2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/debugobjects.h>
52 #include <linux/kmemleak.h>
53 #include <linux/compaction.h>
54 #include <trace/events/kmem.h>
55 #include <linux/prefetch.h>
56 #include <linux/mm_inline.h>
57 #include <linux/migrate.h>
58 #include <linux/page-debug-flags.h>
59 #include <linux/hugetlb.h>
60 #include <linux/sched/rt.h>
62 #include <asm/sections.h>
63 #include <asm/tlbflush.h>
64 #include <asm/div64.h>
67 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
68 static DEFINE_MUTEX(pcp_batch_high_lock
);
69 #define MIN_PERCPU_PAGELIST_FRACTION (8)
71 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
72 DEFINE_PER_CPU(int, numa_node
);
73 EXPORT_PER_CPU_SYMBOL(numa_node
);
76 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
78 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
79 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
80 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
81 * defined in <linux/topology.h>.
83 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
84 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
85 int _node_numa_mem_
[MAX_NUMNODES
];
89 * Array of node states.
91 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
92 [N_POSSIBLE
] = NODE_MASK_ALL
,
93 [N_ONLINE
] = { { [0] = 1UL } },
95 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
97 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
99 #ifdef CONFIG_MOVABLE_NODE
100 [N_MEMORY
] = { { [0] = 1UL } },
102 [N_CPU
] = { { [0] = 1UL } },
105 EXPORT_SYMBOL(node_states
);
107 /* Protect totalram_pages and zone->managed_pages */
108 static DEFINE_SPINLOCK(managed_page_count_lock
);
110 unsigned long totalram_pages __read_mostly
;
111 unsigned long totalreserve_pages __read_mostly
;
113 * When calculating the number of globally allowed dirty pages, there
114 * is a certain number of per-zone reserves that should not be
115 * considered dirtyable memory. This is the sum of those reserves
116 * over all existing zones that contribute dirtyable memory.
118 unsigned long dirty_balance_reserve __read_mostly
;
120 int percpu_pagelist_fraction
;
121 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
123 #ifdef CONFIG_PM_SLEEP
125 * The following functions are used by the suspend/hibernate code to temporarily
126 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
127 * while devices are suspended. To avoid races with the suspend/hibernate code,
128 * they should always be called with pm_mutex held (gfp_allowed_mask also should
129 * only be modified with pm_mutex held, unless the suspend/hibernate code is
130 * guaranteed not to run in parallel with that modification).
133 static gfp_t saved_gfp_mask
;
135 void pm_restore_gfp_mask(void)
137 WARN_ON(!mutex_is_locked(&pm_mutex
));
138 if (saved_gfp_mask
) {
139 gfp_allowed_mask
= saved_gfp_mask
;
144 void pm_restrict_gfp_mask(void)
146 WARN_ON(!mutex_is_locked(&pm_mutex
));
147 WARN_ON(saved_gfp_mask
);
148 saved_gfp_mask
= gfp_allowed_mask
;
149 gfp_allowed_mask
&= ~GFP_IOFS
;
152 bool pm_suspended_storage(void)
154 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
158 #endif /* CONFIG_PM_SLEEP */
160 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
161 int pageblock_order __read_mostly
;
164 static void __free_pages_ok(struct page
*page
, unsigned int order
);
167 * results with 256, 32 in the lowmem_reserve sysctl:
168 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
169 * 1G machine -> (16M dma, 784M normal, 224M high)
170 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
171 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
172 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
174 * TBD: should special case ZONE_DMA32 machines here - in those we normally
175 * don't need any ZONE_NORMAL reservation
177 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
178 #ifdef CONFIG_ZONE_DMA
181 #ifdef CONFIG_ZONE_DMA32
184 #ifdef CONFIG_HIGHMEM
190 EXPORT_SYMBOL(totalram_pages
);
192 static char * const zone_names
[MAX_NR_ZONES
] = {
193 #ifdef CONFIG_ZONE_DMA
196 #ifdef CONFIG_ZONE_DMA32
200 #ifdef CONFIG_HIGHMEM
206 int min_free_kbytes
= 1024;
207 int user_min_free_kbytes
= -1;
209 static unsigned long __meminitdata nr_kernel_pages
;
210 static unsigned long __meminitdata nr_all_pages
;
211 static unsigned long __meminitdata dma_reserve
;
213 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
214 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
215 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
216 static unsigned long __initdata required_kernelcore
;
217 static unsigned long __initdata required_movablecore
;
218 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
220 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
222 EXPORT_SYMBOL(movable_zone
);
223 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
226 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
227 int nr_online_nodes __read_mostly
= 1;
228 EXPORT_SYMBOL(nr_node_ids
);
229 EXPORT_SYMBOL(nr_online_nodes
);
232 int page_group_by_mobility_disabled __read_mostly
;
234 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
236 if (unlikely(page_group_by_mobility_disabled
&&
237 migratetype
< MIGRATE_PCPTYPES
))
238 migratetype
= MIGRATE_UNMOVABLE
;
240 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
241 PB_migrate
, PB_migrate_end
);
244 bool oom_killer_disabled __read_mostly
;
246 #ifdef CONFIG_DEBUG_VM
247 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
251 unsigned long pfn
= page_to_pfn(page
);
252 unsigned long sp
, start_pfn
;
255 seq
= zone_span_seqbegin(zone
);
256 start_pfn
= zone
->zone_start_pfn
;
257 sp
= zone
->spanned_pages
;
258 if (!zone_spans_pfn(zone
, pfn
))
260 } while (zone_span_seqretry(zone
, seq
));
263 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
264 pfn
, zone_to_nid(zone
), zone
->name
,
265 start_pfn
, start_pfn
+ sp
);
270 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
272 if (!pfn_valid_within(page_to_pfn(page
)))
274 if (zone
!= page_zone(page
))
280 * Temporary debugging check for pages not lying within a given zone.
282 static int bad_range(struct zone
*zone
, struct page
*page
)
284 if (page_outside_zone_boundaries(zone
, page
))
286 if (!page_is_consistent(zone
, page
))
292 static inline int bad_range(struct zone
*zone
, struct page
*page
)
298 static void bad_page(struct page
*page
, const char *reason
,
299 unsigned long bad_flags
)
301 static unsigned long resume
;
302 static unsigned long nr_shown
;
303 static unsigned long nr_unshown
;
305 /* Don't complain about poisoned pages */
306 if (PageHWPoison(page
)) {
307 page_mapcount_reset(page
); /* remove PageBuddy */
312 * Allow a burst of 60 reports, then keep quiet for that minute;
313 * or allow a steady drip of one report per second.
315 if (nr_shown
== 60) {
316 if (time_before(jiffies
, resume
)) {
322 "BUG: Bad page state: %lu messages suppressed\n",
329 resume
= jiffies
+ 60 * HZ
;
331 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
332 current
->comm
, page_to_pfn(page
));
333 dump_page_badflags(page
, reason
, bad_flags
);
338 /* Leave bad fields for debug, except PageBuddy could make trouble */
339 page_mapcount_reset(page
); /* remove PageBuddy */
340 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
344 * Higher-order pages are called "compound pages". They are structured thusly:
346 * The first PAGE_SIZE page is called the "head page".
348 * The remaining PAGE_SIZE pages are called "tail pages".
350 * All pages have PG_compound set. All tail pages have their ->first_page
351 * pointing at the head page.
353 * The first tail page's ->lru.next holds the address of the compound page's
354 * put_page() function. Its ->lru.prev holds the order of allocation.
355 * This usage means that zero-order pages may not be compound.
358 static void free_compound_page(struct page
*page
)
360 __free_pages_ok(page
, compound_order(page
));
363 void prep_compound_page(struct page
*page
, unsigned long order
)
366 int nr_pages
= 1 << order
;
368 set_compound_page_dtor(page
, free_compound_page
);
369 set_compound_order(page
, order
);
371 for (i
= 1; i
< nr_pages
; i
++) {
372 struct page
*p
= page
+ i
;
373 set_page_count(p
, 0);
374 p
->first_page
= page
;
375 /* Make sure p->first_page is always valid for PageTail() */
381 /* update __split_huge_page_refcount if you change this function */
382 static int destroy_compound_page(struct page
*page
, unsigned long order
)
385 int nr_pages
= 1 << order
;
388 if (unlikely(compound_order(page
) != order
)) {
389 bad_page(page
, "wrong compound order", 0);
393 __ClearPageHead(page
);
395 for (i
= 1; i
< nr_pages
; i
++) {
396 struct page
*p
= page
+ i
;
398 if (unlikely(!PageTail(p
))) {
399 bad_page(page
, "PageTail not set", 0);
401 } else if (unlikely(p
->first_page
!= page
)) {
402 bad_page(page
, "first_page not consistent", 0);
411 static inline void prep_zero_page(struct page
*page
, unsigned int order
,
417 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
418 * and __GFP_HIGHMEM from hard or soft interrupt context.
420 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
421 for (i
= 0; i
< (1 << order
); i
++)
422 clear_highpage(page
+ i
);
425 #ifdef CONFIG_DEBUG_PAGEALLOC
426 unsigned int _debug_guardpage_minorder
;
428 static int __init
debug_guardpage_minorder_setup(char *buf
)
432 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
433 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
436 _debug_guardpage_minorder
= res
;
437 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
440 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
442 static inline void set_page_guard_flag(struct page
*page
)
444 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
447 static inline void clear_page_guard_flag(struct page
*page
)
449 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
452 static inline void set_page_guard_flag(struct page
*page
) { }
453 static inline void clear_page_guard_flag(struct page
*page
) { }
456 static inline void set_page_order(struct page
*page
, unsigned int order
)
458 set_page_private(page
, order
);
459 __SetPageBuddy(page
);
462 static inline void rmv_page_order(struct page
*page
)
464 __ClearPageBuddy(page
);
465 set_page_private(page
, 0);
469 * This function checks whether a page is free && is the buddy
470 * we can do coalesce a page and its buddy if
471 * (a) the buddy is not in a hole &&
472 * (b) the buddy is in the buddy system &&
473 * (c) a page and its buddy have the same order &&
474 * (d) a page and its buddy are in the same zone.
476 * For recording whether a page is in the buddy system, we set ->_mapcount
477 * PAGE_BUDDY_MAPCOUNT_VALUE.
478 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
479 * serialized by zone->lock.
481 * For recording page's order, we use page_private(page).
483 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
486 if (!pfn_valid_within(page_to_pfn(buddy
)))
489 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
490 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
492 if (page_zone_id(page
) != page_zone_id(buddy
))
498 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
499 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
502 * zone check is done late to avoid uselessly
503 * calculating zone/node ids for pages that could
506 if (page_zone_id(page
) != page_zone_id(buddy
))
515 * Freeing function for a buddy system allocator.
517 * The concept of a buddy system is to maintain direct-mapped table
518 * (containing bit values) for memory blocks of various "orders".
519 * The bottom level table contains the map for the smallest allocatable
520 * units of memory (here, pages), and each level above it describes
521 * pairs of units from the levels below, hence, "buddies".
522 * At a high level, all that happens here is marking the table entry
523 * at the bottom level available, and propagating the changes upward
524 * as necessary, plus some accounting needed to play nicely with other
525 * parts of the VM system.
526 * At each level, we keep a list of pages, which are heads of continuous
527 * free pages of length of (1 << order) and marked with _mapcount
528 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
530 * So when we are allocating or freeing one, we can derive the state of the
531 * other. That is, if we allocate a small block, and both were
532 * free, the remainder of the region must be split into blocks.
533 * If a block is freed, and its buddy is also free, then this
534 * triggers coalescing into a block of larger size.
539 static inline void __free_one_page(struct page
*page
,
541 struct zone
*zone
, unsigned int order
,
544 unsigned long page_idx
;
545 unsigned long combined_idx
;
546 unsigned long uninitialized_var(buddy_idx
);
548 int max_order
= MAX_ORDER
;
550 VM_BUG_ON(!zone_is_initialized(zone
));
552 if (unlikely(PageCompound(page
)))
553 if (unlikely(destroy_compound_page(page
, order
)))
556 VM_BUG_ON(migratetype
== -1);
557 if (is_migrate_isolate(migratetype
)) {
559 * We restrict max order of merging to prevent merge
560 * between freepages on isolate pageblock and normal
561 * pageblock. Without this, pageblock isolation
562 * could cause incorrect freepage accounting.
564 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
566 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
569 page_idx
= pfn
& ((1 << max_order
) - 1);
571 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
572 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
574 while (order
< max_order
- 1) {
575 buddy_idx
= __find_buddy_index(page_idx
, order
);
576 buddy
= page
+ (buddy_idx
- page_idx
);
577 if (!page_is_buddy(page
, buddy
, order
))
580 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
581 * merge with it and move up one order.
583 if (page_is_guard(buddy
)) {
584 clear_page_guard_flag(buddy
);
585 set_page_private(buddy
, 0);
586 if (!is_migrate_isolate(migratetype
)) {
587 __mod_zone_freepage_state(zone
, 1 << order
,
591 list_del(&buddy
->lru
);
592 zone
->free_area
[order
].nr_free
--;
593 rmv_page_order(buddy
);
595 combined_idx
= buddy_idx
& page_idx
;
596 page
= page
+ (combined_idx
- page_idx
);
597 page_idx
= combined_idx
;
600 set_page_order(page
, order
);
603 * If this is not the largest possible page, check if the buddy
604 * of the next-highest order is free. If it is, it's possible
605 * that pages are being freed that will coalesce soon. In case,
606 * that is happening, add the free page to the tail of the list
607 * so it's less likely to be used soon and more likely to be merged
608 * as a higher order page
610 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
611 struct page
*higher_page
, *higher_buddy
;
612 combined_idx
= buddy_idx
& page_idx
;
613 higher_page
= page
+ (combined_idx
- page_idx
);
614 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
615 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
616 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
617 list_add_tail(&page
->lru
,
618 &zone
->free_area
[order
].free_list
[migratetype
]);
623 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
625 zone
->free_area
[order
].nr_free
++;
628 static inline int free_pages_check(struct page
*page
)
630 const char *bad_reason
= NULL
;
631 unsigned long bad_flags
= 0;
633 if (unlikely(page_mapcount(page
)))
634 bad_reason
= "nonzero mapcount";
635 if (unlikely(page
->mapping
!= NULL
))
636 bad_reason
= "non-NULL mapping";
637 if (unlikely(atomic_read(&page
->_count
) != 0))
638 bad_reason
= "nonzero _count";
639 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
640 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
641 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
644 if (unlikely(page
->mem_cgroup
))
645 bad_reason
= "page still charged to cgroup";
647 if (unlikely(bad_reason
)) {
648 bad_page(page
, bad_reason
, bad_flags
);
651 page_cpupid_reset_last(page
);
652 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
653 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
658 * Frees a number of pages from the PCP lists
659 * Assumes all pages on list are in same zone, and of same order.
660 * count is the number of pages to free.
662 * If the zone was previously in an "all pages pinned" state then look to
663 * see if this freeing clears that state.
665 * And clear the zone's pages_scanned counter, to hold off the "all pages are
666 * pinned" detection logic.
668 static void free_pcppages_bulk(struct zone
*zone
, int count
,
669 struct per_cpu_pages
*pcp
)
674 unsigned long nr_scanned
;
676 spin_lock(&zone
->lock
);
677 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
679 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
683 struct list_head
*list
;
686 * Remove pages from lists in a round-robin fashion. A
687 * batch_free count is maintained that is incremented when an
688 * empty list is encountered. This is so more pages are freed
689 * off fuller lists instead of spinning excessively around empty
694 if (++migratetype
== MIGRATE_PCPTYPES
)
696 list
= &pcp
->lists
[migratetype
];
697 } while (list_empty(list
));
699 /* This is the only non-empty list. Free them all. */
700 if (batch_free
== MIGRATE_PCPTYPES
)
701 batch_free
= to_free
;
704 int mt
; /* migratetype of the to-be-freed page */
706 page
= list_entry(list
->prev
, struct page
, lru
);
707 /* must delete as __free_one_page list manipulates */
708 list_del(&page
->lru
);
709 mt
= get_freepage_migratetype(page
);
710 if (unlikely(has_isolate_pageblock(zone
)))
711 mt
= get_pageblock_migratetype(page
);
713 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
714 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
715 trace_mm_page_pcpu_drain(page
, 0, mt
);
716 } while (--to_free
&& --batch_free
&& !list_empty(list
));
718 spin_unlock(&zone
->lock
);
721 static void free_one_page(struct zone
*zone
,
722 struct page
*page
, unsigned long pfn
,
726 unsigned long nr_scanned
;
727 spin_lock(&zone
->lock
);
728 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
730 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
732 if (unlikely(has_isolate_pageblock(zone
) ||
733 is_migrate_isolate(migratetype
))) {
734 migratetype
= get_pfnblock_migratetype(page
, pfn
);
736 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
737 spin_unlock(&zone
->lock
);
740 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
745 VM_BUG_ON_PAGE(PageTail(page
), page
);
746 VM_BUG_ON_PAGE(PageHead(page
) && compound_order(page
) != order
, page
);
748 trace_mm_page_free(page
, order
);
749 kmemcheck_free_shadow(page
, order
);
752 page
->mapping
= NULL
;
753 for (i
= 0; i
< (1 << order
); i
++)
754 bad
+= free_pages_check(page
+ i
);
758 if (!PageHighMem(page
)) {
759 debug_check_no_locks_freed(page_address(page
),
761 debug_check_no_obj_freed(page_address(page
),
764 arch_free_page(page
, order
);
765 kernel_map_pages(page
, 1 << order
, 0);
770 static void __free_pages_ok(struct page
*page
, unsigned int order
)
774 unsigned long pfn
= page_to_pfn(page
);
776 if (!free_pages_prepare(page
, order
))
779 migratetype
= get_pfnblock_migratetype(page
, pfn
);
780 local_irq_save(flags
);
781 __count_vm_events(PGFREE
, 1 << order
);
782 set_freepage_migratetype(page
, migratetype
);
783 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
784 local_irq_restore(flags
);
787 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
789 unsigned int nr_pages
= 1 << order
;
790 struct page
*p
= page
;
794 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
796 __ClearPageReserved(p
);
797 set_page_count(p
, 0);
799 __ClearPageReserved(p
);
800 set_page_count(p
, 0);
802 page_zone(page
)->managed_pages
+= nr_pages
;
803 set_page_refcounted(page
);
804 __free_pages(page
, order
);
808 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
809 void __init
init_cma_reserved_pageblock(struct page
*page
)
811 unsigned i
= pageblock_nr_pages
;
812 struct page
*p
= page
;
815 __ClearPageReserved(p
);
816 set_page_count(p
, 0);
819 set_pageblock_migratetype(page
, MIGRATE_CMA
);
821 if (pageblock_order
>= MAX_ORDER
) {
822 i
= pageblock_nr_pages
;
825 set_page_refcounted(p
);
826 __free_pages(p
, MAX_ORDER
- 1);
827 p
+= MAX_ORDER_NR_PAGES
;
828 } while (i
-= MAX_ORDER_NR_PAGES
);
830 set_page_refcounted(page
);
831 __free_pages(page
, pageblock_order
);
834 adjust_managed_page_count(page
, pageblock_nr_pages
);
839 * The order of subdivision here is critical for the IO subsystem.
