2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/prefetch.h>
57 #include <linux/mm_inline.h>
58 #include <linux/migrate.h>
59 #include <linux/page-debug-flags.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
63 #include <asm/sections.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
69 static DEFINE_MUTEX(pcp_batch_high_lock
);
70 #define MIN_PERCPU_PAGELIST_FRACTION (8)
72 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
73 DEFINE_PER_CPU(int, numa_node
);
74 EXPORT_PER_CPU_SYMBOL(numa_node
);
77 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
79 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
80 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
81 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
82 * defined in <linux/topology.h>.
84 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
85 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
86 int _node_numa_mem_
[MAX_NUMNODES
];
90 * Array of node states.
92 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
93 [N_POSSIBLE
] = NODE_MASK_ALL
,
94 [N_ONLINE
] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY
] = { { [0] = 1UL } },
103 [N_CPU
] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states
);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock
);
111 unsigned long totalram_pages __read_mostly
;
112 unsigned long totalreserve_pages __read_mostly
;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly
;
121 int percpu_pagelist_fraction
;
122 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask
;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex
));
139 if (saved_gfp_mask
) {
140 gfp_allowed_mask
= saved_gfp_mask
;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex
));
148 WARN_ON(saved_gfp_mask
);
149 saved_gfp_mask
= gfp_allowed_mask
;
150 gfp_allowed_mask
&= ~GFP_IOFS
;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly
;
165 static void __free_pages_ok(struct page
*page
, unsigned int order
);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages
);
193 static char * const zone_names
[MAX_NR_ZONES
] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes
= 1024;
208 int user_min_free_kbytes
= -1;
210 static unsigned long __meminitdata nr_kernel_pages
;
211 static unsigned long __meminitdata nr_all_pages
;
212 static unsigned long __meminitdata dma_reserve
;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
217 static unsigned long __initdata required_kernelcore
;
218 static unsigned long __initdata required_movablecore
;
219 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone
);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
228 int nr_online_nodes __read_mostly
= 1;
229 EXPORT_SYMBOL(nr_node_ids
);
230 EXPORT_SYMBOL(nr_online_nodes
);
233 int page_group_by_mobility_disabled __read_mostly
;
235 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
237 if (unlikely(page_group_by_mobility_disabled
&&
238 migratetype
< MIGRATE_PCPTYPES
))
239 migratetype
= MIGRATE_UNMOVABLE
;
241 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
242 PB_migrate
, PB_migrate_end
);
245 bool oom_killer_disabled __read_mostly
;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
252 unsigned long pfn
= page_to_pfn(page
);
253 unsigned long sp
, start_pfn
;
256 seq
= zone_span_seqbegin(zone
);
257 start_pfn
= zone
->zone_start_pfn
;
258 sp
= zone
->spanned_pages
;
259 if (!zone_spans_pfn(zone
, pfn
))
261 } while (zone_span_seqretry(zone
, seq
));
264 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
265 pfn
, zone_to_nid(zone
), zone
->name
,
266 start_pfn
, start_pfn
+ sp
);
271 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
273 if (!pfn_valid_within(page_to_pfn(page
)))
275 if (zone
!= page_zone(page
))
281 * Temporary debugging check for pages not lying within a given zone.
283 static int bad_range(struct zone
*zone
, struct page
*page
)
285 if (page_outside_zone_boundaries(zone
, page
))
287 if (!page_is_consistent(zone
, page
))
293 static inline int bad_range(struct zone
*zone
, struct page
*page
)
299 static void bad_page(struct page
*page
, const char *reason
,
300 unsigned long bad_flags
)
302 static unsigned long resume
;
303 static unsigned long nr_shown
;
304 static unsigned long nr_unshown
;
306 /* Don't complain about poisoned pages */
307 if (PageHWPoison(page
)) {
308 page_mapcount_reset(page
); /* remove PageBuddy */
313 * Allow a burst of 60 reports, then keep quiet for that minute;
314 * or allow a steady drip of one report per second.
316 if (nr_shown
== 60) {
317 if (time_before(jiffies
, resume
)) {
323 "BUG: Bad page state: %lu messages suppressed\n",
330 resume
= jiffies
+ 60 * HZ
;
332 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
333 current
->comm
, page_to_pfn(page
));
334 dump_page_badflags(page
, reason
, bad_flags
);
339 /* Leave bad fields for debug, except PageBuddy could make trouble */
340 page_mapcount_reset(page
); /* remove PageBuddy */
341 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
345 * Higher-order pages are called "compound pages". They are structured thusly:
347 * The first PAGE_SIZE page is called the "head page".
349 * The remaining PAGE_SIZE pages are called "tail pages".
351 * All pages have PG_compound set. All tail pages have their ->first_page
352 * pointing at the head page.
354 * The first tail page's ->lru.next holds the address of the compound page's
355 * put_page() function. Its ->lru.prev holds the order of allocation.
356 * This usage means that zero-order pages may not be compound.
359 static void free_compound_page(struct page
*page
)
361 __free_pages_ok(page
, compound_order(page
));
364 void prep_compound_page(struct page
*page
, unsigned long order
)
367 int nr_pages
= 1 << order
;
369 set_compound_page_dtor(page
, free_compound_page
);
370 set_compound_order(page
, order
);
372 for (i
= 1; i
< nr_pages
; i
++) {
373 struct page
*p
= page
+ i
;
374 set_page_count(p
, 0);
375 p
->first_page
= page
;
376 /* Make sure p->first_page is always valid for PageTail() */
382 /* update __split_huge_page_refcount if you change this function */
383 static int destroy_compound_page(struct page
*page
, unsigned long order
)
386 int nr_pages
= 1 << order
;
389 if (unlikely(compound_order(page
) != order
)) {
390 bad_page(page
, "wrong compound order", 0);
394 __ClearPageHead(page
);
396 for (i
= 1; i
< nr_pages
; i
++) {
397 struct page
*p
= page
+ i
;
399 if (unlikely(!PageTail(p
))) {
400 bad_page(page
, "PageTail not set", 0);
402 } else if (unlikely(p
->first_page
!= page
)) {
403 bad_page(page
, "first_page not consistent", 0);
412 static inline void prep_zero_page(struct page
*page
, unsigned int order
,
418 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
419 * and __GFP_HIGHMEM from hard or soft interrupt context.
421 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
422 for (i
= 0; i
< (1 << order
); i
++)
423 clear_highpage(page
+ i
);
426 #ifdef CONFIG_DEBUG_PAGEALLOC
427 unsigned int _debug_guardpage_minorder
;
429 static int __init
debug_guardpage_minorder_setup(char *buf
)
433 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
434 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
437 _debug_guardpage_minorder
= res
;
438 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
441 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
443 static inline void set_page_guard_flag(struct page
*page
)
445 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
448 static inline void clear_page_guard_flag(struct page
*page
)
450 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
453 static inline void set_page_guard_flag(struct page
*page
) { }
454 static inline void clear_page_guard_flag(struct page
*page
) { }
457 static inline void set_page_order(struct page
*page
, unsigned int order
)
459 set_page_private(page
, order
);
460 __SetPageBuddy(page
);
463 static inline void rmv_page_order(struct page
*page
)
465 __ClearPageBuddy(page
);
466 set_page_private(page
, 0);
470 * This function checks whether a page is free && is the buddy
471 * we can do coalesce a page and its buddy if
472 * (a) the buddy is not in a hole &&
473 * (b) the buddy is in the buddy system &&
474 * (c) a page and its buddy have the same order &&
475 * (d) a page and its buddy are in the same zone.
477 * For recording whether a page is in the buddy system, we set ->_mapcount
478 * PAGE_BUDDY_MAPCOUNT_VALUE.
479 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
480 * serialized by zone->lock.
482 * For recording page's order, we use page_private(page).
484 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
487 if (!pfn_valid_within(page_to_pfn(buddy
)))
490 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
491 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
493 if (page_zone_id(page
) != page_zone_id(buddy
))
499 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
500 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
503 * zone check is done late to avoid uselessly
504 * calculating zone/node ids for pages that could
507 if (page_zone_id(page
) != page_zone_id(buddy
))
516 * Freeing function for a buddy system allocator.
518 * The concept of a buddy system is to maintain direct-mapped table
519 * (containing bit values) for memory blocks of various "orders".
