock_set_node(), the size of each
5638 * zone in each node and their holes is calculated. If the maximum PFN
5639 * between two adjacent zones match, it is assumed that the zone is empty.
5640 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5641 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5642 * starts where the previous one ended. For example, ZONE_DMA32 starts
5643 * at arch_max_dma_pfn.
5645 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5647 unsigned long start_pfn
, end_pfn
;
5650 /* Record where the zone boundaries are */
5651 memset(arch_zone_lowest_possible_pfn
, 0,
5652 sizeof(arch_zone_lowest_possible_pfn
));
5653 memset(arch_zone_highest_possible_pfn
, 0,
5654 sizeof(arch_zone_highest_possible_pfn
));
5655 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5656 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5657 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5658 if (i
== ZONE_MOVABLE
)
5660 arch_zone_lowest_possible_pfn
[i
] =
5661 arch_zone_highest_possible_pfn
[i
-1];
5662 arch_zone_highest_possible_pfn
[i
] =
5663 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5665 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5666 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5668 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5669 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5670 find_zone_movable_pfns_for_nodes();
5672 /* Print out the zone ranges */
5673 pr_info("Zone ranges:\n");
5674 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5675 if (i
== ZONE_MOVABLE
)
5677 pr_info(" %-8s ", zone_names
[i
]);
5678 if (arch_zone_lowest_possible_pfn
[i
] ==
5679 arch_zone_highest_possible_pfn
[i
])
5682 pr_cont("[mem %#018Lx-%#018Lx]\n",
5683 (u64
)arch_zone_lowest_possible_pfn
[i
]
5685 ((u64
)arch_zone_highest_possible_pfn
[i
]
5686 << PAGE_SHIFT
) - 1);
5689 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5690 pr_info("Movable zone start for each node\n");
5691 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5692 if (zone_movable_pfn
[i
])
5693 pr_info(" Node %d: %#018Lx\n", i
,
5694 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
5697 /* Print out the early node map */
5698 pr_info("Early memory node ranges\n");
5699 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5700 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
5701 (u64
)start_pfn
<< PAGE_SHIFT
,
5702 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5704 /* Initialise every node */
5705 mminit_verify_pageflags_layout();
5706 setup_nr_node_ids();
5707 for_each_online_node(nid
) {
5708 pg_data_t
*pgdat
= NODE_DATA(nid
);
5709 free_area_init_node(nid
, NULL
,
5710 find_min_pfn_for_node(nid
), NULL
);
5712 /* Any memory on that node */
5713 if (pgdat
->node_present_pages
)
5714 node_set_state(nid
, N_MEMORY
);
5715 check_for_memory(pgdat
, nid
);
5719 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5721 unsigned long long coremem
;
5725 coremem
= memparse(p
, &p
);
5726 *core
= coremem
>> PAGE_SHIFT
;
5728 /* Paranoid check that UL is enough for the coremem value */
5729 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5735 * kernelcore=size sets the amount of memory for use for allocations that
5736 * cannot be reclaimed or migrated.
5738 static int __init
cmdline_parse_kernelcore(char *p
)
5740 return cmdline_parse_core(p
, &required_kernelcore
);
5744 * movablecore=size sets the amount of memory for use for allocations that
5745 * can be reclaimed or migrated.
5747 static int __init
cmdline_parse_movablecore(char *p
)
5749 return cmdline_parse_core(p
, &required_movablecore
);
5752 early_param("kernelcore", cmdline_parse_kernelcore
);
5753 early_param("movablecore", cmdline_parse_movablecore
);
5755 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5757 void adjust_managed_page_count(struct page
*page
, long count
)
5759 spin_lock(&managed_page_count_lock
);
5760 page_zone(page
)->managed_pages
+= count
;
5761 totalram_pages
+= count
;
5762 #ifdef CONFIG_HIGHMEM
5763 if (PageHighMem(page
))
5764 totalhigh_pages
+= count
;
5766 spin_unlock(&managed_page_count_lock
);
5768 EXPORT_SYMBOL(adjust_managed_page_count
);
5770 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5773 unsigned long pages
= 0;
5775 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5776 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5777 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5778 if ((unsigned int)poison
<= 0xFF)
5779 memset(pos
, poison
, PAGE_SIZE
);
5780 free_reserved_page(virt_to_page(pos
));
5784 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5785 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5789 EXPORT_SYMBOL(free_reserved_area
);
5791 #ifdef CONFIG_HIGHMEM
5792 void free_highmem_page(struct page
*page
)
5794 __free_reserved_page(page
);
5796 page_zone(page
)->managed_pages
++;
5802 void __init
mem_init_print_info(const char *str
)
5804 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5805 unsigned long init_code_size
, init_data_size
;
5807 physpages
= get_num_physpages();
5808 codesize
= _etext
- _stext
;
5809 datasize
= _edata
- _sdata
;
5810 rosize
= __end_rodata
- __start_rodata
;
5811 bss_size
= __bss_stop
- __bss_start
;
5812 init_data_size
= __init_end
- __init_begin
;
5813 init_code_size
= _einittext
- _sinittext
;
5816 * Detect special cases and adjust section sizes accordingly:
5817 * 1) .init.* may be embedded into .data sections
5818 * 2) .init.text.* may be out of [__init_begin, __init_end],
5819 * please refer to arch/tile/kernel/vmlinux.lds.S.