840 * Please do not alter this order without good reasons and regression
841 * testing. Specifically, as large blocks of memory are subdivided,
842 * the order in which smaller blocks are delivered depends on the order
843 * they're subdivided in this function. This is the primary factor
844 * influencing the order in which pages are delivered to the IO
845 * subsystem according to empirical testing, and this is also justified
846 * by considering the behavior of a buddy system containing a single
847 * large block of memory acted on by a series of small allocations.
848 * This behavior is a critical factor in sglist merging's success.
852 static inline void expand(struct zone
*zone
, struct page
*page
,
853 int low
, int high
, struct free_area
*area
,
856 unsigned long size
= 1 << high
;
862 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
864 #ifdef CONFIG_DEBUG_PAGEALLOC
865 if (high
< debug_guardpage_minorder()) {
867 * Mark as guard pages (or page), that will allow to
868 * merge back to allocator when buddy will be freed.
869 * Corresponding page table entries will not be touched,
870 * pages will stay not present in virtual address space
872 INIT_LIST_HEAD(&page
[size
].lru
);
873 set_page_guard_flag(&page
[size
]);
874 set_page_private(&page
[size
], high
);
875 /* Guard pages are not available for any usage */
876 __mod_zone_freepage_state(zone
, -(1 << high
),
881 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
883 set_page_order(&page
[size
], high
);
888 * This page is about to be returned from the page allocator
890 static inline int check_new_page(struct page
*page
)
892 const char *bad_reason
= NULL
;
893 unsigned long bad_flags
= 0;
895 if (unlikely(page_mapcount(page
)))
896 bad_reason
= "nonzero mapcount";
897 if (unlikely(page
->mapping
!= NULL
))
898 bad_reason
= "non-NULL mapping";
899 if (unlikely(atomic_read(&page
->_count
) != 0))
900 bad_reason
= "nonzero _count";
901 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
902 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
903 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
906 if (unlikely(page
->mem_cgroup
))
907 bad_reason
= "page still charged to cgroup";
909 if (unlikely(bad_reason
)) {
910 bad_page(page
, bad_reason
, bad_flags
);
916 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
)
920 for (i
= 0; i
< (1 << order
); i
++) {
921 struct page
*p
= page
+ i
;
922 if (unlikely(check_new_page(p
)))
926 set_page_private(page
, 0);
927 set_page_refcounted(page
);
929 arch_alloc_page(page
, order
);
930 kernel_map_pages(page
, 1 << order
, 1);
932 if (gfp_flags
& __GFP_ZERO
)
933 prep_zero_page(page
, order
, gfp_flags
);
935 if (order
&& (gfp_flags
& __GFP_COMP
))
936 prep_compound_page(page
, order
);
942 * Go through the free lists for the given migratetype and remove
943 * the smallest available page from the freelists
946 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
949 unsigned int current_order
;
950 struct free_area
*area
;
953 /* Find a page of the appropriate size in the preferred list */
954 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
955 area
= &(zone
->free_area
[current_order
]);
956 if (list_empty(&area
->free_list
[migratetype
]))
959 page
= list_entry(area
->free_list
[migratetype
].next
,
961 list_del(&page
->lru
);
962 rmv_page_order(page
);
964 expand(zone
, page
, order
, current_order
, area
, migratetype
);
965 set_freepage_migratetype(page
, migratetype
);
974 * This array describes the order lists are fallen back to when
975 * the free lists for the desirable migrate type are depleted
977 static int fallbacks
[MIGRATE_TYPES
][4] = {
978 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
979 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
981 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
982 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
984 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
986 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
987 #ifdef CONFIG_MEMORY_ISOLATION
988 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
993 * Move the free pages in a range to the free lists of the requested type.
994 * Note that start_page and end_pages are not aligned on a pageblock
995 * boundary. If alignment is required, use move_freepages_block()
997 int move_freepages(struct zone
*zone
,
998 struct page
*start_page
, struct page
*end_page
,
1002 unsigned long order
;
1003 int pages_moved
= 0;
1005 #ifndef CONFIG_HOLES_IN_ZONE
1007 * page_zone is not safe to call in this context when
1008 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1009 * anyway as we check zone boundaries in move_freepages_block().
1010 * Remove at a later date when no bug reports exist related to
1011 * grouping pages by mobility
1013 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1016 for (page
= start_page
; page
<= end_page
;) {
1017 /* Make sure we are not inadvertently changing nodes */
1018 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1020 if (!pfn_valid_within(page_to_pfn(page
))) {
1025 if (!PageBuddy(page
)) {
1030 order
= page_order(page
);
1031 list_move(&page
->lru
,
1032 &zone
->free_area
[order
].free_list
[migratetype
]);
1033 set_freepage_migratetype(page
, migratetype
);
1035 pages_moved
+= 1 << order
;
1041 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1044 unsigned long start_pfn
, end_pfn
;
1045 struct page
*start_page
, *end_page
;
1047 start_pfn
= page_to_pfn(page
);
1048 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1049 start_page
= pfn_to_page(start_pfn
);
1050 end_page
= start_page
+ pageblock_nr_pages
- 1;
1051 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1053 /* Do not cross zone boundaries */
1054 if (!zone_spans_pfn(zone
, start_pfn
))
1056 if (!zone_spans_pfn(zone
, end_pfn
))
1059 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1062 static void change_pageblock_range(struct page
*pageblock_page
,
1063 int start_order
, int migratetype
)
1065 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1067 while (nr_pageblocks
--) {
1068 set_pageblock_migratetype(pageblock_page
, migratetype
);
1069 pageblock_page
+= pageblock_nr_pages
;
1074 * If breaking a large block of pages, move all free pages to the preferred
1075 * allocation list. If falling back for a reclaimable kernel allocation, be
1076 * more aggressive about taking ownership of free pages.
1078 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1079 * nor move CMA pages to different free lists. We don't want unmovable pages
1080 * to be allocated from MIGRATE_CMA areas.
1082 * Returns the new migratetype of the pageblock (or the same old migratetype
1083 * if it was unchanged).
1085 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1086 int start_type
, int fallback_type
)
1088 int current_order
= page_order(page
);
1091 * When borrowing from MIGRATE_CMA, we need to release the excess
1092 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1093 * is set to CMA so it is returned to the correct freelist in case
1094 * the page ends up being not actually allocated from the pcp lists.
1096 if (is_migrate_cma(fallback_type
))
1097 return fallback_type
;
1099 /* Take ownership for orders >= pageblock_order */
1100 if (current_order
>= pageblock_order
) {
1101 change_pageblock_range(page
, current_order
, start_type
);
1105 if (current_order
>= pageblock_order
/ 2 ||
1106 start_type
== MIGRATE_RECLAIMABLE
||
1107 page_group_by_mobility_disabled
) {
1110 pages
= move_freepages_block(zone
, page
, start_type
);
1112 /* Claim the whole block if over half of it is free */
1113 if (pages
>= (1 << (pageblock_order
-1)) ||
1114 page_group_by_mobility_disabled
) {
1116 set_pageblock_migratetype(page
, start_type
);
1122 return fallback_type
;
1125 /* Remove an element from the buddy allocator from the fallback list */
1126 static inline struct page
*
1127 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1129 struct free_area
*area
;
1130 unsigned int current_order
;
1132 int migratetype
, new_type
, i
;
1134 /* Find the largest possible block of pages in the other list */
1135 for (current_order
= MAX_ORDER
-1;
1136 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1139 migratetype
= fallbacks
[start_migratetype
][i
];
1141 /* MIGRATE_RESERVE handled later if necessary */
1142 if (migratetype
== MIGRATE_RESERVE
)
1145 area
= &(zone
->free_area
[current_order
]);
1146 if (list_empty(&area
->free_list
[migratetype
]))
1149 page
= list_entry(area
->free_list
[migratetype
].next
,
1153 new_type
= try_to_steal_freepages(zone
, page
,
1157 /* Remove the page from the freelists */
1158 list_del(&page
->lru
);
1159 rmv_page_order(page
);
1161 expand(zone
, page
, order
, current_order
, area
,
1163 /* The freepage_migratetype may differ from pageblock's
1164 * migratetype depending on the decisions in
1165 * try_to_steal_freepages. This is OK as long as it does
1166 * not differ for MIGRATE_CMA type.
1168 set_freepage_migratetype(page
, new_type
);
1170 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1171 start_migratetype
, migratetype
, new_type
);
1181 * Do the hard work of removing an element from the buddy allocator.
1182 * Call me with the zone->lock already held.
1184 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1190 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1192 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1193 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1196 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1197 * is used because __rmqueue_smallest is an inline function
1198 * and we want just one call site
1201 migratetype
= MIGRATE_RESERVE
;
1206 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1211 * Obtain a specified number of elements from the buddy allocator, all under
1212 * a single hold of the lock, for efficiency. Add them to the supplied list.
1213 * Returns the number of new pages which were placed at *list.
1215 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1216 unsigned long count
, struct list_head
*list
,
1217 int migratetype
, bool cold
)
1221 spin_lock(&zone
->lock
);
1222 for (i
= 0; i
< count
; ++i
) {
1223 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1224 if (unlikely(page
== NULL
))
1228 * Split buddy pages returned by expand() are received here
1229 * in physical page order. The page is added to the callers and
1230 * list and the list head then moves forward. From the callers
1231 * perspective, the linked list is ordered by page number in
1232 * some conditions. This is useful for IO devices that can
1233 * merge IO requests if the physical pages are ordered
1237 list_add(&page
->lru
, list
);
1239 list_add_tail(&page
->lru
, list
);
1241 if (is_migrate_cma(get_freepage_migratetype(page
)))
1242 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1245 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1246 spin_unlock(&zone
->lock
);
1252 * Called from the vmstat counter updater to drain pagesets of this
1253 * currently executing processor on remote nodes after they have
1256 * Note that this function must be called with the thread pinned to
1257 * a single processor.
1259 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1261 unsigned long flags
;
1262 int to_drain
, batch
;
1264 local_irq_save(flags
);
1265 batch
= ACCESS_ONCE(pcp
->batch
);
1266 to_drain
= min(pcp
->count
, batch
);
1268 free_pcppages_bulk(zone
, to_drain
, pcp
);
1269 pcp
->count
-= to_drain
;
1271 local_irq_restore(flags
);
1276 * Drain pcplists of the indicated processor and zone.
1278 * The processor must either be the current processor and the
1279 * thread pinned to the current processor or a processor that
1282 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1284 unsigned long flags
;
1285 struct per_cpu_pageset
*pset
;
1286 struct per_cpu_pages
*pcp
;
1288 local_irq_save(flags
);
1289 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1293 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1296 local_irq_restore(flags
);
1300 * Drain pcplists of all zones on the indicated processor.
1302 * The processor must either be the current processor and the
1303 * thread pinned to the current processor or a processor that
1306 static void drain_pages(unsigned int cpu
)
1310 for_each_populated_zone(zone
) {
1311 drain_pages_zone(cpu
, zone
);
1316 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1318 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1319 * the single zone's pages.
1321 void drain_local_pages(struct zone
*zone
)
1323 int cpu
= smp_processor_id();
1326 drain_pages_zone(cpu
, zone
);
1332 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1334 * When zone parameter is non-NULL, spill just the single zone's pages.
1336 * Note that this code is protected against sending an IPI to an offline
1337 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1338 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1339 * nothing keeps CPUs from showing up after we populated the cpumask and
1340 * before the call to on_each_cpu_mask().
1342 void drain_all_pages(struct zone
*zone
)
1347 * Allocate in the BSS so we wont require allocation in
1348 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1350 static cpumask_t cpus_with_pcps
;
1353 * We don't care about racing with CPU hotplug event
1354 * as offline notification will cause the notified
1355 * cpu to drain that CPU pcps and on_each_cpu_mask
1356 * disables preemption as part of its processing
1358 for_each_online_cpu(cpu
) {
1359 struct per_cpu_pageset
*pcp
;
1361 bool has_pcps
= false;
1364 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1368 for_each_populated_zone(z
) {
1369 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1370 if (pcp
->pcp
.count
) {
1378 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1380 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1382 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1386 #ifdef CONFIG_HIBERNATION
1388 void mark_free_pages(struct zone
*zone
)
1390 unsigned long pfn
, max_zone_pfn
;
1391 unsigned long flags
;
1392 unsigned int order
, t
;
1393 struct list_head
*curr
;
1395 if (zone_is_empty(zone
))
1398 spin_lock_irqsave(&zone
->lock
, flags
);
1400 max_zone_pfn
= zone_end_pfn(zone
);
1401 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1402 if (pfn_valid(pfn
)) {
1403 struct page
*page
= pfn_to_page(pfn
);
1405 if (!swsusp_page_is_forbidden(page
))
1406 swsusp_unset_page_free(page
);
1409 for_each_migratetype_order(order
, t
) {
1410 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1413 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1414 for (i
= 0; i
< (1UL << order
); i
++)
1415 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1418 spin_unlock_irqrestore(&zone
->lock
, flags
);
1420 #endif /* CONFIG_PM */
1423 * Free a 0-order page
1424 * cold == true ? free a cold page : free a hot page
1426 void free_hot_cold_page(struct page
*page
, bool cold
)
1428 struct zone
*zone
= page_zone(page
);
1429 struct per_cpu_pages
*pcp
;
1430 unsigned long flags
;
1431 unsigned long pfn
= page_to_pfn(page
);
1434 if (!free_pages_prepare(page
, 0))
1437 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1438 set_freepage_migratetype(page
, migratetype
);
1439 local_irq_save(flags
);
1440 __count_vm_event(PGFREE
);
1443 * We only track unmovable, reclaimable and movable on pcp lists.
1444 * Free ISOLATE pages back to the allocator because they are being
1445 * offlined but treat RESERVE as movable pages so we can get those
1446 * areas back if necessary. Otherwise, we may have to free
1447 * excessively into the page allocator
1449 if (migratetype
>= MIGRATE_PCPTYPES
) {
1450 if (unlikely(is_migrate_isolate(migratetype
))) {
1451 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1454 migratetype
= MIGRATE_MOVABLE
;
1457 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1459 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1461 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1463 if (pcp
->count
>= pcp
->high
) {
1464 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1465 free_pcppages_bulk(zone
, batch
, pcp
);
1466 pcp
->count
-= batch
;
1470 local_irq_restore(flags
);
1474 * Free a list of 0-order pages
1476 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1478 struct page
*page
, *next
;
1480 list_for_each_entry_safe(page
, next
, list
, lru
) {
1481 trace_mm_page_free_batched(page
, cold
);
1482 free_hot_cold_page(page
, cold
);
1487 * split_page takes a non-compound higher-order page, and splits it into
1488 * n (1<<order) sub-pages: page[0..n]
1489 * Each sub-page must be freed individually.
1491 * Note: this is probably too low level an operation for use in drivers.
1492 * Please consult with lkml before using this in your driver.
1494 void split_page(struct page
*page
, unsigned int order
)
1498 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1499 VM_BUG_ON_PAGE(!page_count(page
), page
);
1501 #ifdef CONFIG_KMEMCHECK
1503 * Split shadow pages too, because free(page[0]) would
1504 * otherwise free the whole shadow.
1506 if (kmemcheck_page_is_tracked(page
))
1507 split_page(virt_to_page(page
[0].shadow
), order
);
1510 for (i
= 1; i
< (1 << order
); i
++)
1511 set_page_refcounted(page
+ i
);
1513 EXPORT_SYMBOL_GPL(split_page
);
1515 int __isolate_free_page(struct page
*page
, unsigned int order
)
1517 unsigned long watermark
;
1521 BUG_ON(!PageBuddy(page
));
1523 zone
= page_zone(page
);
1524 mt
= get_pageblock_migratetype(page
);
1526 if (!is_migrate_isolate(mt
)) {
1527 /* Obey watermarks as if the page was being allocated */
1528 watermark
= low_wmark_pages(zone
) + (1 << order
);
1529 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1532 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1535 /* Remove page from free list */
1536 list_del(&page
->lru
);
1537 zone
->free_area
[order
].nr_free
--;
1538 rmv_page_order(page
);
1540 /* Set the pageblock if the isolated page is at least a pageblock */
1541 if (order
>= pageblock_order
- 1) {
1542 struct page
*endpage
= page
+ (1 << order
) - 1;
1543 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1544 int mt
= get_pageblock_migratetype(page
);
1545 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1546 set_pageblock_migratetype(page
,
1551 return 1UL << order
;
1555 * Similar to split_page except the page is already free. As this is only
1556 * being used for migration, the migratetype of the block also changes.
1557 * As this is called with interrupts disabled, the caller is responsible
1558 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1561 * Note: this is probably too low level an operation for use in drivers.
1562 * Please consult with lkml before using this in your driver.
1564 int split_free_page(struct page
*page
)
1569 order
= page_order(page
);
1571 nr_pages
= __isolate_free_page(page
, order
);
1575 /* Split into individual pages */
1576 set_page_refcounted(page
);
1577 split_page(page
, order
);
1582 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1583 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1587 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1588 struct zone
*zone
, unsigned int order
,
1589 gfp_t gfp_flags
, int migratetype
)
1591 unsigned long flags
;
1593 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1596 if (likely(order
== 0)) {
1597 struct per_cpu_pages
*pcp
;
1598 struct list_head
*list
;
1600 local_irq_save(flags
);
1601 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1602 list
= &pcp
->lists
[migratetype
];
1603 if (list_empty(list
)) {
1604 pcp
->count
+= rmqueue_bulk(zone
, 0,
1607 if (unlikely(list_empty(list
)))
1612 page
= list_entry(list
->prev
, struct page
, lru
);
1614 page
= list_entry(list
->next
, struct page
, lru
);
1616 list_del(&page
->lru
);
1619 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1621 * __GFP_NOFAIL is not to be used in new code.
1623 * All __GFP_NOFAIL callers should be fixed so that they
1624 * properly detect and handle allocation failures.