520 * The bottom level table contains the map for the smallest allocatable
521 * units of memory (here, pages), and each level above it describes
522 * pairs of units from the levels below, hence, "buddies".
523 * At a high level, all that happens here is marking the table entry
524 * at the bottom level available, and propagating the changes upward
525 * as necessary, plus some accounting needed to play nicely with other
526 * parts of the VM system.
527 * At each level, we keep a list of pages, which are heads of continuous
528 * free pages of length of (1 << order) and marked with _mapcount
529 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
531 * So when we are allocating or freeing one, we can derive the state of the
532 * other. That is, if we allocate a small block, and both were
533 * free, the remainder of the region must be split into blocks.
534 * If a block is freed, and its buddy is also free, then this
535 * triggers coalescing into a block of larger size.
540 static inline void __free_one_page(struct page
*page
,
542 struct zone
*zone
, unsigned int order
,
545 unsigned long page_idx
;
546 unsigned long combined_idx
;
547 unsigned long uninitialized_var(buddy_idx
);
549 int max_order
= MAX_ORDER
;
551 VM_BUG_ON(!zone_is_initialized(zone
));
553 if (unlikely(PageCompound(page
)))
554 if (unlikely(destroy_compound_page(page
, order
)))
557 VM_BUG_ON(migratetype
== -1);
558 if (is_migrate_isolate(migratetype
)) {
560 * We restrict max order of merging to prevent merge
561 * between freepages on isolate pageblock and normal
562 * pageblock. Without this, pageblock isolation
563 * could cause incorrect freepage accounting.
565 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
567 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
570 page_idx
= pfn
& ((1 << max_order
) - 1);
572 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
573 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
575 while (order
< max_order
- 1) {
576 buddy_idx
= __find_buddy_index(page_idx
, order
);
577 buddy
= page
+ (buddy_idx
- page_idx
);
578 if (!page_is_buddy(page
, buddy
, order
))
581 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
582 * merge with it and move up one order.
584 if (page_is_guard(buddy
)) {
585 clear_page_guard_flag(buddy
);
586 set_page_private(buddy
, 0);
587 if (!is_migrate_isolate(migratetype
)) {
588 __mod_zone_freepage_state(zone
, 1 << order
,
592 list_del(&buddy
->lru
);
593 zone
->free_area
[order
].nr_free
--;
594 rmv_page_order(buddy
);
596 combined_idx
= buddy_idx
& page_idx
;
597 page
= page
+ (combined_idx
- page_idx
);
598 page_idx
= combined_idx
;
601 set_page_order(page
, order
);
604 * If this is not the largest possible page, check if the buddy
605 * of the next-highest order is free. If it is, it's possible
606 * that pages are being freed that will coalesce soon. In case,
607 * that is happening, add the free page to the tail of the list
608 * so it's less likely to be used soon and more likely to be merged
609 * as a higher order page
611 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
612 struct page
*higher_page
, *higher_buddy
;
613 combined_idx
= buddy_idx
& page_idx
;
614 higher_page
= page
+ (combined_idx
- page_idx
);
615 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
616 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
617 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
618 list_add_tail(&page
->lru
,
619 &zone
->free_area
[order
].free_list
[migratetype
]);
624 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
626 zone
->free_area
[order
].nr_free
++;
629 static inline int free_pages_check(struct page
*page
)
631 const char *bad_reason
= NULL
;
632 unsigned long bad_flags
= 0;
634 if (unlikely(page_mapcount(page
)))
635 bad_reason
= "nonzero mapcount";
636 if (unlikely(page
->mapping
!= NULL
))
637 bad_reason
= "non-NULL mapping";
638 if (unlikely(atomic_read(&page
->_count
) != 0))
639 bad_reason
= "nonzero _count";
640 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
641 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
642 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
644 if (unlikely(mem_cgroup_bad_page_check(page
)))
645 bad_reason
= "cgroup check failed";
646 if (unlikely(bad_reason
)) {
647 bad_page(page
, bad_reason
, bad_flags
);
650 page_cpupid_reset_last(page
);
651 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
652 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
657 * Frees a number of pages from the PCP lists
658 * Assumes all pages on list are in same zone, and of same order.
659 * count is the number of pages to free.
661 * If the zone was previously in an "all pages pinned" state then look to
662 * see if this freeing clears that state.
664 * And clear the zone's pages_scanned counter, to hold off the "all pages are
665 * pinned" detection logic.
667 static void free_pcppages_bulk(struct zone
*zone
, int count
,
668 struct per_cpu_pages
*pcp
)
673 unsigned long nr_scanned
;
675 spin_lock(&zone
->lock
);
676 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
678 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
682 struct list_head
*list
;
685 * Remove pages from lists in a round-robin fashion. A
686 * batch_free count is maintained that is incremented when an
687 * empty list is encountered. This is so more pages are freed
688 * off fuller lists instead of spinning excessively around empty
693 if (++migratetype
== MIGRATE_PCPTYPES
)
695 list
= &pcp
->lists
[migratetype
];
696 } while (list_empty(list
));
698 /* This is the only non-empty list. Free them all. */
699 if (batch_free
== MIGRATE_PCPTYPES
)
700 batch_free
= to_free
;
703 int mt
; /* migratetype of the to-be-freed page */
705 page
= list_entry(list
->prev
, struct page
, lru
);
706 /* must delete as __free_one_page list manipulates */
707 list_del(&page
->lru
);
708 mt
= get_freepage_migratetype(page
);
709 if (unlikely(has_isolate_pageblock(zone
)))
710 mt
= get_pageblock_migratetype(page
);
712 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
713 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
714 trace_mm_page_pcpu_drain(page
, 0, mt
);
715 } while (--to_free
&& --batch_free
&& !list_empty(list
));
717 spin_unlock(&zone
->lock
);
720 static void free_one_page(struct zone
*zone
,
721 struct page
*page
, unsigned long pfn
,
725 unsigned long nr_scanned
;
726 spin_lock(&zone
->lock
);
727 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
729 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
731 if (unlikely(has_isolate_pageblock(zone
) ||
732 is_migrate_isolate(migratetype
))) {
733 migratetype
= get_pfnblock_migratetype(page
, pfn
);
735 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
736 spin_unlock(&zone
->lock
);
739 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
744 VM_BUG_ON_PAGE(PageTail(page
), page
);
745 VM_BUG_ON_PAGE(PageHead(page
) && compound_order(page
) != order
, page
);
747 trace_mm_page_free(page
, order
);
748 kmemcheck_free_shadow(page
, order
);
751 page
->mapping
= NULL
;
752 for (i
= 0; i
< (1 << order
); i
++)
753 bad
+= free_pages_check(page
+ i
);
757 if (!PageHighMem(page
)) {
758 debug_check_no_locks_freed(page_address(page
),
760 debug_check_no_obj_freed(page_address(page
),
763 arch_free_page(page
, order
);
764 kernel_map_pages(page
, 1 << order
, 0);
769 static void __free_pages_ok(struct page
*page
, unsigned int order
)
773 unsigned long pfn
= page_to_pfn(page
);
775 if (!free_pages_prepare(page
, order
))
778 migratetype
= get_pfnblock_migratetype(page
, pfn
);
779 local_irq_save(flags
);
780 __count_vm_events(PGFREE
, 1 << order
);
781 set_freepage_migratetype(page
, migratetype
);
782 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
783 local_irq_restore(flags
);
786 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
788 unsigned int nr_pages
= 1 << order
;
789 struct page
*p
= page
;
793 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
795 __ClearPageReserved(p
);
796 set_page_count(p
, 0);
798 __ClearPageReserved(p
);
799 set_page_count(p
, 0);
801 page_zone(page
)->managed_pages
+= nr_pages
;
802 set_page_refcounted(page
);
803 __free_pages(page
, order
);
807 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
808 void __init
init_cma_reserved_pageblock(struct page
*page
)
810 unsigned i
= pageblock_nr_pages
;
811 struct page
*p
= page
;
814 __ClearPageReserved(p
);
815 set_page_count(p
, 0);
818 set_pageblock_migratetype(page
, MIGRATE_CMA
);
820 if (pageblock_order
>= MAX_ORDER
) {
821 i
= pageblock_nr_pages
;
824 set_page_refcounted(p
);
825 __free_pages(p
, MAX_ORDER
- 1);
826 p
+= MAX_ORDER_NR_PAGES
;
827 } while (i
-= MAX_ORDER_NR_PAGES
);
829 set_page_refcounted(page
);
830 __free_pages(page
, pageblock_order
);
833 adjust_managed_page_count(page
, pageblock_nr_pages
);
838 * The order of subdivision here is critical for the IO subsystem.
