2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/kasan.h>
29 #include <linux/module.h>
30 #include <linux/suspend.h>
31 #include <linux/pagevec.h>
32 #include <linux/blkdev.h>
33 #include <linux/slab.h>
34 #include <linux/ratelimit.h>
35 #include <linux/oom.h>
36 #include <linux/notifier.h>
37 #include <linux/topology.h>
38 #include <linux/sysctl.h>
39 #include <linux/cpu.h>
40 #include <linux/cpuset.h>
41 #include <linux/memory_hotplug.h>
42 #include <linux/nodemask.h>
43 #include <linux/vmalloc.h>
44 #include <linux/vmstat.h>
45 #include <linux/mempolicy.h>
46 #include <linux/stop_machine.h>
47 #include <linux/sort.h>
48 #include <linux/pfn.h>
49 #include <linux/backing-dev.h>
50 #include <linux/fault-inject.h>
51 #include <linux/page-isolation.h>
52 #include <linux/page_ext.h>
53 #include <linux/debugobjects.h>
54 #include <linux/kmemleak.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/prefetch.h>
58 #include <linux/mm_inline.h>
59 #include <linux/migrate.h>
60 #include <linux/page_ext.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
63 #include <linux/page_owner.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock
);
72 #define MIN_PERCPU_PAGELIST_FRACTION (8)
74 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
75 DEFINE_PER_CPU(int, numa_node
);
76 EXPORT_PER_CPU_SYMBOL(numa_node
);
79 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
81 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
82 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
83 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
84 * defined in <linux/topology.h>.
86 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
87 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
88 int _node_numa_mem_
[MAX_NUMNODES
];
92 * Array of node states.
94 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
95 [N_POSSIBLE
] = NODE_MASK_ALL
,
96 [N_ONLINE
] = { { [0] = 1UL } },
98 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
100 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
102 #ifdef CONFIG_MOVABLE_NODE
103 [N_MEMORY
] = { { [0] = 1UL } },
105 [N_CPU
] = { { [0] = 1UL } },
108 EXPORT_SYMBOL(node_states
);
110 /* Protect totalram_pages and zone->managed_pages */
111 static DEFINE_SPINLOCK(managed_page_count_lock
);
113 unsigned long totalram_pages __read_mostly
;
114 unsigned long totalreserve_pages __read_mostly
;
115 unsigned long totalcma_pages __read_mostly
;
117 * When calculating the number of globally allowed dirty pages, there
118 * is a certain number of per-zone reserves that should not be
119 * considered dirtyable memory. This is the sum of those reserves
120 * over all existing zones that contribute dirtyable memory.
122 unsigned long dirty_balance_reserve __read_mostly
;
124 int percpu_pagelist_fraction
;
125 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
127 #ifdef CONFIG_PM_SLEEP
129 * The following functions are used by the suspend/hibernate code to temporarily
130 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
131 * while devices are suspended. To avoid races with the suspend/hibernate code,
132 * they should always be called with pm_mutex held (gfp_allowed_mask also should
133 * only be modified with pm_mutex held, unless the suspend/hibernate code is
134 * guaranteed not to run in parallel with that modification).
137 static gfp_t saved_gfp_mask
;
139 void pm_restore_gfp_mask(void)
141 WARN_ON(!mutex_is_locked(&pm_mutex
));
142 if (saved_gfp_mask
) {
143 gfp_allowed_mask
= saved_gfp_mask
;
148 void pm_restrict_gfp_mask(void)
150 WARN_ON(!mutex_is_locked(&pm_mutex
));
151 WARN_ON(saved_gfp_mask
);
152 saved_gfp_mask
= gfp_allowed_mask
;
153 gfp_allowed_mask
&= ~GFP_IOFS
;
156 bool pm_suspended_storage(void)
158 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
162 #endif /* CONFIG_PM_SLEEP */
164 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
165 int pageblock_order __read_mostly
;
168 static void __free_pages_ok(struct page
*page
, unsigned int order
);
171 * results with 256, 32 in the lowmem_reserve sysctl:
172 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
173 * 1G machine -> (16M dma, 784M normal, 224M high)
174 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
175 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
176 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
178 * TBD: should special case ZONE_DMA32 machines here - in those we normally
179 * don't need any ZONE_NORMAL reservation
181 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
182 #ifdef CONFIG_ZONE_DMA
185 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 EXPORT_SYMBOL(totalram_pages
);
196 static char * const zone_names
[MAX_NR_ZONES
] = {
197 #ifdef CONFIG_ZONE_DMA
200 #ifdef CONFIG_ZONE_DMA32
204 #ifdef CONFIG_HIGHMEM
210 int min_free_kbytes
= 1024;
211 int user_min_free_kbytes
= -1;
213 static unsigned long __meminitdata nr_kernel_pages
;
214 static unsigned long __meminitdata nr_all_pages
;
215 static unsigned long __meminitdata dma_reserve
;
217 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
218 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
219 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
220 static unsigned long __initdata required_kernelcore
;
221 static unsigned long __initdata required_movablecore
;
222 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
224 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
226 EXPORT_SYMBOL(movable_zone
);
227 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
230 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
231 int nr_online_nodes __read_mostly
= 1;
232 EXPORT_SYMBOL(nr_node_ids
);
233 EXPORT_SYMBOL(nr_online_nodes
);
236 int page_group_by_mobility_disabled __read_mostly
;
238 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
240 if (unlikely(page_group_by_mobility_disabled
&&
241 migratetype
< MIGRATE_PCPTYPES
))
242 migratetype
= MIGRATE_UNMOVABLE
;
244 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
245 PB_migrate
, PB_migrate_end
);
248 #ifdef CONFIG_DEBUG_VM
249 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
253 unsigned long pfn
= page_to_pfn(page
);
254 unsigned long sp
, start_pfn
;
257 seq
= zone_span_seqbegin(zone
);
258 start_pfn
= zone
->zone_start_pfn
;
259 sp
= zone
->spanned_pages
;
260 if (!zone_spans_pfn(zone
, pfn
))
262 } while (zone_span_seqretry(zone
, seq
));
265 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
266 pfn
, zone_to_nid(zone
), zone
->name
,
267 start_pfn
, start_pfn
+ sp
);
272 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
274 if (!pfn_valid_within(page_to_pfn(page
)))
276 if (zone
!= page_zone(page
))
282 * Temporary debugging check for pages not lying within a given zone.
284 static int bad_range(struct zone
*zone
, struct page
*page
)
286 if (page_outside_zone_boundaries(zone
, page
))
288 if (!page_is_consistent(zone
, page
))
294 static inline int bad_range(struct zone
*zone
, struct page
*page
)
300 static void bad_page(struct page
*page
, const char *reason
,
301 unsigned long bad_flags
)
303 static unsigned long resume
;
304 static unsigned long nr_shown
;
305 static unsigned long nr_unshown
;
307 /* Don't complain about poisoned pages */
308 if (PageHWPoison(page
)) {
309 page_mapcount_reset(page
); /* remove PageBuddy */
314 * Allow a burst of 60 reports, then keep quiet for that minute;
315 * or allow a steady drip of one report per second.
317 if (nr_shown
== 60) {
318 if (time_before(jiffies
, resume
)) {
324 "BUG: Bad page state: %lu messages suppressed\n",
331 resume
= jiffies
+ 60 * HZ
;
333 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
334 current
->comm
, page_to_pfn(page
));
335 dump_page_badflags(page
, reason
, bad_flags
);
340 /* Leave bad fields for debug, except PageBuddy could make trouble */
341 page_mapcount_reset(page
); /* remove PageBuddy */
342 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
346 * Higher-order pages are called "compound pages". They are structured thusly:
348 * The first PAGE_SIZE page is called the "head page".
350 * The remaining PAGE_SIZE pages are called "tail pages".
352 * All pages have PG_compound set. All tail pages have their ->first_page
353 * pointing at the head page.
355 * The first tail page's ->lru.next holds the address of the compound page's
356 * put_page() function. Its ->lru.prev holds the order of allocation.
357 * This usage means that zero-order pages may not be compound.
360 static void free_compound_page(struct page
*page
)
362 __free_pages_ok(page
, compound_order(page
));
365 void prep_compound_page(struct page
*page
, unsigned long order
)
368 int nr_pages
= 1 << order
;
370 set_compound_page_dtor(page
, free_compound_page
);
371 set_compound_order(page
, order
);
373 for (i
= 1; i
< nr_pages
; i
++) {
374 struct page
*p
= page
+ i
;
375 set_page_count(p
, 0);
376 p
->first_page
= page
;
377 /* Make sure p->first_page is always valid for PageTail() */
383 static inline void prep_zero_page(struct page
*page
, unsigned int order
,
389 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
390 * and __GFP_HIGHMEM from hard or soft interrupt context.
392 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
393 for (i
= 0; i
< (1 << order
); i
++)
394 clear_highpage(page
+ i
);
397 #ifdef CONFIG_DEBUG_PAGEALLOC
398 unsigned int _debug_guardpage_minorder
;
399 bool _debug_pagealloc_enabled __read_mostly
;
400 bool _debug_guardpage_enabled __read_mostly
;
402 static int __init
early_debug_pagealloc(char *buf
)
407 if (strcmp(buf
, "on") == 0)
408 _debug_pagealloc_enabled
= true;
412 early_param("debug_pagealloc", early_debug_pagealloc
);
414 static bool need_debug_guardpage(void)
416 /* If we don't use debug_pagealloc, we don't need guard page */
417 if (!debug_pagealloc_enabled())
423 static void init_debug_guardpage(void)
425 if (!debug_pagealloc_enabled())
428 _debug_guardpage_enabled
= true;
431 struct page_ext_operations debug_guardpage_ops
= {
432 .need
= need_debug_guardpage
,
433 .init
= init_debug_guardpage
,
436 static int __init
debug_guardpage_minorder_setup(char *buf
)
440 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
441 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
444 _debug_guardpage_minorder
= res
;
445 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
448 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
450 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
451 unsigned int order
, int migratetype
)
453 struct page_ext
*page_ext
;
455 if (!debug_guardpage_enabled())
458 page_ext
= lookup_page_ext(page
);
459 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
461 INIT_LIST_HEAD(&page
->lru
);
462 set_page_private(page
, order
);
463 /* Guard pages are not available for any usage */
464 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
467 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
468 unsigned int order
, int migratetype
)
470 struct page_ext
*page_ext
;
472 if (!debug_guardpage_enabled())
475 page_ext
= lookup_page_ext(page
);
476 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
478 set_page_private(page
, 0);
479 if (!is_migrate_isolate(migratetype
))
480 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
483 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
484 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
485 unsigned int order
, int migratetype
) {}
486 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
487 unsigned int order
, int migratetype
) {}
490 static inline void set_page_order(struct page
*page
, unsigned int order
)
492 set_page_private(page
, order
);
493 __SetPageBuddy(page
);
496 static inline void rmv_page_order(struct page
*page
)
498 __ClearPageBuddy(page
);
499 set_page_private(page
, 0);
503 * This function checks whether a page is free && is the buddy
504 * we can do coalesce a page and its buddy if
505 * (a) the buddy is not in a hole &&
506 * (b) the buddy is in the buddy system &&
507 * (c) a page and its buddy have the same order &&
508 * (d) a page and its buddy are in the same zone.
510 * For recording whether a page is in the buddy system, we set ->_mapcount
511 * PAGE_BUDDY_MAPCOUNT_VALUE.
512 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
513 * serialized by zone->lock.
515 * For recording page's order, we use page_private(page).
517 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
520 if (!pfn_valid_within(page_to_pfn(buddy
)))
523 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
524 if (page_zone_id(page
) != page_zone_id(buddy
))
527 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
532 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
534 * zone check is done late to avoid uselessly
535 * calculating zone/node ids for pages that could
538 if (page_zone_id(page
) != page_zone_id(buddy
))
541 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
549 * Freeing function for a buddy system allocator.
551 * The concept of a buddy system is to maintain direct-mapped table
552 * (containing bit values) for memory blocks of various "orders".
553 * The bottom level table contains the map for the smallest allocatable
554 * units of memory (here, pages), and each level above it describes
555 * pairs of units from the levels below, hence, "buddies".
556 * At a high level, all that happens here is marking the table entry
557 * at the bottom level available, and propagating the changes upward
558 * as necessary, plus some accounting needed to play nicely with other
559 * parts of the VM system.
560 * At each level, we keep a list of pages, which are heads of continuous
561 * free pages of length of (1 << order) and marked with _mapcount
562 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
564 * So when we are allocating or freeing one, we can derive the state of the
565 * other. That is, if we allocate a small block, and both were
566 * free, the remainder of the region must be split into blocks.
567 * If a block is freed, and its buddy is also free, then this
568 * triggers coalescing into a block of larger size.
573 static inline void __free_one_page(struct page
*page
,
575 struct zone
*zone
, unsigned int order
,
578 unsigned long page_idx
;
579 unsigned long combined_idx
;
580 unsigned long uninitialized_var(buddy_idx
);
582 int max_order
= MAX_ORDER
;
584 VM_BUG_ON(!zone_is_initialized(zone
));
585 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
587 VM_BUG_ON(migratetype
== -1);
588 if (is_migrate_isolate(migratetype
)) {
590 * We restrict max order of merging to prevent merge
591 * between freepages on isolate pageblock and normal
592 * pageblock. Without this, pageblock isolation
593 * could cause incorrect freepage accounting.
595 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
597 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
600 page_idx
= pfn
& ((1 << max_order
) - 1);
602 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
603 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
605 while (order
< max_order
- 1) {
606 buddy_idx
= __find_buddy_index(page_idx
, order
);
607 buddy
= page
+ (buddy_idx
- page_idx
);
608 if (!page_is_buddy(page
, buddy
, order
))
611 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
612 * merge with it and move up one order.
614 if (page_is_guard(buddy
)) {
615 clear_page_guard(zone
, buddy
, order
, migratetype
);
617 list_del(&buddy
->lru
);
618 zone
->free_area
[order
].nr_free
--;
619 rmv_page_order(buddy
);
621 combined_idx
= buddy_idx
& page_idx
;
622 page
= page
+ (combined_idx
- page_idx
);
623 page_idx
= combined_idx
;
626 set_page_order(page
, order
);
629 * If this is not the largest possible page, check if the buddy
630 * of the next-highest order is free. If it is, it's possible
631 * that pages are being freed that will coalesce soon. In case,
632 * that is happening, add the free page to the tail of the list
633 * so it's less likely to be used soon and more likely to be merged
634 * as a higher order page
636 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
637 struct page
*higher_page
, *higher_buddy
;
638 combined_idx
= buddy_idx
& page_idx
;
639 higher_page
= page
+ (combined_idx
- page_idx
);
640 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
641 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
642 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
643 list_add_tail(&page
->lru
,
644 &zone
->free_area
[order
].free_list
[migratetype
]);
649 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
651 zone
->free_area
[order
].nr_free
++;
654 static inline int free_pages_check(struct page
*page
)
656 const char *bad_reason
= NULL
;
657 unsigned long bad_flags
= 0;
659 if (unlikely(page_mapcount(page
)))
660 bad_reason
= "nonzero mapcount";
661 if (unlikely(page
->mapping
!= NULL
))
662 bad_reason
= "non-NULL mapping";
663 if (unlikely(atomic_read(&page
->_count
) != 0))
664 bad_reason
= "nonzero _count";
665 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
666 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
667 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
670 if (unlikely(page
->mem_cgroup
))
671 bad_reason
= "page still charged to cgroup";
673 if (unlikely(bad_reason
)) {
674 bad_page(page
, bad_reason
, bad_flags
);
677 page_cpupid_reset_last(page
);
678 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
679 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
684 * Frees a number of pages from the PCP lists
685 * Assumes all pages on list are in same zone, and of same order.
686 * count is the number of pages to free.
688 * If the zone was previously in an "all pages pinned" state then look to
689 * see if this freeing clears that state.
691 * And clear the zone's pages_scanned counter, to hold off the "all pages are
692 * pinned" detection logic.
694 static void free_pcppages_bulk(struct zone
*zone
, int count
,
695 struct per_cpu_pages
*pcp
)
700 unsigned long nr_scanned
;
702 spin_lock(&zone
->lock
);
703 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
705 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
709 struct list_head
*list
;
712 * Remove pages from lists in a round-robin fashion. A
713 * batch_free count is maintained that is incremented when an
714 * empty list is encountered. This is so more pages are freed
715 * off fuller lists instead of spinning excessively around empty
720 if (++migratetype
== MIGRATE_PCPTYPES
)
722 list
= &pcp
->lists
[migratetype
];
723 } while (list_empty(list
));
725 /* This is the only non-empty list. Free them all. */
726 if (batch_free
== MIGRATE_PCPTYPES
)
727 batch_free
= to_free
;
730 int mt
; /* migratetype of the to-be-freed page */
732 page
= list_entry(list
->prev
, struct page
, lru
);
733 /* must delete as __free_one_page list manipulates */
734 list_del(&page
->lru
);
735 mt
= get_freepage_migratetype(page
);
736 if (unlikely(has_isolate_pageblock(zone
)))
737 mt
= get_pageblock_migratetype(page
);
739 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
740 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
741 trace_mm_page_pcpu_drain(page
, 0, mt
);
742 } while (--to_free
&& --batch_free
&& !list_empty(list
));
744 spin_unlock(&zone
->lock
);
747 static void free_one_page(struct zone
*zone
,
748 struct page
*page
, unsigned long pfn
,
752 unsigned long nr_scanned
;
753 spin_lock(&zone
->lock
);
754 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
756 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
758 if (unlikely(has_isolate_pageblock(zone
) ||
759 is_migrate_isolate(migratetype
))) {
760 migratetype
= get_pfnblock_migratetype(page
, pfn
);
762 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
763 spin_unlock(&zone
->lock
);
766 static int free_tail_pages_check(struct page
*head_page
, struct page
*page
)
768 if (!IS_ENABLED(CONFIG_DEBUG_VM
))
770 if (unlikely(!PageTail(page
))) {
771 bad_page(page
, "PageTail not set", 0);
774 if (unlikely(page
->first_page
!= head_page
)) {
775 bad_page(page
, "first_page not consistent", 0);
781 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
783 bool compound
= PageCompound(page
);
786 VM_BUG_ON_PAGE(PageTail(page
), page
);
787 VM_BUG_ON_PAGE(compound
&& compound_order(page
) != order
, page
);
789 trace_mm_page_free(page
, order
);
790 kmemcheck_free_shadow(page
, order
);
791 kasan_free_pages(page
, order
);
794 page
->mapping
= NULL
;
795 bad
+= free_pages_check(page
);
796 for (i
= 1; i
< (1 << order
); i
++) {
798 bad
+= free_tail_pages_check(page
, page
+ i
);
799 bad
+= free_pages_check(page
+ i
);
804 reset_page_owner(page
, order
);
806 if (!PageHighMem(page
)) {
807 debug_check_no_locks_freed(page_address(page
),
809 debug_check_no_obj_freed(page_address(page
),
812 arch_free_page(page
, order
);
813 kernel_map_pages(page
, 1 << order
, 0);
818 static void __free_pages_ok(struct page
*page
, unsigned int order
)
822 unsigned long pfn
= page_to_pfn(page
);
824 if (!free_pages_prepare(page
, order
))
827 migratetype
= get_pfnblock_migratetype(page
, pfn
);
828 local_irq_save(flags
);
829 __count_vm_events(PGFREE
, 1 << order
);
830 set_freepage_migratetype(page
, migratetype
);
831 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
832 local_irq_restore(flags
);
835 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
837 unsigned int nr_pages
= 1 << order
;
838 struct page
*p
= page
;
842 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
844 __ClearPageReserved(p
);
845 set_page_count(p
, 0);
847 __ClearPageReserved(p
);
848 set_page_count(p
, 0);
850 page_zone(page
)->managed_pages
+= nr_pages
;
851 set_page_refcounted(page
);
852 __free_pages(page
, order
);
856 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
857 void __init
init_cma_reserved_pageblock(struct page
*page
)
859 unsigned i
= pageblock_nr_pages
;
860 struct page
*p
= page
;
863 __ClearPageReserved(p
);
864 set_page_count(p
, 0);
867 set_pageblock_migratetype(page
, MIGRATE_CMA
);
869 if (pageblock_order
>= MAX_ORDER
) {
870 i
= pageblock_nr_pages
;
873 set_page_refcounted(p
);
874 __free_pages(p
, MAX_ORDER
- 1);
875 p
+= MAX_ORDER_NR_PAGES
;
876 } while (i
-= MAX_ORDER_NR_PAGES
);
878 set_page_refcounted(page
);
879 __free_pages(page
, pageblock_order
);
882 adjust_managed_page_count(page
, pageblock_nr_pages
);
887 * The order of subdivision here is critical for the IO subsystem.