5820 * 3) .rodata.* may be embedded into .text or .data sections.
5822 #define adj_init_size(start, end, size, pos, adj) \
5824 if (start <= pos && pos < end && size > adj) \
5828 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5829 _sinittext
, init_code_size
);
5830 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5831 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5832 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5833 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5835 #undef adj_init_size
5837 pr_info("Memory: %luK/%luK available "
5838 "(%luK kernel code, %luK rwdata, %luK rodata, "
5839 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5840 #ifdef CONFIG_HIGHMEM
5844 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5845 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5846 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5847 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
-10),
5848 totalcma_pages
<< (PAGE_SHIFT
-10),
5849 #ifdef CONFIG_HIGHMEM
5850 totalhigh_pages
<< (PAGE_SHIFT
-10),
5852 str
? ", " : "", str
? str
: "");
5856 * set_dma_reserve - set the specified number of pages reserved in the first zone
5857 * @new_dma_reserve: The number of pages to mark reserved
5859 * The per-cpu batchsize and zone watermarks are determined by managed_pages.
5860 * In the DMA zone, a significant percentage may be consumed by kernel image
5861 * and other unfreeable allocations which can skew the watermarks badly. This
5862 * function may optionally be used to account for unfreeable pages in the
5863 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5864 * smaller per-cpu batchsize.
5866 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5868 dma_reserve
= new_dma_reserve
;
5871 void __init
free_area_init(unsigned long *zones_size
)
5873 free_area_init_node(0, zones_size
,
5874 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5877 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5878 unsigned long action
, void *hcpu
)
5880 int cpu
= (unsigned long)hcpu
;
5882 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5883 lru_add_drain_cpu(cpu
);
5887 * Spill the event counters of the dead processor
5888 * into the current processors event counters.
5889 * This artificially elevates the count of the current
5892 vm_events_fold_cpu(cpu
);
5895 * Zero the differential counters of the dead processor
5896 * so that the vm statistics are consistent.
5898 * This is only okay since the processor is dead and cannot
5899 * race with what we are doing.
5901 cpu_vm_stats_fold(cpu
);
5906 void __init
page_alloc_init(void)
5908 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5912 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5913 * or min_free_kbytes changes.
5915 static void calculate_totalreserve_pages(void)
5917 struct pglist_data
*pgdat
;
5918 unsigned long reserve_pages
= 0;
5919 enum zone_type i
, j
;
5921 for_each_online_pgdat(pgdat
) {
5922 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5923 struct zone
*zone
= pgdat
->node_zones
+ i
;
5926 /* Find valid and maximum lowmem_reserve in the zone */
5927 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5928 if (zone
->lowmem_reserve
[j
] > max
)
5929 max
= zone
->lowmem_reserve
[j
];
5932 /* we treat the high watermark as reserved pages. */
5933 max
+= high_wmark_pages(zone
);
5935 if (max
> zone
->managed_pages
)
5936 max
= zone
->managed_pages
;
5938 zone
->totalreserve_pages
= max
;
5940 reserve_pages
+= max
;
5943 totalreserve_pages
= reserve_pages
;
5947 * setup_per_zone_lowmem_reserve - called whenever
5948 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5949 * has a correct pages reserved value, so an adequate number of
5950 * pages are left in the zone after a successful __alloc_pages().