839 * Please do not alter this order without good reasons and regression
840 * testing. Specifically, as large blocks of memory are subdivided,
841 * the order in which smaller blocks are delivered depends on the order
842 * they're subdivided in this function. This is the primary factor
843 * influencing the order in which pages are delivered to the IO
844 * subsystem according to empirical testing, and this is also justified
845 * by considering the behavior of a buddy system containing a single
846 * large block of memory acted on by a series of small allocations.
847 * This behavior is a critical factor in sglist merging's success.
851 static inline void expand(struct zone
*zone
, struct page
*page
,
852 int low
, int high
, struct free_area
*area
,
855 unsigned long size
= 1 << high
;
861 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
863 #ifdef CONFIG_DEBUG_PAGEALLOC
864 if (high
< debug_guardpage_minorder()) {
866 * Mark as guard pages (or page), that will allow to
867 * merge back to allocator when buddy will be freed.
868 * Corresponding page table entries will not be touched,
869 * pages will stay not present in virtual address space
871 INIT_LIST_HEAD(&page
[size
].lru
);
872 set_page_guard_flag(&page
[size
]);
873 set_page_private(&page
[size
], high
);
874 /* Guard pages are not available for any usage */
875 __mod_zone_freepage_state(zone
, -(1 << high
),
880 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
882 set_page_order(&page
[size
], high
);
887 * This page is about to be returned from the page allocator
889 static inline int check_new_page(struct page
*page
)
891 const char *bad_reason
= NULL
;
892 unsigned long bad_flags
= 0;
894 if (unlikely(page_mapcount(page
)))
895 bad_reason
= "nonzero mapcount";
896 if (unlikely(page
->mapping
!= NULL
))
897 bad_reason
= "non-NULL mapping";
898 if (unlikely(atomic_read(&page
->_count
) != 0))
899 bad_reason
= "nonzero _count";
900 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
901 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
902 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
904 if (unlikely(mem_cgroup_bad_page_check(page
)))
905 bad_reason
= "cgroup check failed";
906 if (unlikely(bad_reason
)) {
907 bad_page(page
, bad_reason
, bad_flags
);
913 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
)
917 for (i
= 0; i
< (1 << order
); i
++) {
918 struct page
*p
= page
+ i
;
919 if (unlikely(check_new_page(p
)))
923 set_page_private(page
, 0);
924 set_page_refcounted(page
);
926 arch_alloc_page(page
, order
);
927 kernel_map_pages(page
, 1 << order
, 1);
929 if (gfp_flags
& __GFP_ZERO
)
930 prep_zero_page(page
, order
, gfp_flags
);
932 if (order
&& (gfp_flags
& __GFP_COMP
))
933 prep_compound_page(page
, order
);
939 * Go through the free lists for the given migratetype and remove
940 * the smallest available page from the freelists
943 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
946 unsigned int current_order
;
947 struct free_area
*area
;
950 /* Find a page of the appropriate size in the preferred list */
951 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
952 area
= &(zone
->free_area
[current_order
]);
953 if (list_empty(&area
->free_list
[migratetype
]))
956 page
= list_entry(area
->free_list
[migratetype
].next
,
958 list_del(&page
->lru
);
959 rmv_page_order(page
);
961 expand(zone
, page
, order
, current_order
, area
, migratetype
);
962 set_freepage_migratetype(page
, migratetype
);
971 * This array describes the order lists are fallen back to when
972 * the free lists for the desirable migrate type are depleted
974 static int fallbacks
[MIGRATE_TYPES
][4] = {
975 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
976 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
978 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
979 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
981 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
983 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
984 #ifdef CONFIG_MEMORY_ISOLATION
985 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
990 * Move the free pages in a range to the free lists of the requested type.
991 * Note that start_page and end_pages are not aligned on a pageblock
992 * boundary. If alignment is required, use move_freepages_block()
994 int move_freepages(struct zone
*zone
,
995 struct page
*start_page
, struct page
*end_page
,
1000 int pages_moved
= 0;
1002 #ifndef CONFIG_HOLES_IN_ZONE
1004 * page_zone is not safe to call in this context when
1005 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1006 * anyway as we check zone boundaries in move_freepages_block().
1007 * Remove at a later date when no bug reports exist related to
1008 * grouping pages by mobility
1010 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1013 for (page
= start_page
; page
<= end_page
;) {
1014 /* Make sure we are not inadvertently changing nodes */
1015 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1017 if (!pfn_valid_within(page_to_pfn(page
))) {
1022 if (!PageBuddy(page
)) {
1027 order
= page_order(page
);
1028 list_move(&page
->lru
,
1029 &zone
->free_area
[order
].free_list
[migratetype
]);
1030 set_freepage_migratetype(page
, migratetype
);
1032 pages_moved
+= 1 << order
;
1038 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1041 unsigned long start_pfn
, end_pfn
;
1042 struct page
*start_page
, *end_page
;
1044 start_pfn
= page_to_pfn(page
);
1045 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1046 start_page
= pfn_to_page(start_pfn
);
1047 end_page
= start_page
+ pageblock_nr_pages
- 1;
1048 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1050 /* Do not cross zone boundaries */
1051 if (!zone_spans_pfn(zone
, start_pfn
))
1053 if (!zone_spans_pfn(zone
, end_pfn
))
1056 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1059 static void change_pageblock_range(struct page
*pageblock_page
,
1060 int start_order
, int migratetype
)
1062 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1064 while (nr_pageblocks
--) {
1065 set_pageblock_migratetype(pageblock_page
, migratetype
);
1066 pageblock_page
+= pageblock_nr_pages
;
1071 * If breaking a large block of pages, move all free pages to the preferred
1072 * allocation list. If falling back for a reclaimable kernel allocation, be
1073 * more aggressive about taking ownership of free pages.
1075 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1076 * nor move CMA pages to different free lists. We don't want unmovable pages
1077 * to be allocated from MIGRATE_CMA areas.
1079 * Returns the new migratetype of the pageblock (or the same old migratetype
1080 * if it was unchanged).
1082 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1083 int start_type
, int fallback_type
)
1085 int current_order
= page_order(page
);
1088 * When borrowing from MIGRATE_CMA, we need to release the excess
1089 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1090 * is set to CMA so it is returned to the correct freelist in case
1091 * the page ends up being not actually allocated from the pcp lists.
1093 if (is_migrate_cma(fallback_type
))
1094 return fallback_type
;
1096 /* Take ownership for orders >= pageblock_order */
1097 if (current_order
>= pageblock_order
) {
1098 change_pageblock_range(page
, current_order
, start_type
);
1102 if (current_order
>= pageblock_order
/ 2 ||
1103 start_type
== MIGRATE_RECLAIMABLE
||
1104 page_group_by_mobility_disabled
) {
1107 pages
= move_freepages_block(zone
, page
, start_type
);
1109 /* Claim the whole block if over half of it is free */
1110 if (pages
>= (1 << (pageblock_order
-1)) ||
1111 page_group_by_mobility_disabled
) {
1113 set_pageblock_migratetype(page
, start_type
);
1119 return fallback_type
;
1122 /* Remove an element from the buddy allocator from the fallback list */
1123 static inline struct page
*
1124 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1126 struct free_area
*area
;
1127 unsigned int current_order
;
1129 int migratetype
, new_type
, i
;
1131 /* Find the largest possible block of pages in the other list */
1132 for (current_order
= MAX_ORDER
-1;
1133 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1136 migratetype
= fallbacks
[start_migratetype
][i
];
1138 /* MIGRATE_RESERVE handled later if necessary */
1139 if (migratetype
== MIGRATE_RESERVE
)
1142 area
= &(zone
->free_area
[current_order
]);
1143 if (list_empty(&area
->free_list
[migratetype
]))
1146 page
= list_entry(area
->free_list
[migratetype
].next
,
1150 new_type
= try_to_steal_freepages(zone
, page
,
1154 /* Remove the page from the freelists */
1155 list_del(&page
->lru
);
1156 rmv_page_order(page
);
1158 expand(zone
, page
, order
, current_order
, area
,
1160 /* The freepage_migratetype may differ from pageblock's
1161 * migratetype depending on the decisions in
1162 * try_to_steal_freepages. This is OK as long as it does
1163 * not differ for MIGRATE_CMA type.