888 * Please do not alter this order without good reasons and regression
889 * testing. Specifically, as large blocks of memory are subdivided,
890 * the order in which smaller blocks are delivered depends on the order
891 * they're subdivided in this function. This is the primary factor
892 * influencing the order in which pages are delivered to the IO
893 * subsystem according to empirical testing, and this is also justified
894 * by considering the behavior of a buddy system containing a single
895 * large block of memory acted on by a series of small allocations.
896 * This behavior is a critical factor in sglist merging's success.
900 static inline void expand(struct zone
*zone
, struct page
*page
,
901 int low
, int high
, struct free_area
*area
,
904 unsigned long size
= 1 << high
;
910 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
912 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
913 debug_guardpage_enabled() &&
914 high
< debug_guardpage_minorder()) {
916 * Mark as guard pages (or page), that will allow to
917 * merge back to allocator when buddy will be freed.
918 * Corresponding page table entries will not be touched,
919 * pages will stay not present in virtual address space
921 set_page_guard(zone
, &page
[size
], high
, migratetype
);
924 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
926 set_page_order(&page
[size
], high
);
931 * This page is about to be returned from the page allocator
933 static inline int check_new_page(struct page
*page
)
935 const char *bad_reason
= NULL
;
936 unsigned long bad_flags
= 0;
938 if (unlikely(page_mapcount(page
)))
939 bad_reason
= "nonzero mapcount";
940 if (unlikely(page
->mapping
!= NULL
))
941 bad_reason
= "non-NULL mapping";
942 if (unlikely(atomic_read(&page
->_count
) != 0))
943 bad_reason
= "nonzero _count";
944 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
945 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
946 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
949 if (unlikely(page
->mem_cgroup
))
950 bad_reason
= "page still charged to cgroup";
952 if (unlikely(bad_reason
)) {
953 bad_page(page
, bad_reason
, bad_flags
);
959 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
964 for (i
= 0; i
< (1 << order
); i
++) {
965 struct page
*p
= page
+ i
;
966 if (unlikely(check_new_page(p
)))
970 set_page_private(page
, 0);
971 set_page_refcounted(page
);
973 arch_alloc_page(page
, order
);
974 kernel_map_pages(page
, 1 << order
, 1);
975 kasan_alloc_pages(page
, order
);
977 if (gfp_flags
& __GFP_ZERO
)
978 prep_zero_page(page
, order
, gfp_flags
);
980 if (order
&& (gfp_flags
& __GFP_COMP
))
981 prep_compound_page(page
, order
);
983 set_page_owner(page
, order
, gfp_flags
);
986 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
987 * allocate the page. The expectation is that the caller is taking
988 * steps that will free more memory. The caller should avoid the page
989 * being used for !PFMEMALLOC purposes.
991 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
997 * Go through the free lists for the given migratetype and remove
998 * the smallest available page from the freelists
1001 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
1004 unsigned int current_order
;
1005 struct free_area
*area
;
1008 /* Find a page of the appropriate size in the preferred list */
1009 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
1010 area
= &(zone
->free_area
[current_order
]);
1011 if (list_empty(&area
->free_list
[migratetype
]))
1014 page
= list_entry(area
->free_list
[migratetype
].next
,
1016 list_del(&page
->lru
);
1017 rmv_page_order(page
);
1019 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1020 set_freepage_migratetype(page
, migratetype
);
1029 * This array describes the order lists are fallen back to when
1030 * the free lists for the desirable migrate type are depleted
1032 static int fallbacks
[MIGRATE_TYPES
][4] = {
1033 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1034 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1035 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
1037 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
1039 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
1040 #ifdef CONFIG_MEMORY_ISOLATION
1041 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
1046 static struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1049 return __rmqueue_smallest(zone
, order
, MIGRATE_CMA
);
1052 static inline struct page
*__rmqueue_cma_fallback(struct zone
*zone
,
1053 unsigned int order
) { return NULL
; }
1057 * Move the free pages in a range to the free lists of the requested type.
1058 * Note that start_page and end_pages are not aligned on a pageblock
1059 * boundary. If alignment is required, use move_freepages_block()
1061 int move_freepages(struct zone
*zone
,
1062 struct page
*start_page
, struct page
*end_page
,
1066 unsigned long order
;
1067 int pages_moved
= 0;
1069 #ifndef CONFIG_HOLES_IN_ZONE
1071 * page_zone is not safe to call in this context when
1072 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1073 * anyway as we check zone boundaries in move_freepages_block().
1074 * Remove at a later date when no bug reports exist related to
1075 * grouping pages by mobility
1077 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1080 for (page
= start_page
; page
<= end_page
;) {
1081 /* Make sure we are not inadvertently changing nodes */
1082 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1084 if (!pfn_valid_within(page_to_pfn(page
))) {
1089 if (!PageBuddy(page
)) {
1094 order
= page_order(page
);
1095 list_move(&page
->lru
,
1096 &zone
->free_area
[order
].free_list
[migratetype
]);
1097 set_freepage_migratetype(page
, migratetype
);
1099 pages_moved
+= 1 << order
;
1105 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1108 unsigned long start_pfn
, end_pfn
;
1109 struct page
*start_page
, *end_page
;
1111 start_pfn
= page_to_pfn(page
);
1112 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1113 start_page
= pfn_to_page(start_pfn
);
1114 end_page
= start_page
+ pageblock_nr_pages
- 1;
1115 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1117 /* Do not cross zone boundaries */
1118 if (!zone_spans_pfn(zone
, start_pfn
))
1120 if (!zone_spans_pfn(zone
, end_pfn
))
1123 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1126 static void change_pageblock_range(struct page
*pageblock_page
,
1127 int start_order
, int migratetype
)
1129 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1131 while (nr_pageblocks
--) {
1132 set_pageblock_migratetype(pageblock_page
, migratetype
);
1133 pageblock_page
+= pageblock_nr_pages
;
1138 * When we are falling back to another migratetype during allocation, try to
1139 * steal extra free pages from the same pageblocks to satisfy further
1140 * allocations, instead of polluting multiple pageblocks.
1142 * If we are stealing a relatively large buddy page, it is likely there will
1143 * be more free pages in the pageblock, so try to steal them all. For
1144 * reclaimable and unmovable allocations, we steal regardless of page size,
1145 * as fragmentation caused by those allocations polluting movable pageblocks
1146 * is worse than movable allocations stealing from unmovable and reclaimable
1149 static bool can_steal_fallback(unsigned int order
, int start_mt
)
1152 * Leaving this order check is intended, although there is
1153 * relaxed order check in next check. The reason is that
1154 * we can actually steal whole pageblock if this condition met,
1155 * but, below check doesn't guarantee it and that is just heuristic
1156 * so could be changed anytime.
1158 if (order
>= pageblock_order
)
1161 if (order
>= pageblock_order
/ 2 ||
1162 start_mt
== MIGRATE_RECLAIMABLE
||
1163 start_mt
== MIGRATE_UNMOVABLE
||
1164 page_group_by_mobility_disabled
)
1171 * This function implements actual steal behaviour. If order is large enough,
1172 * we can steal whole pageblock. If not, we first move freepages in this
1173 * pageblock and check whether half of pages are moved or not. If half of
1174 * pages are moved, we can change migratetype of pageblock and permanently
1175 * use it's pages as requested migratetype in the future.
1177 static void steal_suitable_fallback(struct zone
*zone
, struct page
*page
,
1180 int current_order
= page_order(page
);
1183 /* Take ownership for orders >= pageblock_order */
1184 if (current_order
>= pageblock_order
) {
1185 change_pageblock_range(page
, current_order
, start_type
);
1189 pages
= move_freepages_block(zone
, page
, start_type
);
1191 /* Claim the whole block if over half of it is free */
1192 if (pages
>= (1 << (pageblock_order
-1)) ||
1193 page_group_by_mobility_disabled
)
1194 set_pageblock_migratetype(page
, start_type
);
1197 /* Check whether there is a suitable fallback freepage with requested order. */
1198 static int find_suitable_fallback(struct free_area
*area
, unsigned int order
,
1199 int migratetype
, bool *can_steal
)
1204 if (area
->nr_free
== 0)
1209 fallback_mt
= fallbacks
[migratetype
][i
];
1210 if (fallback_mt
== MIGRATE_RESERVE
)
1213 if (list_empty(&area
->free_list
[fallback_mt
]))
1216 if (can_steal_fallback(order
, migratetype
))
1225 /* Remove an element from the buddy allocator from the fallback list */
1226 static inline struct page
*
1227 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1229 struct free_area
*area
;
1230 unsigned int current_order
;
1235 /* Find the largest possible block of pages in the other list */
1236 for (current_order
= MAX_ORDER
-1;
1237 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1239 area
= &(zone
->free_area
[current_order
]);
1240 fallback_mt
= find_suitable_fallback(area
, current_order
,
1241 start_migratetype
, &can_steal
);
1242 if (fallback_mt
== -1)
1245 page
= list_entry(area
->free_list
[fallback_mt
].next
,
1248 steal_suitable_fallback(zone
, page
, start_migratetype
);
1250 /* Remove the page from the freelists */
1252 list_del(&page
->lru
);
1253 rmv_page_order(page
);
1255 expand(zone
, page
, order
, current_order
, area
,
1258 * The freepage_migratetype may differ from pageblock's
1259 * migratetype depending on the decisions in
1260 * try_to_steal_freepages(). This is OK as long as it
1261 * does not differ for MIGRATE_CMA pageblocks. For CMA
1262 * we need to make sure unallocated pages flushed from
1263 * pcp lists are returned to the correct freelist.
1265 set_freepage_migratetype(page
, start_migratetype
);
1267 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1268 start_migratetype
, fallback_mt
);
1277 * Do the hard work of removing an element from the buddy allocator.
1278 * Call me with the zone->lock already held.
1280 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1286 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1288 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1289 if (migratetype
== MIGRATE_MOVABLE
)
1290 page
= __rmqueue_cma_fallback(zone
, order
);
1293 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1296 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1297 * is used because __rmqueue_smallest is an inline function
1298 * and we want just one call site
1301 migratetype
= MIGRATE_RESERVE
;
1306 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1311 * Obtain a specified number of elements from the buddy allocator, all under
1312 * a single hold of the lock, for efficiency. Add them to the supplied list.
1313 * Returns the number of new pages which were placed at *list.
1315 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1316 unsigned long count
, struct list_head
*list
,
1317 int migratetype
, bool cold
)
1321 spin_lock(&zone
->lock
);
1322 for (i
= 0; i
< count
; ++i
) {
1323 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1324 if (unlikely(page
== NULL
))
1328 * Split buddy pages returned by expand() are received here
1329 * in physical page order. The page is added to the callers and
1330 * list and the list head then moves forward. From the callers
1331 * perspective, the linked list is ordered by page number in
1332 * some conditions. This is useful for IO devices that can
1333 * merge IO requests if the physical pages are ordered
1337 list_add(&page
->lru
, list
);
1339 list_add_tail(&page
->lru
, list
);
1341 if (is_migrate_cma(get_freepage_migratetype(page
)))
1342 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1345 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1346 spin_unlock(&zone
->lock
);
1352 * Called from the vmstat counter updater to drain pagesets of this
1353 * currently executing processor on remote nodes after they have
1356 * Note that this function must be called with the thread pinned to
1357 * a single processor.
1359 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1361 unsigned long flags
;
1362 int to_drain
, batch
;
1364 local_irq_save(flags
);
1365 batch
= ACCESS_ONCE(pcp
->batch
);
1366 to_drain
= min(pcp
->count
, batch
);
1368 free_pcppages_bulk(zone
, to_drain
, pcp
);
1369 pcp
->count
-= to_drain
;
1371 local_irq_restore(flags
);
1376 * Drain pcplists of the indicated processor and zone.
1378 * The processor must either be the current processor and the
1379 * thread pinned to the current processor or a processor that
1382 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1384 unsigned long flags
;
1385 struct per_cpu_pageset
*pset
;
1386 struct per_cpu_pages
*pcp
;
1388 local_irq_save(flags
);
1389 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1393 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1396 local_irq_restore(flags
);
1400 * Drain pcplists of all zones on the indicated processor.
1402 * The processor must either be the current processor and the
1403 * thread pinned to the current processor or a processor that
1406 static void drain_pages(unsigned int cpu
)
1410 for_each_populated_zone(zone
) {
1411 drain_pages_zone(cpu
, zone
);
1416 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1418 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1419 * the single zone's pages.
1421 void drain_local_pages(struct zone
*zone
)
1423 int cpu
= smp_processor_id();
1426 drain_pages_zone(cpu
, zone
);
1432 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1434 * When zone parameter is non-NULL, spill just the single zone's pages.
1436 * Note that this code is protected against sending an IPI to an offline
1437 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1438 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1439 * nothing keeps CPUs from showing up after we populated the cpumask and
1440 * before the call to on_each_cpu_mask().
1442 void drain_all_pages(struct zone
*zone
)
1447 * Allocate in the BSS so we wont require allocation in
1448 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1450 static cpumask_t cpus_with_pcps
;
1453 * We don't care about racing with CPU hotplug event
1454 * as offline notification will cause the notified
1455 * cpu to drain that CPU pcps and on_each_cpu_mask
1456 * disables preemption as part of its processing
1458 for_each_online_cpu(cpu
) {
1459 struct per_cpu_pageset
*pcp
;
1461 bool has_pcps
= false;
1464 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1468 for_each_populated_zone(z
) {
1469 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1470 if (pcp
->pcp
.count
) {
1478 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1480 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1482 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1486 #ifdef CONFIG_HIBERNATION
1488 void mark_free_pages(struct zone
*zone
)
1490 unsigned long pfn
, max_zone_pfn
;
1491 unsigned long flags
;
1492 unsigned int order
, t
;
1493 struct list_head
*curr
;
1495 if (zone_is_empty(zone
))
1498 spin_lock_irqsave(&zone
->lock
, flags
);
1500 max_zone_pfn
= zone_end_pfn(zone
);
1501 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1502 if (pfn_valid(pfn
)) {
1503 struct page
*page
= pfn_to_page(pfn
);
1505 if (!swsusp_page_is_forbidden(page
))
1506 swsusp_unset_page_free(page
);
1509 for_each_migratetype_order(order
, t
) {
1510 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1513 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1514 for (i
= 0; i
< (1UL << order
); i
++)
1515 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1518 spin_unlock_irqrestore(&zone
->lock
, flags
);
1520 #endif /* CONFIG_PM */
1523 * Free a 0-order page
1524 * cold == true ? free a cold page : free a hot page
1526 void free_hot_cold_page(struct page
*page
, bool cold
)
1528 struct zone
*zone
= page_zone(page
);
1529 struct per_cpu_pages
*pcp
;
1530 unsigned long flags
;
1531 unsigned long pfn
= page_to_pfn(page
);
1534 if (!free_pages_prepare(page
, 0))
1537 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1538 set_freepage_migratetype(page
, migratetype
);
1539 local_irq_save(flags
);
1540 __count_vm_event(PGFREE
);
1543 * We only track unmovable, reclaimable and movable on pcp lists.
1544 * Free ISOLATE pages back to the allocator because they are being
1545 * offlined but treat RESERVE as movable pages so we can get those
1546 * areas back if necessary. Otherwise, we may have to free
1547 * excessively into the page allocator
1549 if (migratetype
>= MIGRATE_PCPTYPES
) {
1550 if (unlikely(is_migrate_isolate(migratetype
))) {
1551 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1554 migratetype
= MIGRATE_MOVABLE
;
1557 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1559 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1561 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1563 if (pcp
->count
>= pcp
->high
) {
1564 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1565 free_pcppages_bulk(zone
, batch
, pcp
);
1566 pcp
->count
-= batch
;
1570 local_irq_restore(flags
);
1574 * Free a list of 0-order pages
1576 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1578 struct page
*page
, *next
;
1580 list_for_each_entry_safe(page
, next
, list
, lru
) {
1581 trace_mm_page_free_batched(page
, cold
);
1582 free_hot_cold_page(page
, cold
);
1587 * split_page takes a non-compound higher-order page, and splits it into
1588 * n (1<<order) sub-pages: page[0..n]
1589 * Each sub-page must be freed individually.
1591 * Note: this is probably too low level an operation for use in drivers.
1592 * Please consult with lkml before using this in your driver.
1594 void split_page(struct page
*page
, unsigned int order
)
1598 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1599 VM_BUG_ON_PAGE(!page_count(page
), page
);
1601 #ifdef CONFIG_KMEMCHECK
1603 * Split shadow pages too, because free(page[0]) would
1604 * otherwise free the whole shadow.
1606 if (kmemcheck_page_is_tracked(page
))
1607 split_page(virt_to_page(page
[0].shadow
), order
);
1610 set_page_owner(page
, 0, 0);
1611 for (i
= 1; i
< (1 << order
); i
++) {
1612 set_page_refcounted(page
+ i
);
1613 set_page_owner(page
+ i
, 0, 0);
1616 EXPORT_SYMBOL_GPL(split_page
);
1618 int __isolate_free_page(struct page
*page
, unsigned int order
)
1620 unsigned long watermark
;
1624 BUG_ON(!PageBuddy(page
));
1626 zone
= page_zone(page
);
1627 mt
= get_pageblock_migratetype(page
);
1629 if (!is_migrate_isolate(mt
)) {
1630 /* Obey watermarks as if the page was being allocated */
1631 watermark
= low_wmark_pages(zone
) + (1 << order
);
1632 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1635 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1638 /* Remove page from free list */
1639 list_del(&page
->lru
);
1640 zone
->free_area
[order
].nr_free
--;
1641 rmv_page_order(page
);
1643 /* Set the pageblock if the isolated page is at least a pageblock */
1644 if (order
>= pageblock_order
- 1) {
1645 struct page
*endpage
= page
+ (1 << order
) - 1;
1646 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1647 int mt
= get_pageblock_migratetype(page
);
1648 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1649 set_pageblock_migratetype(page
,
1654 set_page_owner(page
, order
, 0);
1655 return 1UL << order
;
1659 * Similar to split_page except the page is already free. As this is only
1660 * being used for migration, the migratetype of the block also changes.