5952 static void setup_per_zone_lowmem_reserve(void)
5954 struct pglist_data
*pgdat
;
5955 enum zone_type j
, idx
;
5957 for_each_online_pgdat(pgdat
) {
5958 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5959 struct zone
*zone
= pgdat
->node_zones
+ j
;
5960 unsigned long managed_pages
= zone
->managed_pages
;
5962 zone
->lowmem_reserve
[j
] = 0;
5966 struct zone
*lower_zone
;
5970 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5971 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5973 lower_zone
= pgdat
->node_zones
+ idx
;
5974 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5975 sysctl_lowmem_reserve_ratio
[idx
];
5976 managed_pages
+= lower_zone
->managed_pages
;
5981 /* update totalreserve_pages */
5982 calculate_totalreserve_pages();
5985 static void __setup_per_zone_wmarks(void)
5987 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5988 unsigned long lowmem_pages
= 0;
5990 unsigned long flags
;
5992 /* Calculate total number of !ZONE_HIGHMEM pages */
5993 for_each_zone(zone
) {
5994 if (!is_highmem(zone
))
5995 lowmem_pages
+= zone
->managed_pages
;
5998 for_each_zone(zone
) {
6001 spin_lock_irqsave(&zone
->lock
, flags
);
6002 tmp
= (u64
)pages_min
* zone
->managed_pages
;
6003 do_div(tmp
, lowmem_pages
);
6004 if (is_highmem(zone
)) {
6006 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
6007 * need highmem pages, so cap pages_min to a small
6010 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
6011 * deltas control asynch page reclaim, and so should
6012 * not be capped for highmem.
6014 unsigned long min_pages
;
6016 min_pages
= zone
->managed_pages
/ 1024;
6017 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
6018 zone
->watermark
[WMARK_MIN
] = min_pages
;
6021 * If it's a lowmem zone, reserve a number of pages
6022 * proportionate to the zone's size.
6024 zone
->watermark
[WMARK_MIN
] = tmp
;
6027 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
6028 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
6030 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
6031 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
6032 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
6034 spin_unlock_irqrestore(&zone
->lock
, flags
);
6037 /* update totalreserve_pages */
6038 calculate_totalreserve_pages();
6042 * setup_per_zone_wmarks - called when min_free_kbytes changes
6043 * or when memory is hot-{added|removed}
6045 * Ensures that the watermark[min,low,high] values for each zone are set
6046 * correctly with respect to min_free_kbytes.
6048 void setup_per_zone_wmarks(void)
6050 mutex_lock(&zonelists_mutex
);
6051 __setup_per_zone_wmarks();
6052 mutex_unlock(&zonelists_mutex
);
6056 * The inactive anon list should be small enough that the VM never has to
6057 * do too much work, but large enough that each inactive page has a chance
6058 * to be referenced again before it is swapped out.
6060 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
6061 * INACTIVE_ANON pages on this zone's LRU, maintained by the
6062 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
6063 * the anonymous pages are kept on the inactive list.
6066 * memory ratio inactive anon
6067 * -------------------------------------
6076 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
6078 unsigned int gb
, ratio
;
6080 /* Zone size in gigabytes */
6081 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
6083 ratio
= int_sqrt(10 * gb
);
6087 zone
->inactive_ratio
= ratio
;
6090 static void __meminit
setup_per_zone_inactive_ratio(void)
6095 calculate_zone_inactive_ratio(zone
);
6099 * Initialise min_free_kbytes.
6101 * For small machines we want it small (128k min). For large machines
6102 * we want it large (64MB max). But it is not linear, because network
6103 * bandwidth does not increase linearly with machine size. We use
6105 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
6106 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
6122 int __meminit
init_per_zone_wmark_min(void)
6124 unsigned long lowmem_kbytes
;
6125 int new_min_free_kbytes
;
6127 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
6128 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
6130 if (new_min_free_kbytes
> user_min_free_kbytes
) {
6131 min_free_kbytes
= new_min_free_kbytes
;
6132 if (min_free_kbytes
< 128)
6133 min_free_kbytes
= 128;
6134 if (min_free_kbytes
> 65536)
6135 min_free_kbytes
= 65536;
6137 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6138 new_min_free_kbytes
, user_min_free_kbytes
);
6140 setup_per_zone_wmarks();
6141 refresh_zone_stat_thresholds();
6142 setup_per_zone_lowmem_reserve();
6143 setup_per_zone_inactive_ratio();
6146 module_init(init_per_zone_wmark_min
)
6149 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6150 * that we can call two helper functions whenever min_free_kbytes
6153 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
6154 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6158 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6163 user_min_free_kbytes
= min_free_kbytes
;
6164 setup_per_zone_wmarks();
6170 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6171 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6176 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6181 zone
->min_unmapped_pages
= (zone
->managed_pages
*
6182 sysctl_min_unmapped_ratio
) / 100;
6186 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6187 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6192 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6197 zone
->min_slab_pages
= (zone
->managed_pages
*
6198 sysctl_min_slab_ratio
) / 100;
6204 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6205 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6206 * whenever sysctl_lowmem_reserve_ratio changes.