1165 set_freepage_migratetype(page
, new_type
);
1167 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1168 start_migratetype
, migratetype
, new_type
);
1178 * Do the hard work of removing an element from the buddy allocator.
1179 * Call me with the zone->lock already held.
1181 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1187 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1189 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1190 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1193 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1194 * is used because __rmqueue_smallest is an inline function
1195 * and we want just one call site
1198 migratetype
= MIGRATE_RESERVE
;
1203 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1208 * Obtain a specified number of elements from the buddy allocator, all under
1209 * a single hold of the lock, for efficiency. Add them to the supplied list.
1210 * Returns the number of new pages which were placed at *list.
1212 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1213 unsigned long count
, struct list_head
*list
,
1214 int migratetype
, bool cold
)
1218 spin_lock(&zone
->lock
);
1219 for (i
= 0; i
< count
; ++i
) {
1220 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1221 if (unlikely(page
== NULL
))
1225 * Split buddy pages returned by expand() are received here
1226 * in physical page order. The page is added to the callers and
1227 * list and the list head then moves forward. From the callers
1228 * perspective, the linked list is ordered by page number in
1229 * some conditions. This is useful for IO devices that can
1230 * merge IO requests if the physical pages are ordered
1234 list_add(&page
->lru
, list
);
1236 list_add_tail(&page
->lru
, list
);
1238 if (is_migrate_cma(get_freepage_migratetype(page
)))
1239 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1242 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1243 spin_unlock(&zone
->lock
);
1249 * Called from the vmstat counter updater to drain pagesets of this
1250 * currently executing processor on remote nodes after they have
1253 * Note that this function must be called with the thread pinned to
1254 * a single processor.
1256 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1258 unsigned long flags
;
1259 int to_drain
, batch
;
1261 local_irq_save(flags
);
1262 batch
= ACCESS_ONCE(pcp
->batch
);
1263 to_drain
= min(pcp
->count
, batch
);
1265 free_pcppages_bulk(zone
, to_drain
, pcp
);
1266 pcp
->count
-= to_drain
;
1268 local_irq_restore(flags
);
1273 * Drain pcplists of the indicated processor and zone.
1275 * The processor must either be the current processor and the
1276 * thread pinned to the current processor or a processor that
1279 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1281 unsigned long flags
;
1282 struct per_cpu_pageset
*pset
;
1283 struct per_cpu_pages
*pcp
;
1285 local_irq_save(flags
);
1286 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1290 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1293 local_irq_restore(flags
);
1297 * Drain pcplists of all zones on the indicated processor.
1299 * The processor must either be the current processor and the
1300 * thread pinned to the current processor or a processor that
1303 static void drain_pages(unsigned int cpu
)
1307 for_each_populated_zone(zone
) {
1308 drain_pages_zone(cpu
, zone
);
1313 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1315 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1316 * the single zone's pages.
1318 void drain_local_pages(struct zone
*zone
)
1320 int cpu
= smp_processor_id();
1323 drain_pages_zone(cpu
, zone
);
1329 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1331 * When zone parameter is non-NULL, spill just the single zone's pages.
1333 * Note that this code is protected against sending an IPI to an offline
1334 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1335 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1336 * nothing keeps CPUs from showing up after we populated the cpumask and
1337 * before the call to on_each_cpu_mask().
1339 void drain_all_pages(struct zone
*zone
)
1344 * Allocate in the BSS so we wont require allocation in
1345 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1347 static cpumask_t cpus_with_pcps
;
1350 * We don't care about racing with CPU hotplug event
1351 * as offline notification will cause the notified
1352 * cpu to drain that CPU pcps and on_each_cpu_mask
1353 * disables preemption as part of its processing
1355 for_each_online_cpu(cpu
) {
1356 struct per_cpu_pageset
*pcp
;
1358 bool has_pcps
= false;
1361 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1365 for_each_populated_zone(z
) {
1366 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1367 if (pcp
->pcp
.count
) {
1375 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1377 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1379 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1383 #ifdef CONFIG_HIBERNATION
1385 void mark_free_pages(struct zone
*zone
)
1387 unsigned long pfn
, max_zone_pfn
;
1388 unsigned long flags
;
1389 unsigned int order
, t
;
1390 struct list_head
*curr
;
1392 if (zone_is_empty(zone
))
1395 spin_lock_irqsave(&zone
->lock
, flags
);
1397 max_zone_pfn
= zone_end_pfn(zone
);
1398 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1399 if (pfn_valid(pfn
)) {
1400 struct page
*page
= pfn_to_page(pfn
);
1402 if (!swsusp_page_is_forbidden(page
))
1403 swsusp_unset_page_free(page
);
1406 for_each_migratetype_order(order
, t
) {
1407 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1410 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1411 for (i
= 0; i
< (1UL << order
); i
++)
1412 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1415 spin_unlock_irqrestore(&zone
->lock
, flags
);
1417 #endif /* CONFIG_PM */
1420 * Free a 0-order page
1421 * cold == true ? free a cold page : free a hot page
1423 void free_hot_cold_page(struct page
*page
, bool cold
)
1425 struct zone
*zone
= page_zone(page
);
1426 struct per_cpu_pages
*pcp
;
1427 unsigned long flags
;
1428 unsigned long pfn
= page_to_pfn(page
);
1431 if (!free_pages_prepare(page
, 0))
1434 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1435 set_freepage_migratetype(page
, migratetype
);
1436 local_irq_save(flags
);
1437 __count_vm_event(PGFREE
);
1440 * We only track unmovable, reclaimable and movable on pcp lists.
1441 * Free ISOLATE pages back to the allocator because they are being
1442 * offlined but treat RESERVE as movable pages so we can get those
1443 * areas back if necessary. Otherwise, we may have to free
1444 * excessively into the page allocator
1446 if (migratetype
>= MIGRATE_PCPTYPES
) {
1447 if (unlikely(is_migrate_isolate(migratetype
))) {
1448 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1451 migratetype
= MIGRATE_MOVABLE
;
1454 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1456 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1458 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1460 if (pcp
->count
>= pcp
->high
) {
1461 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1462 free_pcppages_bulk(zone
, batch
, pcp
);
1463 pcp
->count
-= batch
;
1467 local_irq_restore(flags
);
1471 * Free a list of 0-order pages
1473 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1475 struct page
*page
, *next
;
1477 list_for_each_entry_safe(page
, next
, list
, lru
) {
1478 trace_mm_page_free_batched(page
, cold
);
1479 free_hot_cold_page(page
, cold
);
1484 * split_page takes a non-compound higher-order page, and splits it into
1485 * n (1<<order) sub-pages: page[0..n]
1486 * Each sub-page must be freed individually.
1488 * Note: this is probably too low level an operation for use in drivers.
1489 * Please consult with lkml before using this in your driver.
1491 void split_page(struct page
*page
, unsigned int order
)
1495 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1496 VM_BUG_ON_PAGE(!page_count(page
), page
);
1498 #ifdef CONFIG_KMEMCHECK
1500 * Split shadow pages too, because free(page[0]) would
1501 * otherwise free the whole shadow.
1503 if (kmemcheck_page_is_tracked(page
))
1504 split_page(virt_to_page(page
[0].shadow
), order
);
1507 for (i
= 1; i
< (1 << order
); i
++)
1508 set_page_refcounted(page
+ i
);
1510 EXPORT_SYMBOL_GPL(split_page
);
1512 int __isolate_free_page(struct page
*page
, unsigned int order
)
1514 unsigned long watermark
;
1518 BUG_ON(!PageBuddy(page
));
1520 zone
= page_zone(page
);
1521 mt
= get_pageblock_migratetype(page
);
1523 if (!is_migrate_isolate(mt
)) {
1524 /* Obey watermarks as if the page was being allocated */
1525 watermark
= low_wmark_pages(zone
) + (1 << order
);
1526 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1529 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1532 /* Remove page from free list */
1533 list_del(&page
->lru
);
1534 zone
->free_area
[order
].nr_free
--;
1535 rmv_page_order(page
);
1537 /* Set the pageblock if the isolated page is at least a pageblock */
1538 if (order
>= pageblock_order
- 1) {
1539 struct page
*endpage
= page
+ (1 << order
) - 1;
1540 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1541 int mt
= get_pageblock_migratetype(page
);
1542 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1543 set_pageblock_migratetype(page
,
1548 return 1UL << order
;
1552 * Similar to split_page except the page is already free. As this is only
1553 * being used for migration, the migratetype of the block also changes.