1661 * As this is called with interrupts disabled, the caller is responsible
1662 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1665 * Note: this is probably too low level an operation for use in drivers.
1666 * Please consult with lkml before using this in your driver.
1668 int split_free_page(struct page
*page
)
1673 order
= page_order(page
);
1675 nr_pages
= __isolate_free_page(page
, order
);
1679 /* Split into individual pages */
1680 set_page_refcounted(page
);
1681 split_page(page
, order
);
1686 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1689 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1690 struct zone
*zone
, unsigned int order
,
1691 gfp_t gfp_flags
, int migratetype
)
1693 unsigned long flags
;
1695 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1697 if (likely(order
== 0)) {
1698 struct per_cpu_pages
*pcp
;
1699 struct list_head
*list
;
1701 local_irq_save(flags
);
1702 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1703 list
= &pcp
->lists
[migratetype
];
1704 if (list_empty(list
)) {
1705 pcp
->count
+= rmqueue_bulk(zone
, 0,
1708 if (unlikely(list_empty(list
)))
1713 page
= list_entry(list
->prev
, struct page
, lru
);
1715 page
= list_entry(list
->next
, struct page
, lru
);
1717 list_del(&page
->lru
);
1720 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1722 * __GFP_NOFAIL is not to be used in new code.
1724 * All __GFP_NOFAIL callers should be fixed so that they
1725 * properly detect and handle allocation failures.
1727 * We most definitely don't want callers attempting to
1728 * allocate greater than order-1 page units with
1731 WARN_ON_ONCE(order
> 1);
1733 spin_lock_irqsave(&zone
->lock
, flags
);
1734 page
= __rmqueue(zone
, order
, migratetype
);
1735 spin_unlock(&zone
->lock
);
1738 __mod_zone_freepage_state(zone
, -(1 << order
),
1739 get_freepage_migratetype(page
));
1742 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1743 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1744 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1745 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1747 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1748 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1749 local_irq_restore(flags
);
1751 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1755 local_irq_restore(flags
);
1759 #ifdef CONFIG_FAIL_PAGE_ALLOC
1762 struct fault_attr attr
;
1764 u32 ignore_gfp_highmem
;
1765 u32 ignore_gfp_wait
;
1767 } fail_page_alloc
= {
1768 .attr
= FAULT_ATTR_INITIALIZER
,
1769 .ignore_gfp_wait
= 1,
1770 .ignore_gfp_highmem
= 1,
1774 static int __init
setup_fail_page_alloc(char *str
)
1776 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1778 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1780 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1782 if (order
< fail_page_alloc
.min_order
)
1784 if (gfp_mask
& __GFP_NOFAIL
)
1786 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1788 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1791 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1794 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1796 static int __init
fail_page_alloc_debugfs(void)
1798 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1801 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1802 &fail_page_alloc
.attr
);
1804 return PTR_ERR(dir
);
1806 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1807 &fail_page_alloc
.ignore_gfp_wait
))
1809 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1810 &fail_page_alloc
.ignore_gfp_highmem
))
1812 if (!debugfs_create_u32("min-order", mode
, dir
,
1813 &fail_page_alloc
.min_order
))
1818 debugfs_remove_recursive(dir
);
1823 late_initcall(fail_page_alloc_debugfs
);
1825 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1827 #else /* CONFIG_FAIL_PAGE_ALLOC */
1829 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1834 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1837 * Return true if free pages are above 'mark'. This takes into account the order
1838 * of the allocation.
1840 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1841 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1844 /* free_pages may go negative - that's OK */
1849 free_pages
-= (1 << order
) - 1;
1850 if (alloc_flags
& ALLOC_HIGH
)
1852 if (alloc_flags
& ALLOC_HARDER
)
1855 /* If allocation can't use CMA areas don't use free CMA pages */
1856 if (!(alloc_flags
& ALLOC_CMA
))
1857 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1860 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1862 for (o
= 0; o
< order
; o
++) {
1863 /* At the next order, this order's pages become unavailable */
1864 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1866 /* Require fewer higher order pages to be free */
1869 if (free_pages
<= min
)
1875 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1876 int classzone_idx
, int alloc_flags
)
1878 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1879 zone_page_state(z
, NR_FREE_PAGES
));
1882 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1883 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1885 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1887 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1888 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1890 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1896 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1897 * skip over zones that are not allowed by the cpuset, or that have
1898 * been recently (in last second) found to be nearly full. See further
1899 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1900 * that have to skip over a lot of full or unallowed zones.
1902 * If the zonelist cache is present in the passed zonelist, then
1903 * returns a pointer to the allowed node mask (either the current
1904 * tasks mems_allowed, or node_states[N_MEMORY].)
1906 * If the zonelist cache is not available for this zonelist, does
1907 * nothing and returns NULL.
1909 * If the fullzones BITMAP in the zonelist cache is stale (more than
1910 * a second since last zap'd) then we zap it out (clear its bits.)
1912 * We hold off even calling zlc_setup, until after we've checked the
1913 * first zone in the zonelist, on the theory that most allocations will
1914 * be satisfied from that first zone, so best to examine that zone as
1915 * quickly as we can.
1917 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1919 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1920 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1922 zlc
= zonelist
->zlcache_ptr
;
1926 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1927 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1928 zlc
->last_full_zap
= jiffies
;
1931 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1932 &cpuset_current_mems_allowed
:
1933 &node_states
[N_MEMORY
];
1934 return allowednodes
;
1938 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1939 * if it is worth looking at further for free memory:
1940 * 1) Check that the zone isn't thought to be full (doesn't have its
1941 * bit set in the zonelist_cache fullzones BITMAP).
1942 * 2) Check that the zones node (obtained from the zonelist_cache
1943 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1944 * Return true (non-zero) if zone is worth looking at further, or
1945 * else return false (zero) if it is not.
1947 * This check -ignores- the distinction between various watermarks,
1948 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1949 * found to be full for any variation of these watermarks, it will
1950 * be considered full for up to one second by all requests, unless
1951 * we are so low on memory on all allowed nodes that we are forced
1952 * into the second scan of the zonelist.
1954 * In the second scan we ignore this zonelist cache and exactly
1955 * apply the watermarks to all zones, even it is slower to do so.
1956 * We are low on memory in the second scan, and should leave no stone
1957 * unturned looking for a free page.
1959 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1960 nodemask_t
*allowednodes
)
1962 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1963 int i
; /* index of *z in zonelist zones */
1964 int n
; /* node that zone *z is on */
1966 zlc
= zonelist
->zlcache_ptr
;
1970 i
= z
- zonelist
->_zonerefs
;
1973 /* This zone is worth trying if it is allowed but not full */
1974 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1978 * Given 'z' scanning a zonelist, set the corresponding bit in
1979 * zlc->fullzones, so that subsequent attempts to allocate a page
1980 * from that zone don't waste time re-examining it.
1982 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1984 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1985 int i
; /* index of *z in zonelist zones */
1987 zlc
= zonelist
->zlcache_ptr
;
1991 i
= z
- zonelist
->_zonerefs
;
1993 set_bit(i
, zlc
->fullzones
);
1997 * clear all zones full, called after direct reclaim makes progress so that
1998 * a zone that was recently full is not skipped over for up to a second
2000 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
2002 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
2004 zlc
= zonelist
->zlcache_ptr
;
2008 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
2011 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2013 return local_zone
->node
== zone
->node
;
2016 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2018 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
2022 #else /* CONFIG_NUMA */
2024 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
2029 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
2030 nodemask_t
*allowednodes
)
2035 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
2039 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
2043 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
2048 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
2053 #endif /* CONFIG_NUMA */
2055 static void reset_alloc_batches(struct zone
*preferred_zone
)
2057 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
2060 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2061 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
2062 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
2063 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
2064 } while (zone
++ != preferred_zone
);
2068 * get_page_from_freelist goes through the zonelist trying to allocate
2071 static struct page
*
2072 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2073 const struct alloc_context
*ac
)
2075 struct zonelist
*zonelist
= ac
->zonelist
;
2077 struct page
*page
= NULL
;
2079 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
2080 int zlc_active
= 0; /* set if using zonelist_cache */
2081 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
2082 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
2083 (gfp_mask
& __GFP_WRITE
);
2084 int nr_fair_skipped
= 0;
2085 bool zonelist_rescan
;
2088 zonelist_rescan
= false;
2091 * Scan zonelist, looking for a zone with enough free.
2092 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2094 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2098 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2099 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2101 if (cpusets_enabled() &&
2102 (alloc_flags
& ALLOC_CPUSET
) &&
2103 !cpuset_zone_allowed(zone
, gfp_mask
))
2106 * Distribute pages in proportion to the individual
2107 * zone size to ensure fair page aging. The zone a
2108 * page was allocated in should have no effect on the
2109 * time the page has in memory before being reclaimed.
2111 if (alloc_flags
& ALLOC_FAIR
) {
2112 if (!zone_local(ac
->preferred_zone
, zone
))
2114 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2120 * When allocating a page cache page for writing, we
2121 * want to get it from a zone that is within its dirty
2122 * limit, such that no single zone holds more than its
2123 * proportional share of globally allowed dirty pages.
2124 * The dirty limits take into account the zone's
2125 * lowmem reserves and high watermark so that kswapd
2126 * should be able to balance it without having to
2127 * write pages from its LRU list.
2129 * This may look like it could increase pressure on
2130 * lower zones by failing allocations in higher zones
2131 * before they are full. But the pages that do spill
2132 * over are limited as the lower zones are protected
2133 * by this very same mechanism. It should not become
2134 * a practical burden to them.
2136 * XXX: For now, allow allocations to potentially
2137 * exceed the per-zone dirty limit in the slowpath
2138 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2139 * which is important when on a NUMA setup the allowed
2140 * zones are together not big enough to reach the
2141 * global limit. The proper fix for these situations
2142 * will require awareness of zones in the
2143 * dirty-throttling and the flusher threads.
2145 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2148 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2149 if (!zone_watermark_ok(zone
, order
, mark
,
2150 ac
->classzone_idx
, alloc_flags
)) {
2153 /* Checked here to keep the fast path fast */
2154 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2155 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2158 if (IS_ENABLED(CONFIG_NUMA
) &&
2159 !did_zlc_setup
&& nr_online_nodes
> 1) {
2161 * we do zlc_setup if there are multiple nodes
2162 * and before considering the first zone allowed
2165 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2170 if (zone_reclaim_mode
== 0 ||
2171 !zone_allows_reclaim(ac
->preferred_zone
, zone
))
2172 goto this_zone_full
;
2175 * As we may have just activated ZLC, check if the first
2176 * eligible zone has failed zone_reclaim recently.
2178 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2179 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2182 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2184 case ZONE_RECLAIM_NOSCAN
:
2187 case ZONE_RECLAIM_FULL
:
2188 /* scanned but unreclaimable */
2191 /* did we reclaim enough */
2192 if (zone_watermark_ok(zone
, order
, mark
,
2193 ac
->classzone_idx
, alloc_flags
))
2197 * Failed to reclaim enough to meet watermark.
2198 * Only mark the zone full if checking the min
2199 * watermark or if we failed to reclaim just
2200 * 1<<order pages or else the page allocator
2201 * fastpath will prematurely mark zones full
2202 * when the watermark is between the low and
2205 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2206 ret
== ZONE_RECLAIM_SOME
)
2207 goto this_zone_full
;
2214 page
= buffered_rmqueue(ac
->preferred_zone
, zone
, order
,
2215 gfp_mask
, ac
->migratetype
);
2217 if (prep_new_page(page
, order
, gfp_mask
, alloc_flags
))
2222 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2223 zlc_mark_zone_full(zonelist
, z
);
2227 * The first pass makes sure allocations are spread fairly within the
2228 * local node. However, the local node might have free pages left
2229 * after the fairness batches are exhausted, and remote zones haven't
2230 * even been considered yet. Try once more without fairness, and
2231 * include remote zones now, before entering the slowpath and waking
2232 * kswapd: prefer spilling to a remote zone over swapping locally.
2234 if (alloc_flags
& ALLOC_FAIR
) {
2235 alloc_flags
&= ~ALLOC_FAIR
;
2236 if (nr_fair_skipped
) {
2237 zonelist_rescan
= true;
2238 reset_alloc_batches(ac
->preferred_zone
);
2240 if (nr_online_nodes
> 1)
2241 zonelist_rescan
= true;
2244 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2245 /* Disable zlc cache for second zonelist scan */
2247 zonelist_rescan
= true;
2250 if (zonelist_rescan
)
2257 * Large machines with many possible nodes should not always dump per-node
2258 * meminfo in irq context.
2260 static inline bool should_suppress_show_mem(void)
2265 ret
= in_interrupt();
2270 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2271 DEFAULT_RATELIMIT_INTERVAL
,
2272 DEFAULT_RATELIMIT_BURST
);
2274 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2276 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2278 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2279 debug_guardpage_minorder() > 0)
2283 * This documents exceptions given to allocations in certain
2284 * contexts that are allowed to allocate outside current's set
2287 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2288 if (test_thread_flag(TIF_MEMDIE
) ||
2289 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2290 filter
&= ~SHOW_MEM_FILTER_NODES
;
2291 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2292 filter
&= ~SHOW_MEM_FILTER_NODES
;
2295 struct va_format vaf
;
2298 va_start(args
, fmt
);
2303 pr_warn("%pV", &vaf
);
2308 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2309 current
->comm
, order
, gfp_mask
);
2312 if (!should_suppress_show_mem())
2317 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2318 unsigned long did_some_progress
,
2319 unsigned long pages_reclaimed
)
2321 /* Do not loop if specifically requested */
2322 if (gfp_mask
& __GFP_NORETRY
)
2325 /* Always retry if specifically requested */
2326 if (gfp_mask
& __GFP_NOFAIL
)
2330 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2331 * making forward progress without invoking OOM. Suspend also disables
2332 * storage devices so kswapd will not help. Bail if we are suspending.
2334 if (!did_some_progress
&& pm_suspended_storage())
2338 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2339 * means __GFP_NOFAIL, but that may not be true in other
2342 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2346 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2347 * specified, then we retry until we no longer reclaim any pages
2348 * (above), or we've reclaimed an order of pages at least as
2349 * large as the allocation's order. In both cases, if the
2350 * allocation still fails, we stop retrying.
2352 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2358 static inline struct page
*
2359 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2360 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
2364 *did_some_progress
= 0;
2367 * Acquire the per-zone oom lock for each zone. If that
2368 * fails, somebody else is making progress for us.
2370 if (!oom_zonelist_trylock(ac
->zonelist
, gfp_mask
)) {
2371 *did_some_progress
= 1;
2372 schedule_timeout_uninterruptible(1);
2377 * Go through the zonelist yet one more time, keep very high watermark
2378 * here, this is only to catch a parallel oom killing, we must fail if
2379 * we're still under heavy pressure.
2381 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
2382 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
2386 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2387 /* Coredumps can quickly deplete all memory reserves */
2388 if (current
->flags
& PF_DUMPCORE
)
2390 /* The OOM killer will not help higher order allocs */
2391 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2393 /* The OOM killer does not needlessly kill tasks for lowmem */
2394 if (ac
->high_zoneidx
< ZONE_NORMAL
)
2396 /* The OOM killer does not compensate for light reclaim */
2397 if (!(gfp_mask
& __GFP_FS
)) {
2399 * XXX: Page reclaim didn't yield anything,
2400 * and the OOM killer can't be invoked, but
2401 * keep looping as per should_alloc_retry().
2403 *did_some_progress
= 1;
2407 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2408 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2409 * The caller should handle page allocation failure by itself if
2410 * it specifies __GFP_THISNODE.
2411 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2413 if (gfp_mask
& __GFP_THISNODE
)
2416 /* Exhausted what can be done so it's blamo time */
2417 if (out_of_memory(ac
->zonelist
, gfp_mask
, order
, ac
->nodemask
, false)
2418 || WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
))
2419 *did_some_progress
= 1;
2421 oom_zonelist_unlock(ac
->zonelist
, gfp_mask
);
2425 #ifdef CONFIG_COMPACTION
2426 /* Try memory compaction for high-order allocations before reclaim */
2427 static struct page
*
2428 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2429 int alloc_flags
, const struct alloc_context
*ac
,
2430 enum migrate_mode mode
, int *contended_compaction
,
2431 bool *deferred_compaction
)
2433 unsigned long compact_result
;
2439 current
->flags
|= PF_MEMALLOC
;
2440 compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
2441 mode
, contended_compaction
);
2442 current
->flags
&= ~PF_MEMALLOC
;
2444 switch (compact_result
) {
2445 case COMPACT_DEFERRED
:
2446 *deferred_compaction
= true;
2448 case COMPACT_SKIPPED
:
2455 * At least in one zone compaction wasn't deferred or skipped, so let's
2456 * count a compaction stall
2458 count_vm_event(COMPACTSTALL
);
2460 page
= get_page_from_freelist(gfp_mask
, order
,
2461 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2464 struct zone
*zone
= page_zone(page
);
2466 zone
->compact_blockskip_flush
= false;
2467 compaction_defer_reset(zone
, order
, true);
2468 count_vm_event(COMPACTSUCCESS
);
2473 * It's bad if compaction run occurs and fails. The most likely reason
2474 * is that pages exist, but not enough to satisfy watermarks.