6208 * The reserve ratio obviously has absolutely no relation with the
6209 * minimum watermarks. The lowmem reserve ratio can only make sense
6210 * if in function of the boot time zone sizes.
6212 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
6213 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6215 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6216 setup_per_zone_lowmem_reserve();
6221 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
6222 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6223 * pagelist can have before it gets flushed back to buddy allocator.
6225 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
6226 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
6229 int old_percpu_pagelist_fraction
;
6232 mutex_lock(&pcp_batch_high_lock
);
6233 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
6235 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
6236 if (!write
|| ret
< 0)
6239 /* Sanity checking to avoid pcp imbalance */
6240 if (percpu_pagelist_fraction
&&
6241 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
6242 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
6248 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6251 for_each_populated_zone(zone
) {
6254 for_each_possible_cpu(cpu
)
6255 pageset_set_high_and_batch(zone
,
6256 per_cpu_ptr(zone
->pageset
, cpu
));
6259 mutex_unlock(&pcp_batch_high_lock
);
6264 int hashdist
= HASHDIST_DEFAULT
;
6266 static int __init
set_hashdist(char *str
)
6270 hashdist
= simple_strtoul(str
, &str
, 0);
6273 __setup("hashdist=", set_hashdist
);
6277 * allocate a large system hash table from bootmem
6278 * - it is assumed that the hash table must contain an exact power-of-2
6279 * quantity of entries
6280 * - limit is the number of hash buckets, not the total allocation size
6282 void *__init
alloc_large_system_hash(const char *tablename
,
6283 unsigned long bucketsize
,
6284 unsigned long numentries
,
6287 unsigned int *_hash_shift
,
6288 unsigned int *_hash_mask
,
6289 unsigned long low_limit
,
6290 unsigned long high_limit
)
6292 unsigned long long max
= high_limit
;
6293 unsigned long log2qty
, size
;
6296 /* allow the kernel cmdline to have a say */
6298 /* round applicable memory size up to nearest megabyte */
6299 numentries
= nr_kernel_pages
;
6301 /* It isn't necessary when PAGE_SIZE >= 1MB */
6302 if (PAGE_SHIFT
< 20)
6303 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6305 /* limit to 1 bucket per 2^scale bytes of low memory */
6306 if (scale
> PAGE_SHIFT
)
6307 numentries
>>= (scale
- PAGE_SHIFT
);
6309 numentries
<<= (PAGE_SHIFT
- scale
);
6311 /* Make sure we've got at least a 0-order allocation.. */
6312 if (unlikely(flags
& HASH_SMALL
)) {
6313 /* Makes no sense without HASH_EARLY */
6314 WARN_ON(!(flags
& HASH_EARLY
));
6315 if (!(numentries
>> *_hash_shift
)) {
6316 numentries
= 1UL << *_hash_shift
;
6317 BUG_ON(!numentries
);
6319 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6320 numentries
= PAGE_SIZE
/ bucketsize
;
6322 numentries
= roundup_pow_of_two(numentries
);
6324 /* limit allocation size to 1/16 total memory by default */
6326 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6327 do_div(max
, bucketsize
);
6329 max
= min(max
, 0x80000000ULL
);
6331 if (numentries
< low_limit
)
6332 numentries
= low_limit
;
6333 if (numentries
> max
)
6336 log2qty
= ilog2(numentries
);
6339 size
= bucketsize
<< log2qty
;
6340 if (flags
& HASH_EARLY
)
6341 table
= memblock_virt_alloc_nopanic(size
, 0);
6343 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6346 * If bucketsize is not a power-of-two, we may free
6347 * some pages at the end of hash table which
6348 * alloc_pages_exact() automatically does
6350 if (get_order(size
) < MAX_ORDER
) {
6351 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6352 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6355 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6358 panic("Failed to allocate %s hash table\n", tablename
);
6360 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6363 ilog2(size
) - PAGE_SHIFT
,
6367 *_hash_shift
= log2qty
;
6369 *_hash_mask
= (1 << log2qty
) - 1;
6374 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6375 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6378 #ifdef CONFIG_SPARSEMEM
6379 return __pfn_to_section(pfn
)->pageblock_flags
;
6381 return zone
->pageblock_flags
;
6382 #endif /* CONFIG_SPARSEMEM */
6385 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6387 #ifdef CONFIG_SPARSEMEM