1626 * We most definitely don't want callers attempting to
1627 * allocate greater than order-1 page units with
1630 WARN_ON_ONCE(order
> 1);
1632 spin_lock_irqsave(&zone
->lock
, flags
);
1633 page
= __rmqueue(zone
, order
, migratetype
);
1634 spin_unlock(&zone
->lock
);
1637 __mod_zone_freepage_state(zone
, -(1 << order
),
1638 get_freepage_migratetype(page
));
1641 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1642 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1643 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1644 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1646 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1647 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1648 local_irq_restore(flags
);
1650 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1651 if (prep_new_page(page
, order
, gfp_flags
))
1656 local_irq_restore(flags
);
1660 #ifdef CONFIG_FAIL_PAGE_ALLOC
1663 struct fault_attr attr
;
1665 u32 ignore_gfp_highmem
;
1666 u32 ignore_gfp_wait
;
1668 } fail_page_alloc
= {
1669 .attr
= FAULT_ATTR_INITIALIZER
,
1670 .ignore_gfp_wait
= 1,
1671 .ignore_gfp_highmem
= 1,
1675 static int __init
setup_fail_page_alloc(char *str
)
1677 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1679 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1681 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1683 if (order
< fail_page_alloc
.min_order
)
1685 if (gfp_mask
& __GFP_NOFAIL
)
1687 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1689 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1692 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1695 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1697 static int __init
fail_page_alloc_debugfs(void)
1699 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1702 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1703 &fail_page_alloc
.attr
);
1705 return PTR_ERR(dir
);
1707 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1708 &fail_page_alloc
.ignore_gfp_wait
))
1710 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1711 &fail_page_alloc
.ignore_gfp_highmem
))
1713 if (!debugfs_create_u32("min-order", mode
, dir
,
1714 &fail_page_alloc
.min_order
))
1719 debugfs_remove_recursive(dir
);
1724 late_initcall(fail_page_alloc_debugfs
);
1726 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1728 #else /* CONFIG_FAIL_PAGE_ALLOC */
1730 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1735 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1738 * Return true if free pages are above 'mark'. This takes into account the order
1739 * of the allocation.
1741 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1742 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1745 /* free_pages may go negative - that's OK */
1750 free_pages
-= (1 << order
) - 1;
1751 if (alloc_flags
& ALLOC_HIGH
)
1753 if (alloc_flags
& ALLOC_HARDER
)
1756 /* If allocation can't use CMA areas don't use free CMA pages */
1757 if (!(alloc_flags
& ALLOC_CMA
))
1758 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1761 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1763 for (o
= 0; o
< order
; o
++) {
1764 /* At the next order, this order's pages become unavailable */
1765 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1767 /* Require fewer higher order pages to be free */
1770 if (free_pages
<= min
)
1776 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1777 int classzone_idx
, int alloc_flags
)
1779 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1780 zone_page_state(z
, NR_FREE_PAGES
));
1783 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1784 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1786 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1788 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1789 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1791 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1797 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1798 * skip over zones that are not allowed by the cpuset, or that have
1799 * been recently (in last second) found to be nearly full. See further
1800 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1801 * that have to skip over a lot of full or unallowed zones.
1803 * If the zonelist cache is present in the passed zonelist, then
1804 * returns a pointer to the allowed node mask (either the current
1805 * tasks mems_allowed, or node_states[N_MEMORY].)
1807 * If the zonelist cache is not available for this zonelist, does
1808 * nothing and returns NULL.
1810 * If the fullzones BITMAP in the zonelist cache is stale (more than
1811 * a second since last zap'd) then we zap it out (clear its bits.)
1813 * We hold off even calling zlc_setup, until after we've checked the
1814 * first zone in the zonelist, on the theory that most allocations will
1815 * be satisfied from that first zone, so best to examine that zone as
1816 * quickly as we can.
1818 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1820 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1821 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1823 zlc
= zonelist
->zlcache_ptr
;
1827 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1828 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1829 zlc
->last_full_zap
= jiffies
;
1832 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1833 &cpuset_current_mems_allowed
:
1834 &node_states
[N_MEMORY
];
1835 return allowednodes
;
1839 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1840 * if it is worth looking at further for free memory:
1841 * 1) Check that the zone isn't thought to be full (doesn't have its
1842 * bit set in the zonelist_cache fullzones BITMAP).
1843 * 2) Check that the zones node (obtained from the zonelist_cache
1844 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1845 * Return true (non-zero) if zone is worth looking at further, or
1846 * else return false (zero) if it is not.
1848 * This check -ignores- the distinction between various watermarks,
1849 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1850 * found to be full for any variation of these watermarks, it will
1851 * be considered full for up to one second by all requests, unless
1852 * we are so low on memory on all allowed nodes that we are forced
1853 * into the second scan of the zonelist.
1855 * In the second scan we ignore this zonelist cache and exactly
1856 * apply the watermarks to all zones, even it is slower to do so.
1857 * We are low on memory in the second scan, and should leave no stone
1858 * unturned looking for a free page.
1860 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1861 nodemask_t
*allowednodes
)
1863 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1864 int i
; /* index of *z in zonelist zones */
1865 int n
; /* node that zone *z is on */
1867 zlc
= zonelist
->zlcache_ptr
;
1871 i
= z
- zonelist
->_zonerefs
;
1874 /* This zone is worth trying if it is allowed but not full */
1875 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1879 * Given 'z' scanning a zonelist, set the corresponding bit in
1880 * zlc->fullzones, so that subsequent attempts to allocate a page
1881 * from that zone don't waste time re-examining it.
1883 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1885 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1886 int i
; /* index of *z in zonelist zones */
1888 zlc
= zonelist
->zlcache_ptr
;
1892 i
= z
- zonelist
->_zonerefs
;
1894 set_bit(i
, zlc
->fullzones
);
1898 * clear all zones full, called after direct reclaim makes progress so that
1899 * a zone that was recently full is not skipped over for up to a second
1901 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1903 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1905 zlc
= zonelist
->zlcache_ptr
;
1909 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1912 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1914 return local_zone
->node
== zone
->node
;
1917 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1919 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1923 #else /* CONFIG_NUMA */
1925 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1930 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1931 nodemask_t
*allowednodes
)
1936 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1940 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1944 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1949 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1954 #endif /* CONFIG_NUMA */
1956 static void reset_alloc_batches(struct zone
*preferred_zone
)
1958 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
1961 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
1962 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
1963 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
1964 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1965 } while (zone
++ != preferred_zone
);
1969 * get_page_from_freelist goes through the zonelist trying to allocate
1972 static struct page
*
1973 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1974 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1975 struct zone
*preferred_zone
, int classzone_idx
, int migratetype
)
1978 struct page
*page
= NULL
;
1980 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1981 int zlc_active
= 0; /* set if using zonelist_cache */
1982 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1983 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
1984 (gfp_mask
& __GFP_WRITE
);
1985 int nr_fair_skipped
= 0;
1986 bool zonelist_rescan
;
1989 zonelist_rescan
= false;
1992 * Scan zonelist, looking for a zone with enough free.
1993 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1995 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1996 high_zoneidx
, nodemask
) {
1999 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2000 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2002 if (cpusets_enabled() &&
2003 (alloc_flags
& ALLOC_CPUSET
) &&
2004 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
2007 * Distribute pages in proportion to the individual
2008 * zone size to ensure fair page aging. The zone a
2009 * page was allocated in should have no effect on the
2010 * time the page has in memory before being reclaimed.
2012 if (alloc_flags
& ALLOC_FAIR
) {
2013 if (!zone_local(preferred_zone
, zone
))
2015 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2021 * When allocating a page cache page for writing, we
2022 * want to get it from a zone that is within its dirty
2023 * limit, such that no single zone holds more than its
2024 * proportional share of globally allowed dirty pages.
2025 * The dirty limits take into account the zone's
2026 * lowmem reserves and high watermark so that kswapd
2027 * should be able to balance it without having to
2028 * write pages from its LRU list.
2030 * This may look like it could increase pressure on
2031 * lower zones by failing allocations in higher zones
2032 * before they are full. But the pages that do spill
2033 * over are limited as the lower zones are protected
2034 * by this very same mechanism. It should not become
2035 * a practical burden to them.
2037 * XXX: For now, allow allocations to potentially
2038 * exceed the per-zone dirty limit in the slowpath
2039 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2040 * which is important when on a NUMA setup the allowed
2041 * zones are together not big enough to reach the
2042 * global limit. The proper fix for these situations
2043 * will require awareness of zones in the
2044 * dirty-throttling and the flusher threads.
2046 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2049 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2050 if (!zone_watermark_ok(zone
, order
, mark
,
2051 classzone_idx
, alloc_flags
)) {
2054 /* Checked here to keep the fast path fast */
2055 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2056 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2059 if (IS_ENABLED(CONFIG_NUMA
) &&
2060 !did_zlc_setup
&& nr_online_nodes
> 1) {
2062 * we do zlc_setup if there are multiple nodes
2063 * and before considering the first zone allowed
2066 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2071 if (zone_reclaim_mode
== 0 ||
2072 !zone_allows_reclaim(preferred_zone
, zone
))
2073 goto this_zone_full
;
2076 * As we may have just activated ZLC, check if the first
2077 * eligible zone has failed zone_reclaim recently.
2079 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2080 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2083 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2085 case ZONE_RECLAIM_NOSCAN
:
2088 case ZONE_RECLAIM_FULL
:
2089 /* scanned but unreclaimable */
2092 /* did we reclaim enough */
2093 if (zone_watermark_ok(zone
, order
, mark
,
2094 classzone_idx
, alloc_flags
))
2098 * Failed to reclaim enough to meet watermark.
2099 * Only mark the zone full if checking the min
2100 * watermark or if we failed to reclaim just
2101 * 1<<order pages or else the page allocator
2102 * fastpath will prematurely mark zones full
2103 * when the watermark is between the low and
2106 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2107 ret
== ZONE_RECLAIM_SOME
)
2108 goto this_zone_full
;
2115 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2116 gfp_mask
, migratetype
);
2120 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2121 zlc_mark_zone_full(zonelist
, z
);
2126 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2127 * necessary to allocate the page. The expectation is
2128 * that the caller is taking steps that will free more
2129 * memory. The caller should avoid the page being used
2130 * for !PFMEMALLOC purposes.
2132 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2137 * The first pass makes sure allocations are spread fairly within the
2138 * local node. However, the local node might have free pages left
2139 * after the fairness batches are exhausted, and remote zones haven't
2140 * even been considered yet. Try once more without fairness, and
2141 * include remote zones now, before entering the slowpath and waking
2142 * kswapd: prefer spilling to a remote zone over swapping locally.
2144 if (alloc_flags
& ALLOC_FAIR
) {
2145 alloc_flags
&= ~ALLOC_FAIR
;
2146 if (nr_fair_skipped
) {
2147 zonelist_rescan
= true;
2148 reset_alloc_batches(preferred_zone
);
2150 if (nr_online_nodes
> 1)
2151 zonelist_rescan
= true;
2154 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2155 /* Disable zlc cache for second zonelist scan */
2157 zonelist_rescan
= true;
2160 if (zonelist_rescan
)
2167 * Large machines with many possible nodes should not always dump per-node
2168 * meminfo in irq context.
2170 static inline bool should_suppress_show_mem(void)
2175 ret
= in_interrupt();
2180 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2181 DEFAULT_RATELIMIT_INTERVAL
,
2182 DEFAULT_RATELIMIT_BURST
);
2184 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2186 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2188 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2189 debug_guardpage_minorder() > 0)
2193 * This documents exceptions given to allocations in certain
2194 * contexts that are allowed to allocate outside current's set
2197 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2198 if (test_thread_flag(TIF_MEMDIE
) ||
2199 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2200 filter
&= ~SHOW_MEM_FILTER_NODES
;
2201 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2202 filter
&= ~SHOW_MEM_FILTER_NODES
;
2205 struct va_format vaf
;
2208 va_start(args
, fmt
);
2213 pr_warn("%pV", &vaf
);
2218 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2219 current
->comm
, order
, gfp_mask
);
2222 if (!should_suppress_show_mem())
2227 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2228 unsigned long did_some_progress
,
2229 unsigned long pages_reclaimed
)
2231 /* Do not loop if specifically requested */
2232 if (gfp_mask
& __GFP_NORETRY
)
2235 /* Always retry if specifically requested */
2236 if (gfp_mask
& __GFP_NOFAIL
)
2240 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2241 * making forward progress without invoking OOM. Suspend also disables
2242 * storage devices so kswapd will not help. Bail if we are suspending.
2244 if (!did_some_progress
&& pm_suspended_storage())
2248 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2249 * means __GFP_NOFAIL, but that may not be true in other
2252 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2256 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2257 * specified, then we retry until we no longer reclaim any pages
2258 * (above), or we've reclaimed an order of pages at least as
2259 * large as the allocation's order. In both cases, if the
2260 * allocation still fails, we stop retrying.
2262 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2268 static inline struct page
*
2269 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2270 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2271 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2272 int classzone_idx
, int migratetype
)
2276 /* Acquire the per-zone oom lock for each zone */
2277 if (!oom_zonelist_trylock(zonelist
, gfp_mask
)) {
2278 schedule_timeout_uninterruptible(1);
2283 * PM-freezer should be notified that there might be an OOM killer on
2284 * its way to kill and wake somebody up. This is too early and we might
2285 * end up not killing anything but false positives are acceptable.
2286 * See freeze_processes.
2291 * Go through the zonelist yet one more time, keep very high watermark
2292 * here, this is only to catch a parallel oom killing, we must fail if
2293 * we're still under heavy pressure.
2295 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2296 order
, zonelist
, high_zoneidx
,
2297 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2298 preferred_zone
, classzone_idx
, migratetype
);
2302 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2303 /* The OOM killer will not help higher order allocs */
2304 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2306 /* The OOM killer does not needlessly kill tasks for lowmem */
2307 if (high_zoneidx
< ZONE_NORMAL
)
2310 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2311 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2312 * The caller should handle page allocation failure by itself if
2313 * it specifies __GFP_THISNODE.
2314 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2316 if (gfp_mask
& __GFP_THISNODE
)
2319 /* Exhausted what can be done so it's blamo time */
2320 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2323 oom_zonelist_unlock(zonelist
, gfp_mask
);
2327 #ifdef CONFIG_COMPACTION
2328 /* Try memory compaction for high-order allocations before reclaim */
2329 static struct page
*
2330 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2331 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2332 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2333 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2334 int *contended_compaction
, bool *deferred_compaction
)
2336 unsigned long compact_result
;
2342 current
->flags
|= PF_MEMALLOC
;
2343 compact_result
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2345 contended_compaction
,
2346 alloc_flags
, classzone_idx
);
2347 current
->flags
&= ~PF_MEMALLOC
;
2349 switch (compact_result
) {
2350 case COMPACT_DEFERRED
:
2351 *deferred_compaction
= true;
2353 case COMPACT_SKIPPED
:
2360 * At least in one zone compaction wasn't deferred or skipped, so let's
2361 * count a compaction stall
2363 count_vm_event(COMPACTSTALL
);
2365 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2366 order
, zonelist
, high_zoneidx
,
2367 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2368 preferred_zone
, classzone_idx
, migratetype
);
2371 struct zone
*zone
= page_zone(page
);
2373 zone
->compact_blockskip_flush
= false;
2374 compaction_defer_reset(zone
, order
, true);
2375 count_vm_event(COMPACTSUCCESS
);
2380 * It's bad if compaction run occurs and fails. The most likely reason
2381 * is that pages exist, but not enough to satisfy watermarks.
2383 count_vm_event(COMPACTFAIL
);
2390 static inline struct page
*
2391 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2392 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2393 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2394 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2395 int *contended_compaction
, bool *deferred_compaction
)
2399 #endif /* CONFIG_COMPACTION */
2401 /* Perform direct synchronous page reclaim */
2403 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2404 nodemask_t
*nodemask
)
2406 struct reclaim_state reclaim_state
;
2411 /* We now go into synchronous reclaim */
2412 cpuset_memory_pressure_bump();
2413 current
->flags
|= PF_MEMALLOC
;
2414 lockdep_set_current_reclaim_state(gfp_mask
);
2415 reclaim_state
.reclaimed_slab
= 0;
2416 current
->reclaim_state
= &reclaim_state
;
2418 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2420 current
->reclaim_state
= NULL
;
2421 lockdep_clear_current_reclaim_state();
2422 current
->flags
&= ~PF_MEMALLOC
;
2429 /* The really slow allocator path where we enter direct reclaim */
2430 static inline struct page
*
2431 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2432 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2433 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2434 int classzone_idx
, int migratetype
, unsigned long *did_some_progress
)
2436 struct page
*page
= NULL
;
2437 bool drained
= false;
2439 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2441 if (unlikely(!(*did_some_progress
)))
2444 /* After successful reclaim, reconsider all zones for allocation */
2445 if (IS_ENABLED(CONFIG_NUMA
))
2446 zlc_clear_zones_full(zonelist
);
2449 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2450 zonelist
, high_zoneidx
,
2451 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2452 preferred_zone
, classzone_idx
,
2456 * If an allocation failed after direct reclaim, it could be because
2457 * pages are pinned on the per-cpu lists. Drain them and try again
2459 if (!page
&& !drained
) {
2460 drain_all_pages(NULL
);
2469 * This is called in the allocator slow-path if the allocation request is of
2470 * sufficient urgency to ignore watermarks and take other desperate measures
2472 static inline struct page
*
2473 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2474 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2475 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2476 int classzone_idx
, int migratetype
)
2481 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2482 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2483 preferred_zone
, classzone_idx
, migratetype
);
2485 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2486 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2487 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2492 static void wake_all_kswapds(unsigned int order
,
2493 struct zonelist
*zonelist
,
2494 enum zone_type high_zoneidx
,
2495 struct zone
*preferred_zone
,
2496 nodemask_t
*nodemask
)
2501 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2502 high_zoneidx
, nodemask
)
2503 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2507 gfp_to_alloc_flags(gfp_t gfp_mask
)
2509 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2510 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2512 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2513 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2516 * The caller may dip into page reserves a bit more if the caller
2517 * cannot run direct reclaim, or if the caller has realtime scheduling
2518 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2519 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2521 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2525 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2526 * if it can't schedule.
2528 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2529 alloc_flags
|= ALLOC_HARDER
;
2531 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2532 * comment for __cpuset_node_allowed_softwall().