2476 count_vm_event(COMPACTFAIL
);
2483 static inline struct page
*
2484 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2485 int alloc_flags
, const struct alloc_context
*ac
,
2486 enum migrate_mode mode
, int *contended_compaction
,
2487 bool *deferred_compaction
)
2491 #endif /* CONFIG_COMPACTION */
2493 /* Perform direct synchronous page reclaim */
2495 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
2496 const struct alloc_context
*ac
)
2498 struct reclaim_state reclaim_state
;
2503 /* We now go into synchronous reclaim */
2504 cpuset_memory_pressure_bump();
2505 current
->flags
|= PF_MEMALLOC
;
2506 lockdep_set_current_reclaim_state(gfp_mask
);
2507 reclaim_state
.reclaimed_slab
= 0;
2508 current
->reclaim_state
= &reclaim_state
;
2510 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
2513 current
->reclaim_state
= NULL
;
2514 lockdep_clear_current_reclaim_state();
2515 current
->flags
&= ~PF_MEMALLOC
;
2522 /* The really slow allocator path where we enter direct reclaim */
2523 static inline struct page
*
2524 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2525 int alloc_flags
, const struct alloc_context
*ac
,
2526 unsigned long *did_some_progress
)
2528 struct page
*page
= NULL
;
2529 bool drained
= false;
2531 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
2532 if (unlikely(!(*did_some_progress
)))
2535 /* After successful reclaim, reconsider all zones for allocation */
2536 if (IS_ENABLED(CONFIG_NUMA
))
2537 zlc_clear_zones_full(ac
->zonelist
);
2540 page
= get_page_from_freelist(gfp_mask
, order
,
2541 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2544 * If an allocation failed after direct reclaim, it could be because
2545 * pages are pinned on the per-cpu lists. Drain them and try again
2547 if (!page
&& !drained
) {
2548 drain_all_pages(NULL
);
2557 * This is called in the allocator slow-path if the allocation request is of
2558 * sufficient urgency to ignore watermarks and take other desperate measures
2560 static inline struct page
*
2561 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2562 const struct alloc_context
*ac
)
2567 page
= get_page_from_freelist(gfp_mask
, order
,
2568 ALLOC_NO_WATERMARKS
, ac
);
2570 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2571 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
,
2573 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2578 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
2583 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2584 ac
->high_zoneidx
, ac
->nodemask
)
2585 wakeup_kswapd(zone
, order
, zone_idx(ac
->preferred_zone
));
2589 gfp_to_alloc_flags(gfp_t gfp_mask
)
2591 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2592 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2594 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2595 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2598 * The caller may dip into page reserves a bit more if the caller
2599 * cannot run direct reclaim, or if the caller has realtime scheduling
2600 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2601 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2603 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2607 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2608 * if it can't schedule.
2610 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2611 alloc_flags
|= ALLOC_HARDER
;
2613 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2614 * comment for __cpuset_node_allowed().
2616 alloc_flags
&= ~ALLOC_CPUSET
;
2617 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2618 alloc_flags
|= ALLOC_HARDER
;
2620 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2621 if (gfp_mask
& __GFP_MEMALLOC
)
2622 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2623 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2624 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2625 else if (!in_interrupt() &&
2626 ((current
->flags
& PF_MEMALLOC
) ||
2627 unlikely(test_thread_flag(TIF_MEMDIE
))))
2628 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2631 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2632 alloc_flags
|= ALLOC_CMA
;
2637 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2639 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2642 static inline struct page
*
2643 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2644 struct alloc_context
*ac
)
2646 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2647 struct page
*page
= NULL
;
2649 unsigned long pages_reclaimed
= 0;
2650 unsigned long did_some_progress
;
2651 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2652 bool deferred_compaction
= false;
2653 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2656 * In the slowpath, we sanity check order to avoid ever trying to
2657 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2658 * be using allocators in order of preference for an area that is
2661 if (order
>= MAX_ORDER
) {
2662 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2667 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2668 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2669 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2670 * using a larger set of nodes after it has established that the
2671 * allowed per node queues are empty and that nodes are
2674 if (IS_ENABLED(CONFIG_NUMA
) &&
2675 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2679 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2680 wake_all_kswapds(order
, ac
);
2683 * OK, we're below the kswapd watermark and have kicked background
2684 * reclaim. Now things get more complex, so set up alloc_flags according
2685 * to how we want to proceed.
2687 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2690 * Find the true preferred zone if the allocation is unconstrained by
2693 if (!(alloc_flags
& ALLOC_CPUSET
) && !ac
->nodemask
) {
2694 struct zoneref
*preferred_zoneref
;
2695 preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
2696 ac
->high_zoneidx
, NULL
, &ac
->preferred_zone
);
2697 ac
->classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2700 /* This is the last chance, in general, before the goto nopage. */
2701 page
= get_page_from_freelist(gfp_mask
, order
,
2702 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2706 /* Allocate without watermarks if the context allows */
2707 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2709 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2710 * the allocation is high priority and these type of
2711 * allocations are system rather than user orientated
2713 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2715 page
= __alloc_pages_high_priority(gfp_mask
, order
, ac
);
2722 /* Atomic allocations - we can't balance anything */
2725 * All existing users of the deprecated __GFP_NOFAIL are
2726 * blockable, so warn of any new users that actually allow this
2727 * type of allocation to fail.
2729 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2733 /* Avoid recursion of direct reclaim */
2734 if (current
->flags
& PF_MEMALLOC
)
2737 /* Avoid allocations with no watermarks from looping endlessly */
2738 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2742 * Try direct compaction. The first pass is asynchronous. Subsequent
2743 * attempts after direct reclaim are synchronous
2745 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
2747 &contended_compaction
,
2748 &deferred_compaction
);
2752 /* Checks for THP-specific high-order allocations */
2753 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2755 * If compaction is deferred for high-order allocations, it is
2756 * because sync compaction recently failed. If this is the case
2757 * and the caller requested a THP allocation, we do not want
2758 * to heavily disrupt the system, so we fail the allocation
2759 * instead of entering direct reclaim.
2761 if (deferred_compaction
)
2765 * In all zones where compaction was attempted (and not
2766 * deferred or skipped), lock contention has been detected.
2767 * For THP allocation we do not want to disrupt the others
2768 * so we fallback to base pages instead.
2770 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2774 * If compaction was aborted due to need_resched(), we do not
2775 * want to further increase allocation latency, unless it is
2776 * khugepaged trying to collapse.
2778 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2779 && !(current
->flags
& PF_KTHREAD
))
2784 * It can become very expensive to allocate transparent hugepages at
2785 * fault, so use asynchronous memory compaction for THP unless it is
2786 * khugepaged trying to collapse.
2788 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2789 (current
->flags
& PF_KTHREAD
))
2790 migration_mode
= MIGRATE_SYNC_LIGHT
;
2792 /* Try direct reclaim and then allocating */
2793 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
2794 &did_some_progress
);
2798 /* Check if we should retry the allocation */
2799 pages_reclaimed
+= did_some_progress
;
2800 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2803 * If we fail to make progress by freeing individual
2804 * pages, but the allocation wants us to keep going,
2805 * start OOM killing tasks.
2807 if (!did_some_progress
) {
2808 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
,
2809 &did_some_progress
);
2812 if (!did_some_progress
)
2815 /* Wait for some write requests to complete then retry */
2816 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2820 * High-order allocations do not necessarily loop after
2821 * direct reclaim and reclaim/compaction depends on compaction
2822 * being called after reclaim so call directly if necessary
2824 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2825 alloc_flags
, ac
, migration_mode
,
2826 &contended_compaction
,
2827 &deferred_compaction
);
2833 warn_alloc_failed(gfp_mask
, order
, NULL
);
2839 * This is the 'heart' of the zoned buddy allocator.
2842 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2843 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2845 struct zoneref
*preferred_zoneref
;
2846 struct page
*page
= NULL
;
2847 unsigned int cpuset_mems_cookie
;
2848 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2849 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
2850 struct alloc_context ac
= {
2851 .high_zoneidx
= gfp_zone(gfp_mask
),
2852 .nodemask
= nodemask
,
2853 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
2856 gfp_mask
&= gfp_allowed_mask
;
2858 lockdep_trace_alloc(gfp_mask
);
2860 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2862 if (should_fail_alloc_page(gfp_mask
, order
))
2866 * Check the zones suitable for the gfp_mask contain at least one
2867 * valid zone. It's possible to have an empty zonelist as a result
2868 * of GFP_THISNODE and a memoryless node
2870 if (unlikely(!zonelist
->_zonerefs
->zone
))
2873 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
2874 alloc_flags
|= ALLOC_CMA
;
2877 cpuset_mems_cookie
= read_mems_allowed_begin();
2879 /* We set it here, as __alloc_pages_slowpath might have changed it */
2880 ac
.zonelist
= zonelist
;
2881 /* The preferred zone is used for statistics later */
2882 preferred_zoneref
= first_zones_zonelist(ac
.zonelist
, ac
.high_zoneidx
,
2883 ac
.nodemask
? : &cpuset_current_mems_allowed
,
2884 &ac
.preferred_zone
);
2885 if (!ac
.preferred_zone
)
2887 ac
.classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2889 /* First allocation attempt */
2890 alloc_mask
= gfp_mask
|__GFP_HARDWALL
;
2891 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
2892 if (unlikely(!page
)) {
2894 * Runtime PM, block IO and its error handling path
2895 * can deadlock because I/O on the device might not
2898 alloc_mask
= memalloc_noio_flags(gfp_mask
);
2900 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
2903 if (kmemcheck_enabled
&& page
)
2904 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2906 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
2910 * When updating a task's mems_allowed, it is possible to race with
2911 * parallel threads in such a way that an allocation can fail while
2912 * the mask is being updated. If a page allocation is about to fail,
2913 * check if the cpuset changed during allocation and if so, retry.
2915 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2920 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2923 * Common helper functions.
2925 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2930 * __get_free_pages() returns a 32-bit address, which cannot represent
2933 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2935 page
= alloc_pages(gfp_mask
, order
);
2938 return (unsigned long) page_address(page
);
2940 EXPORT_SYMBOL(__get_free_pages
);
2942 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2944 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2946 EXPORT_SYMBOL(get_zeroed_page
);
2948 void __free_pages(struct page
*page
, unsigned int order
)
2950 if (put_page_testzero(page
)) {
2952 free_hot_cold_page(page
, false);
2954 __free_pages_ok(page
, order
);
2958 EXPORT_SYMBOL(__free_pages
);
2960 void free_pages(unsigned long addr
, unsigned int order
)
2963 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2964 __free_pages(virt_to_page((void *)addr
), order
);
2968 EXPORT_SYMBOL(free_pages
);
2971 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2972 * of the current memory cgroup.
2974 * It should be used when the caller would like to use kmalloc, but since the
2975 * allocation is large, it has to fall back to the page allocator.
2977 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2980 struct mem_cgroup
*memcg
= NULL
;
2982 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2984 page
= alloc_pages(gfp_mask
, order
);
2985 memcg_kmem_commit_charge(page
, memcg
, order
);
2989 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2992 struct mem_cgroup
*memcg
= NULL
;
2994 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2996 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2997 memcg_kmem_commit_charge(page
, memcg
, order
);
3002 * __free_kmem_pages and free_kmem_pages will free pages allocated with
3005 void __free_kmem_pages(struct page
*page
, unsigned int order
)
3007 memcg_kmem_uncharge_pages(page
, order
);
3008 __free_pages(page
, order
);
3011 void free_kmem_pages(unsigned long addr
, unsigned int order
)
3014 VM_BUG_ON(!virt_addr_valid((void *)addr
));
3015 __free_kmem_pages(virt_to_page((void *)addr
), order
);
3019 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
3022 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
3023 unsigned long used
= addr
+ PAGE_ALIGN(size
);
3025 split_page(virt_to_page((void *)addr
), order
);
3026 while (used
< alloc_end
) {
3031 return (void *)addr
;
3035 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
3036 * @size: the number of bytes to allocate
3037 * @gfp_mask: GFP flags for the allocation
3039 * This function is similar to alloc_pages(), except that it allocates the
3040 * minimum number of pages to satisfy the request. alloc_pages() can only
3041 * allocate memory in power-of-two pages.
3043 * This function is also limited by MAX_ORDER.
3045 * Memory allocated by this function must be released by free_pages_exact().
3047 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3049 unsigned int order
= get_order(size
);
3052 addr
= __get_free_pages(gfp_mask
, order
);
3053 return make_alloc_exact(addr
, order
, size
);
3055 EXPORT_SYMBOL(alloc_pages_exact
);
3058 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3060 * @nid: the preferred node ID where memory should be allocated
3061 * @size: the number of bytes to allocate
3062 * @gfp_mask: GFP flags for the allocation
3064 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3066 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3069 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3071 unsigned order
= get_order(size
);
3072 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3075 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3079 * free_pages_exact - release memory allocated via alloc_pages_exact()
3080 * @virt: the value returned by alloc_pages_exact.
3081 * @size: size of allocation, same value as passed to alloc_pages_exact().
3083 * Release the memory allocated by a previous call to alloc_pages_exact.
3085 void free_pages_exact(void *virt
, size_t size
)
3087 unsigned long addr
= (unsigned long)virt
;
3088 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3090 while (addr
< end
) {
3095 EXPORT_SYMBOL(free_pages_exact
);
3098 * nr_free_zone_pages - count number of pages beyond high watermark
3099 * @offset: The zone index of the highest zone
3101 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3102 * high watermark within all zones at or below a given zone index. For each
3103 * zone, the number of pages is calculated as:
3104 * managed_pages - high_pages
3106 static unsigned long nr_free_zone_pages(int offset
)
3111 /* Just pick one node, since fallback list is circular */
3112 unsigned long sum
= 0;
3114 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3116 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3117 unsigned long size
= zone
->managed_pages
;
3118 unsigned long high
= high_wmark_pages(zone
);
3127 * nr_free_buffer_pages - count number of pages beyond high watermark
3129 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3130 * watermark within ZONE_DMA and ZONE_NORMAL.
3132 unsigned long nr_free_buffer_pages(void)
3134 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3136 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3139 * nr_free_pagecache_pages - count number of pages beyond high watermark
3141 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3142 * high watermark within all zones.
3144 unsigned long nr_free_pagecache_pages(void)
3146 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3149 static inline void show_node(struct zone
*zone
)
3151 if (IS_ENABLED(CONFIG_NUMA
))
3152 printk("Node %d ", zone_to_nid(zone
));
3155 void si_meminfo(struct sysinfo
*val
)
3157 val
->totalram
= totalram_pages
;
3158 val
->sharedram
= global_page_state(NR_SHMEM
);
3159 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3160 val
->bufferram
= nr_blockdev_pages();
3161 val
->totalhigh
= totalhigh_pages
;
3162 val
->freehigh
= nr_free_highpages();
3163 val
->mem_unit
= PAGE_SIZE
;
3166 EXPORT_SYMBOL(si_meminfo
);
3169 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3171 int zone_type
; /* needs to be signed */
3172 unsigned long managed_pages
= 0;
3173 pg_data_t
*pgdat
= NODE_DATA(nid
);
3175 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3176 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3177 val
->totalram
= managed_pages
;
3178 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3179 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3180 #ifdef CONFIG_HIGHMEM
3181 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3182 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3188 val
->mem_unit
= PAGE_SIZE
;
3193 * Determine whether the node should be displayed or not, depending on whether
3194 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3196 bool skip_free_areas_node(unsigned int flags
, int nid
)
3199 unsigned int cpuset_mems_cookie
;
3201 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3205 cpuset_mems_cookie
= read_mems_allowed_begin();
3206 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3207 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3212 #define K(x) ((x) << (PAGE_SHIFT-10))
3214 static void show_migration_types(unsigned char type
)
3216 static const char types
[MIGRATE_TYPES
] = {
3217 [MIGRATE_UNMOVABLE
] = 'U',
3218 [MIGRATE_RECLAIMABLE
] = 'E',
3219 [MIGRATE_MOVABLE
] = 'M',
3220 [MIGRATE_RESERVE
] = 'R',
3222 [MIGRATE_CMA
] = 'C',
3224 #ifdef CONFIG_MEMORY_ISOLATION
3225 [MIGRATE_ISOLATE
] = 'I',
3228 char tmp
[MIGRATE_TYPES
+ 1];
3232 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3233 if (type
& (1 << i
))
3238 printk("(%s) ", tmp
);
3242 * Show free area list (used inside shift_scroll-lock stuff)
3243 * We also calculate the percentage fragmentation. We do this by counting the
3244 * memory on each free list with the exception of the first item on the list.
3245 * Suppresses nodes that are not allowed by current's cpuset if
3246 * SHOW_MEM_FILTER_NODES is passed.
3248 void show_free_areas(unsigned int filter
)
3253 for_each_populated_zone(zone
) {
3254 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3257 printk("%s per-cpu:\n", zone
->name
);
3259 for_each_online_cpu(cpu
) {
3260 struct per_cpu_pageset
*pageset
;
3262 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3264 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3265 cpu
, pageset
->pcp
.high
,
3266 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3270 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3271 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3273 " dirty:%lu writeback:%lu unstable:%lu\n"
3274 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3275 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3277 global_page_state(NR_ACTIVE_ANON
),
3278 global_page_state(NR_INACTIVE_ANON
),
3279 global_page_state(NR_ISOLATED_ANON
),
3280 global_page_state(NR_ACTIVE_FILE
),
3281 global_page_state(NR_INACTIVE_FILE
),
3282 global_page_state(NR_ISOLATED_FILE
),
3283 global_page_state(NR_UNEVICTABLE
),
3284 global_page_state(NR_FILE_DIRTY
),
3285 global_page_state(NR_WRITEBACK
),
3286 global_page_state(NR_UNSTABLE_NFS
),
3287 global_page_state(NR_FREE_PAGES
),
3288 global_page_state(NR_SLAB_RECLAIMABLE
),
3289 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3290 global_page_state(NR_FILE_MAPPED
),
3291 global_page_state(NR_SHMEM
),
3292 global_page_state(NR_PAGETABLE
),
3293 global_page_state(NR_BOUNCE
),
3294 global_page_state(NR_FREE_CMA_PAGES
));
3296 for_each_populated_zone(zone
) {
3299 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3307 " active_anon:%lukB"
3308 " inactive_anon:%lukB"
3309 " active_file:%lukB"
3310 " inactive_file:%lukB"
3311 " unevictable:%lukB"
3312 " isolated(anon):%lukB"
3313 " isolated(file):%lukB"
3321 " slab_reclaimable:%lukB"
3322 " slab_unreclaimable:%lukB"
3323 " kernel_stack:%lukB"
3328 " writeback_tmp:%lukB"
3329 " pages_scanned:%lu"
3330 " all_unreclaimable? %s"
3333 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3334 K(min_wmark_pages(zone
)),
3335 K(low_wmark_pages(zone
)),
3336 K(high_wmark_pages(zone
)),
3337 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3338 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3339 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3340 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3341 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3342 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3343 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3344 K(zone
->present_pages
),
3345 K(zone
->managed_pages
),
3346 K(zone_page_state(zone
, NR_MLOCK
)),
3347 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3348 K(zone_page_state(zone
, NR_WRITEBACK
)),
3349 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3350 K(zone_page_state(zone
, NR_SHMEM
)),
3351 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3352 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3353 zone_page_state(zone
, NR_KERNEL_STACK
) *
3355 K(zone_page_state(zone
, NR_PAGETABLE
)),
3356 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3357 K(zone_page_state(zone
, NR_BOUNCE
)),
3358 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3359 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3360 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3361 (!zone_reclaimable(zone
) ? "yes" : "no")
3363 printk("lowmem_reserve[]:");
3364 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3365 printk(" %ld", zone
->lowmem_reserve
[i
]);
3369 for_each_populated_zone(zone
) {
3370 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3371 unsigned char types
[MAX_ORDER
];
3373 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3376 printk("%s: ", zone
->name
);
3378 spin_lock_irqsave(&zone
->lock
, flags
);
3379 for (order
= 0; order
< MAX_ORDER
; order
++) {
3380 struct free_area
*area
= &zone
->free_area
[order
];
3383 nr
[order
] = area
->nr_free
;
3384 total
+= nr
[order
] << order
;
3387 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3388 if (!list_empty(&area
->free_list
[type
]))
3389 types
[order
] |= 1 << type
;
3392 spin_unlock_irqrestore(&zone
->lock
, flags
);
3393 for (order
= 0; order
< MAX_ORDER
; order
++) {
3394 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3396 show_migration_types(types
[order
]);
3398 printk("= %lukB\n", K(total
));
3401 hugetlb_show_meminfo();
3403 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3405 show_swap_cache_info();
3408 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3410 zoneref
->zone
= zone
;
3411 zoneref
->zone_idx
= zone_idx(zone
);
3415 * Builds allocation fallback zone lists.