1554 * As this is called with interrupts disabled, the caller is responsible
1555 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1558 * Note: this is probably too low level an operation for use in drivers.
1559 * Please consult with lkml before using this in your driver.
1561 int split_free_page(struct page
*page
)
1566 order
= page_order(page
);
1568 nr_pages
= __isolate_free_page(page
, order
);
1572 /* Split into individual pages */
1573 set_page_refcounted(page
);
1574 split_page(page
, order
);
1579 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1580 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1584 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1585 struct zone
*zone
, unsigned int order
,
1586 gfp_t gfp_flags
, int migratetype
)
1588 unsigned long flags
;
1590 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1593 if (likely(order
== 0)) {
1594 struct per_cpu_pages
*pcp
;
1595 struct list_head
*list
;
1597 local_irq_save(flags
);
1598 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1599 list
= &pcp
->lists
[migratetype
];
1600 if (list_empty(list
)) {
1601 pcp
->count
+= rmqueue_bulk(zone
, 0,
1604 if (unlikely(list_empty(list
)))
1609 page
= list_entry(list
->prev
, struct page
, lru
);
1611 page
= list_entry(list
->next
, struct page
, lru
);
1613 list_del(&page
->lru
);
1616 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1618 * __GFP_NOFAIL is not to be used in new code.
1620 * All __GFP_NOFAIL callers should be fixed so that they
1621 * properly detect and handle allocation failures.
1623 * We most definitely don't want callers attempting to
1624 * allocate greater than order-1 page units with
1627 WARN_ON_ONCE(order
> 1);
1629 spin_lock_irqsave(&zone
->lock
, flags
);
1630 page
= __rmqueue(zone
, order
, migratetype
);
1631 spin_unlock(&zone
->lock
);
1634 __mod_zone_freepage_state(zone
, -(1 << order
),
1635 get_freepage_migratetype(page
));
1638 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1639 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1640 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1641 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1643 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1644 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1645 local_irq_restore(flags
);
1647 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1648 if (prep_new_page(page
, order
, gfp_flags
))
1653 local_irq_restore(flags
);
1657 #ifdef CONFIG_FAIL_PAGE_ALLOC
1660 struct fault_attr attr
;
1662 u32 ignore_gfp_highmem
;
1663 u32 ignore_gfp_wait
;
1665 } fail_page_alloc
= {
1666 .attr
= FAULT_ATTR_INITIALIZER
,
1667 .ignore_gfp_wait
= 1,
1668 .ignore_gfp_highmem
= 1,
1672 static int __init
setup_fail_page_alloc(char *str
)
1674 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1676 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1678 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1680 if (order
< fail_page_alloc
.min_order
)
1682 if (gfp_mask
& __GFP_NOFAIL
)
1684 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1686 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1689 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1692 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1694 static int __init
fail_page_alloc_debugfs(void)
1696 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1699 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1700 &fail_page_alloc
.attr
);
1702 return PTR_ERR(dir
);
1704 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1705 &fail_page_alloc
.ignore_gfp_wait
))
1707 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1708 &fail_page_alloc
.ignore_gfp_highmem
))
1710 if (!debugfs_create_u32("min-order", mode
, dir
,
1711 &fail_page_alloc
.min_order
))
1716 debugfs_remove_recursive(dir
);
1721 late_initcall(fail_page_alloc_debugfs
);
1723 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1725 #else /* CONFIG_FAIL_PAGE_ALLOC */
1727 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1732 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1735 * Return true if free pages are above 'mark'. This takes into account the order
1736 * of the allocation.
1738 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1739 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1742 /* free_pages may go negative - that's OK */
1747 free_pages
-= (1 << order
) - 1;
1748 if (alloc_flags
& ALLOC_HIGH
)
1750 if (alloc_flags
& ALLOC_HARDER
)
1753 /* If allocation can't use CMA areas don't use free CMA pages */
1754 if (!(alloc_flags
& ALLOC_CMA
))
1755 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1758 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1760 for (o
= 0; o
< order
; o
++) {
1761 /* At the next order, this order's pages become unavailable */
1762 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1764 /* Require fewer higher order pages to be free */
1767 if (free_pages
<= min
)
1773 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1774 int classzone_idx
, int alloc_flags
)
1776 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1777 zone_page_state(z
, NR_FREE_PAGES
));
1780 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1781 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1783 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1785 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1786 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1788 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1794 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1795 * skip over zones that are not allowed by the cpuset, or that have
1796 * been recently (in last second) found to be nearly full. See further
1797 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1798 * that have to skip over a lot of full or unallowed zones.
1800 * If the zonelist cache is present in the passed zonelist, then
1801 * returns a pointer to the allowed node mask (either the current
1802 * tasks mems_allowed, or node_states[N_MEMORY].)
1804 * If the zonelist cache is not available for this zonelist, does
1805 * nothing and returns NULL.
1807 * If the fullzones BITMAP in the zonelist cache is stale (more than
1808 * a second since last zap'd) then we zap it out (clear its bits.)
1810 * We hold off even calling zlc_setup, until after we've checked the
1811 * first zone in the zonelist, on the theory that most allocations will
1812 * be satisfied from that first zone, so best to examine that zone as
1813 * quickly as we can.
1815 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1817 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1818 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1820 zlc
= zonelist
->zlcache_ptr
;
1824 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1825 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1826 zlc
->last_full_zap
= jiffies
;
1829 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1830 &cpuset_current_mems_allowed
:
1831 &node_states
[N_MEMORY
];
1832 return allowednodes
;
1836 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1837 * if it is worth looking at further for free memory:
1838 * 1) Check that the zone isn't thought to be full (doesn't have its
1839 * bit set in the zonelist_cache fullzones BITMAP).
1840 * 2) Check that the zones node (obtained from the zonelist_cache
1841 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1842 * Return true (non-zero) if zone is worth looking at further, or
1843 * else return false (zero) if it is not.
1845 * This check -ignores- the distinction between various watermarks,
1846 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1847 * found to be full for any variation of these watermarks, it will
1848 * be considered full for up to one second by all requests, unless
1849 * we are so low on memory on all allowed nodes that we are forced
1850 * into the second scan of the zonelist.
1852 * In the second scan we ignore this zonelist cache and exactly
1853 * apply the watermarks to all zones, even it is slower to do so.
1854 * We are low on memory in the second scan, and should leave no stone
1855 * unturned looking for a free page.
1857 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1858 nodemask_t
*allowednodes
)
1860 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1861 int i
; /* index of *z in zonelist zones */
1862 int n
; /* node that zone *z is on */
1864 zlc
= zonelist
->zlcache_ptr
;
1868 i
= z
- zonelist
->_zonerefs
;
1871 /* This zone is worth trying if it is allowed but not full */
1872 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1876 * Given 'z' scanning a zonelist, set the corresponding bit in
1877 * zlc->fullzones, so that subsequent attempts to allocate a page
1878 * from that zone don't waste time re-examining it.
1880 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1882 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1883 int i
; /* index of *z in zonelist zones */
1885 zlc
= zonelist
->zlcache_ptr
;
1889 i
= z
- zonelist
->_zonerefs
;
1891 set_bit(i
, zlc
->fullzones
);
1895 * clear all zones full, called after direct reclaim makes progress so that
1896 * a zone that was recently full is not skipped over for up to a second
1898 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1900 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1902 zlc
= zonelist
->zlcache_ptr
;
1906 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1909 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1911 return local_zone
->node
== zone
->node
;
1914 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1916 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1920 #else /* CONFIG_NUMA */
1922 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1927 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1928 nodemask_t
*allowednodes
)
1933 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1937 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1941 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1946 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1951 #endif /* CONFIG_NUMA */
1953 static void reset_alloc_batches(struct zone
*preferred_zone
)
1955 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
1958 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
1959 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
1960 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
1961 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1962 } while (zone
++ != preferred_zone
);
1966 * get_page_from_freelist goes through the zonelist trying to allocate
1969 static struct page
*
1970 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1971 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1972 struct zone
*preferred_zone
, int classzone_idx
, int migratetype
)
1975 struct page
*page
= NULL
;
1977 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1978 int zlc_active
= 0; /* set if using zonelist_cache */
1979 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1980 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
1981 (gfp_mask
& __GFP_WRITE
);
1982 int nr_fair_skipped
= 0;
1983 bool zonelist_rescan
;
1986 zonelist_rescan
= false;
1989 * Scan zonelist, looking for a zone with enough free.