2534 alloc_flags
&= ~ALLOC_CPUSET
;
2535 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2536 alloc_flags
|= ALLOC_HARDER
;
2538 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2539 if (gfp_mask
& __GFP_MEMALLOC
)
2540 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2541 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2542 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2543 else if (!in_interrupt() &&
2544 ((current
->flags
& PF_MEMALLOC
) ||
2545 unlikely(test_thread_flag(TIF_MEMDIE
))))
2546 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2549 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2550 alloc_flags
|= ALLOC_CMA
;
2555 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2557 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2560 static inline struct page
*
2561 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2562 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2563 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2564 int classzone_idx
, int migratetype
)
2566 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2567 struct page
*page
= NULL
;
2569 unsigned long pages_reclaimed
= 0;
2570 unsigned long did_some_progress
;
2571 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2572 bool deferred_compaction
= false;
2573 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2576 * In the slowpath, we sanity check order to avoid ever trying to
2577 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2578 * be using allocators in order of preference for an area that is
2581 if (order
>= MAX_ORDER
) {
2582 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2587 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2588 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2589 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2590 * using a larger set of nodes after it has established that the
2591 * allowed per node queues are empty and that nodes are
2594 if (IS_ENABLED(CONFIG_NUMA
) &&
2595 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2599 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2600 wake_all_kswapds(order
, zonelist
, high_zoneidx
,
2601 preferred_zone
, nodemask
);
2604 * OK, we're below the kswapd watermark and have kicked background
2605 * reclaim. Now things get more complex, so set up alloc_flags according
2606 * to how we want to proceed.
2608 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2611 * Find the true preferred zone if the allocation is unconstrained by
2614 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
) {
2615 struct zoneref
*preferred_zoneref
;
2616 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2617 NULL
, &preferred_zone
);
2618 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2622 /* This is the last chance, in general, before the goto nopage. */
2623 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2624 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2625 preferred_zone
, classzone_idx
, migratetype
);
2629 /* Allocate without watermarks if the context allows */
2630 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2632 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2633 * the allocation is high priority and these type of
2634 * allocations are system rather than user orientated
2636 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2638 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2639 zonelist
, high_zoneidx
, nodemask
,
2640 preferred_zone
, classzone_idx
, migratetype
);
2646 /* Atomic allocations - we can't balance anything */
2649 * All existing users of the deprecated __GFP_NOFAIL are
2650 * blockable, so warn of any new users that actually allow this
2651 * type of allocation to fail.
2653 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2657 /* Avoid recursion of direct reclaim */
2658 if (current
->flags
& PF_MEMALLOC
)
2661 /* Avoid allocations with no watermarks from looping endlessly */
2662 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2666 * Try direct compaction. The first pass is asynchronous. Subsequent
2667 * attempts after direct reclaim are synchronous
2669 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2670 high_zoneidx
, nodemask
, alloc_flags
,
2672 classzone_idx
, migratetype
,
2673 migration_mode
, &contended_compaction
,
2674 &deferred_compaction
);
2678 /* Checks for THP-specific high-order allocations */
2679 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2681 * If compaction is deferred for high-order allocations, it is
2682 * because sync compaction recently failed. If this is the case
2683 * and the caller requested a THP allocation, we do not want
2684 * to heavily disrupt the system, so we fail the allocation
2685 * instead of entering direct reclaim.
2687 if (deferred_compaction
)
2691 * In all zones where compaction was attempted (and not
2692 * deferred or skipped), lock contention has been detected.
2693 * For THP allocation we do not want to disrupt the others
2694 * so we fallback to base pages instead.
2696 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2700 * If compaction was aborted due to need_resched(), we do not
2701 * want to further increase allocation latency, unless it is
2702 * khugepaged trying to collapse.
2704 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2705 && !(current
->flags
& PF_KTHREAD
))
2710 * It can become very expensive to allocate transparent hugepages at
2711 * fault, so use asynchronous memory compaction for THP unless it is
2712 * khugepaged trying to collapse.
2714 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2715 (current
->flags
& PF_KTHREAD
))
2716 migration_mode
= MIGRATE_SYNC_LIGHT
;
2718 /* Try direct reclaim and then allocating */
2719 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2720 zonelist
, high_zoneidx
,
2722 alloc_flags
, preferred_zone
,
2723 classzone_idx
, migratetype
,
2724 &did_some_progress
);
2729 * If we failed to make any progress reclaiming, then we are
2730 * running out of options and have to consider going OOM
2732 if (!did_some_progress
) {
2733 if (oom_gfp_allowed(gfp_mask
)) {
2734 if (oom_killer_disabled
)
2736 /* Coredumps can quickly deplete all memory reserves */
2737 if ((current
->flags
& PF_DUMPCORE
) &&
2738 !(gfp_mask
& __GFP_NOFAIL
))
2740 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2741 zonelist
, high_zoneidx
,
2742 nodemask
, preferred_zone
,
2743 classzone_idx
, migratetype
);
2747 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2749 * The oom killer is not called for high-order
2750 * allocations that may fail, so if no progress
2751 * is being made, there are no other options and
2752 * retrying is unlikely to help.
2754 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2757 * The oom killer is not called for lowmem
2758 * allocations to prevent needlessly killing
2761 if (high_zoneidx
< ZONE_NORMAL
)
2769 /* Check if we should retry the allocation */
2770 pages_reclaimed
+= did_some_progress
;
2771 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2773 /* Wait for some write requests to complete then retry */
2774 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2778 * High-order allocations do not necessarily loop after
2779 * direct reclaim and reclaim/compaction depends on compaction
2780 * being called after reclaim so call directly if necessary
2782 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2783 high_zoneidx
, nodemask
, alloc_flags
,
2785 classzone_idx
, migratetype
,
2786 migration_mode
, &contended_compaction
,
2787 &deferred_compaction
);
2793 warn_alloc_failed(gfp_mask
, order
, NULL
);
2796 if (kmemcheck_enabled
)
2797 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2803 * This is the 'heart' of the zoned buddy allocator.
2806 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2807 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2809 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2810 struct zone
*preferred_zone
;
2811 struct zoneref
*preferred_zoneref
;
2812 struct page
*page
= NULL
;
2813 int migratetype
= gfpflags_to_migratetype(gfp_mask
);
2814 unsigned int cpuset_mems_cookie
;
2815 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2818 gfp_mask
&= gfp_allowed_mask
;
2820 lockdep_trace_alloc(gfp_mask
);
2822 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2824 if (should_fail_alloc_page(gfp_mask
, order
))
2828 * Check the zones suitable for the gfp_mask contain at least one
2829 * valid zone. It's possible to have an empty zonelist as a result
2830 * of GFP_THISNODE and a memoryless node
2832 if (unlikely(!zonelist
->_zonerefs
->zone
))
2835 if (IS_ENABLED(CONFIG_CMA
) && migratetype
== MIGRATE_MOVABLE
)
2836 alloc_flags
|= ALLOC_CMA
;
2839 cpuset_mems_cookie
= read_mems_allowed_begin();
2841 /* The preferred zone is used for statistics later */
2842 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2843 nodemask
? : &cpuset_current_mems_allowed
,
2845 if (!preferred_zone
)
2847 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2849 /* First allocation attempt */
2850 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2851 zonelist
, high_zoneidx
, alloc_flags
,
2852 preferred_zone
, classzone_idx
, migratetype
);
2853 if (unlikely(!page
)) {
2855 * Runtime PM, block IO and its error handling path
2856 * can deadlock because I/O on the device might not
2859 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2860 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2861 zonelist
, high_zoneidx
, nodemask
,
2862 preferred_zone
, classzone_idx
, migratetype
);
2865 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2869 * When updating a task's mems_allowed, it is possible to race with
2870 * parallel threads in such a way that an allocation can fail while
2871 * the mask is being updated. If a page allocation is about to fail,
2872 * check if the cpuset changed during allocation and if so, retry.
2874 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2879 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2882 * Common helper functions.
2884 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2889 * __get_free_pages() returns a 32-bit address, which cannot represent
2892 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2894 page
= alloc_pages(gfp_mask
, order
);
2897 return (unsigned long) page_address(page
);
2899 EXPORT_SYMBOL(__get_free_pages
);
2901 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2903 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2905 EXPORT_SYMBOL(get_zeroed_page
);
2907 void __free_pages(struct page
*page
, unsigned int order
)
2909 if (put_page_testzero(page
)) {
2911 free_hot_cold_page(page
, false);
2913 __free_pages_ok(page
, order
);
2917 EXPORT_SYMBOL(__free_pages
);
2919 void free_pages(unsigned long addr
, unsigned int order
)
2922 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2923 __free_pages(virt_to_page((void *)addr
), order
);
2927 EXPORT_SYMBOL(free_pages
);
2930 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2931 * of the current memory cgroup.
2933 * It should be used when the caller would like to use kmalloc, but since the
2934 * allocation is large, it has to fall back to the page allocator.
2936 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2939 struct mem_cgroup
*memcg
= NULL
;
2941 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2943 page
= alloc_pages(gfp_mask
, order
);
2944 memcg_kmem_commit_charge(page
, memcg
, order
);
2948 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2951 struct mem_cgroup
*memcg
= NULL
;
2953 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2955 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2956 memcg_kmem_commit_charge(page
, memcg
, order
);
2961 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2964 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2966 memcg_kmem_uncharge_pages(page
, order
);
2967 __free_pages(page
, order
);
2970 void free_kmem_pages(unsigned long addr
, unsigned int order
)
2973 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2974 __free_kmem_pages(virt_to_page((void *)addr
), order
);
2978 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2981 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2982 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2984 split_page(virt_to_page((void *)addr
), order
);
2985 while (used
< alloc_end
) {
2990 return (void *)addr
;
2994 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2995 * @size: the number of bytes to allocate
2996 * @gfp_mask: GFP flags for the allocation
2998 * This function is similar to alloc_pages(), except that it allocates the
2999 * minimum number of pages to satisfy the request. alloc_pages() can only
3000 * allocate memory in power-of-two pages.
3002 * This function is also limited by MAX_ORDER.
3004 * Memory allocated by this function must be released by free_pages_exact().
3006 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3008 unsigned int order
= get_order(size
);
3011 addr
= __get_free_pages(gfp_mask
, order
);
3012 return make_alloc_exact(addr
, order
, size
);
3014 EXPORT_SYMBOL(alloc_pages_exact
);
3017 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3019 * @nid: the preferred node ID where memory should be allocated
3020 * @size: the number of bytes to allocate
3021 * @gfp_mask: GFP flags for the allocation
3023 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3025 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3028 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3030 unsigned order
= get_order(size
);
3031 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3034 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3038 * free_pages_exact - release memory allocated via alloc_pages_exact()
3039 * @virt: the value returned by alloc_pages_exact.
3040 * @size: size of allocation, same value as passed to alloc_pages_exact().
3042 * Release the memory allocated by a previous call to alloc_pages_exact.
3044 void free_pages_exact(void *virt
, size_t size
)
3046 unsigned long addr
= (unsigned long)virt
;
3047 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3049 while (addr
< end
) {
3054 EXPORT_SYMBOL(free_pages_exact
);
3057 * nr_free_zone_pages - count number of pages beyond high watermark
3058 * @offset: The zone index of the highest zone
3060 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3061 * high watermark within all zones at or below a given zone index. For each
3062 * zone, the number of pages is calculated as:
3063 * managed_pages - high_pages
3065 static unsigned long nr_free_zone_pages(int offset
)
3070 /* Just pick one node, since fallback list is circular */
3071 unsigned long sum
= 0;
3073 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3075 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3076 unsigned long size
= zone
->managed_pages
;
3077 unsigned long high
= high_wmark_pages(zone
);
3086 * nr_free_buffer_pages - count number of pages beyond high watermark
3088 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3089 * watermark within ZONE_DMA and ZONE_NORMAL.
3091 unsigned long nr_free_buffer_pages(void)
3093 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3095 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3098 * nr_free_pagecache_pages - count number of pages beyond high watermark
3100 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3101 * high watermark within all zones.
3103 unsigned long nr_free_pagecache_pages(void)
3105 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3108 static inline void show_node(struct zone
*zone
)
3110 if (IS_ENABLED(CONFIG_NUMA
))
3111 printk("Node %d ", zone_to_nid(zone
));
3114 void si_meminfo(struct sysinfo
*val
)
3116 val
->totalram
= totalram_pages
;
3117 val
->sharedram
= global_page_state(NR_SHMEM
);
3118 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3119 val
->bufferram
= nr_blockdev_pages();
3120 val
->totalhigh
= totalhigh_pages
;
3121 val
->freehigh
= nr_free_highpages();
3122 val
->mem_unit
= PAGE_SIZE
;
3125 EXPORT_SYMBOL(si_meminfo
);
3128 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3130 int zone_type
; /* needs to be signed */
3131 unsigned long managed_pages
= 0;
3132 pg_data_t
*pgdat
= NODE_DATA(nid
);
3134 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3135 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3136 val
->totalram
= managed_pages
;
3137 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3138 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3139 #ifdef CONFIG_HIGHMEM
3140 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3141 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3147 val
->mem_unit
= PAGE_SIZE
;
3152 * Determine whether the node should be displayed or not, depending on whether
3153 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3155 bool skip_free_areas_node(unsigned int flags
, int nid
)
3158 unsigned int cpuset_mems_cookie
;
3160 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3164 cpuset_mems_cookie
= read_mems_allowed_begin();
3165 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3166 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3171 #define K(x) ((x) << (PAGE_SHIFT-10))
3173 static void show_migration_types(unsigned char type
)
3175 static const char types
[MIGRATE_TYPES
] = {
3176 [MIGRATE_UNMOVABLE
] = 'U',
3177 [MIGRATE_RECLAIMABLE
] = 'E',
3178 [MIGRATE_MOVABLE
] = 'M',
3179 [MIGRATE_RESERVE
] = 'R',
3181 [MIGRATE_CMA
] = 'C',
3183 #ifdef CONFIG_MEMORY_ISOLATION
3184 [MIGRATE_ISOLATE
] = 'I',
3187 char tmp
[MIGRATE_TYPES
+ 1];
3191 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3192 if (type
& (1 << i
))
3197 printk("(%s) ", tmp
);
3201 * Show free area list (used inside shift_scroll-lock stuff)
3202 * We also calculate the percentage fragmentation. We do this by counting the
3203 * memory on each free list with the exception of the first item on the list.
3204 * Suppresses nodes that are not allowed by current's cpuset if
3205 * SHOW_MEM_FILTER_NODES is passed.
3207 void show_free_areas(unsigned int filter
)
3212 for_each_populated_zone(zone
) {
3213 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3216 printk("%s per-cpu:\n", zone
->name
);
3218 for_each_online_cpu(cpu
) {
3219 struct per_cpu_pageset
*pageset
;
3221 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3223 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3224 cpu
, pageset
->pcp
.high
,
3225 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3229 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3230 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3232 " dirty:%lu writeback:%lu unstable:%lu\n"
3233 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3234 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3236 global_page_state(NR_ACTIVE_ANON
),
3237 global_page_state(NR_INACTIVE_ANON
),
3238 global_page_state(NR_ISOLATED_ANON
),
3239 global_page_state(NR_ACTIVE_FILE
),
3240 global_page_state(NR_INACTIVE_FILE
),
3241 global_page_state(NR_ISOLATED_FILE
),
3242 global_page_state(NR_UNEVICTABLE
),
3243 global_page_state(NR_FILE_DIRTY
),
3244 global_page_state(NR_WRITEBACK
),
3245 global_page_state(NR_UNSTABLE_NFS
),
3246 global_page_state(NR_FREE_PAGES
),
3247 global_page_state(NR_SLAB_RECLAIMABLE
),
3248 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3249 global_page_state(NR_FILE_MAPPED
),
3250 global_page_state(NR_SHMEM
),
3251 global_page_state(NR_PAGETABLE
),
3252 global_page_state(NR_BOUNCE
),
3253 global_page_state(NR_FREE_CMA_PAGES
));
3255 for_each_populated_zone(zone
) {
3258 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3266 " active_anon:%lukB"
3267 " inactive_anon:%lukB"
3268 " active_file:%lukB"
3269 " inactive_file:%lukB"
3270 " unevictable:%lukB"
3271 " isolated(anon):%lukB"
3272 " isolated(file):%lukB"
3280 " slab_reclaimable:%lukB"
3281 " slab_unreclaimable:%lukB"
3282 " kernel_stack:%lukB"
3287 " writeback_tmp:%lukB"
3288 " pages_scanned:%lu"
3289 " all_unreclaimable? %s"
3292 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3293 K(min_wmark_pages(zone
)),
3294 K(low_wmark_pages(zone
)),
3295 K(high_wmark_pages(zone
)),
3296 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3297 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3298 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3299 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3300 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3301 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3302 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3303 K(zone
->present_pages
),
3304 K(zone
->managed_pages
),
3305 K(zone_page_state(zone
, NR_MLOCK
)),
3306 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3307 K(zone_page_state(zone
, NR_WRITEBACK
)),
3308 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3309 K(zone_page_state(zone
, NR_SHMEM
)),
3310 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3311 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3312 zone_page_state(zone
, NR_KERNEL_STACK
) *
3314 K(zone_page_state(zone
, NR_PAGETABLE
)),
3315 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3316 K(zone_page_state(zone
, NR_BOUNCE
)),
3317 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3318 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3319 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3320 (!zone_reclaimable(zone
) ? "yes" : "no")
3322 printk("lowmem_reserve[]:");
3323 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3324 printk(" %ld", zone
->lowmem_reserve
[i
]);
3328 for_each_populated_zone(zone
) {
3329 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3330 unsigned char types
[MAX_ORDER
];
3332 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3335 printk("%s: ", zone
->name
);
3337 spin_lock_irqsave(&zone
->lock
, flags
);
3338 for (order
= 0; order
< MAX_ORDER
; order
++) {
3339 struct free_area
*area
= &zone
->free_area
[order
];
3342 nr
[order
] = area
->nr_free
;
3343 total
+= nr
[order
] << order
;
3346 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3347 if (!list_empty(&area
->free_list
[type
]))
3348 types
[order
] |= 1 << type
;
3351 spin_unlock_irqrestore(&zone
->lock
, flags
);
3352 for (order
= 0; order
< MAX_ORDER
; order
++) {
3353 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3355 show_migration_types(types
[order
]);
3357 printk("= %lukB\n", K(total
));
3360 hugetlb_show_meminfo();
3362 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3364 show_swap_cache_info();
3367 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3369 zoneref
->zone
= zone
;
3370 zoneref
->zone_idx
= zone_idx(zone
);
3374 * Builds allocation fallback zone lists.
3376 * Add all populated zones of a node to the zonelist.
3378 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3382 enum zone_type zone_type
= MAX_NR_ZONES
;
3386 zone
= pgdat
->node_zones
+ zone_type
;
3387 if (populated_zone(zone
)) {
3388 zoneref_set_zone(zone
,
3389 &zonelist
->_zonerefs
[nr_zones
++]);
3390 check_highest_zone(zone_type
);
3392 } while (zone_type
);
3400 * 0 = automatic detection of better ordering.