3417 * Add all populated zones of a node to the zonelist.
3419 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3423 enum zone_type zone_type
= MAX_NR_ZONES
;
3427 zone
= pgdat
->node_zones
+ zone_type
;
3428 if (populated_zone(zone
)) {
3429 zoneref_set_zone(zone
,
3430 &zonelist
->_zonerefs
[nr_zones
++]);
3431 check_highest_zone(zone_type
);
3433 } while (zone_type
);
3441 * 0 = automatic detection of better ordering.
3442 * 1 = order by ([node] distance, -zonetype)
3443 * 2 = order by (-zonetype, [node] distance)
3445 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3446 * the same zonelist. So only NUMA can configure this param.
3448 #define ZONELIST_ORDER_DEFAULT 0
3449 #define ZONELIST_ORDER_NODE 1
3450 #define ZONELIST_ORDER_ZONE 2
3452 /* zonelist order in the kernel.
3453 * set_zonelist_order() will set this to NODE or ZONE.
3455 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3456 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3460 /* The value user specified ....changed by config */
3461 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3462 /* string for sysctl */
3463 #define NUMA_ZONELIST_ORDER_LEN 16
3464 char numa_zonelist_order
[16] = "default";
3467 * interface for configure zonelist ordering.
3468 * command line option "numa_zonelist_order"
3469 * = "[dD]efault - default, automatic configuration.
3470 * = "[nN]ode - order by node locality, then by zone within node
3471 * = "[zZ]one - order by zone, then by locality within zone
3474 static int __parse_numa_zonelist_order(char *s
)
3476 if (*s
== 'd' || *s
== 'D') {
3477 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3478 } else if (*s
== 'n' || *s
== 'N') {
3479 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3480 } else if (*s
== 'z' || *s
== 'Z') {
3481 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3484 "Ignoring invalid numa_zonelist_order value: "
3491 static __init
int setup_numa_zonelist_order(char *s
)
3498 ret
= __parse_numa_zonelist_order(s
);
3500 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3504 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3507 * sysctl handler for numa_zonelist_order
3509 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3510 void __user
*buffer
, size_t *length
,
3513 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3515 static DEFINE_MUTEX(zl_order_mutex
);
3517 mutex_lock(&zl_order_mutex
);
3519 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3523 strcpy(saved_string
, (char *)table
->data
);
3525 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3529 int oldval
= user_zonelist_order
;
3531 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3534 * bogus value. restore saved string
3536 strncpy((char *)table
->data
, saved_string
,
3537 NUMA_ZONELIST_ORDER_LEN
);
3538 user_zonelist_order
= oldval
;
3539 } else if (oldval
!= user_zonelist_order
) {
3540 mutex_lock(&zonelists_mutex
);
3541 build_all_zonelists(NULL
, NULL
);
3542 mutex_unlock(&zonelists_mutex
);
3546 mutex_unlock(&zl_order_mutex
);
3551 #define MAX_NODE_LOAD (nr_online_nodes)
3552 static int node_load
[MAX_NUMNODES
];
3555 * find_next_best_node - find the next node that should appear in a given node's fallback list
3556 * @node: node whose fallback list we're appending
3557 * @used_node_mask: nodemask_t of already used nodes
3559 * We use a number of factors to determine which is the next node that should
3560 * appear on a given node's fallback list. The node should not have appeared
3561 * already in @node's fallback list, and it should be the next closest node
3562 * according to the distance array (which contains arbitrary distance values
3563 * from each node to each node in the system), and should also prefer nodes
3564 * with no CPUs, since presumably they'll have very little allocation pressure
3565 * on them otherwise.
3566 * It returns -1 if no node is found.
3568 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3571 int min_val
= INT_MAX
;
3572 int best_node
= NUMA_NO_NODE
;
3573 const struct cpumask
*tmp
= cpumask_of_node(0);
3575 /* Use the local node if we haven't already */
3576 if (!node_isset(node
, *used_node_mask
)) {
3577 node_set(node
, *used_node_mask
);
3581 for_each_node_state(n
, N_MEMORY
) {
3583 /* Don't want a node to appear more than once */
3584 if (node_isset(n
, *used_node_mask
))
3587 /* Use the distance array to find the distance */
3588 val
= node_distance(node
, n
);
3590 /* Penalize nodes under us ("prefer the next node") */
3593 /* Give preference to headless and unused nodes */
3594 tmp
= cpumask_of_node(n
);
3595 if (!cpumask_empty(tmp
))
3596 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3598 /* Slight preference for less loaded node */
3599 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3600 val
+= node_load
[n
];
3602 if (val
< min_val
) {
3609 node_set(best_node
, *used_node_mask
);
3616 * Build zonelists ordered by node and zones within node.
3617 * This results in maximum locality--normal zone overflows into local
3618 * DMA zone, if any--but risks exhausting DMA zone.
3620 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3623 struct zonelist
*zonelist
;
3625 zonelist
= &pgdat
->node_zonelists
[0];
3626 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3628 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3629 zonelist
->_zonerefs
[j
].zone
= NULL
;
3630 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3634 * Build gfp_thisnode zonelists
3636 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3639 struct zonelist
*zonelist
;
3641 zonelist
= &pgdat
->node_zonelists
[1];
3642 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3643 zonelist
->_zonerefs
[j
].zone
= NULL
;
3644 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3648 * Build zonelists ordered by zone and nodes within zones.
3649 * This results in conserving DMA zone[s] until all Normal memory is
3650 * exhausted, but results in overflowing to remote node while memory
3651 * may still exist in local DMA zone.
3653 static int node_order
[MAX_NUMNODES
];
3655 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3658 int zone_type
; /* needs to be signed */
3660 struct zonelist
*zonelist
;
3662 zonelist
= &pgdat
->node_zonelists
[0];
3664 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3665 for (j
= 0; j
< nr_nodes
; j
++) {
3666 node
= node_order
[j
];
3667 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3668 if (populated_zone(z
)) {
3670 &zonelist
->_zonerefs
[pos
++]);
3671 check_highest_zone(zone_type
);
3675 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3676 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3679 #if defined(CONFIG_64BIT)
3681 * Devices that require DMA32/DMA are relatively rare and do not justify a
3682 * penalty to every machine in case the specialised case applies. Default
3683 * to Node-ordering on 64-bit NUMA machines
3685 static int default_zonelist_order(void)
3687 return ZONELIST_ORDER_NODE
;
3691 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3692 * by the kernel. If processes running on node 0 deplete the low memory zone
3693 * then reclaim will occur more frequency increasing stalls and potentially
3694 * be easier to OOM if a large percentage of the zone is under writeback or
3695 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3696 * Hence, default to zone ordering on 32-bit.
3698 static int default_zonelist_order(void)
3700 return ZONELIST_ORDER_ZONE
;
3702 #endif /* CONFIG_64BIT */
3704 static void set_zonelist_order(void)
3706 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3707 current_zonelist_order
= default_zonelist_order();
3709 current_zonelist_order
= user_zonelist_order
;
3712 static void build_zonelists(pg_data_t
*pgdat
)
3716 nodemask_t used_mask
;
3717 int local_node
, prev_node
;
3718 struct zonelist
*zonelist
;
3719 int order
= current_zonelist_order
;
3721 /* initialize zonelists */
3722 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3723 zonelist
= pgdat
->node_zonelists
+ i
;
3724 zonelist
->_zonerefs
[0].zone
= NULL
;
3725 zonelist
->_zonerefs
[0].zone_idx
= 0;
3728 /* NUMA-aware ordering of nodes */
3729 local_node
= pgdat
->node_id
;
3730 load
= nr_online_nodes
;
3731 prev_node
= local_node
;
3732 nodes_clear(used_mask
);
3734 memset(node_order
, 0, sizeof(node_order
));
3737 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3739 * We don't want to pressure a particular node.
3740 * So adding penalty to the first node in same
3741 * distance group to make it round-robin.
3743 if (node_distance(local_node
, node
) !=
3744 node_distance(local_node
, prev_node
))
3745 node_load
[node
] = load
;
3749 if (order
== ZONELIST_ORDER_NODE
)
3750 build_zonelists_in_node_order(pgdat
, node
);
3752 node_order
[j
++] = node
; /* remember order */
3755 if (order
== ZONELIST_ORDER_ZONE
) {
3756 /* calculate node order -- i.e., DMA last! */
3757 build_zonelists_in_zone_order(pgdat
, j
);
3760 build_thisnode_zonelists(pgdat
);
3763 /* Construct the zonelist performance cache - see further mmzone.h */
3764 static void build_zonelist_cache(pg_data_t
*pgdat
)
3766 struct zonelist
*zonelist
;
3767 struct zonelist_cache
*zlc
;
3770 zonelist
= &pgdat
->node_zonelists
[0];
3771 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3772 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3773 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3774 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3777 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3779 * Return node id of node used for "local" allocations.
3780 * I.e., first node id of first zone in arg node's generic zonelist.
3781 * Used for initializing percpu 'numa_mem', which is used primarily
3782 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3784 int local_memory_node(int node
)
3788 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3789 gfp_zone(GFP_KERNEL
),
3796 #else /* CONFIG_NUMA */
3798 static void set_zonelist_order(void)
3800 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3803 static void build_zonelists(pg_data_t
*pgdat
)
3805 int node
, local_node
;
3807 struct zonelist
*zonelist
;
3809 local_node
= pgdat
->node_id
;
3811 zonelist
= &pgdat
->node_zonelists
[0];
3812 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3815 * Now we build the zonelist so that it contains the zones
3816 * of all the other nodes.
3817 * We don't want to pressure a particular node, so when
3818 * building the zones for node N, we make sure that the
3819 * zones coming right after the local ones are those from
3820 * node N+1 (modulo N)
3822 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3823 if (!node_online(node
))
3825 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3827 for (node
= 0; node
< local_node
; node
++) {
3828 if (!node_online(node
))
3830 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3833 zonelist
->_zonerefs
[j
].zone
= NULL
;
3834 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3837 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3838 static void build_zonelist_cache(pg_data_t
*pgdat
)
3840 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3843 #endif /* CONFIG_NUMA */
3846 * Boot pageset table. One per cpu which is going to be used for all
3847 * zones and all nodes. The parameters will be set in such a way
3848 * that an item put on a list will immediately be handed over to
3849 * the buddy list. This is safe since pageset manipulation is done
3850 * with interrupts disabled.
3852 * The boot_pagesets must be kept even after bootup is complete for
3853 * unused processors and/or zones. They do play a role for bootstrapping
3854 * hotplugged processors.
3856 * zoneinfo_show() and maybe other functions do
3857 * not check if the processor is online before following the pageset pointer.
3858 * Other parts of the kernel may not check if the zone is available.
3860 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3861 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3862 static void setup_zone_pageset(struct zone
*zone
);
3865 * Global mutex to protect against size modification of zonelists
3866 * as well as to serialize pageset setup for the new populated zone.
3868 DEFINE_MUTEX(zonelists_mutex
);
3870 /* return values int ....just for stop_machine() */
3871 static int __build_all_zonelists(void *data
)
3875 pg_data_t
*self
= data
;
3878 memset(node_load
, 0, sizeof(node_load
));
3881 if (self
&& !node_online(self
->node_id
)) {
3882 build_zonelists(self
);
3883 build_zonelist_cache(self
);
3886 for_each_online_node(nid
) {
3887 pg_data_t
*pgdat
= NODE_DATA(nid
);
3889 build_zonelists(pgdat
);
3890 build_zonelist_cache(pgdat
);
3894 * Initialize the boot_pagesets that are going to be used
3895 * for bootstrapping processors. The real pagesets for
3896 * each zone will be allocated later when the per cpu
3897 * allocator is available.
3899 * boot_pagesets are used also for bootstrapping offline
3900 * cpus if the system is already booted because the pagesets
3901 * are needed to initialize allocators on a specific cpu too.
3902 * F.e. the percpu allocator needs the page allocator which
3903 * needs the percpu allocator in order to allocate its pagesets
3904 * (a chicken-egg dilemma).
3906 for_each_possible_cpu(cpu
) {
3907 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3909 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3911 * We now know the "local memory node" for each node--
3912 * i.e., the node of the first zone in the generic zonelist.
3913 * Set up numa_mem percpu variable for on-line cpus. During
3914 * boot, only the boot cpu should be on-line; we'll init the
3915 * secondary cpus' numa_mem as they come on-line. During
3916 * node/memory hotplug, we'll fixup all on-line cpus.
3918 if (cpu_online(cpu
))
3919 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3926 static noinline
void __init
3927 build_all_zonelists_init(void)
3929 __build_all_zonelists(NULL
);
3930 mminit_verify_zonelist();
3931 cpuset_init_current_mems_allowed();
3935 * Called with zonelists_mutex held always
3936 * unless system_state == SYSTEM_BOOTING.
3938 * __ref due to (1) call of __meminit annotated setup_zone_pageset
3939 * [we're only called with non-NULL zone through __meminit paths] and
3940 * (2) call of __init annotated helper build_all_zonelists_init
3941 * [protected by SYSTEM_BOOTING].
3943 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3945 set_zonelist_order();
3947 if (system_state
== SYSTEM_BOOTING
) {
3948 build_all_zonelists_init();
3950 #ifdef CONFIG_MEMORY_HOTPLUG
3952 setup_zone_pageset(zone
);
3954 /* we have to stop all cpus to guarantee there is no user
3956 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3957 /* cpuset refresh routine should be here */
3959 vm_total_pages
= nr_free_pagecache_pages();
3961 * Disable grouping by mobility if the number of pages in the
3962 * system is too low to allow the mechanism to work. It would be
3963 * more accurate, but expensive to check per-zone. This check is
3964 * made on memory-hotadd so a system can start with mobility
3965 * disabled and enable it later
3967 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3968 page_group_by_mobility_disabled
= 1;
3970 page_group_by_mobility_disabled
= 0;
3972 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3973 "Total pages: %ld\n",
3975 zonelist_order_name
[current_zonelist_order
],
3976 page_group_by_mobility_disabled
? "off" : "on",
3979 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
3984 * Helper functions to size the waitqueue hash table.
3985 * Essentially these want to choose hash table sizes sufficiently
3986 * large so that collisions trying to wait on pages are rare.
3987 * But in fact, the number of active page waitqueues on typical
3988 * systems is ridiculously low, less than 200. So this is even
3989 * conservative, even though it seems large.
3991 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3992 * waitqueues, i.e. the size of the waitq table given the number of pages.
3994 #define PAGES_PER_WAITQUEUE 256
3996 #ifndef CONFIG_MEMORY_HOTPLUG
3997 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3999 unsigned long size
= 1;
4001 pages
/= PAGES_PER_WAITQUEUE
;
4003 while (size
< pages
)
4007 * Once we have dozens or even hundreds of threads sleeping
4008 * on IO we've got bigger problems than wait queue collision.
4009 * Limit the size of the wait table to a reasonable size.
4011 size
= min(size
, 4096UL);
4013 return max(size
, 4UL);
4017 * A zone's size might be changed by hot-add, so it is not possible to determine
4018 * a suitable size for its wait_table. So we use the maximum size now.
4020 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
4022 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
4023 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
4024 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
4026 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
4027 * or more by the traditional way. (See above). It equals:
4029 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
4030 * ia64(16K page size) : = ( 8G + 4M)byte.
4031 * powerpc (64K page size) : = (32G +16M)byte.
4033 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
4040 * This is an integer logarithm so that shifts can be used later
4041 * to extract the more random high bits from the multiplicative
4042 * hash function before the remainder is taken.
4044 static inline unsigned long wait_table_bits(unsigned long size
)
4050 * Check if a pageblock contains reserved pages
4052 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
4056 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4057 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
4064 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4065 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4066 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4067 * higher will lead to a bigger reserve which will get freed as contiguous
4068 * blocks as reclaim kicks in
4070 static void setup_zone_migrate_reserve(struct zone
*zone
)
4072 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4074 unsigned long block_migratetype
;
4079 * Get the start pfn, end pfn and the number of blocks to reserve
4080 * We have to be careful to be aligned to pageblock_nr_pages to
4081 * make sure that we always check pfn_valid for the first page in
4084 start_pfn
= zone
->zone_start_pfn
;
4085 end_pfn
= zone_end_pfn(zone
);
4086 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4087 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4091 * Reserve blocks are generally in place to help high-order atomic
4092 * allocations that are short-lived. A min_free_kbytes value that
4093 * would result in more than 2 reserve blocks for atomic allocations
4094 * is assumed to be in place to help anti-fragmentation for the
4095 * future allocation of hugepages at runtime.
4097 reserve
= min(2, reserve
);
4098 old_reserve
= zone
->nr_migrate_reserve_block
;
4100 /* When memory hot-add, we almost always need to do nothing */
4101 if (reserve
== old_reserve
)
4103 zone
->nr_migrate_reserve_block
= reserve
;
4105 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4106 if (!pfn_valid(pfn
))
4108 page
= pfn_to_page(pfn
);
4110 /* Watch out for overlapping nodes */
4111 if (page_to_nid(page
) != zone_to_nid(zone
))
4114 block_migratetype
= get_pageblock_migratetype(page
);
4116 /* Only test what is necessary when the reserves are not met */
4119 * Blocks with reserved pages will never free, skip
4122 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4123 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4126 /* If this block is reserved, account for it */
4127 if (block_migratetype
== MIGRATE_RESERVE
) {
4132 /* Suitable for reserving if this block is movable */
4133 if (block_migratetype
== MIGRATE_MOVABLE
) {
4134 set_pageblock_migratetype(page
,
4136 move_freepages_block(zone
, page
,
4141 } else if (!old_reserve
) {
4143 * At boot time we don't need to scan the whole zone
4144 * for turning off MIGRATE_RESERVE.
4150 * If the reserve is met and this is a previous reserved block,
4153 if (block_migratetype
== MIGRATE_RESERVE
) {
4154 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4155 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4161 * Initially all pages are reserved - free ones are freed
4162 * up by free_all_bootmem() once the early boot process is
4163 * done. Non-atomic initialization, single-pass.
4165 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4166 unsigned long start_pfn
, enum memmap_context context
)
4169 unsigned long end_pfn
= start_pfn
+ size
;
4173 if (highest_memmap_pfn
< end_pfn
- 1)
4174 highest_memmap_pfn
= end_pfn
- 1;
4176 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4177 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4179 * There can be holes in boot-time mem_map[]s
4180 * handed to this function. They do not
4181 * exist on hotplugged memory.