6388 pfn
&= (PAGES_PER_SECTION
-1);
6389 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6391 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6392 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6393 #endif /* CONFIG_SPARSEMEM */
6397 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6398 * @page: The page within the block of interest
6399 * @pfn: The target page frame number
6400 * @end_bitidx: The last bit of interest to retrieve
6401 * @mask: mask of bits that the caller is interested in
6403 * Return: pageblock_bits flags
6405 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6406 unsigned long end_bitidx
,
6410 unsigned long *bitmap
;
6411 unsigned long bitidx
, word_bitidx
;
6414 zone
= page_zone(page
);
6415 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6416 bitidx
= pfn_to_bitidx(zone
, pfn
);
6417 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6418 bitidx
&= (BITS_PER_LONG
-1);
6420 word
= bitmap
[word_bitidx
];
6421 bitidx
+= end_bitidx
;
6422 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6426 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6427 * @page: The page within the block of interest
6428 * @flags: The flags to set
6429 * @pfn: The target page frame number
6430 * @end_bitidx: The last bit of interest
6431 * @mask: mask of bits that the caller is interested in
6433 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6435 unsigned long end_bitidx
,
6439 unsigned long *bitmap
;
6440 unsigned long bitidx
, word_bitidx
;
6441 unsigned long old_word
, word
;
6443 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6445 zone
= page_zone(page
);
6446 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6447 bitidx
= pfn_to_bitidx(zone
, pfn
);
6448 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6449 bitidx
&= (BITS_PER_LONG
-1);
6451 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6453 bitidx
+= end_bitidx
;
6454 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6455 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6457 word
= READ_ONCE(bitmap
[word_bitidx
]);
6459 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6460 if (word
== old_word
)
6467 * This function checks whether pageblock includes unmovable pages or not.
6468 * If @count is not zero, it is okay to include less @count unmovable pages
6470 * PageLRU check without isolation or lru_lock could race so that
6471 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6472 * expect this function should be exact.
6474 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6475 bool skip_hwpoisoned_pages
)
6477 unsigned long pfn
, iter
, found
;
6481 * For avoiding noise data, lru_add_drain_all() should be called
6482 * If ZONE_MOVABLE, the zone never contains unmovable pages
6484 if (zone_idx(zone
) == ZONE_MOVABLE
)
6486 mt
= get_pageblock_migratetype(page
);
6487 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6490 pfn
= page_to_pfn(page
);
6491 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6492 unsigned long check
= pfn
+ iter
;
6494 if (!pfn_valid_within(check
))
6497 page
= pfn_to_page(check
);
6500 * Hugepages are not in LRU lists, but they're movable.
6501 * We need not scan over tail pages bacause we don't
6502 * handle each tail page individually in migration.
6504 if (PageHuge(page
)) {
6505 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6510 * We can't use page_count without pin a page
6511 * because another CPU can free compound page.
6512 * This check already skips compound tails of THP
6513 * because their page->_count is zero at all time.
6515 if (!atomic_read(&page
->_count
)) {
6516 if (PageBuddy(page
))
6517 iter
+= (1 << page_order(page
)) - 1;
6522 * The HWPoisoned page may be not in buddy system, and
6523 * page_count() is not 0.
6525 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6531 * If there are RECLAIMABLE pages, we need to check
6532 * it. But now, memory offline itself doesn't call
6533 * shrink_node_slabs() and it still to be fixed.
6536 * If the page is not RAM, page_count()should be 0.
6537 * we don't need more check. This is an _used_ not-movable page.
6539 * The problematic thing here is PG_reserved pages. PG_reserved
6540 * is set to both of a memory hole page and a _used_ kernel
6549 bool is_pageblock_removable_nolock(struct page
*page
)
6555 * We have to be careful here because we are iterating over memory
6556 * sections which are not zone aware so we might end up outside of
6557 * the zone but still within the section.