3401 * 1 = order by ([node] distance, -zonetype)
3402 * 2 = order by (-zonetype, [node] distance)
3404 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3405 * the same zonelist. So only NUMA can configure this param.
3407 #define ZONELIST_ORDER_DEFAULT 0
3408 #define ZONELIST_ORDER_NODE 1
3409 #define ZONELIST_ORDER_ZONE 2
3411 /* zonelist order in the kernel.
3412 * set_zonelist_order() will set this to NODE or ZONE.
3414 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3415 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3419 /* The value user specified ....changed by config */
3420 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3421 /* string for sysctl */
3422 #define NUMA_ZONELIST_ORDER_LEN 16
3423 char numa_zonelist_order
[16] = "default";
3426 * interface for configure zonelist ordering.
3427 * command line option "numa_zonelist_order"
3428 * = "[dD]efault - default, automatic configuration.
3429 * = "[nN]ode - order by node locality, then by zone within node
3430 * = "[zZ]one - order by zone, then by locality within zone
3433 static int __parse_numa_zonelist_order(char *s
)
3435 if (*s
== 'd' || *s
== 'D') {
3436 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3437 } else if (*s
== 'n' || *s
== 'N') {
3438 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3439 } else if (*s
== 'z' || *s
== 'Z') {
3440 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3443 "Ignoring invalid numa_zonelist_order value: "
3450 static __init
int setup_numa_zonelist_order(char *s
)
3457 ret
= __parse_numa_zonelist_order(s
);
3459 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3463 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3466 * sysctl handler for numa_zonelist_order
3468 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3469 void __user
*buffer
, size_t *length
,
3472 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3474 static DEFINE_MUTEX(zl_order_mutex
);
3476 mutex_lock(&zl_order_mutex
);
3478 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3482 strcpy(saved_string
, (char *)table
->data
);
3484 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3488 int oldval
= user_zonelist_order
;
3490 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3493 * bogus value. restore saved string
3495 strncpy((char *)table
->data
, saved_string
,
3496 NUMA_ZONELIST_ORDER_LEN
);
3497 user_zonelist_order
= oldval
;
3498 } else if (oldval
!= user_zonelist_order
) {
3499 mutex_lock(&zonelists_mutex
);
3500 build_all_zonelists(NULL
, NULL
);
3501 mutex_unlock(&zonelists_mutex
);
3505 mutex_unlock(&zl_order_mutex
);
3510 #define MAX_NODE_LOAD (nr_online_nodes)
3511 static int node_load
[MAX_NUMNODES
];
3514 * find_next_best_node - find the next node that should appear in a given node's fallback list
3515 * @node: node whose fallback list we're appending
3516 * @used_node_mask: nodemask_t of already used nodes
3518 * We use a number of factors to determine which is the next node that should
3519 * appear on a given node's fallback list. The node should not have appeared
3520 * already in @node's fallback list, and it should be the next closest node
3521 * according to the distance array (which contains arbitrary distance values
3522 * from each node to each node in the system), and should also prefer nodes
3523 * with no CPUs, since presumably they'll have very little allocation pressure
3524 * on them otherwise.
3525 * It returns -1 if no node is found.
3527 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3530 int min_val
= INT_MAX
;
3531 int best_node
= NUMA_NO_NODE
;
3532 const struct cpumask
*tmp
= cpumask_of_node(0);
3534 /* Use the local node if we haven't already */
3535 if (!node_isset(node
, *used_node_mask
)) {
3536 node_set(node
, *used_node_mask
);
3540 for_each_node_state(n
, N_MEMORY
) {
3542 /* Don't want a node to appear more than once */
3543 if (node_isset(n
, *used_node_mask
))
3546 /* Use the distance array to find the distance */
3547 val
= node_distance(node
, n
);
3549 /* Penalize nodes under us ("prefer the next node") */
3552 /* Give preference to headless and unused nodes */
3553 tmp
= cpumask_of_node(n
);
3554 if (!cpumask_empty(tmp
))
3555 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3557 /* Slight preference for less loaded node */
3558 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3559 val
+= node_load
[n
];
3561 if (val
< min_val
) {
3568 node_set(best_node
, *used_node_mask
);
3575 * Build zonelists ordered by node and zones within node.
3576 * This results in maximum locality--normal zone overflows into local
3577 * DMA zone, if any--but risks exhausting DMA zone.
3579 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3582 struct zonelist
*zonelist
;
3584 zonelist
= &pgdat
->node_zonelists
[0];
3585 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3587 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3588 zonelist
->_zonerefs
[j
].zone
= NULL
;
3589 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3593 * Build gfp_thisnode zonelists
3595 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3598 struct zonelist
*zonelist
;
3600 zonelist
= &pgdat
->node_zonelists
[1];
3601 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3602 zonelist
->_zonerefs
[j
].zone
= NULL
;
3603 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3607 * Build zonelists ordered by zone and nodes within zones.
3608 * This results in conserving DMA zone[s] until all Normal memory is
3609 * exhausted, but results in overflowing to remote node while memory
3610 * may still exist in local DMA zone.
3612 static int node_order
[MAX_NUMNODES
];
3614 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3617 int zone_type
; /* needs to be signed */
3619 struct zonelist
*zonelist
;
3621 zonelist
= &pgdat
->node_zonelists
[0];
3623 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3624 for (j
= 0; j
< nr_nodes
; j
++) {
3625 node
= node_order
[j
];
3626 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3627 if (populated_zone(z
)) {
3629 &zonelist
->_zonerefs
[pos
++]);
3630 check_highest_zone(zone_type
);
3634 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3635 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3638 #if defined(CONFIG_64BIT)
3640 * Devices that require DMA32/DMA are relatively rare and do not justify a
3641 * penalty to every machine in case the specialised case applies. Default
3642 * to Node-ordering on 64-bit NUMA machines
3644 static int default_zonelist_order(void)
3646 return ZONELIST_ORDER_NODE
;
3650 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3651 * by the kernel. If processes running on node 0 deplete the low memory zone
3652 * then reclaim will occur more frequency increasing stalls and potentially
3653 * be easier to OOM if a large percentage of the zone is under writeback or
3654 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3655 * Hence, default to zone ordering on 32-bit.
3657 static int default_zonelist_order(void)
3659 return ZONELIST_ORDER_ZONE
;
3661 #endif /* CONFIG_64BIT */
3663 static void set_zonelist_order(void)
3665 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3666 current_zonelist_order
= default_zonelist_order();
3668 current_zonelist_order
= user_zonelist_order
;
3671 static void build_zonelists(pg_data_t
*pgdat
)
3675 nodemask_t used_mask
;
3676 int local_node
, prev_node
;
3677 struct zonelist
*zonelist
;
3678 int order
= current_zonelist_order
;
3680 /* initialize zonelists */
3681 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3682 zonelist
= pgdat
->node_zonelists
+ i
;
3683 zonelist
->_zonerefs
[0].zone
= NULL
;
3684 zonelist
->_zonerefs
[0].zone_idx
= 0;
3687 /* NUMA-aware ordering of nodes */
3688 local_node
= pgdat
->node_id
;
3689 load
= nr_online_nodes
;
3690 prev_node
= local_node
;
3691 nodes_clear(used_mask
);
3693 memset(node_order
, 0, sizeof(node_order
));
3696 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3698 * We don't want to pressure a particular node.
3699 * So adding penalty to the first node in same
3700 * distance group to make it round-robin.
3702 if (node_distance(local_node
, node
) !=
3703 node_distance(local_node
, prev_node
))
3704 node_load
[node
] = load
;
3708 if (order
== ZONELIST_ORDER_NODE
)
3709 build_zonelists_in_node_order(pgdat
, node
);
3711 node_order
[j
++] = node
; /* remember order */
3714 if (order
== ZONELIST_ORDER_ZONE
) {
3715 /* calculate node order -- i.e., DMA last! */
3716 build_zonelists_in_zone_order(pgdat
, j
);
3719 build_thisnode_zonelists(pgdat
);
3722 /* Construct the zonelist performance cache - see further mmzone.h */
3723 static void build_zonelist_cache(pg_data_t
*pgdat
)
3725 struct zonelist
*zonelist
;
3726 struct zonelist_cache
*zlc
;
3729 zonelist
= &pgdat
->node_zonelists
[0];
3730 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3731 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3732 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3733 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3736 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3738 * Return node id of node used for "local" allocations.
3739 * I.e., first node id of first zone in arg node's generic zonelist.
3740 * Used for initializing percpu 'numa_mem', which is used primarily
3741 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3743 int local_memory_node(int node
)
3747 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3748 gfp_zone(GFP_KERNEL
),
3755 #else /* CONFIG_NUMA */
3757 static void set_zonelist_order(void)
3759 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3762 static void build_zonelists(pg_data_t
*pgdat
)
3764 int node
, local_node
;
3766 struct zonelist
*zonelist
;
3768 local_node
= pgdat
->node_id
;
3770 zonelist
= &pgdat
->node_zonelists
[0];
3771 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3774 * Now we build the zonelist so that it contains the zones
3775 * of all the other nodes.
3776 * We don't want to pressure a particular node, so when
3777 * building the zones for node N, we make sure that the
3778 * zones coming right after the local ones are those from
3779 * node N+1 (modulo N)
3781 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3782 if (!node_online(node
))
3784 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3786 for (node
= 0; node
< local_node
; node
++) {
3787 if (!node_online(node
))
3789 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3792 zonelist
->_zonerefs
[j
].zone
= NULL
;
3793 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3796 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3797 static void build_zonelist_cache(pg_data_t
*pgdat
)
3799 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3802 #endif /* CONFIG_NUMA */
3805 * Boot pageset table. One per cpu which is going to be used for all
3806 * zones and all nodes. The parameters will be set in such a way
3807 * that an item put on a list will immediately be handed over to
3808 * the buddy list. This is safe since pageset manipulation is done
3809 * with interrupts disabled.
3811 * The boot_pagesets must be kept even after bootup is complete for
3812 * unused processors and/or zones. They do play a role for bootstrapping
3813 * hotplugged processors.
3815 * zoneinfo_show() and maybe other functions do
3816 * not check if the processor is online before following the pageset pointer.
3817 * Other parts of the kernel may not check if the zone is available.
3819 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3820 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3821 static void setup_zone_pageset(struct zone
*zone
);
3824 * Global mutex to protect against size modification of zonelists
3825 * as well as to serialize pageset setup for the new populated zone.
3827 DEFINE_MUTEX(zonelists_mutex
);
3829 /* return values int ....just for stop_machine() */
3830 static int __build_all_zonelists(void *data
)
3834 pg_data_t
*self
= data
;
3837 memset(node_load
, 0, sizeof(node_load
));
3840 if (self
&& !node_online(self
->node_id
)) {
3841 build_zonelists(self
);
3842 build_zonelist_cache(self
);
3845 for_each_online_node(nid
) {
3846 pg_data_t
*pgdat
= NODE_DATA(nid
);
3848 build_zonelists(pgdat
);
3849 build_zonelist_cache(pgdat
);
3853 * Initialize the boot_pagesets that are going to be used
3854 * for bootstrapping processors. The real pagesets for
3855 * each zone will be allocated later when the per cpu
3856 * allocator is available.
3858 * boot_pagesets are used also for bootstrapping offline
3859 * cpus if the system is already booted because the pagesets
3860 * are needed to initialize allocators on a specific cpu too.
3861 * F.e. the percpu allocator needs the page allocator which
3862 * needs the percpu allocator in order to allocate its pagesets
3863 * (a chicken-egg dilemma).
3865 for_each_possible_cpu(cpu
) {
3866 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3868 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3870 * We now know the "local memory node" for each node--
3871 * i.e., the node of the first zone in the generic zonelist.
3872 * Set up numa_mem percpu variable for on-line cpus. During
3873 * boot, only the boot cpu should be on-line; we'll init the
3874 * secondary cpus' numa_mem as they come on-line. During
3875 * node/memory hotplug, we'll fixup all on-line cpus.
3877 if (cpu_online(cpu
))
3878 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3886 * Called with zonelists_mutex held always
3887 * unless system_state == SYSTEM_BOOTING.
3889 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3891 set_zonelist_order();
3893 if (system_state
== SYSTEM_BOOTING
) {
3894 __build_all_zonelists(NULL
);
3895 mminit_verify_zonelist();
3896 cpuset_init_current_mems_allowed();
3898 #ifdef CONFIG_MEMORY_HOTPLUG
3900 setup_zone_pageset(zone
);
3902 /* we have to stop all cpus to guarantee there is no user
3904 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3905 /* cpuset refresh routine should be here */
3907 vm_total_pages
= nr_free_pagecache_pages();
3909 * Disable grouping by mobility if the number of pages in the
3910 * system is too low to allow the mechanism to work. It would be
3911 * more accurate, but expensive to check per-zone. This check is
3912 * made on memory-hotadd so a system can start with mobility
3913 * disabled and enable it later
3915 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3916 page_group_by_mobility_disabled
= 1;
3918 page_group_by_mobility_disabled
= 0;
3920 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3921 "Total pages: %ld\n",
3923 zonelist_order_name
[current_zonelist_order
],
3924 page_group_by_mobility_disabled
? "off" : "on",
3927 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
3932 * Helper functions to size the waitqueue hash table.
3933 * Essentially these want to choose hash table sizes sufficiently
3934 * large so that collisions trying to wait on pages are rare.
3935 * But in fact, the number of active page waitqueues on typical
3936 * systems is ridiculously low, less than 200. So this is even
3937 * conservative, even though it seems large.
3939 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3940 * waitqueues, i.e. the size of the waitq table given the number of pages.
3942 #define PAGES_PER_WAITQUEUE 256
3944 #ifndef CONFIG_MEMORY_HOTPLUG
3945 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3947 unsigned long size
= 1;
3949 pages
/= PAGES_PER_WAITQUEUE
;
3951 while (size
< pages
)
3955 * Once we have dozens or even hundreds of threads sleeping
3956 * on IO we've got bigger problems than wait queue collision.
3957 * Limit the size of the wait table to a reasonable size.
3959 size
= min(size
, 4096UL);
3961 return max(size
, 4UL);
3965 * A zone's size might be changed by hot-add, so it is not possible to determine
3966 * a suitable size for its wait_table. So we use the maximum size now.
3968 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3970 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3971 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3972 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3974 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3975 * or more by the traditional way. (See above). It equals:
3977 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3978 * ia64(16K page size) : = ( 8G + 4M)byte.
3979 * powerpc (64K page size) : = (32G +16M)byte.
3981 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3988 * This is an integer logarithm so that shifts can be used later
3989 * to extract the more random high bits from the multiplicative
3990 * hash function before the remainder is taken.
3992 static inline unsigned long wait_table_bits(unsigned long size
)
3998 * Check if a pageblock contains reserved pages
4000 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
4004 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4005 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
4012 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4013 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4014 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4015 * higher will lead to a bigger reserve which will get freed as contiguous
4016 * blocks as reclaim kicks in
4018 static void setup_zone_migrate_reserve(struct zone
*zone
)
4020 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4022 unsigned long block_migratetype
;
4027 * Get the start pfn, end pfn and the number of blocks to reserve
4028 * We have to be careful to be aligned to pageblock_nr_pages to
4029 * make sure that we always check pfn_valid for the first page in
4032 start_pfn
= zone
->zone_start_pfn
;
4033 end_pfn
= zone_end_pfn(zone
);
4034 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4035 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4039 * Reserve blocks are generally in place to help high-order atomic
4040 * allocations that are short-lived. A min_free_kbytes value that
4041 * would result in more than 2 reserve blocks for atomic allocations
4042 * is assumed to be in place to help anti-fragmentation for the
4043 * future allocation of hugepages at runtime.
4045 reserve
= min(2, reserve
);
4046 old_reserve
= zone
->nr_migrate_reserve_block
;
4048 /* When memory hot-add, we almost always need to do nothing */
4049 if (reserve
== old_reserve
)
4051 zone
->nr_migrate_reserve_block
= reserve
;
4053 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4054 if (!pfn_valid(pfn
))
4056 page
= pfn_to_page(pfn
);
4058 /* Watch out for overlapping nodes */
4059 if (page_to_nid(page
) != zone_to_nid(zone
))
4062 block_migratetype
= get_pageblock_migratetype(page
);
4064 /* Only test what is necessary when the reserves are not met */
4067 * Blocks with reserved pages will never free, skip
4070 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4071 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4074 /* If this block is reserved, account for it */
4075 if (block_migratetype
== MIGRATE_RESERVE
) {
4080 /* Suitable for reserving if this block is movable */
4081 if (block_migratetype
== MIGRATE_MOVABLE
) {
4082 set_pageblock_migratetype(page
,
4084 move_freepages_block(zone
, page
,
4089 } else if (!old_reserve
) {
4091 * At boot time we don't need to scan the whole zone
4092 * for turning off MIGRATE_RESERVE.
4098 * If the reserve is met and this is a previous reserved block,
4101 if (block_migratetype
== MIGRATE_RESERVE
) {
4102 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4103 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4109 * Initially all pages are reserved - free ones are freed
4110 * up by free_all_bootmem() once the early boot process is
4111 * done. Non-atomic initialization, single-pass.
4113 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4114 unsigned long start_pfn
, enum memmap_context context
)
4117 unsigned long end_pfn
= start_pfn
+ size
;
4121 if (highest_memmap_pfn
< end_pfn
- 1)
4122 highest_memmap_pfn
= end_pfn
- 1;
4124 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4125 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4127 * There can be holes in boot-time mem_map[]s
4128 * handed to this function. They do not
4129 * exist on hotplugged memory.
4131 if (context
== MEMMAP_EARLY
) {
4132 if (!early_pfn_valid(pfn
))
4134 if (!early_pfn_in_nid(pfn
, nid
))
4137 page
= pfn_to_page(pfn
);
4138 set_page_links(page
, zone
, nid
, pfn
);
4139 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4140 init_page_count(page
);
4141 page_mapcount_reset(page
);
4142 page_cpupid_reset_last(page
);
4143 SetPageReserved(page
);
4145 * Mark the block movable so that blocks are reserved for
4146 * movable at startup. This will force kernel allocations
4147 * to reserve their blocks rather than leaking throughout
4148 * the address space during boot when many long-lived
4149 * kernel allocations are made. Later some blocks near
4150 * the start are marked MIGRATE_RESERVE by
4151 * setup_zone_migrate_reserve()
4153 * bitmap is created for zone's valid pfn range. but memmap
4154 * can be created for invalid pages (for alignment)
4155 * check here not to call set_pageblock_migratetype() against
4158 if ((z
->zone_start_pfn
<= pfn
)
4159 && (pfn
< zone_end_pfn(z
))
4160 && !(pfn
& (pageblock_nr_pages
- 1)))
4161 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4163 INIT_LIST_HEAD(&page
->lru
);
4164 #ifdef WANT_PAGE_VIRTUAL
4165 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4166 if (!is_highmem_idx(zone
))
4167 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4172 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4174 unsigned int order
, t
;
4175 for_each_migratetype_order(order
, t
) {
4176 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4177 zone
->free_area
[order
].nr_free
= 0;
4181 #ifndef __HAVE_ARCH_MEMMAP_INIT
4182 #define memmap_init(size, nid, zone, start_pfn) \
4183 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4186 static int zone_batchsize(struct zone
*zone
)
4192 * The per-cpu-pages pools are set to around 1000th of the
4193 * size of the zone. But no more than 1/2 of a meg.