4183 if (context
== MEMMAP_EARLY
) {
4184 if (!early_pfn_valid(pfn
))
4186 if (!early_pfn_in_nid(pfn
, nid
))
4189 page
= pfn_to_page(pfn
);
4190 set_page_links(page
, zone
, nid
, pfn
);
4191 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4192 init_page_count(page
);
4193 page_mapcount_reset(page
);
4194 page_cpupid_reset_last(page
);
4195 SetPageReserved(page
);
4197 * Mark the block movable so that blocks are reserved for
4198 * movable at startup. This will force kernel allocations
4199 * to reserve their blocks rather than leaking throughout
4200 * the address space during boot when many long-lived
4201 * kernel allocations are made. Later some blocks near
4202 * the start are marked MIGRATE_RESERVE by
4203 * setup_zone_migrate_reserve()
4205 * bitmap is created for zone's valid pfn range. but memmap
4206 * can be created for invalid pages (for alignment)
4207 * check here not to call set_pageblock_migratetype() against
4210 if ((z
->zone_start_pfn
<= pfn
)
4211 && (pfn
< zone_end_pfn(z
))
4212 && !(pfn
& (pageblock_nr_pages
- 1)))
4213 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4215 INIT_LIST_HEAD(&page
->lru
);
4216 #ifdef WANT_PAGE_VIRTUAL
4217 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4218 if (!is_highmem_idx(zone
))
4219 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4224 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4226 unsigned int order
, t
;
4227 for_each_migratetype_order(order
, t
) {
4228 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4229 zone
->free_area
[order
].nr_free
= 0;
4233 #ifndef __HAVE_ARCH_MEMMAP_INIT
4234 #define memmap_init(size, nid, zone, start_pfn) \
4235 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4238 static int zone_batchsize(struct zone
*zone
)
4244 * The per-cpu-pages pools are set to around 1000th of the
4245 * size of the zone. But no more than 1/2 of a meg.
4247 * OK, so we don't know how big the cache is. So guess.
4249 batch
= zone
->managed_pages
/ 1024;
4250 if (batch
* PAGE_SIZE
> 512 * 1024)
4251 batch
= (512 * 1024) / PAGE_SIZE
;
4252 batch
/= 4; /* We effectively *= 4 below */
4257 * Clamp the batch to a 2^n - 1 value. Having a power
4258 * of 2 value was found to be more likely to have
4259 * suboptimal cache aliasing properties in some cases.
4261 * For example if 2 tasks are alternately allocating
4262 * batches of pages, one task can end up with a lot
4263 * of pages of one half of the possible page colors
4264 * and the other with pages of the other colors.
4266 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4271 /* The deferral and batching of frees should be suppressed under NOMMU
4274 * The problem is that NOMMU needs to be able to allocate large chunks
4275 * of contiguous memory as there's no hardware page translation to
4276 * assemble apparent contiguous memory from discontiguous pages.
4278 * Queueing large contiguous runs of pages for batching, however,
4279 * causes the pages to actually be freed in smaller chunks. As there
4280 * can be a significant delay between the individual batches being
4281 * recycled, this leads to the once large chunks of space being
4282 * fragmented and becoming unavailable for high-order allocations.
4289 * pcp->high and pcp->batch values are related and dependent on one another:
4290 * ->batch must never be higher then ->high.
4291 * The following function updates them in a safe manner without read side
4294 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4295 * those fields changing asynchronously (acording the the above rule).
4297 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4298 * outside of boot time (or some other assurance that no concurrent updaters
4301 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4302 unsigned long batch
)
4304 /* start with a fail safe value for batch */
4308 /* Update high, then batch, in order */
4315 /* a companion to pageset_set_high() */
4316 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4318 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4321 static void pageset_init(struct per_cpu_pageset
*p
)
4323 struct per_cpu_pages
*pcp
;
4326 memset(p
, 0, sizeof(*p
));
4330 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4331 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4334 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4337 pageset_set_batch(p
, batch
);
4341 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4342 * to the value high for the pageset p.
4344 static void pageset_set_high(struct per_cpu_pageset
*p
,
4347 unsigned long batch
= max(1UL, high
/ 4);
4348 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4349 batch
= PAGE_SHIFT
* 8;
4351 pageset_update(&p
->pcp
, high
, batch
);
4354 static void pageset_set_high_and_batch(struct zone
*zone
,
4355 struct per_cpu_pageset
*pcp
)
4357 if (percpu_pagelist_fraction
)
4358 pageset_set_high(pcp
,
4359 (zone
->managed_pages
/
4360 percpu_pagelist_fraction
));
4362 pageset_set_batch(pcp
, zone_batchsize(zone
));
4365 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4367 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4370 pageset_set_high_and_batch(zone
, pcp
);
4373 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4376 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4377 for_each_possible_cpu(cpu
)
4378 zone_pageset_init(zone
, cpu
);
4382 * Allocate per cpu pagesets and initialize them.
4383 * Before this call only boot pagesets were available.
4385 void __init
setup_per_cpu_pageset(void)
4389 for_each_populated_zone(zone
)
4390 setup_zone_pageset(zone
);
4393 static noinline __init_refok
4394 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4400 * The per-page waitqueue mechanism uses hashed waitqueues
4403 zone
->wait_table_hash_nr_entries
=
4404 wait_table_hash_nr_entries(zone_size_pages
);
4405 zone
->wait_table_bits
=
4406 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4407 alloc_size
= zone
->wait_table_hash_nr_entries
4408 * sizeof(wait_queue_head_t
);
4410 if (!slab_is_available()) {
4411 zone
->wait_table
= (wait_queue_head_t
*)
4412 memblock_virt_alloc_node_nopanic(
4413 alloc_size
, zone
->zone_pgdat
->node_id
);
4416 * This case means that a zone whose size was 0 gets new memory
4417 * via memory hot-add.
4418 * But it may be the case that a new node was hot-added. In
4419 * this case vmalloc() will not be able to use this new node's
4420 * memory - this wait_table must be initialized to use this new
4421 * node itself as well.
4422 * To use this new node's memory, further consideration will be
4425 zone
->wait_table
= vmalloc(alloc_size
);
4427 if (!zone
->wait_table
)
4430 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4431 init_waitqueue_head(zone
->wait_table
+ i
);
4436 static __meminit
void zone_pcp_init(struct zone
*zone
)
4439 * per cpu subsystem is not up at this point. The following code
4440 * relies on the ability of the linker to provide the
4441 * offset of a (static) per cpu variable into the per cpu area.
4443 zone
->pageset
= &boot_pageset
;
4445 if (populated_zone(zone
))
4446 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4447 zone
->name
, zone
->present_pages
,
4448 zone_batchsize(zone
));
4451 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4452 unsigned long zone_start_pfn
,
4454 enum memmap_context context
)
4456 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4458 ret
= zone_wait_table_init(zone
, size
);
4461 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4463 zone
->zone_start_pfn
= zone_start_pfn
;
4465 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4466 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4468 (unsigned long)zone_idx(zone
),
4469 zone_start_pfn
, (zone_start_pfn
+ size
));
4471 zone_init_free_lists(zone
);
4476 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4477 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4479 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4481 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4483 unsigned long start_pfn
, end_pfn
;
4486 * NOTE: The following SMP-unsafe globals are only used early in boot
4487 * when the kernel is running single-threaded.
4489 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4490 static int __meminitdata last_nid
;
4492 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4495 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4497 last_start_pfn
= start_pfn
;
4498 last_end_pfn
= end_pfn
;
4504 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4506 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4510 nid
= __early_pfn_to_nid(pfn
);
4513 /* just returns 0 */
4517 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4518 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4522 nid
= __early_pfn_to_nid(pfn
);
4523 if (nid
>= 0 && nid
!= node
)
4530 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4531 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4532 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4534 * If an architecture guarantees that all ranges registered contain no holes
4535 * and may be freed, this this function may be used instead of calling
4536 * memblock_free_early_nid() manually.
4538 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4540 unsigned long start_pfn
, end_pfn
;
4543 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4544 start_pfn
= min(start_pfn
, max_low_pfn
);
4545 end_pfn
= min(end_pfn
, max_low_pfn
);
4547 if (start_pfn
< end_pfn
)
4548 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4549 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4555 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4556 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4558 * If an architecture guarantees that all ranges registered contain no holes and may
4559 * be freed, this function may be used instead of calling memory_present() manually.
4561 void __init
sparse_memory_present_with_active_regions(int nid
)
4563 unsigned long start_pfn
, end_pfn
;
4566 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4567 memory_present(this_nid
, start_pfn
, end_pfn
);
4571 * get_pfn_range_for_nid - Return the start and end page frames for a node
4572 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4573 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4574 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4576 * It returns the start and end page frame of a node based on information
4577 * provided by memblock_set_node(). If called for a node
4578 * with no available memory, a warning is printed and the start and end
4581 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4582 unsigned long *start_pfn
, unsigned long *end_pfn
)
4584 unsigned long this_start_pfn
, this_end_pfn
;
4590 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4591 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4592 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4595 if (*start_pfn
== -1UL)
4600 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4601 * assumption is made that zones within a node are ordered in monotonic
4602 * increasing memory addresses so that the "highest" populated zone is used
4604 static void __init
find_usable_zone_for_movable(void)
4607 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4608 if (zone_index
== ZONE_MOVABLE
)
4611 if (arch_zone_highest_possible_pfn
[zone_index
] >
4612 arch_zone_lowest_possible_pfn
[zone_index
])
4616 VM_BUG_ON(zone_index
== -1);
4617 movable_zone
= zone_index
;
4621 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4622 * because it is sized independent of architecture. Unlike the other zones,
4623 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4624 * in each node depending on the size of each node and how evenly kernelcore
4625 * is distributed. This helper function adjusts the zone ranges
4626 * provided by the architecture for a given node by using the end of the
4627 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4628 * zones within a node are in order of monotonic increases memory addresses
4630 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4631 unsigned long zone_type
,
4632 unsigned long node_start_pfn
,
4633 unsigned long node_end_pfn
,
4634 unsigned long *zone_start_pfn
,
4635 unsigned long *zone_end_pfn
)
4637 /* Only adjust if ZONE_MOVABLE is on this node */
4638 if (zone_movable_pfn
[nid
]) {
4639 /* Size ZONE_MOVABLE */
4640 if (zone_type
== ZONE_MOVABLE
) {
4641 *zone_start_pfn
= zone_movable_pfn
[nid
];
4642 *zone_end_pfn
= min(node_end_pfn
,
4643 arch_zone_highest_possible_pfn
[movable_zone
]);
4645 /* Adjust for ZONE_MOVABLE starting within this range */
4646 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4647 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4648 *zone_end_pfn
= zone_movable_pfn
[nid
];
4650 /* Check if this whole range is within ZONE_MOVABLE */
4651 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4652 *zone_start_pfn
= *zone_end_pfn
;
4657 * Return the number of pages a zone spans in a node, including holes
4658 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4660 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4661 unsigned long zone_type
,
4662 unsigned long node_start_pfn
,
4663 unsigned long node_end_pfn
,
4664 unsigned long *ignored
)
4666 unsigned long zone_start_pfn
, zone_end_pfn
;
4668 /* Get the start and end of the zone */
4669 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4670 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4671 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4672 node_start_pfn
, node_end_pfn
,
4673 &zone_start_pfn
, &zone_end_pfn
);
4675 /* Check that this node has pages within the zone's required range */
4676 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4679 /* Move the zone boundaries inside the node if necessary */
4680 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4681 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4683 /* Return the spanned pages */
4684 return zone_end_pfn
- zone_start_pfn
;
4688 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4689 * then all holes in the requested range will be accounted for.
4691 unsigned long __meminit
__absent_pages_in_range(int nid
,
4692 unsigned long range_start_pfn
,
4693 unsigned long range_end_pfn
)
4695 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4696 unsigned long start_pfn
, end_pfn
;
4699 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4700 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4701 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4702 nr_absent
-= end_pfn
- start_pfn
;
4708 * absent_pages_in_range - Return number of page frames in holes within a range
4709 * @start_pfn: The start PFN to start searching for holes
4710 * @end_pfn: The end PFN to stop searching for holes
4712 * It returns the number of pages frames in memory holes within a range.
4714 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4715 unsigned long end_pfn
)
4717 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4720 /* Return the number of page frames in holes in a zone on a node */
4721 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4722 unsigned long zone_type
,
4723 unsigned long node_start_pfn
,
4724 unsigned long node_end_pfn
,
4725 unsigned long *ignored
)
4727 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4728 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4729 unsigned long zone_start_pfn
, zone_end_pfn
;
4731 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4732 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4734 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4735 node_start_pfn
, node_end_pfn
,
4736 &zone_start_pfn
, &zone_end_pfn
);
4737 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4740 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4741 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4742 unsigned long zone_type
,
4743 unsigned long node_start_pfn
,
4744 unsigned long node_end_pfn
,
4745 unsigned long *zones_size
)
4747 return zones_size
[zone_type
];
4750 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4751 unsigned long zone_type
,
4752 unsigned long node_start_pfn
,
4753 unsigned long node_end_pfn
,
4754 unsigned long *zholes_size
)
4759 return zholes_size
[zone_type
];
4762 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4764 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4765 unsigned long node_start_pfn
,
4766 unsigned long node_end_pfn
,
4767 unsigned long *zones_size
,
4768 unsigned long *zholes_size
)
4770 unsigned long realtotalpages
, totalpages
= 0;
4773 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4774 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4778 pgdat
->node_spanned_pages
= totalpages
;
4780 realtotalpages
= totalpages
;
4781 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4783 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4784 node_start_pfn
, node_end_pfn
,
4786 pgdat
->node_present_pages
= realtotalpages
;
4787 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4791 #ifndef CONFIG_SPARSEMEM
4793 * Calculate the size of the zone->blockflags rounded to an unsigned long
4794 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4795 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4796 * round what is now in bits to nearest long in bits, then return it in
4799 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4801 unsigned long usemapsize
;
4803 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4804 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4805 usemapsize
= usemapsize
>> pageblock_order
;
4806 usemapsize
*= NR_PAGEBLOCK_BITS
;
4807 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4809 return usemapsize
/ 8;
4812 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4814 unsigned long zone_start_pfn
,
4815 unsigned long zonesize
)
4817 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4818 zone
->pageblock_flags
= NULL
;
4820 zone
->pageblock_flags
=
4821 memblock_virt_alloc_node_nopanic(usemapsize
,
4825 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4826 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4827 #endif /* CONFIG_SPARSEMEM */
4829 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4831 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4832 void __paginginit
set_pageblock_order(void)
4836 /* Check that pageblock_nr_pages has not already been setup */
4837 if (pageblock_order
)
4840 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4841 order
= HUGETLB_PAGE_ORDER
;
4843 order
= MAX_ORDER
- 1;
4846 * Assume the largest contiguous order of interest is a huge page.
4847 * This value may be variable depending on boot parameters on IA64 and
4850 pageblock_order
= order
;
4852 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4855 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4856 * is unused as pageblock_order is set at compile-time. See
4857 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4860 void __paginginit
set_pageblock_order(void)
4864 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4866 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4867 unsigned long present_pages
)
4869 unsigned long pages
= spanned_pages
;
4872 * Provide a more accurate estimation if there are holes within
4873 * the zone and SPARSEMEM is in use. If there are holes within the
4874 * zone, each populated memory region may cost us one or two extra
4875 * memmap pages due to alignment because memmap pages for each
4876 * populated regions may not naturally algined on page boundary.
4877 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4879 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4880 IS_ENABLED(CONFIG_SPARSEMEM
))
4881 pages
= present_pages
;
4883 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4887 * Set up the zone data structures:
4888 * - mark all pages reserved
4889 * - mark all memory queues empty
4890 * - clear the memory bitmaps
4892 * NOTE: pgdat should get zeroed by caller.
4894 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4895 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4896 unsigned long *zones_size
, unsigned long *zholes_size
)
4899 int nid
= pgdat
->node_id
;
4900 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4903 pgdat_resize_init(pgdat
);
4904 #ifdef CONFIG_NUMA_BALANCING
4905 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4906 pgdat
->numabalancing_migrate_nr_pages
= 0;
4907 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4909 init_waitqueue_head(&pgdat
->kswapd_wait
);
4910 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4911 pgdat_page_ext_init(pgdat
);
4913 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4914 struct zone
*zone
= pgdat
->node_zones
+ j
;
4915 unsigned long size
, realsize
, freesize
, memmap_pages
;
4917 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4918 node_end_pfn
, zones_size
);
4919 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4925 * Adjust freesize so that it accounts for how much memory
4926 * is used by this zone for memmap. This affects the watermark
4927 * and per-cpu initialisations
4929 memmap_pages
= calc_memmap_size(size
, realsize
);
4930 if (!is_highmem_idx(j
)) {
4931 if (freesize
>= memmap_pages
) {
4932 freesize
-= memmap_pages
;
4935 " %s zone: %lu pages used for memmap\n",
4936 zone_names
[j
], memmap_pages
);
4939 " %s zone: %lu pages exceeds freesize %lu\n",
4940 zone_names
[j
], memmap_pages
, freesize
);
4943 /* Account for reserved pages */
4944 if (j
== 0 && freesize
> dma_reserve
) {
4945 freesize
-= dma_reserve
;
4946 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4947 zone_names
[0], dma_reserve
);
4950 if (!is_highmem_idx(j
))
4951 nr_kernel_pages
+= freesize
;
4952 /* Charge for highmem memmap if there are enough kernel pages */
4953 else if (nr_kernel_pages
> memmap_pages
* 2)
4954 nr_kernel_pages
-= memmap_pages
;
4955 nr_all_pages
+= freesize
;
4957 zone
->spanned_pages
= size
;
4958 zone
->present_pages
= realsize
;
4960 * Set an approximate value for lowmem here, it will be adjusted
4961 * when the bootmem allocator frees pages into the buddy system.
4962 * And all highmem pages will be managed by the buddy system.
4964 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4967 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4969 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4971 zone
->name
= zone_names
[j
];
4972 spin_lock_init(&zone
->lock
);
4973 spin_lock_init(&zone
->lru_lock
);
4974 zone_seqlock_init(zone
);
4975 zone
->zone_pgdat
= pgdat
;
4976 zone_pcp_init(zone
);
4978 /* For bootup, initialized properly in watermark setup */
4979 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4981 lruvec_init(&zone
->lruvec
);
4985 set_pageblock_order();
4986 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4987 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4988 size
, MEMMAP_EARLY
);
4990 memmap_init(size
, nid
, j
, zone_start_pfn
);
4991 zone_start_pfn
+= size
;
4995 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4997 /* Skip empty nodes */
4998 if (!pgdat
->node_spanned_pages
)
5001 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5002 /* ia64 gets its own node_mem_map, before this, without bootmem */
5003 if (!pgdat
->node_mem_map
) {
5004 unsigned long size
, start
, end
;
5008 * The zone's endpoints aren't required to be MAX_ORDER
5009 * aligned but the node_mem_map endpoints must be in order
5010 * for the buddy allocator to function correctly.