6558 * We have to take care about the node as well. If the node is offline
6559 * its NODE_DATA will be NULL - see page_zone.
6561 if (!node_online(page_to_nid(page
)))
6564 zone
= page_zone(page
);
6565 pfn
= page_to_pfn(page
);
6566 if (!zone_spans_pfn(zone
, pfn
))
6569 return !has_unmovable_pages(zone
, page
, 0, true);
6574 static unsigned long pfn_max_align_down(unsigned long pfn
)
6576 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6577 pageblock_nr_pages
) - 1);
6580 static unsigned long pfn_max_align_up(unsigned long pfn
)
6582 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6583 pageblock_nr_pages
));
6586 /* [start, end) must belong to a single zone. */
6587 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6588 unsigned long start
, unsigned long end
)
6590 /* This function is based on compact_zone() from compaction.c. */
6591 unsigned long nr_reclaimed
;
6592 unsigned long pfn
= start
;
6593 unsigned int tries
= 0;
6598 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6599 if (fatal_signal_pending(current
)) {
6604 if (list_empty(&cc
->migratepages
)) {
6605 cc
->nr_migratepages
= 0;
6606 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6612 } else if (++tries
== 5) {
6613 ret
= ret
< 0 ? ret
: -EBUSY
;
6617 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6619 cc
->nr_migratepages
-= nr_reclaimed
;
6621 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6622 NULL
, 0, cc
->mode
, MR_CMA
);
6625 putback_movable_pages(&cc
->migratepages
);
6632 * alloc_contig_range() -- tries to allocate given range of pages
6633 * @start: start PFN to allocate
6634 * @end: one-past-the-last PFN to allocate
6635 * @migratetype: migratetype of the underlaying pageblocks (either
6636 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6637 * in range must have the same migratetype and it must
6638 * be either of the two.
6640 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6641 * aligned, however it's the caller's responsibility to guarantee that
6642 * we are the only thread that changes migrate type of pageblocks the
6645 * The PFN range must belong to a single zone.
6647 * Returns zero on success or negative error code. On success all
6648 * pages which PFN is in [start, end) are allocated for the caller and
6649 * need to be freed with free_contig_range().
6651 int alloc_contig_range(unsigned long start
, unsigned long end
,
6652 unsigned migratetype
)
6654 unsigned long outer_start
, outer_end
;
6658 struct compact_control cc
= {
6659 .nr_migratepages
= 0,
6661 .zone
= page_zone(pfn_to_page(start
)),
6662 .mode
= MIGRATE_SYNC
,
6663 .ignore_skip_hint
= true,
6665 INIT_LIST_HEAD(&cc
.migratepages
);
6668 * What we do here is we mark all pageblocks in range as
6669 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6670 * have different sizes, and due to the way page allocator
6671 * work, we align the range to biggest of the two pages so
6672 * that page allocator won't try to merge buddies from
6673 * different pageblocks and change MIGRATE_ISOLATE to some
6674 * other migration type.
6676 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6677 * migrate the pages from an unaligned range (ie. pages that
6678 * we are interested in). This will put all the pages in
6679 * range back to page allocator as MIGRATE_ISOLATE.
6681 * When this is done, we take the pages in range from page
6682 * allocator removing them from the buddy system. This way
6683 * page allocator will never consider using them.
6685 * This lets us mark the pageblocks back as
6686 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6687 * aligned range but not in the unaligned, original range are
6688 * put back to page allocator so that buddy can use them.
6691 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6692 pfn_max_align_up(end
), migratetype
,
6698 * In case of -EBUSY, we'd like to know which page causes problem.
6699 * So, just fall through. We will check it in test_pages_isolated().
6701 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6702 if (ret
&& ret
!= -EBUSY
)
6706 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6707 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6708 * more, all pages in [start, end) are free in page allocator.
6709 * What we are going to do is to allocate all pages from
6710 * [start, end) (that is remove them from page allocator).
6712 * The only problem is that pages at the beginning and at the
6713 * end of interesting range may be not aligned with pages that
6714 * page allocator holds, ie. they can be part of higher order
6715 * pages. Because of this, we reserve the bigger range and
6716 * once this is done free the pages we are not interested in.
6718 * We don't have to hold zone->lock here because the pages are
6719 * isolated thus they won't get removed from buddy.