4195 * OK, so we don't know how big the cache is. So guess.
4197 batch
= zone
->managed_pages
/ 1024;
4198 if (batch
* PAGE_SIZE
> 512 * 1024)
4199 batch
= (512 * 1024) / PAGE_SIZE
;
4200 batch
/= 4; /* We effectively *= 4 below */
4205 * Clamp the batch to a 2^n - 1 value. Having a power
4206 * of 2 value was found to be more likely to have
4207 * suboptimal cache aliasing properties in some cases.
4209 * For example if 2 tasks are alternately allocating
4210 * batches of pages, one task can end up with a lot
4211 * of pages of one half of the possible page colors
4212 * and the other with pages of the other colors.
4214 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4219 /* The deferral and batching of frees should be suppressed under NOMMU
4222 * The problem is that NOMMU needs to be able to allocate large chunks
4223 * of contiguous memory as there's no hardware page translation to
4224 * assemble apparent contiguous memory from discontiguous pages.
4226 * Queueing large contiguous runs of pages for batching, however,
4227 * causes the pages to actually be freed in smaller chunks. As there
4228 * can be a significant delay between the individual batches being
4229 * recycled, this leads to the once large chunks of space being
4230 * fragmented and becoming unavailable for high-order allocations.
4237 * pcp->high and pcp->batch values are related and dependent on one another:
4238 * ->batch must never be higher then ->high.
4239 * The following function updates them in a safe manner without read side
4242 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4243 * those fields changing asynchronously (acording the the above rule).
4245 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4246 * outside of boot time (or some other assurance that no concurrent updaters
4249 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4250 unsigned long batch
)
4252 /* start with a fail safe value for batch */
4256 /* Update high, then batch, in order */
4263 /* a companion to pageset_set_high() */
4264 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4266 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4269 static void pageset_init(struct per_cpu_pageset
*p
)
4271 struct per_cpu_pages
*pcp
;
4274 memset(p
, 0, sizeof(*p
));
4278 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4279 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4282 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4285 pageset_set_batch(p
, batch
);
4289 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4290 * to the value high for the pageset p.
4292 static void pageset_set_high(struct per_cpu_pageset
*p
,
4295 unsigned long batch
= max(1UL, high
/ 4);
4296 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4297 batch
= PAGE_SHIFT
* 8;
4299 pageset_update(&p
->pcp
, high
, batch
);
4302 static void pageset_set_high_and_batch(struct zone
*zone
,
4303 struct per_cpu_pageset
*pcp
)
4305 if (percpu_pagelist_fraction
)
4306 pageset_set_high(pcp
,
4307 (zone
->managed_pages
/
4308 percpu_pagelist_fraction
));
4310 pageset_set_batch(pcp
, zone_batchsize(zone
));
4313 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4315 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4318 pageset_set_high_and_batch(zone
, pcp
);
4321 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4324 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4325 for_each_possible_cpu(cpu
)
4326 zone_pageset_init(zone
, cpu
);
4330 * Allocate per cpu pagesets and initialize them.
4331 * Before this call only boot pagesets were available.
4333 void __init
setup_per_cpu_pageset(void)
4337 for_each_populated_zone(zone
)
4338 setup_zone_pageset(zone
);
4341 static noinline __init_refok
4342 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4348 * The per-page waitqueue mechanism uses hashed waitqueues
4351 zone
->wait_table_hash_nr_entries
=
4352 wait_table_hash_nr_entries(zone_size_pages
);
4353 zone
->wait_table_bits
=
4354 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4355 alloc_size
= zone
->wait_table_hash_nr_entries
4356 * sizeof(wait_queue_head_t
);
4358 if (!slab_is_available()) {
4359 zone
->wait_table
= (wait_queue_head_t
*)
4360 memblock_virt_alloc_node_nopanic(
4361 alloc_size
, zone
->zone_pgdat
->node_id
);
4364 * This case means that a zone whose size was 0 gets new memory
4365 * via memory hot-add.
4366 * But it may be the case that a new node was hot-added. In
4367 * this case vmalloc() will not be able to use this new node's
4368 * memory - this wait_table must be initialized to use this new
4369 * node itself as well.
4370 * To use this new node's memory, further consideration will be
4373 zone
->wait_table
= vmalloc(alloc_size
);
4375 if (!zone
->wait_table
)
4378 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4379 init_waitqueue_head(zone
->wait_table
+ i
);
4384 static __meminit
void zone_pcp_init(struct zone
*zone
)
4387 * per cpu subsystem is not up at this point. The following code
4388 * relies on the ability of the linker to provide the
4389 * offset of a (static) per cpu variable into the per cpu area.
4391 zone
->pageset
= &boot_pageset
;
4393 if (populated_zone(zone
))
4394 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4395 zone
->name
, zone
->present_pages
,
4396 zone_batchsize(zone
));
4399 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4400 unsigned long zone_start_pfn
,
4402 enum memmap_context context
)
4404 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4406 ret
= zone_wait_table_init(zone
, size
);
4409 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4411 zone
->zone_start_pfn
= zone_start_pfn
;
4413 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4414 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4416 (unsigned long)zone_idx(zone
),
4417 zone_start_pfn
, (zone_start_pfn
+ size
));
4419 zone_init_free_lists(zone
);
4424 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4425 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4427 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4429 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4431 unsigned long start_pfn
, end_pfn
;
4434 * NOTE: The following SMP-unsafe globals are only used early in boot
4435 * when the kernel is running single-threaded.
4437 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4438 static int __meminitdata last_nid
;
4440 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4443 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4445 last_start_pfn
= start_pfn
;
4446 last_end_pfn
= end_pfn
;
4452 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4454 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4458 nid
= __early_pfn_to_nid(pfn
);
4461 /* just returns 0 */
4465 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4466 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4470 nid
= __early_pfn_to_nid(pfn
);
4471 if (nid
>= 0 && nid
!= node
)
4478 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4479 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4480 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4482 * If an architecture guarantees that all ranges registered contain no holes
4483 * and may be freed, this this function may be used instead of calling
4484 * memblock_free_early_nid() manually.
4486 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4488 unsigned long start_pfn
, end_pfn
;
4491 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4492 start_pfn
= min(start_pfn
, max_low_pfn
);
4493 end_pfn
= min(end_pfn
, max_low_pfn
);
4495 if (start_pfn
< end_pfn
)
4496 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4497 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4503 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4504 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4506 * If an architecture guarantees that all ranges registered contain no holes and may
4507 * be freed, this function may be used instead of calling memory_present() manually.
4509 void __init
sparse_memory_present_with_active_regions(int nid
)
4511 unsigned long start_pfn
, end_pfn
;
4514 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4515 memory_present(this_nid
, start_pfn
, end_pfn
);
4519 * get_pfn_range_for_nid - Return the start and end page frames for a node
4520 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4521 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4522 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4524 * It returns the start and end page frame of a node based on information
4525 * provided by memblock_set_node(). If called for a node
4526 * with no available memory, a warning is printed and the start and end
4529 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4530 unsigned long *start_pfn
, unsigned long *end_pfn
)
4532 unsigned long this_start_pfn
, this_end_pfn
;
4538 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4539 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4540 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4543 if (*start_pfn
== -1UL)
4548 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4549 * assumption is made that zones within a node are ordered in monotonic
4550 * increasing memory addresses so that the "highest" populated zone is used
4552 static void __init
find_usable_zone_for_movable(void)
4555 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4556 if (zone_index
== ZONE_MOVABLE
)
4559 if (arch_zone_highest_possible_pfn
[zone_index
] >
4560 arch_zone_lowest_possible_pfn
[zone_index
])
4564 VM_BUG_ON(zone_index
== -1);
4565 movable_zone
= zone_index
;
4569 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4570 * because it is sized independent of architecture. Unlike the other zones,
4571 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4572 * in each node depending on the size of each node and how evenly kernelcore
4573 * is distributed. This helper function adjusts the zone ranges
4574 * provided by the architecture for a given node by using the end of the
4575 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4576 * zones within a node are in order of monotonic increases memory addresses
4578 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4579 unsigned long zone_type
,
4580 unsigned long node_start_pfn
,
4581 unsigned long node_end_pfn
,
4582 unsigned long *zone_start_pfn
,
4583 unsigned long *zone_end_pfn
)
4585 /* Only adjust if ZONE_MOVABLE is on this node */
4586 if (zone_movable_pfn
[nid
]) {
4587 /* Size ZONE_MOVABLE */
4588 if (zone_type
== ZONE_MOVABLE
) {
4589 *zone_start_pfn
= zone_movable_pfn
[nid
];
4590 *zone_end_pfn
= min(node_end_pfn
,
4591 arch_zone_highest_possible_pfn
[movable_zone
]);
4593 /* Adjust for ZONE_MOVABLE starting within this range */
4594 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4595 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4596 *zone_end_pfn
= zone_movable_pfn
[nid
];
4598 /* Check if this whole range is within ZONE_MOVABLE */
4599 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4600 *zone_start_pfn
= *zone_end_pfn
;
4605 * Return the number of pages a zone spans in a node, including holes
4606 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4608 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4609 unsigned long zone_type
,
4610 unsigned long node_start_pfn
,
4611 unsigned long node_end_pfn
,
4612 unsigned long *ignored
)
4614 unsigned long zone_start_pfn
, zone_end_pfn
;
4616 /* Get the start and end of the zone */
4617 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4618 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4619 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4620 node_start_pfn
, node_end_pfn
,
4621 &zone_start_pfn
, &zone_end_pfn
);
4623 /* Check that this node has pages within the zone's required range */
4624 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4627 /* Move the zone boundaries inside the node if necessary */
4628 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4629 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4631 /* Return the spanned pages */
4632 return zone_end_pfn
- zone_start_pfn
;
4636 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4637 * then all holes in the requested range will be accounted for.
4639 unsigned long __meminit
__absent_pages_in_range(int nid
,
4640 unsigned long range_start_pfn
,
4641 unsigned long range_end_pfn
)
4643 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4644 unsigned long start_pfn
, end_pfn
;
4647 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4648 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4649 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4650 nr_absent
-= end_pfn
- start_pfn
;
4656 * absent_pages_in_range - Return number of page frames in holes within a range
4657 * @start_pfn: The start PFN to start searching for holes
4658 * @end_pfn: The end PFN to stop searching for holes
4660 * It returns the number of pages frames in memory holes within a range.
4662 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4663 unsigned long end_pfn
)
4665 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4668 /* Return the number of page frames in holes in a zone on a node */
4669 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4670 unsigned long zone_type
,
4671 unsigned long node_start_pfn
,
4672 unsigned long node_end_pfn
,
4673 unsigned long *ignored
)
4675 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4676 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4677 unsigned long zone_start_pfn
, zone_end_pfn
;
4679 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4680 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4682 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4683 node_start_pfn
, node_end_pfn
,
4684 &zone_start_pfn
, &zone_end_pfn
);
4685 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4688 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4689 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4690 unsigned long zone_type
,
4691 unsigned long node_start_pfn
,
4692 unsigned long node_end_pfn
,
4693 unsigned long *zones_size
)
4695 return zones_size
[zone_type
];
4698 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4699 unsigned long zone_type
,
4700 unsigned long node_start_pfn
,
4701 unsigned long node_end_pfn
,
4702 unsigned long *zholes_size
)
4707 return zholes_size
[zone_type
];
4710 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4712 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4713 unsigned long node_start_pfn
,
4714 unsigned long node_end_pfn
,
4715 unsigned long *zones_size
,
4716 unsigned long *zholes_size
)
4718 unsigned long realtotalpages
, totalpages
= 0;
4721 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4722 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4726 pgdat
->node_spanned_pages
= totalpages
;
4728 realtotalpages
= totalpages
;
4729 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4731 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4732 node_start_pfn
, node_end_pfn
,
4734 pgdat
->node_present_pages
= realtotalpages
;
4735 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4739 #ifndef CONFIG_SPARSEMEM
4741 * Calculate the size of the zone->blockflags rounded to an unsigned long
4742 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4743 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4744 * round what is now in bits to nearest long in bits, then return it in
4747 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4749 unsigned long usemapsize
;
4751 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4752 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4753 usemapsize
= usemapsize
>> pageblock_order
;
4754 usemapsize
*= NR_PAGEBLOCK_BITS
;
4755 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4757 return usemapsize
/ 8;
4760 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4762 unsigned long zone_start_pfn
,
4763 unsigned long zonesize
)
4765 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4766 zone
->pageblock_flags
= NULL
;
4768 zone
->pageblock_flags
=
4769 memblock_virt_alloc_node_nopanic(usemapsize
,
4773 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4774 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4775 #endif /* CONFIG_SPARSEMEM */
4777 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4779 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4780 void __paginginit
set_pageblock_order(void)
4784 /* Check that pageblock_nr_pages has not already been setup */
4785 if (pageblock_order
)
4788 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4789 order
= HUGETLB_PAGE_ORDER
;
4791 order
= MAX_ORDER
- 1;
4794 * Assume the largest contiguous order of interest is a huge page.
4795 * This value may be variable depending on boot parameters on IA64 and
4798 pageblock_order
= order
;
4800 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4803 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4804 * is unused as pageblock_order is set at compile-time. See
4805 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4808 void __paginginit
set_pageblock_order(void)
4812 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4814 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4815 unsigned long present_pages
)
4817 unsigned long pages
= spanned_pages
;
4820 * Provide a more accurate estimation if there are holes within
4821 * the zone and SPARSEMEM is in use. If there are holes within the
4822 * zone, each populated memory region may cost us one or two extra
4823 * memmap pages due to alignment because memmap pages for each
4824 * populated regions may not naturally algined on page boundary.
4825 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4827 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4828 IS_ENABLED(CONFIG_SPARSEMEM
))
4829 pages
= present_pages
;
4831 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4835 * Set up the zone data structures:
4836 * - mark all pages reserved
4837 * - mark all memory queues empty
4838 * - clear the memory bitmaps
4840 * NOTE: pgdat should get zeroed by caller.
4842 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4843 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4844 unsigned long *zones_size
, unsigned long *zholes_size
)
4847 int nid
= pgdat
->node_id
;
4848 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4851 pgdat_resize_init(pgdat
);
4852 #ifdef CONFIG_NUMA_BALANCING
4853 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4854 pgdat
->numabalancing_migrate_nr_pages
= 0;
4855 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4857 init_waitqueue_head(&pgdat
->kswapd_wait
);
4858 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4860 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4861 struct zone
*zone
= pgdat
->node_zones
+ j
;
4862 unsigned long size
, realsize
, freesize
, memmap_pages
;
4864 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4865 node_end_pfn
, zones_size
);
4866 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4872 * Adjust freesize so that it accounts for how much memory
4873 * is used by this zone for memmap. This affects the watermark
4874 * and per-cpu initialisations
4876 memmap_pages
= calc_memmap_size(size
, realsize
);
4877 if (freesize
>= memmap_pages
) {
4878 freesize
-= memmap_pages
;
4881 " %s zone: %lu pages used for memmap\n",
4882 zone_names
[j
], memmap_pages
);
4885 " %s zone: %lu pages exceeds freesize %lu\n",
4886 zone_names
[j
], memmap_pages
, freesize
);
4888 /* Account for reserved pages */
4889 if (j
== 0 && freesize
> dma_reserve
) {
4890 freesize
-= dma_reserve
;
4891 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4892 zone_names
[0], dma_reserve
);
4895 if (!is_highmem_idx(j
))
4896 nr_kernel_pages
+= freesize
;
4897 /* Charge for highmem memmap if there are enough kernel pages */
4898 else if (nr_kernel_pages
> memmap_pages
* 2)
4899 nr_kernel_pages
-= memmap_pages
;
4900 nr_all_pages
+= freesize
;
4902 zone
->spanned_pages
= size
;
4903 zone
->present_pages
= realsize
;
4905 * Set an approximate value for lowmem here, it will be adjusted
4906 * when the bootmem allocator frees pages into the buddy system.
4907 * And all highmem pages will be managed by the buddy system.
4909 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4912 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4914 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4916 zone
->name
= zone_names
[j
];
4917 spin_lock_init(&zone
->lock
);
4918 spin_lock_init(&zone
->lru_lock
);
4919 zone_seqlock_init(zone
);
4920 zone
->zone_pgdat
= pgdat
;
4921 zone_pcp_init(zone
);
4923 /* For bootup, initialized properly in watermark setup */
4924 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4926 lruvec_init(&zone
->lruvec
);
4930 set_pageblock_order();
4931 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4932 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4933 size
, MEMMAP_EARLY
);
4935 memmap_init(size
, nid
, j
, zone_start_pfn
);
4936 zone_start_pfn
+= size
;
4940 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4942 /* Skip empty nodes */
4943 if (!pgdat
->node_spanned_pages
)
4946 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4947 /* ia64 gets its own node_mem_map, before this, without bootmem */
4948 if (!pgdat
->node_mem_map
) {
4949 unsigned long size
, start
, end
;
4953 * The zone's endpoints aren't required to be MAX_ORDER
4954 * aligned but the node_mem_map endpoints must be in order
4955 * for the buddy allocator to function correctly.