5012 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
5013 end
= pgdat_end_pfn(pgdat
);
5014 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
5015 size
= (end
- start
) * sizeof(struct page
);
5016 map
= alloc_remap(pgdat
->node_id
, size
);
5018 map
= memblock_virt_alloc_node_nopanic(size
,
5020 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
5022 #ifndef CONFIG_NEED_MULTIPLE_NODES
5024 * With no DISCONTIG, the global mem_map is just set as node 0's
5026 if (pgdat
== NODE_DATA(0)) {
5027 mem_map
= NODE_DATA(0)->node_mem_map
;
5028 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5029 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
5030 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
5031 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5034 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
5037 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
5038 unsigned long node_start_pfn
, unsigned long *zholes_size
)
5040 pg_data_t
*pgdat
= NODE_DATA(nid
);
5041 unsigned long start_pfn
= 0;
5042 unsigned long end_pfn
= 0;
5044 /* pg_data_t should be reset to zero when it's allocated */
5045 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
5047 pgdat
->node_id
= nid
;
5048 pgdat
->node_start_pfn
= node_start_pfn
;
5049 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5050 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
5051 pr_info("Initmem setup node %d [mem %#018Lx-%#018Lx]\n", nid
,
5052 (u64
)start_pfn
<< PAGE_SHIFT
, ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5054 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5055 zones_size
, zholes_size
);
5057 alloc_node_mem_map(pgdat
);
5058 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5059 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5060 nid
, (unsigned long)pgdat
,
5061 (unsigned long)pgdat
->node_mem_map
);
5064 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
5065 zones_size
, zholes_size
);
5068 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5070 #if MAX_NUMNODES > 1
5072 * Figure out the number of possible node ids.
5074 void __init
setup_nr_node_ids(void)
5077 unsigned int highest
= 0;
5079 for_each_node_mask(node
, node_possible_map
)
5081 nr_node_ids
= highest
+ 1;
5086 * node_map_pfn_alignment - determine the maximum internode alignment
5088 * This function should be called after node map is populated and sorted.
5089 * It calculates the maximum power of two alignment which can distinguish
5092 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5093 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5094 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5095 * shifted, 1GiB is enough and this function will indicate so.
5097 * This is used to test whether pfn -> nid mapping of the chosen memory
5098 * model has fine enough granularity to avoid incorrect mapping for the
5099 * populated node map.
5101 * Returns the determined alignment in pfn's. 0 if there is no alignment
5102 * requirement (single node).
5104 unsigned long __init
node_map_pfn_alignment(void)
5106 unsigned long accl_mask
= 0, last_end
= 0;
5107 unsigned long start
, end
, mask
;
5111 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5112 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5119 * Start with a mask granular enough to pin-point to the
5120 * start pfn and tick off bits one-by-one until it becomes
5121 * too coarse to separate the current node from the last.
5123 mask
= ~((1 << __ffs(start
)) - 1);
5124 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5127 /* accumulate all internode masks */
5131 /* convert mask to number of pages */
5132 return ~accl_mask
+ 1;
5135 /* Find the lowest pfn for a node */
5136 static unsigned long __init
find_min_pfn_for_node(int nid
)
5138 unsigned long min_pfn
= ULONG_MAX
;
5139 unsigned long start_pfn
;
5142 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5143 min_pfn
= min(min_pfn
, start_pfn
);
5145 if (min_pfn
== ULONG_MAX
) {
5147 "Could not find start_pfn for node %d\n", nid
);
5155 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5157 * It returns the minimum PFN based on information provided via
5158 * memblock_set_node().
5160 unsigned long __init
find_min_pfn_with_active_regions(void)
5162 return find_min_pfn_for_node(MAX_NUMNODES
);
5166 * early_calculate_totalpages()
5167 * Sum pages in active regions for movable zone.
5168 * Populate N_MEMORY for calculating usable_nodes.
5170 static unsigned long __init
early_calculate_totalpages(void)
5172 unsigned long totalpages
= 0;
5173 unsigned long start_pfn
, end_pfn
;
5176 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5177 unsigned long pages
= end_pfn
- start_pfn
;
5179 totalpages
+= pages
;
5181 node_set_state(nid
, N_MEMORY
);
5187 * Find the PFN the Movable zone begins in each node. Kernel memory
5188 * is spread evenly between nodes as long as the nodes have enough
5189 * memory. When they don't, some nodes will have more kernelcore than
5192 static void __init
find_zone_movable_pfns_for_nodes(void)
5195 unsigned long usable_startpfn
;
5196 unsigned long kernelcore_node
, kernelcore_remaining
;
5197 /* save the state before borrow the nodemask */
5198 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5199 unsigned long totalpages
= early_calculate_totalpages();
5200 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5201 struct memblock_region
*r
;
5203 /* Need to find movable_zone earlier when movable_node is specified. */
5204 find_usable_zone_for_movable();
5207 * If movable_node is specified, ignore kernelcore and movablecore
5210 if (movable_node_is_enabled()) {
5211 for_each_memblock(memory
, r
) {
5212 if (!memblock_is_hotpluggable(r
))
5217 usable_startpfn
= PFN_DOWN(r
->base
);
5218 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5219 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5227 * If movablecore=nn[KMG] was specified, calculate what size of
5228 * kernelcore that corresponds so that memory usable for
5229 * any allocation type is evenly spread. If both kernelcore
5230 * and movablecore are specified, then the value of kernelcore
5231 * will be used for required_kernelcore if it's greater than
5232 * what movablecore would have allowed.
5234 if (required_movablecore
) {
5235 unsigned long corepages
;
5238 * Round-up so that ZONE_MOVABLE is at least as large as what
5239 * was requested by the user
5241 required_movablecore
=
5242 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5243 corepages
= totalpages
- required_movablecore
;
5245 required_kernelcore
= max(required_kernelcore
, corepages
);
5248 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5249 if (!required_kernelcore
)
5252 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5253 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5256 /* Spread kernelcore memory as evenly as possible throughout nodes */
5257 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5258 for_each_node_state(nid
, N_MEMORY
) {
5259 unsigned long start_pfn
, end_pfn
;
5262 * Recalculate kernelcore_node if the division per node
5263 * now exceeds what is necessary to satisfy the requested
5264 * amount of memory for the kernel
5266 if (required_kernelcore
< kernelcore_node
)
5267 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5270 * As the map is walked, we track how much memory is usable
5271 * by the kernel using kernelcore_remaining. When it is
5272 * 0, the rest of the node is usable by ZONE_MOVABLE
5274 kernelcore_remaining
= kernelcore_node
;
5276 /* Go through each range of PFNs within this node */
5277 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5278 unsigned long size_pages
;
5280 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5281 if (start_pfn
>= end_pfn
)
5284 /* Account for what is only usable for kernelcore */
5285 if (start_pfn
< usable_startpfn
) {
5286 unsigned long kernel_pages
;
5287 kernel_pages
= min(end_pfn
, usable_startpfn
)
5290 kernelcore_remaining
-= min(kernel_pages
,
5291 kernelcore_remaining
);
5292 required_kernelcore
-= min(kernel_pages
,
5293 required_kernelcore
);
5295 /* Continue if range is now fully accounted */
5296 if (end_pfn
<= usable_startpfn
) {
5299 * Push zone_movable_pfn to the end so
5300 * that if we have to rebalance
5301 * kernelcore across nodes, we will
5302 * not double account here
5304 zone_movable_pfn
[nid
] = end_pfn
;
5307 start_pfn
= usable_startpfn
;
5311 * The usable PFN range for ZONE_MOVABLE is from
5312 * start_pfn->end_pfn. Calculate size_pages as the
5313 * number of pages used as kernelcore
5315 size_pages
= end_pfn
- start_pfn
;
5316 if (size_pages
> kernelcore_remaining
)
5317 size_pages
= kernelcore_remaining
;
5318 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5321 * Some kernelcore has been met, update counts and
5322 * break if the kernelcore for this node has been
5325 required_kernelcore
-= min(required_kernelcore
,
5327 kernelcore_remaining
-= size_pages
;
5328 if (!kernelcore_remaining
)
5334 * If there is still required_kernelcore, we do another pass with one
5335 * less node in the count. This will push zone_movable_pfn[nid] further
5336 * along on the nodes that still have memory until kernelcore is
5340 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5344 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5345 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5346 zone_movable_pfn
[nid
] =
5347 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5350 /* restore the node_state */
5351 node_states
[N_MEMORY
] = saved_node_state
;
5354 /* Any regular or high memory on that node ? */
5355 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5357 enum zone_type zone_type
;
5359 if (N_MEMORY
== N_NORMAL_MEMORY
)
5362 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5363 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5364 if (populated_zone(zone
)) {
5365 node_set_state(nid
, N_HIGH_MEMORY
);
5366 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5367 zone_type
<= ZONE_NORMAL
)
5368 node_set_state(nid
, N_NORMAL_MEMORY
);
5375 * free_area_init_nodes - Initialise all pg_data_t and zone data
5376 * @max_zone_pfn: an array of max PFNs for each zone
5378 * This will call free_area_init_node() for each active node in the system.
5379 * Using the page ranges provided by memblock_set_node(), the size of each
5380 * zone in each node and their holes is calculated. If the maximum PFN
5381 * between two adjacent zones match, it is assumed that the zone is empty.
5382 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5383 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5384 * starts where the previous one ended. For example, ZONE_DMA32 starts
5385 * at arch_max_dma_pfn.
5387 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5389 unsigned long start_pfn
, end_pfn
;
5392 /* Record where the zone boundaries are */
5393 memset(arch_zone_lowest_possible_pfn
, 0,
5394 sizeof(arch_zone_lowest_possible_pfn
));
5395 memset(arch_zone_highest_possible_pfn
, 0,
5396 sizeof(arch_zone_highest_possible_pfn
));
5397 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5398 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5399 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5400 if (i
== ZONE_MOVABLE
)
5402 arch_zone_lowest_possible_pfn
[i
] =
5403 arch_zone_highest_possible_pfn
[i
-1];
5404 arch_zone_highest_possible_pfn
[i
] =
5405 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5407 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5408 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5410 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5411 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5412 find_zone_movable_pfns_for_nodes();
5414 /* Print out the zone ranges */
5415 pr_info("Zone ranges:\n");
5416 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5417 if (i
== ZONE_MOVABLE
)
5419 pr_info(" %-8s ", zone_names
[i
]);
5420 if (arch_zone_lowest_possible_pfn
[i
] ==
5421 arch_zone_highest_possible_pfn
[i
])
5424 pr_cont("[mem %#018Lx-%#018Lx]\n",
5425 (u64
)arch_zone_lowest_possible_pfn
[i
]
5427 ((u64
)arch_zone_highest_possible_pfn
[i
]
5428 << PAGE_SHIFT
) - 1);
5431 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5432 pr_info("Movable zone start for each node\n");
5433 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5434 if (zone_movable_pfn
[i
])
5435 pr_info(" Node %d: %#018Lx\n", i
,
5436 (u64
)zone_movable_pfn
[i
] << PAGE_SHIFT
);
5439 /* Print out the early node map */
5440 pr_info("Early memory node ranges\n");
5441 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5442 pr_info(" node %3d: [mem %#018Lx-%#018Lx]\n", nid
,
5443 (u64
)start_pfn
<< PAGE_SHIFT
,
5444 ((u64
)end_pfn
<< PAGE_SHIFT
) - 1);
5446 /* Initialise every node */
5447 mminit_verify_pageflags_layout();
5448 setup_nr_node_ids();
5449 for_each_online_node(nid
) {
5450 pg_data_t
*pgdat
= NODE_DATA(nid
);
5451 free_area_init_node(nid
, NULL
,
5452 find_min_pfn_for_node(nid
), NULL
);
5454 /* Any memory on that node */
5455 if (pgdat
->node_present_pages
)
5456 node_set_state(nid
, N_MEMORY
);
5457 check_for_memory(pgdat
, nid
);
5461 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5463 unsigned long long coremem
;
5467 coremem
= memparse(p
, &p
);
5468 *core
= coremem
>> PAGE_SHIFT
;
5470 /* Paranoid check that UL is enough for the coremem value */
5471 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5477 * kernelcore=size sets the amount of memory for use for allocations that
5478 * cannot be reclaimed or migrated.
5480 static int __init
cmdline_parse_kernelcore(char *p
)
5482 return cmdline_parse_core(p
, &required_kernelcore
);
5486 * movablecore=size sets the amount of memory for use for allocations that
5487 * can be reclaimed or migrated.
5489 static int __init
cmdline_parse_movablecore(char *p
)
5491 return cmdline_parse_core(p
, &required_movablecore
);
5494 early_param("kernelcore", cmdline_parse_kernelcore
);
5495 early_param("movablecore", cmdline_parse_movablecore
);
5497 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5499 void adjust_managed_page_count(struct page
*page
, long count
)
5501 spin_lock(&managed_page_count_lock
);
5502 page_zone(page
)->managed_pages
+= count
;
5503 totalram_pages
+= count
;
5504 #ifdef CONFIG_HIGHMEM
5505 if (PageHighMem(page
))
5506 totalhigh_pages
+= count
;
5508 spin_unlock(&managed_page_count_lock
);
5510 EXPORT_SYMBOL(adjust_managed_page_count
);
5512 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5515 unsigned long pages
= 0;
5517 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5518 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5519 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5520 if ((unsigned int)poison
<= 0xFF)
5521 memset(pos
, poison
, PAGE_SIZE
);
5522 free_reserved_page(virt_to_page(pos
));
5526 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5527 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5531 EXPORT_SYMBOL(free_reserved_area
);
5533 #ifdef CONFIG_HIGHMEM
5534 void free_highmem_page(struct page
*page
)
5536 __free_reserved_page(page
);
5538 page_zone(page
)->managed_pages
++;
5544 void __init
mem_init_print_info(const char *str
)
5546 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5547 unsigned long init_code_size
, init_data_size
;
5549 physpages
= get_num_physpages();
5550 codesize
= _etext
- _stext
;
5551 datasize
= _edata
- _sdata
;
5552 rosize
= __end_rodata
- __start_rodata
;
5553 bss_size
= __bss_stop
- __bss_start
;
5554 init_data_size
= __init_end
- __init_begin
;
5555 init_code_size
= _einittext
- _sinittext
;
5558 * Detect special cases and adjust section sizes accordingly:
5559 * 1) .init.* may be embedded into .data sections
5560 * 2) .init.text.* may be out of [__init_begin, __init_end],
5561 * please refer to arch/tile/kernel/vmlinux.lds.S.
5562 * 3) .rodata.* may be embedded into .text or .data sections.
5564 #define adj_init_size(start, end, size, pos, adj) \
5566 if (start <= pos && pos < end && size > adj) \
5570 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5571 _sinittext
, init_code_size
);
5572 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5573 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5574 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5575 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5577 #undef adj_init_size
5579 pr_info("Memory: %luK/%luK available "
5580 "(%luK kernel code, %luK rwdata, %luK rodata, "
5581 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5582 #ifdef CONFIG_HIGHMEM
5586 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5587 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5588 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5589 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
-10),
5590 totalcma_pages
<< (PAGE_SHIFT
-10),
5591 #ifdef CONFIG_HIGHMEM
5592 totalhigh_pages
<< (PAGE_SHIFT
-10),
5594 str
? ", " : "", str
? str
: "");
5598 * set_dma_reserve - set the specified number of pages reserved in the first zone
5599 * @new_dma_reserve: The number of pages to mark reserved
5601 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5602 * In the DMA zone, a significant percentage may be consumed by kernel image
5603 * and other unfreeable allocations which can skew the watermarks badly. This
5604 * function may optionally be used to account for unfreeable pages in the
5605 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5606 * smaller per-cpu batchsize.
5608 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5610 dma_reserve
= new_dma_reserve
;
5613 void __init
free_area_init(unsigned long *zones_size
)
5615 free_area_init_node(0, zones_size
,
5616 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5619 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5620 unsigned long action
, void *hcpu
)
5622 int cpu
= (unsigned long)hcpu
;
5624 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5625 lru_add_drain_cpu(cpu
);
5629 * Spill the event counters of the dead processor
5630 * into the current processors event counters.
5631 * This artificially elevates the count of the current
5634 vm_events_fold_cpu(cpu
);
5637 * Zero the differential counters of the dead processor
5638 * so that the vm statistics are consistent.
5640 * This is only okay since the processor is dead and cannot
5641 * race with what we are doing.
5643 cpu_vm_stats_fold(cpu
);
5648 void __init
page_alloc_init(void)
5650 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5654 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5655 * or min_free_kbytes changes.
5657 static void calculate_totalreserve_pages(void)
5659 struct pglist_data
*pgdat
;
5660 unsigned long reserve_pages
= 0;
5661 enum zone_type i
, j
;
5663 for_each_online_pgdat(pgdat
) {
5664 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5665 struct zone
*zone
= pgdat
->node_zones
+ i
;
5668 /* Find valid and maximum lowmem_reserve in the zone */
5669 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5670 if (zone
->lowmem_reserve
[j
] > max
)
5671 max
= zone
->lowmem_reserve
[j
];
5674 /* we treat the high watermark as reserved pages. */
5675 max
+= high_wmark_pages(zone
);
5677 if (max
> zone
->managed_pages
)
5678 max
= zone
->managed_pages
;
5679 reserve_pages
+= max
;
5681 * Lowmem reserves are not available to
5682 * GFP_HIGHUSER page cache allocations and
5683 * kswapd tries to balance zones to their high
5684 * watermark. As a result, neither should be
5685 * regarded as dirtyable memory, to prevent a
5686 * situation where reclaim has to clean pages
5687 * in order to balance the zones.
5689 zone
->dirty_balance_reserve
= max
;
5692 dirty_balance_reserve
= reserve_pages
;
5693 totalreserve_pages
= reserve_pages
;
5697 * setup_per_zone_lowmem_reserve - called whenever
5698 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5699 * has a correct pages reserved value, so an adequate number of
5700 * pages are left in the zone after a successful __alloc_pages().
5702 static void setup_per_zone_lowmem_reserve(void)
5704 struct pglist_data
*pgdat
;
5705 enum zone_type j
, idx
;
5707 for_each_online_pgdat(pgdat
) {
5708 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5709 struct zone
*zone
= pgdat
->node_zones
+ j
;
5710 unsigned long managed_pages
= zone
->managed_pages
;
5712 zone
->lowmem_reserve
[j
] = 0;
5716 struct zone
*lower_zone
;
5720 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5721 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5723 lower_zone
= pgdat
->node_zones
+ idx
;
5724 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5725 sysctl_lowmem_reserve_ratio
[idx
];
5726 managed_pages
+= lower_zone
->managed_pages
;
5731 /* update totalreserve_pages */
5732 calculate_totalreserve_pages();
5735 static void __setup_per_zone_wmarks(void)
5737 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5738 unsigned long lowmem_pages
= 0;
5740 unsigned long flags
;
5742 /* Calculate total number of !ZONE_HIGHMEM pages */
5743 for_each_zone(zone
) {
5744 if (!is_highmem(zone
))
5745 lowmem_pages
+= zone
->managed_pages
;
5748 for_each_zone(zone
) {
5751 spin_lock_irqsave(&zone
->lock
, flags
);
5752 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5753 do_div(tmp
, lowmem_pages
);
5754 if (is_highmem(zone
)) {
5756 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5757 * need highmem pages, so cap pages_min to a small
5760 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5761 * deltas controls asynch page reclaim, and so should
5762 * not be capped for highmem.