6722 lru_add_drain_all();
6723 drain_all_pages(cc
.zone
);
6726 outer_start
= start
;
6727 while (!PageBuddy(pfn_to_page(outer_start
))) {
6728 if (++order
>= MAX_ORDER
) {
6729 outer_start
= start
;
6732 outer_start
&= ~0UL << order
;
6735 if (outer_start
!= start
) {
6736 order
= page_order(pfn_to_page(outer_start
));
6739 * outer_start page could be small order buddy page and
6740 * it doesn't include start page. Adjust outer_start
6741 * in this case to report failed page properly
6742 * on tracepoint in test_pages_isolated()
6744 if (outer_start
+ (1UL << order
) <= start
)
6745 outer_start
= start
;
6748 /* Make sure the range is really isolated. */
6749 if (test_pages_isolated(outer_start
, end
, false)) {
6750 pr_info("%s: [%lx, %lx) PFNs busy\n",
6751 __func__
, outer_start
, end
);
6756 /* Grab isolated pages from freelists. */
6757 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6763 /* Free head and tail (if any) */
6764 if (start
!= outer_start
)
6765 free_contig_range(outer_start
, start
- outer_start
);
6766 if (end
!= outer_end
)
6767 free_contig_range(end
, outer_end
- end
);
6770 undo_isolate_page_range(pfn_max_align_down(start
),
6771 pfn_max_align_up(end
), migratetype
);
6775 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6777 unsigned int count
= 0;
6779 for (; nr_pages
--; pfn
++) {
6780 struct page
*page
= pfn_to_page(pfn
);
6782 count
+= page_count(page
) != 1;
6785 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6789 #ifdef CONFIG_MEMORY_HOTPLUG
6791 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6792 * page high values need to be recalulated.
6794 void __meminit
zone_pcp_update(struct zone
*zone
)
6797 mutex_lock(&pcp_batch_high_lock
);
6798 for_each_possible_cpu(cpu
)
6799 pageset_set_high_and_batch(zone
,
6800 per_cpu_ptr(zone
->pageset
, cpu
));
6801 mutex_unlock(&pcp_batch_high_lock
);
6805 void zone_pcp_reset(struct zone
*zone
)
6807 unsigned long flags
;
6809 struct per_cpu_pageset
*pset
;
6811 /* avoid races with drain_pages() */
6812 local_irq_save(flags
);
6813 if (zone
->pageset
!= &boot_pageset
) {
6814 for_each_online_cpu(cpu
) {
6815 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6816 drain_zonestat(zone
, pset
);
6818 free_percpu(zone
->pageset
);
6819 zone
->pageset
= &boot_pageset
;
6821 local_irq_restore(flags
);
6824 #ifdef CONFIG_MEMORY_HOTREMOVE
6826 * All pages in the range must be isolated before calling this.
6829 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6833 unsigned int order
, i
;
6835 unsigned long flags
;
6836 /* find the first valid pfn */
6837 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6842 zone
= page_zone(pfn_to_page(pfn
));
6843 spin_lock_irqsave(&zone
->lock
, flags
);
6845 while (pfn
< end_pfn
) {
6846 if (!pfn_valid(pfn
)) {
6850 page
= pfn_to_page(pfn
);
6852 * The HWPoisoned page may be not in buddy system, and
6853 * page_count() is not 0.
6855 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6857 SetPageReserved(page
);
6861 BUG_ON(page_count(page
));
6862 BUG_ON(!PageBuddy(page
));
6863 order
= page_order(page
);
6864 #ifdef CONFIG_DEBUG_VM
6865 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6866 pfn
, 1 << order
, end_pfn
);
6868 list_del(&page
->lru
);
6869 rmv_page_order(page
);
6870 zone
->free_area
[order
].nr_free
--;
6871 for (i
= 0; i
< (1 << order
); i
++)
6872 SetPageReserved((page
+i
));
6873 pfn
+= (1 << order
);
6875 spin_unlock_irqrestore(&zone
->lock
, flags
);
6879 #ifdef CONFIG_MEMORY_FAILURE
6880 bool is_free_buddy_page(struct page
*page
)
6882 struct zone
*zone
= page_zone(page
);
6883 unsigned long pfn
= page_to_pfn(page
);
6884 unsigned long flags
;
6887 spin_lock_irqsave(&zone
->lock
, flags
);
6888 for (order
= 0; order
< MAX_ORDER
; order
++) {
6889 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6891 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6894 spin_unlock_irqrestore(&zone
->lock
, flags
);
6896 return order
< MAX_ORDER
;