1990 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1992 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1993 high_zoneidx
, nodemask
) {
1996 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1997 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1999 if (cpusets_enabled() &&
2000 (alloc_flags
& ALLOC_CPUSET
) &&
2001 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
2004 * Distribute pages in proportion to the individual
2005 * zone size to ensure fair page aging. The zone a
2006 * page was allocated in should have no effect on the
2007 * time the page has in memory before being reclaimed.
2009 if (alloc_flags
& ALLOC_FAIR
) {
2010 if (!zone_local(preferred_zone
, zone
))
2012 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2018 * When allocating a page cache page for writing, we
2019 * want to get it from a zone that is within its dirty
2020 * limit, such that no single zone holds more than its
2021 * proportional share of globally allowed dirty pages.
2022 * The dirty limits take into account the zone's
2023 * lowmem reserves and high watermark so that kswapd
2024 * should be able to balance it without having to
2025 * write pages from its LRU list.
2027 * This may look like it could increase pressure on
2028 * lower zones by failing allocations in higher zones
2029 * before they are full. But the pages that do spill
2030 * over are limited as the lower zones are protected
2031 * by this very same mechanism. It should not become
2032 * a practical burden to them.
2034 * XXX: For now, allow allocations to potentially
2035 * exceed the per-zone dirty limit in the slowpath
2036 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2037 * which is important when on a NUMA setup the allowed
2038 * zones are together not big enough to reach the
2039 * global limit. The proper fix for these situations
2040 * will require awareness of zones in the
2041 * dirty-throttling and the flusher threads.
2043 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2046 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2047 if (!zone_watermark_ok(zone
, order
, mark
,
2048 classzone_idx
, alloc_flags
)) {
2051 /* Checked here to keep the fast path fast */
2052 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2053 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2056 if (IS_ENABLED(CONFIG_NUMA
) &&
2057 !did_zlc_setup
&& nr_online_nodes
> 1) {
2059 * we do zlc_setup if there are multiple nodes
2060 * and before considering the first zone allowed
2063 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2068 if (zone_reclaim_mode
== 0 ||
2069 !zone_allows_reclaim(preferred_zone
, zone
))
2070 goto this_zone_full
;
2073 * As we may have just activated ZLC, check if the first
2074 * eligible zone has failed zone_reclaim recently.
2076 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2077 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2080 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2082 case ZONE_RECLAIM_NOSCAN
:
2085 case ZONE_RECLAIM_FULL
:
2086 /* scanned but unreclaimable */
2089 /* did we reclaim enough */
2090 if (zone_watermark_ok(zone
, order
, mark
,
2091 classzone_idx
, alloc_flags
))
2095 * Failed to reclaim enough to meet watermark.
2096 * Only mark the zone full if checking the min
2097 * watermark or if we failed to reclaim just
2098 * 1<<order pages or else the page allocator
2099 * fastpath will prematurely mark zones full
2100 * when the watermark is between the low and
2103 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2104 ret
== ZONE_RECLAIM_SOME
)
2105 goto this_zone_full
;
2112 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2113 gfp_mask
, migratetype
);
2117 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2118 zlc_mark_zone_full(zonelist
, z
);
2123 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2124 * necessary to allocate the page. The expectation is
2125 * that the caller is taking steps that will free more
2126 * memory. The caller should avoid the page being used
2127 * for !PFMEMALLOC purposes.
2129 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2134 * The first pass makes sure allocations are spread fairly within the
2135 * local node. However, the local node might have free pages left
2136 * after the fairness batches are exhausted, and remote zones haven't
2137 * even been considered yet. Try once more without fairness, and
2138 * include remote zones now, before entering the slowpath and waking
2139 * kswapd: prefer spilling to a remote zone over swapping locally.
2141 if (alloc_flags
& ALLOC_FAIR
) {
2142 alloc_flags
&= ~ALLOC_FAIR
;
2143 if (nr_fair_skipped
) {
2144 zonelist_rescan
= true;
2145 reset_alloc_batches(preferred_zone
);
2147 if (nr_online_nodes
> 1)
2148 zonelist_rescan
= true;
2151 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2152 /* Disable zlc cache for second zonelist scan */
2154 zonelist_rescan
= true;
2157 if (zonelist_rescan
)
2164 * Large machines with many possible nodes should not always dump per-node
2165 * meminfo in irq context.
2167 static inline bool should_suppress_show_mem(void)
2172 ret
= in_interrupt();
2177 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2178 DEFAULT_RATELIMIT_INTERVAL
,
2179 DEFAULT_RATELIMIT_BURST
);
2181 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2183 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2185 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2186 debug_guardpage_minorder() > 0)
2190 * This documents exceptions given to allocations in certain
2191 * contexts that are allowed to allocate outside current's set
2194 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2195 if (test_thread_flag(TIF_MEMDIE
) ||
2196 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2197 filter
&= ~SHOW_MEM_FILTER_NODES
;
2198 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2199 filter
&= ~SHOW_MEM_FILTER_NODES
;
2202 struct va_format vaf
;
2205 va_start(args
, fmt
);
2210 pr_warn("%pV", &vaf
);
2215 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2216 current
->comm
, order
, gfp_mask
);
2219 if (!should_suppress_show_mem())
2224 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2225 unsigned long did_some_progress
,
2226 unsigned long pages_reclaimed
)
2228 /* Do not loop if specifically requested */
2229 if (gfp_mask
& __GFP_NORETRY
)
2232 /* Always retry if specifically requested */
2233 if (gfp_mask
& __GFP_NOFAIL
)
2237 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2238 * making forward progress without invoking OOM. Suspend also disables
2239 * storage devices so kswapd will not help. Bail if we are suspending.
2241 if (!did_some_progress
&& pm_suspended_storage())
2245 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2246 * means __GFP_NOFAIL, but that may not be true in other
2249 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2253 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2254 * specified, then we retry until we no longer reclaim any pages
2255 * (above), or we've reclaimed an order of pages at least as
2256 * large as the allocation's order. In both cases, if the
2257 * allocation still fails, we stop retrying.
2259 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2265 static inline struct page
*
2266 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2267 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2268 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2269 int classzone_idx
, int migratetype
)
2273 /* Acquire the per-zone oom lock for each zone */
2274 if (!oom_zonelist_trylock(zonelist
, gfp_mask
)) {
2275 schedule_timeout_uninterruptible(1);
2280 * PM-freezer should be notified that there might be an OOM killer on
2281 * its way to kill and wake somebody up. This is too early and we might
2282 * end up not killing anything but false positives are acceptable.
2283 * See freeze_processes.
2288 * Go through the zonelist yet one more time, keep very high watermark
2289 * here, this is only to catch a parallel oom killing, we must fail if
2290 * we're still under heavy pressure.
2292 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2293 order
, zonelist
, high_zoneidx
,
2294 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2295 preferred_zone
, classzone_idx
, migratetype
);
2299 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2300 /* The OOM killer will not help higher order allocs */
2301 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2303 /* The OOM killer does not needlessly kill tasks for lowmem */
2304 if (high_zoneidx
< ZONE_NORMAL
)
2307 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2308 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2309 * The caller should handle page allocation failure by itself if
2310 * it specifies __GFP_THISNODE.
2311 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2313 if (gfp_mask
& __GFP_THISNODE
)
2316 /* Exhausted what can be done so it's blamo time */
2317 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2320 oom_zonelist_unlock(zonelist
, gfp_mask
);
2324 #ifdef CONFIG_COMPACTION
2325 /* Try memory compaction for high-order allocations before reclaim */
2326 static struct page
*
2327 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2328 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2329 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2330 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2331 int *contended_compaction
, bool *deferred_compaction
)
2333 unsigned long compact_result
;
2339 current
->flags
|= PF_MEMALLOC
;
2340 compact_result
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2342 contended_compaction
,
2343 alloc_flags
, classzone_idx
);
2344 current
->flags
&= ~PF_MEMALLOC
;
2346 switch (compact_result
) {
2347 case COMPACT_DEFERRED
:
2348 *deferred_compaction
= true;
2350 case COMPACT_SKIPPED
:
2357 * At least in one zone compaction wasn't deferred or skipped, so let's
2358 * count a compaction stall
2360 count_vm_event(COMPACTSTALL
);
2362 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2363 order
, zonelist
, high_zoneidx
,
2364 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2365 preferred_zone
, classzone_idx
, migratetype
);
2368 struct zone
*zone
= page_zone(page
);
2370 zone
->compact_blockskip_flush
= false;
2371 compaction_defer_reset(zone
, order
, true);
2372 count_vm_event(COMPACTSUCCESS
);
2377 * It's bad if compaction run occurs and fails. The most likely reason
2378 * is that pages exist, but not enough to satisfy watermarks.