5764 unsigned long min_pages
;
5766 min_pages
= zone
->managed_pages
/ 1024;
5767 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5768 zone
->watermark
[WMARK_MIN
] = min_pages
;
5771 * If it's a lowmem zone, reserve a number of pages
5772 * proportionate to the zone's size.
5774 zone
->watermark
[WMARK_MIN
] = tmp
;
5777 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5778 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5780 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5781 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5782 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5784 setup_zone_migrate_reserve(zone
);
5785 spin_unlock_irqrestore(&zone
->lock
, flags
);
5788 /* update totalreserve_pages */
5789 calculate_totalreserve_pages();
5793 * setup_per_zone_wmarks - called when min_free_kbytes changes
5794 * or when memory is hot-{added|removed}
5796 * Ensures that the watermark[min,low,high] values for each zone are set
5797 * correctly with respect to min_free_kbytes.
5799 void setup_per_zone_wmarks(void)
5801 mutex_lock(&zonelists_mutex
);
5802 __setup_per_zone_wmarks();
5803 mutex_unlock(&zonelists_mutex
);
5807 * The inactive anon list should be small enough that the VM never has to
5808 * do too much work, but large enough that each inactive page has a chance
5809 * to be referenced again before it is swapped out.
5811 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5812 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5813 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5814 * the anonymous pages are kept on the inactive list.
5817 * memory ratio inactive anon
5818 * -------------------------------------
5827 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5829 unsigned int gb
, ratio
;
5831 /* Zone size in gigabytes */
5832 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5834 ratio
= int_sqrt(10 * gb
);
5838 zone
->inactive_ratio
= ratio
;
5841 static void __meminit
setup_per_zone_inactive_ratio(void)
5846 calculate_zone_inactive_ratio(zone
);
5850 * Initialise min_free_kbytes.
5852 * For small machines we want it small (128k min). For large machines
5853 * we want it large (64MB max). But it is not linear, because network
5854 * bandwidth does not increase linearly with machine size. We use
5856 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5857 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5873 int __meminit
init_per_zone_wmark_min(void)
5875 unsigned long lowmem_kbytes
;
5876 int new_min_free_kbytes
;
5878 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5879 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5881 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5882 min_free_kbytes
= new_min_free_kbytes
;
5883 if (min_free_kbytes
< 128)
5884 min_free_kbytes
= 128;
5885 if (min_free_kbytes
> 65536)
5886 min_free_kbytes
= 65536;
5888 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5889 new_min_free_kbytes
, user_min_free_kbytes
);
5891 setup_per_zone_wmarks();
5892 refresh_zone_stat_thresholds();
5893 setup_per_zone_lowmem_reserve();
5894 setup_per_zone_inactive_ratio();
5897 module_init(init_per_zone_wmark_min
)
5900 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5901 * that we can call two helper functions whenever min_free_kbytes
5904 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5905 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5909 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5914 user_min_free_kbytes
= min_free_kbytes
;
5915 setup_per_zone_wmarks();
5921 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5922 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5927 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5932 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5933 sysctl_min_unmapped_ratio
) / 100;
5937 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5938 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5943 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5948 zone
->min_slab_pages
= (zone
->managed_pages
*
5949 sysctl_min_slab_ratio
) / 100;
5955 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5956 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5957 * whenever sysctl_lowmem_reserve_ratio changes.
5959 * The reserve ratio obviously has absolutely no relation with the
5960 * minimum watermarks. The lowmem reserve ratio can only make sense
5961 * if in function of the boot time zone sizes.
5963 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5964 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5966 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5967 setup_per_zone_lowmem_reserve();
5972 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5973 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5974 * pagelist can have before it gets flushed back to buddy allocator.
5976 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5977 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5980 int old_percpu_pagelist_fraction
;
5983 mutex_lock(&pcp_batch_high_lock
);
5984 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5986 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5987 if (!write
|| ret
< 0)
5990 /* Sanity checking to avoid pcp imbalance */
5991 if (percpu_pagelist_fraction
&&
5992 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5993 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5999 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
6002 for_each_populated_zone(zone
) {
6005 for_each_possible_cpu(cpu
)
6006 pageset_set_high_and_batch(zone
,
6007 per_cpu_ptr(zone
->pageset
, cpu
));
6010 mutex_unlock(&pcp_batch_high_lock
);
6014 int hashdist
= HASHDIST_DEFAULT
;
6017 static int __init
set_hashdist(char *str
)
6021 hashdist
= simple_strtoul(str
, &str
, 0);
6024 __setup("hashdist=", set_hashdist
);
6028 * allocate a large system hash table from bootmem
6029 * - it is assumed that the hash table must contain an exact power-of-2
6030 * quantity of entries
6031 * - limit is the number of hash buckets, not the total allocation size
6033 void *__init
alloc_large_system_hash(const char *tablename
,
6034 unsigned long bucketsize
,
6035 unsigned long numentries
,
6038 unsigned int *_hash_shift
,
6039 unsigned int *_hash_mask
,
6040 unsigned long low_limit
,
6041 unsigned long high_limit
)
6043 unsigned long long max
= high_limit
;
6044 unsigned long log2qty
, size
;
6047 /* allow the kernel cmdline to have a say */
6049 /* round applicable memory size up to nearest megabyte */
6050 numentries
= nr_kernel_pages
;
6052 /* It isn't necessary when PAGE_SIZE >= 1MB */
6053 if (PAGE_SHIFT
< 20)
6054 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
6056 /* limit to 1 bucket per 2^scale bytes of low memory */
6057 if (scale
> PAGE_SHIFT
)
6058 numentries
>>= (scale
- PAGE_SHIFT
);
6060 numentries
<<= (PAGE_SHIFT
- scale
);
6062 /* Make sure we've got at least a 0-order allocation.. */
6063 if (unlikely(flags
& HASH_SMALL
)) {
6064 /* Makes no sense without HASH_EARLY */
6065 WARN_ON(!(flags
& HASH_EARLY
));
6066 if (!(numentries
>> *_hash_shift
)) {
6067 numentries
= 1UL << *_hash_shift
;
6068 BUG_ON(!numentries
);
6070 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6071 numentries
= PAGE_SIZE
/ bucketsize
;
6073 numentries
= roundup_pow_of_two(numentries
);
6075 /* limit allocation size to 1/16 total memory by default */
6077 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6078 do_div(max
, bucketsize
);
6080 max
= min(max
, 0x80000000ULL
);
6082 if (numentries
< low_limit
)
6083 numentries
= low_limit
;
6084 if (numentries
> max
)
6087 log2qty
= ilog2(numentries
);
6090 size
= bucketsize
<< log2qty
;
6091 if (flags
& HASH_EARLY
)
6092 table
= memblock_virt_alloc_nopanic(size
, 0);
6094 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6097 * If bucketsize is not a power-of-two, we may free
6098 * some pages at the end of hash table which
6099 * alloc_pages_exact() automatically does
6101 if (get_order(size
) < MAX_ORDER
) {
6102 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6103 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6106 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6109 panic("Failed to allocate %s hash table\n", tablename
);
6111 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6114 ilog2(size
) - PAGE_SHIFT
,
6118 *_hash_shift
= log2qty
;
6120 *_hash_mask
= (1 << log2qty
) - 1;
6125 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6126 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6129 #ifdef CONFIG_SPARSEMEM
6130 return __pfn_to_section(pfn
)->pageblock_flags
;
6132 return zone
->pageblock_flags
;
6133 #endif /* CONFIG_SPARSEMEM */
6136 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6138 #ifdef CONFIG_SPARSEMEM
6139 pfn
&= (PAGES_PER_SECTION
-1);
6140 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6142 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6143 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6144 #endif /* CONFIG_SPARSEMEM */
6148 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6149 * @page: The page within the block of interest
6150 * @pfn: The target page frame number
6151 * @end_bitidx: The last bit of interest to retrieve
6152 * @mask: mask of bits that the caller is interested in
6154 * Return: pageblock_bits flags
6156 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6157 unsigned long end_bitidx
,
6161 unsigned long *bitmap
;
6162 unsigned long bitidx
, word_bitidx
;
6165 zone
= page_zone(page
);
6166 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6167 bitidx
= pfn_to_bitidx(zone
, pfn
);
6168 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6169 bitidx
&= (BITS_PER_LONG
-1);
6171 word
= bitmap
[word_bitidx
];
6172 bitidx
+= end_bitidx
;
6173 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6177 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6178 * @page: The page within the block of interest
6179 * @flags: The flags to set
6180 * @pfn: The target page frame number
6181 * @end_bitidx: The last bit of interest
6182 * @mask: mask of bits that the caller is interested in
6184 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6186 unsigned long end_bitidx
,
6190 unsigned long *bitmap
;
6191 unsigned long bitidx
, word_bitidx
;
6192 unsigned long old_word
, word
;
6194 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6196 zone
= page_zone(page
);
6197 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6198 bitidx
= pfn_to_bitidx(zone
, pfn
);
6199 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6200 bitidx
&= (BITS_PER_LONG
-1);
6202 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6204 bitidx
+= end_bitidx
;
6205 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6206 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6208 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6210 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6211 if (word
== old_word
)
6218 * This function checks whether pageblock includes unmovable pages or not.
6219 * If @count is not zero, it is okay to include less @count unmovable pages
6221 * PageLRU check without isolation or lru_lock could race so that
6222 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6223 * expect this function should be exact.
6225 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6226 bool skip_hwpoisoned_pages
)
6228 unsigned long pfn
, iter
, found
;
6232 * For avoiding noise data, lru_add_drain_all() should be called
6233 * If ZONE_MOVABLE, the zone never contains unmovable pages
6235 if (zone_idx(zone
) == ZONE_MOVABLE
)
6237 mt
= get_pageblock_migratetype(page
);
6238 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6241 pfn
= page_to_pfn(page
);
6242 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6243 unsigned long check
= pfn
+ iter
;
6245 if (!pfn_valid_within(check
))
6248 page
= pfn_to_page(check
);
6251 * Hugepages are not in LRU lists, but they're movable.
6252 * We need not scan over tail pages bacause we don't
6253 * handle each tail page individually in migration.
6255 if (PageHuge(page
)) {
6256 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6261 * We can't use page_count without pin a page
6262 * because another CPU can free compound page.
6263 * This check already skips compound tails of THP
6264 * because their page->_count is zero at all time.
6266 if (!atomic_read(&page
->_count
)) {
6267 if (PageBuddy(page
))
6268 iter
+= (1 << page_order(page
)) - 1;
6273 * The HWPoisoned page may be not in buddy system, and
6274 * page_count() is not 0.
6276 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6282 * If there are RECLAIMABLE pages, we need to check
6283 * it. But now, memory offline itself doesn't call
6284 * shrink_node_slabs() and it still to be fixed.
6287 * If the page is not RAM, page_count()should be 0.
6288 * we don't need more check. This is an _used_ not-movable page.
6290 * The problematic thing here is PG_reserved pages. PG_reserved
6291 * is set to both of a memory hole page and a _used_ kernel
6300 bool is_pageblock_removable_nolock(struct page
*page
)
6306 * We have to be careful here because we are iterating over memory
6307 * sections which are not zone aware so we might end up outside of
6308 * the zone but still within the section.
6309 * We have to take care about the node as well. If the node is offline
6310 * its NODE_DATA will be NULL - see page_zone.
6312 if (!node_online(page_to_nid(page
)))
6315 zone
= page_zone(page
);
6316 pfn
= page_to_pfn(page
);
6317 if (!zone_spans_pfn(zone
, pfn
))
6320 return !has_unmovable_pages(zone
, page
, 0, true);
6325 static unsigned long pfn_max_align_down(unsigned long pfn
)
6327 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6328 pageblock_nr_pages
) - 1);
6331 static unsigned long pfn_max_align_up(unsigned long pfn
)
6333 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6334 pageblock_nr_pages
));
6337 /* [start, end) must belong to a single zone. */
6338 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6339 unsigned long start
, unsigned long end
)
6341 /* This function is based on compact_zone() from compaction.c. */
6342 unsigned long nr_reclaimed
;
6343 unsigned long pfn
= start
;
6344 unsigned int tries
= 0;
6349 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6350 if (fatal_signal_pending(current
)) {
6355 if (list_empty(&cc
->migratepages
)) {
6356 cc
->nr_migratepages
= 0;
6357 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6363 } else if (++tries
== 5) {
6364 ret
= ret
< 0 ? ret
: -EBUSY
;
6368 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6370 cc
->nr_migratepages
-= nr_reclaimed
;
6372 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6373 NULL
, 0, cc
->mode
, MR_CMA
);
6376 putback_movable_pages(&cc
->migratepages
);
6383 * alloc_contig_range() -- tries to allocate given range of pages
6384 * @start: start PFN to allocate
6385 * @end: one-past-the-last PFN to allocate
6386 * @migratetype: migratetype of the underlaying pageblocks (either
6387 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6388 * in range must have the same migratetype and it must
6389 * be either of the two.
6391 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6392 * aligned, however it's the caller's responsibility to guarantee that
6393 * we are the only thread that changes migrate type of pageblocks the
6396 * The PFN range must belong to a single zone.
6398 * Returns zero on success or negative error code. On success all
6399 * pages which PFN is in [start, end) are allocated for the caller and
6400 * need to be freed with free_contig_range().
6402 int alloc_contig_range(unsigned long start
, unsigned long end
,
6403 unsigned migratetype
)
6405 unsigned long outer_start
, outer_end
;
6408 struct compact_control cc
= {
6409 .nr_migratepages
= 0,
6411 .zone
= page_zone(pfn_to_page(start
)),
6412 .mode
= MIGRATE_SYNC
,
6413 .ignore_skip_hint
= true,
6415 INIT_LIST_HEAD(&cc
.migratepages
);
6418 * What we do here is we mark all pageblocks in range as
6419 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6420 * have different sizes, and due to the way page allocator
6421 * work, we align the range to biggest of the two pages so
6422 * that page allocator won't try to merge buddies from
6423 * different pageblocks and change MIGRATE_ISOLATE to some
6424 * other migration type.
6426 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6427 * migrate the pages from an unaligned range (ie. pages that
6428 * we are interested in). This will put all the pages in
6429 * range back to page allocator as MIGRATE_ISOLATE.
6431 * When this is done, we take the pages in range from page
6432 * allocator removing them from the buddy system. This way
6433 * page allocator will never consider using them.
6435 * This lets us mark the pageblocks back as
6436 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6437 * aligned range but not in the unaligned, original range are
6438 * put back to page allocator so that buddy can use them.
6441 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6442 pfn_max_align_up(end
), migratetype
,
6447 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6452 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6453 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6454 * more, all pages in [start, end) are free in page allocator.
6455 * What we are going to do is to allocate all pages from
6456 * [start, end) (that is remove them from page allocator).
6458 * The only problem is that pages at the beginning and at the
6459 * end of interesting range may be not aligned with pages that
6460 * page allocator holds, ie. they can be part of higher order
6461 * pages. Because of this, we reserve the bigger range and
6462 * once this is done free the pages we are not interested in.
6464 * We don't have to hold zone->lock here because the pages are
6465 * isolated thus they won't get removed from buddy.
6468 lru_add_drain_all();
6469 drain_all_pages(cc
.zone
);
6472 outer_start
= start
;
6473 while (!PageBuddy(pfn_to_page(outer_start
))) {
6474 if (++order
>= MAX_ORDER
) {
6478 outer_start
&= ~0UL << order
;
6481 /* Make sure the range is really isolated. */
6482 if (test_pages_isolated(outer_start
, end
, false)) {
6483 pr_info("%s: [%lx, %lx) PFNs busy\n",
6484 __func__
, outer_start
, end
);
6489 /* Grab isolated pages from freelists. */
6490 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6496 /* Free head and tail (if any) */
6497 if (start
!= outer_start
)
6498 free_contig_range(outer_start
, start
- outer_start
);
6499 if (end
!= outer_end
)
6500 free_contig_range(end
, outer_end
- end
);
6503 undo_isolate_page_range(pfn_max_align_down(start
),
6504 pfn_max_align_up(end
), migratetype
);
6508 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6510 unsigned int count
= 0;
6512 for (; nr_pages
--; pfn
++) {
6513 struct page
*page
= pfn_to_page(pfn
);
6515 count
+= page_count(page
) != 1;
6518 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6522 #ifdef CONFIG_MEMORY_HOTPLUG
6524 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6525 * page high values need to be recalulated.
6527 void __meminit
zone_pcp_update(struct zone
*zone
)
6530 mutex_lock(&pcp_batch_high_lock
);
6531 for_each_possible_cpu(cpu
)
6532 pageset_set_high_and_batch(zone
,
6533 per_cpu_ptr(zone
->pageset
, cpu
));
6534 mutex_unlock(&pcp_batch_high_lock
);
6538 void zone_pcp_reset(struct zone
*zone
)
6540 unsigned long flags
;
6542 struct per_cpu_pageset
*pset
;
6544 /* avoid races with drain_pages() */
6545 local_irq_save(flags
);
6546 if (zone
->pageset
!= &boot_pageset
) {
6547 for_each_online_cpu(cpu
) {
6548 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6549 drain_zonestat(zone
, pset
);
6551 free_percpu(zone
->pageset
);
6552 zone
->pageset
= &boot_pageset
;
6554 local_irq_restore(flags
);
6557 #ifdef CONFIG_MEMORY_HOTREMOVE
6559 * All pages in the range must be isolated before calling this.
6562 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6566 unsigned int order
, i
;
6568 unsigned long flags
;
6569 /* find the first valid pfn */
6570 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6575 zone
= page_zone(pfn_to_page(pfn
));
6576 spin_lock_irqsave(&zone
->lock
, flags
);
6578 while (pfn
< end_pfn
) {
6579 if (!pfn_valid(pfn
)) {
6583 page
= pfn_to_page(pfn
);
6585 * The HWPoisoned page may be not in buddy system, and
6586 * page_count() is not 0.
6588 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6590 SetPageReserved(page
);
6594 BUG_ON(page_count(page
));
6595 BUG_ON(!PageBuddy(page
));
6596 order
= page_order(page
);
6597 #ifdef CONFIG_DEBUG_VM
6598 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6599 pfn
, 1 << order
, end_pfn
);
6601 list_del(&page
->lru
);
6602 rmv_page_order(page
);
6603 zone
->free_area
[order
].nr_free
--;
6604 for (i
= 0; i
< (1 << order
); i
++)
6605 SetPageReserved((page
+i
));
6606 pfn
+= (1 << order
);
6608 spin_unlock_irqrestore(&zone
->lock
, flags
);
6612 #ifdef CONFIG_MEMORY_FAILURE
6613 bool is_free_buddy_page(struct page
*page
)
6615 struct zone
*zone
= page_zone(page
);
6616 unsigned long pfn
= page_to_pfn(page
);
6617 unsigned long flags
;
6620 spin_lock_irqsave(&zone
->lock
, flags
);
6621 for (order
= 0; order
< MAX_ORDER
; order
++) {
6622 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6624 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6627 spin_unlock_irqrestore(&zone
->lock
, flags
);
6629 return order
< MAX_ORDER
;