4957 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4958 end
= pgdat_end_pfn(pgdat
);
4959 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4960 size
= (end
- start
) * sizeof(struct page
);
4961 map
= alloc_remap(pgdat
->node_id
, size
);
4963 map
= memblock_virt_alloc_node_nopanic(size
,
4965 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4967 #ifndef CONFIG_NEED_MULTIPLE_NODES
4969 * With no DISCONTIG, the global mem_map is just set as node 0's
4971 if (pgdat
== NODE_DATA(0)) {
4972 mem_map
= NODE_DATA(0)->node_mem_map
;
4973 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4974 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4975 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4976 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4979 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4982 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4983 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4985 pg_data_t
*pgdat
= NODE_DATA(nid
);
4986 unsigned long start_pfn
= 0;
4987 unsigned long end_pfn
= 0;
4989 /* pg_data_t should be reset to zero when it's allocated */
4990 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4992 pgdat
->node_id
= nid
;
4993 pgdat
->node_start_pfn
= node_start_pfn
;
4994 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4995 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4996 printk(KERN_INFO
"Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid
,
4997 (u64
) start_pfn
<< PAGE_SHIFT
, (u64
) (end_pfn
<< PAGE_SHIFT
) - 1);
4999 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
5000 zones_size
, zholes_size
);
5002 alloc_node_mem_map(pgdat
);
5003 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5004 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5005 nid
, (unsigned long)pgdat
,
5006 (unsigned long)pgdat
->node_mem_map
);
5009 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
5010 zones_size
, zholes_size
);
5013 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5015 #if MAX_NUMNODES > 1
5017 * Figure out the number of possible node ids.
5019 void __init
setup_nr_node_ids(void)
5022 unsigned int highest
= 0;
5024 for_each_node_mask(node
, node_possible_map
)
5026 nr_node_ids
= highest
+ 1;
5031 * node_map_pfn_alignment - determine the maximum internode alignment
5033 * This function should be called after node map is populated and sorted.
5034 * It calculates the maximum power of two alignment which can distinguish
5037 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5038 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5039 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5040 * shifted, 1GiB is enough and this function will indicate so.
5042 * This is used to test whether pfn -> nid mapping of the chosen memory
5043 * model has fine enough granularity to avoid incorrect mapping for the
5044 * populated node map.
5046 * Returns the determined alignment in pfn's. 0 if there is no alignment
5047 * requirement (single node).
5049 unsigned long __init
node_map_pfn_alignment(void)
5051 unsigned long accl_mask
= 0, last_end
= 0;
5052 unsigned long start
, end
, mask
;
5056 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5057 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5064 * Start with a mask granular enough to pin-point to the
5065 * start pfn and tick off bits one-by-one until it becomes
5066 * too coarse to separate the current node from the last.
5068 mask
= ~((1 << __ffs(start
)) - 1);
5069 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5072 /* accumulate all internode masks */
5076 /* convert mask to number of pages */
5077 return ~accl_mask
+ 1;
5080 /* Find the lowest pfn for a node */
5081 static unsigned long __init
find_min_pfn_for_node(int nid
)
5083 unsigned long min_pfn
= ULONG_MAX
;
5084 unsigned long start_pfn
;
5087 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5088 min_pfn
= min(min_pfn
, start_pfn
);
5090 if (min_pfn
== ULONG_MAX
) {
5092 "Could not find start_pfn for node %d\n", nid
);
5100 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5102 * It returns the minimum PFN based on information provided via
5103 * memblock_set_node().
5105 unsigned long __init
find_min_pfn_with_active_regions(void)
5107 return find_min_pfn_for_node(MAX_NUMNODES
);
5111 * early_calculate_totalpages()
5112 * Sum pages in active regions for movable zone.
5113 * Populate N_MEMORY for calculating usable_nodes.
5115 static unsigned long __init
early_calculate_totalpages(void)
5117 unsigned long totalpages
= 0;
5118 unsigned long start_pfn
, end_pfn
;
5121 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5122 unsigned long pages
= end_pfn
- start_pfn
;
5124 totalpages
+= pages
;
5126 node_set_state(nid
, N_MEMORY
);
5132 * Find the PFN the Movable zone begins in each node. Kernel memory
5133 * is spread evenly between nodes as long as the nodes have enough
5134 * memory. When they don't, some nodes will have more kernelcore than
5137 static void __init
find_zone_movable_pfns_for_nodes(void)
5140 unsigned long usable_startpfn
;
5141 unsigned long kernelcore_node
, kernelcore_remaining
;
5142 /* save the state before borrow the nodemask */
5143 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5144 unsigned long totalpages
= early_calculate_totalpages();
5145 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5146 struct memblock_region
*r
;
5148 /* Need to find movable_zone earlier when movable_node is specified. */
5149 find_usable_zone_for_movable();
5152 * If movable_node is specified, ignore kernelcore and movablecore
5155 if (movable_node_is_enabled()) {
5156 for_each_memblock(memory
, r
) {
5157 if (!memblock_is_hotpluggable(r
))
5162 usable_startpfn
= PFN_DOWN(r
->base
);
5163 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5164 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5172 * If movablecore=nn[KMG] was specified, calculate what size of
5173 * kernelcore that corresponds so that memory usable for
5174 * any allocation type is evenly spread. If both kernelcore
5175 * and movablecore are specified, then the value of kernelcore
5176 * will be used for required_kernelcore if it's greater than
5177 * what movablecore would have allowed.
5179 if (required_movablecore
) {
5180 unsigned long corepages
;
5183 * Round-up so that ZONE_MOVABLE is at least as large as what
5184 * was requested by the user
5186 required_movablecore
=
5187 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5188 corepages
= totalpages
- required_movablecore
;
5190 required_kernelcore
= max(required_kernelcore
, corepages
);
5193 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5194 if (!required_kernelcore
)
5197 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5198 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5201 /* Spread kernelcore memory as evenly as possible throughout nodes */
5202 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5203 for_each_node_state(nid
, N_MEMORY
) {
5204 unsigned long start_pfn
, end_pfn
;
5207 * Recalculate kernelcore_node if the division per node
5208 * now exceeds what is necessary to satisfy the requested
5209 * amount of memory for the kernel
5211 if (required_kernelcore
< kernelcore_node
)
5212 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5215 * As the map is walked, we track how much memory is usable
5216 * by the kernel using kernelcore_remaining. When it is
5217 * 0, the rest of the node is usable by ZONE_MOVABLE
5219 kernelcore_remaining
= kernelcore_node
;
5221 /* Go through each range of PFNs within this node */
5222 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5223 unsigned long size_pages
;
5225 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5226 if (start_pfn
>= end_pfn
)
5229 /* Account for what is only usable for kernelcore */
5230 if (start_pfn
< usable_startpfn
) {
5231 unsigned long kernel_pages
;
5232 kernel_pages
= min(end_pfn
, usable_startpfn
)
5235 kernelcore_remaining
-= min(kernel_pages
,
5236 kernelcore_remaining
);
5237 required_kernelcore
-= min(kernel_pages
,
5238 required_kernelcore
);
5240 /* Continue if range is now fully accounted */
5241 if (end_pfn
<= usable_startpfn
) {
5244 * Push zone_movable_pfn to the end so
5245 * that if we have to rebalance
5246 * kernelcore across nodes, we will
5247 * not double account here
5249 zone_movable_pfn
[nid
] = end_pfn
;
5252 start_pfn
= usable_startpfn
;
5256 * The usable PFN range for ZONE_MOVABLE is from
5257 * start_pfn->end_pfn. Calculate size_pages as the
5258 * number of pages used as kernelcore
5260 size_pages
= end_pfn
- start_pfn
;
5261 if (size_pages
> kernelcore_remaining
)
5262 size_pages
= kernelcore_remaining
;
5263 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5266 * Some kernelcore has been met, update counts and
5267 * break if the kernelcore for this node has been
5270 required_kernelcore
-= min(required_kernelcore
,
5272 kernelcore_remaining
-= size_pages
;
5273 if (!kernelcore_remaining
)
5279 * If there is still required_kernelcore, we do another pass with one
5280 * less node in the count. This will push zone_movable_pfn[nid] further
5281 * along on the nodes that still have memory until kernelcore is
5285 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5289 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5290 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5291 zone_movable_pfn
[nid
] =
5292 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5295 /* restore the node_state */
5296 node_states
[N_MEMORY
] = saved_node_state
;
5299 /* Any regular or high memory on that node ? */
5300 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5302 enum zone_type zone_type
;
5304 if (N_MEMORY
== N_NORMAL_MEMORY
)
5307 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5308 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5309 if (populated_zone(zone
)) {
5310 node_set_state(nid
, N_HIGH_MEMORY
);
5311 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5312 zone_type
<= ZONE_NORMAL
)
5313 node_set_state(nid
, N_NORMAL_MEMORY
);
5320 * free_area_init_nodes - Initialise all pg_data_t and zone data
5321 * @max_zone_pfn: an array of max PFNs for each zone
5323 * This will call free_area_init_node() for each active node in the system.
5324 * Using the page ranges provided by memblock_set_node(), the size of each
5325 * zone in each node and their holes is calculated. If the maximum PFN
5326 * between two adjacent zones match, it is assumed that the zone is empty.
5327 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5328 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5329 * starts where the previous one ended. For example, ZONE_DMA32 starts
5330 * at arch_max_dma_pfn.
5332 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5334 unsigned long start_pfn
, end_pfn
;
5337 /* Record where the zone boundaries are */
5338 memset(arch_zone_lowest_possible_pfn
, 0,
5339 sizeof(arch_zone_lowest_possible_pfn
));
5340 memset(arch_zone_highest_possible_pfn
, 0,
5341 sizeof(arch_zone_highest_possible_pfn
));
5342 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5343 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5344 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5345 if (i
== ZONE_MOVABLE
)
5347 arch_zone_lowest_possible_pfn
[i
] =
5348 arch_zone_highest_possible_pfn
[i
-1];
5349 arch_zone_highest_possible_pfn
[i
] =
5350 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5352 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5353 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5355 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5356 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5357 find_zone_movable_pfns_for_nodes();
5359 /* Print out the zone ranges */
5360 pr_info("Zone ranges:\n");
5361 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5362 if (i
== ZONE_MOVABLE
)
5364 pr_info(" %-8s ", zone_names
[i
]);
5365 if (arch_zone_lowest_possible_pfn
[i
] ==
5366 arch_zone_highest_possible_pfn
[i
])
5369 pr_cont("[mem %0#10lx-%0#10lx]\n",
5370 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5371 (arch_zone_highest_possible_pfn
[i
]
5372 << PAGE_SHIFT
) - 1);
5375 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5376 pr_info("Movable zone start for each node\n");
5377 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5378 if (zone_movable_pfn
[i
])
5379 pr_info(" Node %d: %#010lx\n", i
,
5380 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5383 /* Print out the early node map */
5384 pr_info("Early memory node ranges\n");
5385 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5386 pr_info(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5387 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5389 /* Initialise every node */
5390 mminit_verify_pageflags_layout();
5391 setup_nr_node_ids();
5392 for_each_online_node(nid
) {
5393 pg_data_t
*pgdat
= NODE_DATA(nid
);
5394 free_area_init_node(nid
, NULL
,
5395 find_min_pfn_for_node(nid
), NULL
);
5397 /* Any memory on that node */
5398 if (pgdat
->node_present_pages
)
5399 node_set_state(nid
, N_MEMORY
);
5400 check_for_memory(pgdat
, nid
);
5404 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5406 unsigned long long coremem
;
5410 coremem
= memparse(p
, &p
);
5411 *core
= coremem
>> PAGE_SHIFT
;
5413 /* Paranoid check that UL is enough for the coremem value */
5414 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5420 * kernelcore=size sets the amount of memory for use for allocations that
5421 * cannot be reclaimed or migrated.
5423 static int __init
cmdline_parse_kernelcore(char *p
)
5425 return cmdline_parse_core(p
, &required_kernelcore
);
5429 * movablecore=size sets the amount of memory for use for allocations that
5430 * can be reclaimed or migrated.
5432 static int __init
cmdline_parse_movablecore(char *p
)
5434 return cmdline_parse_core(p
, &required_movablecore
);
5437 early_param("kernelcore", cmdline_parse_kernelcore
);
5438 early_param("movablecore", cmdline_parse_movablecore
);
5440 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5442 void adjust_managed_page_count(struct page
*page
, long count
)
5444 spin_lock(&managed_page_count_lock
);
5445 page_zone(page
)->managed_pages
+= count
;
5446 totalram_pages
+= count
;
5447 #ifdef CONFIG_HIGHMEM
5448 if (PageHighMem(page
))
5449 totalhigh_pages
+= count
;
5451 spin_unlock(&managed_page_count_lock
);
5453 EXPORT_SYMBOL(adjust_managed_page_count
);
5455 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5458 unsigned long pages
= 0;
5460 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5461 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5462 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5463 if ((unsigned int)poison
<= 0xFF)
5464 memset(pos
, poison
, PAGE_SIZE
);
5465 free_reserved_page(virt_to_page(pos
));
5469 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5470 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5474 EXPORT_SYMBOL(free_reserved_area
);
5476 #ifdef CONFIG_HIGHMEM
5477 void free_highmem_page(struct page
*page
)
5479 __free_reserved_page(page
);
5481 page_zone(page
)->managed_pages
++;
5487 void __init
mem_init_print_info(const char *str
)
5489 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5490 unsigned long init_code_size
, init_data_size
;
5492 physpages
= get_num_physpages();
5493 codesize
= _etext
- _stext
;
5494 datasize
= _edata
- _sdata
;
5495 rosize
= __end_rodata
- __start_rodata
;
5496 bss_size
= __bss_stop
- __bss_start
;
5497 init_data_size
= __init_end
- __init_begin
;
5498 init_code_size
= _einittext
- _sinittext
;
5501 * Detect special cases and adjust section sizes accordingly:
5502 * 1) .init.* may be embedded into .data sections
5503 * 2) .init.text.* may be out of [__init_begin, __init_end],
5504 * please refer to arch/tile/kernel/vmlinux.lds.S.
5505 * 3) .rodata.* may be embedded into .text or .data sections.
5507 #define adj_init_size(start, end, size, pos, adj) \
5509 if (start <= pos && pos < end && size > adj) \
5513 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5514 _sinittext
, init_code_size
);
5515 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5516 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5517 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5518 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5520 #undef adj_init_size
5522 pr_info("Memory: %luK/%luK available "
5523 "(%luK kernel code, %luK rwdata, %luK rodata, "
5524 "%luK init, %luK bss, %luK reserved"
5525 #ifdef CONFIG_HIGHMEM
5529 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5530 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5531 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5532 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5533 #ifdef CONFIG_HIGHMEM
5534 totalhigh_pages
<< (PAGE_SHIFT
-10),
5536 str
? ", " : "", str
? str
: "");
5540 * set_dma_reserve - set the specified number of pages reserved in the first zone
5541 * @new_dma_reserve: The number of pages to mark reserved
5543 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5544 * In the DMA zone, a significant percentage may be consumed by kernel image
5545 * and other unfreeable allocations which can skew the watermarks badly. This
5546 * function may optionally be used to account for unfreeable pages in the
5547 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5548 * smaller per-cpu batchsize.
5550 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5552 dma_reserve
= new_dma_reserve
;
5555 void __init
free_area_init(unsigned long *zones_size
)
5557 free_area_init_node(0, zones_size
,
5558 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5561 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5562 unsigned long action
, void *hcpu
)
5564 int cpu
= (unsigned long)hcpu
;
5566 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5567 lru_add_drain_cpu(cpu
);
5571 * Spill the event counters of the dead processor
5572 * into the current processors event counters.
5573 * This artificially elevates the count of the current
5576 vm_events_fold_cpu(cpu
);
5579 * Zero the differential counters of the dead processor
5580 * so that the vm statistics are consistent.
5582 * This is only okay since the processor is dead and cannot
5583 * race with what we are doing.
5585 cpu_vm_stats_fold(cpu
);
5590 void __init
page_alloc_init(void)
5592 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5596 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5597 * or min_free_kbytes changes.
5599 static void calculate_totalreserve_pages(void)
5601 struct pglist_data
*pgdat
;
5602 unsigned long reserve_pages
= 0;
5603 enum zone_type i
, j
;
5605 for_each_online_pgdat(pgdat
) {
5606 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5607 struct zone
*zone
= pgdat
->node_zones
+ i
;
5610 /* Find valid and maximum lowmem_reserve in the zone */
5611 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5612 if (zone
->lowmem_reserve
[j
] > max
)
5613 max
= zone
->lowmem_reserve
[j
];
5616 /* we treat the high watermark as reserved pages. */
5617 max
+= high_wmark_pages(zone
);
5619 if (max
> zone
->managed_pages
)
5620 max
= zone
->managed_pages
;
5621 reserve_pages
+= max
;
5623 * Lowmem reserves are not available to
5624 * GFP_HIGHUSER page cache allocations and
5625 * kswapd tries to balance zones to their high
5626 * watermark. As a result, neither should be
5627 * regarded as dirtyable memory, to prevent a
5628 * situation where reclaim has to clean pages
5629 * in order to balance the zones.
5631 zone
->dirty_balance_reserve
= max
;
5634 dirty_balance_reserve
= reserve_pages
;
5635 totalreserve_pages
= reserve_pages
;
5639 * setup_per_zone_lowmem_reserve - called whenever
5640 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5641 * has a correct pages reserved value, so an adequate number of
5642 * pages are left in the zone after a successful __alloc_pages().
5644 static void setup_per_zone_lowmem_reserve(void)
5646 struct pglist_data
*pgdat
;
5647 enum zone_type j
, idx
;
5649 for_each_online_pgdat(pgdat
) {
5650 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5651 struct zone
*zone
= pgdat
->node_zones
+ j
;
5652 unsigned long managed_pages
= zone
->managed_pages
;
5654 zone
->lowmem_reserve
[j
] = 0;
5658 struct zone
*lower_zone
;
5662 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5663 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5665 lower_zone
= pgdat
->node_zones
+ idx
;
5666 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5667 sysctl_lowmem_reserve_ratio
[idx
];
5668 managed_pages
+= lower_zone
->managed_pages
;
5673 /* update totalreserve_pages */
5674 calculate_totalreserve_pages();
5677 static void __setup_per_zone_wmarks(void)
5679 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5680 unsigned long lowmem_pages
= 0;
5682 unsigned long flags
;
5684 /* Calculate total number of !ZONE_HIGHMEM pages */
5685 for_each_zone(zone
) {
5686 if (!is_highmem(zone
))
5687 lowmem_pages
+= zone
->managed_pages
;
5690 for_each_zone(zone
) {
5693 spin_lock_irqsave(&zone
->lock
, flags
);
5694 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5695 do_div(tmp
, lowmem_pages
);
5696 if (is_highmem(zone
)) {
5698 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5699 * need highmem pages, so cap pages_min to a small
5702 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5703 * deltas controls asynch page reclaim, and so should
5704 * not be capped for highmem.
5706 unsigned long min_pages
;
5708 min_pages
= zone
->managed_pages
/ 1024;
5709 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5710 zone
->watermark
[WMARK_MIN
] = min_pages
;
5713 * If it's a lowmem zone, reserve a number of pages
5714 * proportionate to the zone's size.