2380 count_vm_event(COMPACTFAIL
);
2387 static inline struct page
*
2388 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2389 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2390 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2391 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2392 int *contended_compaction
, bool *deferred_compaction
)
2396 #endif /* CONFIG_COMPACTION */
2398 /* Perform direct synchronous page reclaim */
2400 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2401 nodemask_t
*nodemask
)
2403 struct reclaim_state reclaim_state
;
2408 /* We now go into synchronous reclaim */
2409 cpuset_memory_pressure_bump();
2410 current
->flags
|= PF_MEMALLOC
;
2411 lockdep_set_current_reclaim_state(gfp_mask
);
2412 reclaim_state
.reclaimed_slab
= 0;
2413 current
->reclaim_state
= &reclaim_state
;
2415 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2417 current
->reclaim_state
= NULL
;
2418 lockdep_clear_current_reclaim_state();
2419 current
->flags
&= ~PF_MEMALLOC
;
2426 /* The really slow allocator path where we enter direct reclaim */
2427 static inline struct page
*
2428 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2429 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2430 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2431 int classzone_idx
, int migratetype
, unsigned long *did_some_progress
)
2433 struct page
*page
= NULL
;
2434 bool drained
= false;
2436 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2438 if (unlikely(!(*did_some_progress
)))
2441 /* After successful reclaim, reconsider all zones for allocation */
2442 if (IS_ENABLED(CONFIG_NUMA
))
2443 zlc_clear_zones_full(zonelist
);
2446 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2447 zonelist
, high_zoneidx
,
2448 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2449 preferred_zone
, classzone_idx
,
2453 * If an allocation failed after direct reclaim, it could be because
2454 * pages are pinned on the per-cpu lists. Drain them and try again
2456 if (!page
&& !drained
) {
2457 drain_all_pages(NULL
);
2466 * This is called in the allocator slow-path if the allocation request is of
2467 * sufficient urgency to ignore watermarks and take other desperate measures
2469 static inline struct page
*
2470 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2471 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2472 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2473 int classzone_idx
, int migratetype
)
2478 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2479 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2480 preferred_zone
, classzone_idx
, migratetype
);
2482 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2483 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2484 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2489 static void wake_all_kswapds(unsigned int order
,
2490 struct zonelist
*zonelist
,
2491 enum zone_type high_zoneidx
,
2492 struct zone
*preferred_zone
,
2493 nodemask_t
*nodemask
)
2498 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2499 high_zoneidx
, nodemask
)
2500 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2504 gfp_to_alloc_flags(gfp_t gfp_mask
)
2506 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2507 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2509 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2510 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2513 * The caller may dip into page reserves a bit more if the caller
2514 * cannot run direct reclaim, or if the caller has realtime scheduling
2515 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2516 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2518 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2522 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2523 * if it can't schedule.
2525 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2526 alloc_flags
|= ALLOC_HARDER
;
2528 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2529 * comment for __cpuset_node_allowed_softwall().
2531 alloc_flags
&= ~ALLOC_CPUSET
;
2532 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2533 alloc_flags
|= ALLOC_HARDER
;
2535 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2536 if (gfp_mask
& __GFP_MEMALLOC
)
2537 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2538 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2539 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2540 else if (!in_interrupt() &&
2541 ((current
->flags
& PF_MEMALLOC
) ||
2542 unlikely(test_thread_flag(TIF_MEMDIE
))))
2543 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2546 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2547 alloc_flags
|= ALLOC_CMA
;
2552 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2554 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2557 static inline struct page
*
2558 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2559 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2560 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2561 int classzone_idx
, int migratetype
)
2563 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2564 struct page
*page
= NULL
;
2566 unsigned long pages_reclaimed
= 0;
2567 unsigned long did_some_progress
;
2568 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2569 bool deferred_compaction
= false;
2570 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2573 * In the slowpath, we sanity check order to avoid ever trying to
2574 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2575 * be using allocators in order of preference for an area that is
2578 if (order
>= MAX_ORDER
) {
2579 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2584 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2585 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2586 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2587 * using a larger set of nodes after it has established that the
2588 * allowed per node queues are empty and that nodes are
2591 if (IS_ENABLED(CONFIG_NUMA
) &&
2592 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2596 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2597 wake_all_kswapds(order
, zonelist
, high_zoneidx
,
2598 preferred_zone
, nodemask
);
2601 * OK, we're below the kswapd watermark and have kicked background
2602 * reclaim. Now things get more complex, so set up alloc_flags according
2603 * to how we want to proceed.
2605 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2608 * Find the true preferred zone if the allocation is unconstrained by
2611 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
) {
2612 struct zoneref
*preferred_zoneref
;
2613 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2614 NULL
, &preferred_zone
);
2615 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2619 /* This is the last chance, in general, before the goto nopage. */
2620 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2621 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2622 preferred_zone
, classzone_idx
, migratetype
);
2626 /* Allocate without watermarks if the context allows */
2627 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2629 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2630 * the allocation is high priority and these type of
2631 * allocations are system rather than user orientated
2633 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2635 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2636 zonelist
, high_zoneidx
, nodemask
,
2637 preferred_zone
, classzone_idx
, migratetype
);
2643 /* Atomic allocations - we can't balance anything */
2646 * All existing users of the deprecated __GFP_NOFAIL are
2647 * blockable, so warn of any new users that actually allow this
2648 * type of allocation to fail.
2650 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2654 /* Avoid recursion of direct reclaim */
2655 if (current
->flags
& PF_MEMALLOC
)
2658 /* Avoid allocations with no watermarks from looping endlessly */
2659 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2663 * Try direct compaction. The first pass is asynchronous. Subsequent
2664 * attempts after direct reclaim are synchronous
2666 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2667 high_zoneidx
, nodemask
, alloc_flags
,
2669 classzone_idx
, migratetype
,
2670 migration_mode
, &contended_compaction
,
2671 &deferred_compaction
);
2675 /* Checks for THP-specific high-order allocations */
2676 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2678 * If compaction is deferred for high-order allocations, it is
2679 * because sync compaction recently failed. If this is the case
2680 * and the caller requested a THP allocation, we do not want
2681 * to heavily disrupt the system, so we fail the allocation
2682 * instead of entering direct reclaim.
2684 if (deferred_compaction
)
2688 * In all zones where compaction was attempted (and not
2689 * deferred or skipped), lock contention has been detected.
2690 * For THP allocation we do not want to disrupt the others
2691 * so we fallback to base pages instead.
2693 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2697 * If compaction was aborted due to need_resched(), we do not
2698 * want to further increase allocation latency, unless it is
2699 * khugepaged trying to collapse.
2701 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2702 && !(current
->flags
& PF_KTHREAD
))
2707 * It can become very expensive to allocate transparent hugepages at
2708 * fault, so use asynchronous memory compaction for THP unless it is
2709 * khugepaged trying to collapse.
2711 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2712 (current
->flags
& PF_KTHREAD
))
2713 migration_mode
= MIGRATE_SYNC_LIGHT
;
2715 /* Try direct reclaim and then allocating */
2716 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2717 zonelist
, high_zoneidx
,
2719 alloc_flags
, preferred_zone
,
2720 classzone_idx
, migratetype
,
2721 &did_some_progress
);
2726 * If we failed to make any progress reclaiming, then we are
2727 * running out of options and have to consider going OOM
2729 if (!did_some_progress
) {
2730 if (oom_gfp_allowed(gfp_mask
)) {
2731 if (oom_killer_disabled
)
2733 /* Coredumps can quickly deplete all memory reserves */
2734 if ((current
->flags
& PF_DUMPCORE
) &&
2735 !(gfp_mask
& __GFP_NOFAIL
))
2737 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2738 zonelist
, high_zoneidx
,
2739 nodemask
, preferred_zone
,
2740 classzone_idx
, migratetype
);
2744 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2746 * The oom killer is not called for high-order
2747 * allocations that may fail, so if no progress
2748 * is being made, there are no other options and
2749 * retrying is unlikely to help.