5716 zone
->watermark
[WMARK_MIN
] = tmp
;
5719 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5720 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5722 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5723 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5724 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5726 setup_zone_migrate_reserve(zone
);
5727 spin_unlock_irqrestore(&zone
->lock
, flags
);
5730 /* update totalreserve_pages */
5731 calculate_totalreserve_pages();
5735 * setup_per_zone_wmarks - called when min_free_kbytes changes
5736 * or when memory is hot-{added|removed}
5738 * Ensures that the watermark[min,low,high] values for each zone are set
5739 * correctly with respect to min_free_kbytes.
5741 void setup_per_zone_wmarks(void)
5743 mutex_lock(&zonelists_mutex
);
5744 __setup_per_zone_wmarks();
5745 mutex_unlock(&zonelists_mutex
);
5749 * The inactive anon list should be small enough that the VM never has to
5750 * do too much work, but large enough that each inactive page has a chance
5751 * to be referenced again before it is swapped out.
5753 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5754 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5755 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5756 * the anonymous pages are kept on the inactive list.
5759 * memory ratio inactive anon
5760 * -------------------------------------
5769 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5771 unsigned int gb
, ratio
;
5773 /* Zone size in gigabytes */
5774 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5776 ratio
= int_sqrt(10 * gb
);
5780 zone
->inactive_ratio
= ratio
;
5783 static void __meminit
setup_per_zone_inactive_ratio(void)
5788 calculate_zone_inactive_ratio(zone
);
5792 * Initialise min_free_kbytes.
5794 * For small machines we want it small (128k min). For large machines
5795 * we want it large (64MB max). But it is not linear, because network
5796 * bandwidth does not increase linearly with machine size. We use
5798 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5799 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5815 int __meminit
init_per_zone_wmark_min(void)
5817 unsigned long lowmem_kbytes
;
5818 int new_min_free_kbytes
;
5820 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5821 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5823 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5824 min_free_kbytes
= new_min_free_kbytes
;
5825 if (min_free_kbytes
< 128)
5826 min_free_kbytes
= 128;
5827 if (min_free_kbytes
> 65536)
5828 min_free_kbytes
= 65536;
5830 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5831 new_min_free_kbytes
, user_min_free_kbytes
);
5833 setup_per_zone_wmarks();
5834 refresh_zone_stat_thresholds();
5835 setup_per_zone_lowmem_reserve();
5836 setup_per_zone_inactive_ratio();
5839 module_init(init_per_zone_wmark_min
)
5842 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5843 * that we can call two helper functions whenever min_free_kbytes
5846 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5847 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5851 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5856 user_min_free_kbytes
= min_free_kbytes
;
5857 setup_per_zone_wmarks();
5863 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5864 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5869 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5874 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5875 sysctl_min_unmapped_ratio
) / 100;
5879 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5880 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5885 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5890 zone
->min_slab_pages
= (zone
->managed_pages
*
5891 sysctl_min_slab_ratio
) / 100;
5897 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5898 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5899 * whenever sysctl_lowmem_reserve_ratio changes.
5901 * The reserve ratio obviously has absolutely no relation with the
5902 * minimum watermarks. The lowmem reserve ratio can only make sense
5903 * if in function of the boot time zone sizes.
5905 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5906 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5908 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5909 setup_per_zone_lowmem_reserve();
5914 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5915 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5916 * pagelist can have before it gets flushed back to buddy allocator.
5918 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5919 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5922 int old_percpu_pagelist_fraction
;
5925 mutex_lock(&pcp_batch_high_lock
);
5926 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5928 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5929 if (!write
|| ret
< 0)
5932 /* Sanity checking to avoid pcp imbalance */
5933 if (percpu_pagelist_fraction
&&
5934 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5935 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5941 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
5944 for_each_populated_zone(zone
) {
5947 for_each_possible_cpu(cpu
)
5948 pageset_set_high_and_batch(zone
,
5949 per_cpu_ptr(zone
->pageset
, cpu
));
5952 mutex_unlock(&pcp_batch_high_lock
);
5956 int hashdist
= HASHDIST_DEFAULT
;
5959 static int __init
set_hashdist(char *str
)
5963 hashdist
= simple_strtoul(str
, &str
, 0);
5966 __setup("hashdist=", set_hashdist
);
5970 * allocate a large system hash table from bootmem
5971 * - it is assumed that the hash table must contain an exact power-of-2
5972 * quantity of entries
5973 * - limit is the number of hash buckets, not the total allocation size
5975 void *__init
alloc_large_system_hash(const char *tablename
,
5976 unsigned long bucketsize
,
5977 unsigned long numentries
,
5980 unsigned int *_hash_shift
,
5981 unsigned int *_hash_mask
,
5982 unsigned long low_limit
,
5983 unsigned long high_limit
)
5985 unsigned long long max
= high_limit
;
5986 unsigned long log2qty
, size
;
5989 /* allow the kernel cmdline to have a say */
5991 /* round applicable memory size up to nearest megabyte */
5992 numentries
= nr_kernel_pages
;
5994 /* It isn't necessary when PAGE_SIZE >= 1MB */
5995 if (PAGE_SHIFT
< 20)
5996 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
5998 /* limit to 1 bucket per 2^scale bytes of low memory */
5999 if (scale
> PAGE_SHIFT
)
6000 numentries
>>= (scale
- PAGE_SHIFT
);
6002 numentries
<<= (PAGE_SHIFT
- scale
);
6004 /* Make sure we've got at least a 0-order allocation.. */
6005 if (unlikely(flags
& HASH_SMALL
)) {
6006 /* Makes no sense without HASH_EARLY */
6007 WARN_ON(!(flags
& HASH_EARLY
));
6008 if (!(numentries
>> *_hash_shift
)) {
6009 numentries
= 1UL << *_hash_shift
;
6010 BUG_ON(!numentries
);
6012 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6013 numentries
= PAGE_SIZE
/ bucketsize
;
6015 numentries
= roundup_pow_of_two(numentries
);
6017 /* limit allocation size to 1/16 total memory by default */
6019 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6020 do_div(max
, bucketsize
);
6022 max
= min(max
, 0x80000000ULL
);
6024 if (numentries
< low_limit
)
6025 numentries
= low_limit
;
6026 if (numentries
> max
)
6029 log2qty
= ilog2(numentries
);
6032 size
= bucketsize
<< log2qty
;
6033 if (flags
& HASH_EARLY
)
6034 table
= memblock_virt_alloc_nopanic(size
, 0);
6036 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6039 * If bucketsize is not a power-of-two, we may free
6040 * some pages at the end of hash table which
6041 * alloc_pages_exact() automatically does
6043 if (get_order(size
) < MAX_ORDER
) {
6044 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6045 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6048 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6051 panic("Failed to allocate %s hash table\n", tablename
);
6053 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6056 ilog2(size
) - PAGE_SHIFT
,
6060 *_hash_shift
= log2qty
;
6062 *_hash_mask
= (1 << log2qty
) - 1;
6067 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6068 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6071 #ifdef CONFIG_SPARSEMEM
6072 return __pfn_to_section(pfn
)->pageblock_flags
;
6074 return zone
->pageblock_flags
;
6075 #endif /* CONFIG_SPARSEMEM */
6078 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6080 #ifdef CONFIG_SPARSEMEM
6081 pfn
&= (PAGES_PER_SECTION
-1);
6082 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6084 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6085 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6086 #endif /* CONFIG_SPARSEMEM */
6090 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6091 * @page: The page within the block of interest
6092 * @pfn: The target page frame number
6093 * @end_bitidx: The last bit of interest to retrieve
6094 * @mask: mask of bits that the caller is interested in
6096 * Return: pageblock_bits flags
6098 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6099 unsigned long end_bitidx
,
6103 unsigned long *bitmap
;
6104 unsigned long bitidx
, word_bitidx
;
6107 zone
= page_zone(page
);
6108 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6109 bitidx
= pfn_to_bitidx(zone
, pfn
);
6110 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6111 bitidx
&= (BITS_PER_LONG
-1);
6113 word
= bitmap
[word_bitidx
];
6114 bitidx
+= end_bitidx
;
6115 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6119 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6120 * @page: The page within the block of interest
6121 * @flags: The flags to set
6122 * @pfn: The target page frame number
6123 * @end_bitidx: The last bit of interest
6124 * @mask: mask of bits that the caller is interested in
6126 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6128 unsigned long end_bitidx
,
6132 unsigned long *bitmap
;
6133 unsigned long bitidx
, word_bitidx
;
6134 unsigned long old_word
, word
;
6136 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6138 zone
= page_zone(page
);
6139 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6140 bitidx
= pfn_to_bitidx(zone
, pfn
);
6141 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6142 bitidx
&= (BITS_PER_LONG
-1);
6144 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6146 bitidx
+= end_bitidx
;
6147 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6148 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6150 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6152 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6153 if (word
== old_word
)
6160 * This function checks whether pageblock includes unmovable pages or not.
6161 * If @count is not zero, it is okay to include less @count unmovable pages
6163 * PageLRU check without isolation or lru_lock could race so that
6164 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6165 * expect this function should be exact.
6167 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6168 bool skip_hwpoisoned_pages
)
6170 unsigned long pfn
, iter
, found
;
6174 * For avoiding noise data, lru_add_drain_all() should be called
6175 * If ZONE_MOVABLE, the zone never contains unmovable pages
6177 if (zone_idx(zone
) == ZONE_MOVABLE
)
6179 mt
= get_pageblock_migratetype(page
);
6180 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6183 pfn
= page_to_pfn(page
);
6184 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6185 unsigned long check
= pfn
+ iter
;
6187 if (!pfn_valid_within(check
))
6190 page
= pfn_to_page(check
);
6193 * Hugepages are not in LRU lists, but they're movable.
6194 * We need not scan over tail pages bacause we don't
6195 * handle each tail page individually in migration.
6197 if (PageHuge(page
)) {
6198 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6203 * We can't use page_count without pin a page
6204 * because another CPU can free compound page.
6205 * This check already skips compound tails of THP
6206 * because their page->_count is zero at all time.
6208 if (!atomic_read(&page
->_count
)) {
6209 if (PageBuddy(page
))
6210 iter
+= (1 << page_order(page
)) - 1;
6215 * The HWPoisoned page may be not in buddy system, and
6216 * page_count() is not 0.
6218 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6224 * If there are RECLAIMABLE pages, we need to check it.
6225 * But now, memory offline itself doesn't call shrink_slab()
6226 * and it still to be fixed.
6229 * If the page is not RAM, page_count()should be 0.
6230 * we don't need more check. This is an _used_ not-movable page.
6232 * The problematic thing here is PG_reserved pages. PG_reserved
6233 * is set to both of a memory hole page and a _used_ kernel
6242 bool is_pageblock_removable_nolock(struct page
*page
)
6248 * We have to be careful here because we are iterating over memory
6249 * sections which are not zone aware so we might end up outside of
6250 * the zone but still within the section.
6251 * We have to take care about the node as well. If the node is offline
6252 * its NODE_DATA will be NULL - see page_zone.
6254 if (!node_online(page_to_nid(page
)))
6257 zone
= page_zone(page
);
6258 pfn
= page_to_pfn(page
);
6259 if (!zone_spans_pfn(zone
, pfn
))
6262 return !has_unmovable_pages(zone
, page
, 0, true);
6267 static unsigned long pfn_max_align_down(unsigned long pfn
)
6269 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6270 pageblock_nr_pages
) - 1);
6273 static unsigned long pfn_max_align_up(unsigned long pfn
)
6275 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6276 pageblock_nr_pages
));
6279 /* [start, end) must belong to a single zone. */
6280 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6281 unsigned long start
, unsigned long end
)
6283 /* This function is based on compact_zone() from compaction.c. */
6284 unsigned long nr_reclaimed
;
6285 unsigned long pfn
= start
;
6286 unsigned int tries
= 0;
6291 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6292 if (fatal_signal_pending(current
)) {
6297 if (list_empty(&cc
->migratepages
)) {
6298 cc
->nr_migratepages
= 0;
6299 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6305 } else if (++tries
== 5) {
6306 ret
= ret
< 0 ? ret
: -EBUSY
;
6310 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6312 cc
->nr_migratepages
-= nr_reclaimed
;
6314 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6315 NULL
, 0, cc
->mode
, MR_CMA
);
6318 putback_movable_pages(&cc
->migratepages
);
6325 * alloc_contig_range() -- tries to allocate given range of pages
6326 * @start: start PFN to allocate
6327 * @end: one-past-the-last PFN to allocate
6328 * @migratetype: migratetype of the underlaying pageblocks (either
6329 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6330 * in range must have the same migratetype and it must
6331 * be either of the two.
6333 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6334 * aligned, however it's the caller's responsibility to guarantee that
6335 * we are the only thread that changes migrate type of pageblocks the
6338 * The PFN range must belong to a single zone.
6340 * Returns zero on success or negative error code. On success all
6341 * pages which PFN is in [start, end) are allocated for the caller and
6342 * need to be freed with free_contig_range().
6344 int alloc_contig_range(unsigned long start
, unsigned long end
,
6345 unsigned migratetype
)
6347 unsigned long outer_start
, outer_end
;
6350 struct compact_control cc
= {
6351 .nr_migratepages
= 0,
6353 .zone
= page_zone(pfn_to_page(start
)),
6354 .mode
= MIGRATE_SYNC
,
6355 .ignore_skip_hint
= true,
6357 INIT_LIST_HEAD(&cc
.migratepages
);
6360 * What we do here is we mark all pageblocks in range as
6361 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6362 * have different sizes, and due to the way page allocator
6363 * work, we align the range to biggest of the two pages so
6364 * that page allocator won't try to merge buddies from
6365 * different pageblocks and change MIGRATE_ISOLATE to some
6366 * other migration type.
6368 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6369 * migrate the pages from an unaligned range (ie. pages that
6370 * we are interested in). This will put all the pages in
6371 * range back to page allocator as MIGRATE_ISOLATE.
6373 * When this is done, we take the pages in range from page
6374 * allocator removing them from the buddy system. This way
6375 * page allocator will never consider using them.
6377 * This lets us mark the pageblocks back as
6378 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6379 * aligned range but not in the unaligned, original range are
6380 * put back to page allocator so that buddy can use them.
6383 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6384 pfn_max_align_up(end
), migratetype
,
6389 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6394 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6395 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6396 * more, all pages in [start, end) are free in page allocator.
6397 * What we are going to do is to allocate all pages from
6398 * [start, end) (that is remove them from page allocator).
6400 * The only problem is that pages at the beginning and at the
6401 * end of interesting range may be not aligned with pages that
6402 * page allocator holds, ie. they can be part of higher order
6403 * pages. Because of this, we reserve the bigger range and
6404 * once this is done free the pages we are not interested in.
6406 * We don't have to hold zone->lock here because the pages are
6407 * isolated thus they won't get removed from buddy.
6410 lru_add_drain_all();
6411 drain_all_pages(cc
.zone
);
6414 outer_start
= start
;
6415 while (!PageBuddy(pfn_to_page(outer_start
))) {
6416 if (++order
>= MAX_ORDER
) {
6420 outer_start
&= ~0UL << order
;
6423 /* Make sure the range is really isolated. */
6424 if (test_pages_isolated(outer_start
, end
, false)) {
6425 pr_info("%s: [%lx, %lx) PFNs busy\n",
6426 __func__
, outer_start
, end
);
6431 /* Grab isolated pages from freelists. */
6432 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6438 /* Free head and tail (if any) */
6439 if (start
!= outer_start
)
6440 free_contig_range(outer_start
, start
- outer_start
);
6441 if (end
!= outer_end
)
6442 free_contig_range(end
, outer_end
- end
);
6445 undo_isolate_page_range(pfn_max_align_down(start
),
6446 pfn_max_align_up(end
), migratetype
);
6450 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6452 unsigned int count
= 0;
6454 for (; nr_pages
--; pfn
++) {
6455 struct page
*page
= pfn_to_page(pfn
);
6457 count
+= page_count(page
) != 1;
6460 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6464 #ifdef CONFIG_MEMORY_HOTPLUG
6466 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6467 * page high values need to be recalulated.
6469 void __meminit
zone_pcp_update(struct zone
*zone
)
6472 mutex_lock(&pcp_batch_high_lock
);
6473 for_each_possible_cpu(cpu
)
6474 pageset_set_high_and_batch(zone
,
6475 per_cpu_ptr(zone
->pageset
, cpu
));
6476 mutex_unlock(&pcp_batch_high_lock
);
6480 void zone_pcp_reset(struct zone
*zone
)
6482 unsigned long flags
;
6484 struct per_cpu_pageset
*pset
;
6486 /* avoid races with drain_pages() */
6487 local_irq_save(flags
);
6488 if (zone
->pageset
!= &boot_pageset
) {
6489 for_each_online_cpu(cpu
) {
6490 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6491 drain_zonestat(zone
, pset
);
6493 free_percpu(zone
->pageset
);
6494 zone
->pageset
= &boot_pageset
;
6496 local_irq_restore(flags
);
6499 #ifdef CONFIG_MEMORY_HOTREMOVE
6501 * All pages in the range must be isolated before calling this.
6504 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6508 unsigned int order
, i
;
6510 unsigned long flags
;
6511 /* find the first valid pfn */
6512 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6517 zone
= page_zone(pfn_to_page(pfn
));
6518 spin_lock_irqsave(&zone
->lock
, flags
);
6520 while (pfn
< end_pfn
) {
6521 if (!pfn_valid(pfn
)) {
6525 page
= pfn_to_page(pfn
);
6527 * The HWPoisoned page may be not in buddy system, and
6528 * page_count() is not 0.
6530 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6532 SetPageReserved(page
);
6536 BUG_ON(page_count(page
));
6537 BUG_ON(!PageBuddy(page
));
6538 order
= page_order(page
);
6539 #ifdef CONFIG_DEBUG_VM
6540 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6541 pfn
, 1 << order
, end_pfn
);
6543 list_del(&page
->lru
);
6544 rmv_page_order(page
);
6545 zone
->free_area
[order
].nr_free
--;
6546 for (i
= 0; i
< (1 << order
); i
++)
6547 SetPageReserved((page
+i
));
6548 pfn
+= (1 << order
);
6550 spin_unlock_irqrestore(&zone
->lock
, flags
);
6554 #ifdef CONFIG_MEMORY_FAILURE
6555 bool is_free_buddy_page(struct page
*page
)
6557 struct zone
*zone
= page_zone(page
);
6558 unsigned long pfn
= page_to_pfn(page
);
6559 unsigned long flags
;
6562 spin_lock_irqsave(&zone
->lock
, flags
);
6563 for (order
= 0; order
< MAX_ORDER
; order
++) {
6564 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6566 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6569 spin_unlock_irqrestore(&zone
->lock
, flags
);
6571 return order
< MAX_ORDER
;