2751 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2754 * The oom killer is not called for lowmem
2755 * allocations to prevent needlessly killing
2758 if (high_zoneidx
< ZONE_NORMAL
)
2766 /* Check if we should retry the allocation */
2767 pages_reclaimed
+= did_some_progress
;
2768 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2770 /* Wait for some write requests to complete then retry */
2771 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2775 * High-order allocations do not necessarily loop after
2776 * direct reclaim and reclaim/compaction depends on compaction
2777 * being called after reclaim so call directly if necessary
2779 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2780 high_zoneidx
, nodemask
, alloc_flags
,
2782 classzone_idx
, migratetype
,
2783 migration_mode
, &contended_compaction
,
2784 &deferred_compaction
);
2790 warn_alloc_failed(gfp_mask
, order
, NULL
);
2793 if (kmemcheck_enabled
)
2794 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2800 * This is the 'heart' of the zoned buddy allocator.
2803 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2804 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2806 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2807 struct zone
*preferred_zone
;
2808 struct zoneref
*preferred_zoneref
;
2809 struct page
*page
= NULL
;
2810 int migratetype
= gfpflags_to_migratetype(gfp_mask
);
2811 unsigned int cpuset_mems_cookie
;
2812 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2815 gfp_mask
&= gfp_allowed_mask
;
2817 lockdep_trace_alloc(gfp_mask
);
2819 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2821 if (should_fail_alloc_page(gfp_mask
, order
))
2825 * Check the zones suitable for the gfp_mask contain at least one
2826 * valid zone. It's possible to have an empty zonelist as a result
2827 * of GFP_THISNODE and a memoryless node
2829 if (unlikely(!zonelist
->_zonerefs
->zone
))
2832 if (IS_ENABLED(CONFIG_CMA
) && migratetype
== MIGRATE_MOVABLE
)
2833 alloc_flags
|= ALLOC_CMA
;
2836 cpuset_mems_cookie
= read_mems_allowed_begin();
2838 /* The preferred zone is used for statistics later */
2839 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2840 nodemask
? : &cpuset_current_mems_allowed
,
2842 if (!preferred_zone
)
2844 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2846 /* First allocation attempt */
2847 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2848 zonelist
, high_zoneidx
, alloc_flags
,
2849 preferred_zone
, classzone_idx
, migratetype
);
2850 if (unlikely(!page
)) {
2852 * Runtime PM, block IO and its error handling path
2853 * can deadlock because I/O on the device might not
2856 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2857 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2858 zonelist
, high_zoneidx
, nodemask
,
2859 preferred_zone
, classzone_idx
, migratetype
);
2862 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2866 * When updating a task's mems_allowed, it is possible to race with
2867 * parallel threads in such a way that an allocation can fail while
2868 * the mask is being updated. If a page allocation is about to fail,
2869 * check if the cpuset changed during allocation and if so, retry.
2871 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2876 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2879 * Common helper functions.
2881 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2886 * __get_free_pages() returns a 32-bit address, which cannot represent
2889 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2891 page
= alloc_pages(gfp_mask
, order
);
2894 return (unsigned long) page_address(page
);
2896 EXPORT_SYMBOL(__get_free_pages
);
2898 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2900 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2902 EXPORT_SYMBOL(get_zeroed_page
);
2904 void __free_pages(struct page
*page
, unsigned int order
)
2906 if (put_page_testzero(page
)) {
2908 free_hot_cold_page(page
, false);
2910 __free_pages_ok(page
, order
);
2914 EXPORT_SYMBOL(__free_pages
);
2916 void free_pages(unsigned long addr
, unsigned int order
)
2919 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2920 __free_pages(virt_to_page((void *)addr
), order
);
2924 EXPORT_SYMBOL(free_pages
);
2927 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2928 * of the current memory cgroup.
2930 * It should be used when the caller would like to use kmalloc, but since the
2931 * allocation is large, it has to fall back to the page allocator.
2933 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2936 struct mem_cgroup
*memcg
= NULL
;
2938 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2940 page
= alloc_pages(gfp_mask
, order
);
2941 memcg_kmem_commit_charge(page
, memcg
, order
);
2945 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2948 struct mem_cgroup
*memcg
= NULL
;
2950 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2952 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2953 memcg_kmem_commit_charge(page
, memcg
, order
);
2958 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2961 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2963 memcg_kmem_uncharge_pages(page
, order
);
2964 __free_pages(page
, order
);
2967 void free_kmem_pages(unsigned long addr
, unsigned int order
)
2970 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2971 __free_kmem_pages(virt_to_page((void *)addr
), order
);
2975 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2978 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2979 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2981 split_page(virt_to_page((void *)addr
), order
);
2982 while (used
< alloc_end
) {
2987 return (void *)addr
;
2991 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2992 * @size: the number of bytes to allocate
2993 * @gfp_mask: GFP flags for the allocation
2995 * This function is similar to alloc_pages(), except that it allocates the
2996 * minimum number of pages to satisfy the request. alloc_pages() can only
2997 * allocate memory in power-of-two pages.
2999 * This function is also limited by MAX_ORDER.
3001 * Memory allocated by this function must be released by free_pages_exact().
3003 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3005 unsigned int order
= get_order(size
);
3008 addr
= __get_free_pages(gfp_mask
, order
);
3009 return make_alloc_exact(addr
, order
, size
);
3011 EXPORT_SYMBOL(alloc_pages_exact
);
3014 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3016 * @nid: the preferred node ID where memory should be allocated
3017 * @size: the number of bytes to allocate
3018 * @gfp_mask: GFP flags for the allocation
3020 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3022 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3025 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3027 unsigned order
= get_order(size
);
3028 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3031 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3035 * free_pages_exact - release memory allocated via alloc_pages_exact()
3036 * @virt: the value returned by alloc_pages_exact.
3037 * @size: size of allocation, same value as passed to alloc_pages_exact().
3039 * Release the memory allocated by a previous call to alloc_pages_exact.
3041 void free_pages_exact(void *virt
, size_t size
)
3043 unsigned long addr
= (unsigned long)virt
;
3044 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3046 while (addr
< end
) {
3051 EXPORT_SYMBOL(free_pages_exact
);
3054 * nr_free_zone_pages - count number of pages beyond high watermark
3055 * @offset: The zone index of the highest zone
3057 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3058 * high watermark within all zones at or below a given zone index. For each
3059 * zone, the number of pages is calculated as:
3060 * managed_pages - high_pages
3062 static unsigned long nr_free_zone_pages(int offset
)
3067 /* Just pick one node, since fallback list is circular */
3068 unsigned long sum
= 0;
3070 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3072 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3073 unsigned long size
= zone
->managed_pages
;
3074 unsigned long high
= high_wmark_pages(zone
);
3083 * nr_free_buffer_pages - count number of pages beyond high watermark
3085 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3086 * watermark within ZONE_DMA and ZONE_NORMAL.
3088 unsigned long nr_free_buffer_pages(void)
3090 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3092 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3095 * nr_free_pagecache_pages - count number of pages beyond high watermark
3097 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3098 * high watermark within all zones.
3100 unsigned long nr_free_pagecache_pages(void)
3102 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3105 static inline void show_node(struct zone
*zone
)
3107 if (IS_ENABLED(CONFIG_NUMA
))
3108 printk("Node %d ", zone_to_nid(zone
));
3111 void si_meminfo(struct sysinfo
*val
)
3113 val
->totalram
= totalram_pages
;
3114 val
->sharedram
= global_page_state(NR_SHMEM
);
3115 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3116 val
->bufferram
= nr_blockdev_pages();
3117 val
->totalhigh
= totalhigh_pages
;
3118 val
->freehigh
= nr_free_highpages();
3119 val
->mem_unit
= PAGE_SIZE
;
3122 EXPORT_SYMBOL(si_meminfo
);
3125 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3127 int zone_type
; /* needs to be signed */
3128 unsigned long managed_pages
= 0;
3129 pg_data_t
*pgdat
= NODE_DATA(nid
);
3131 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3132 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3133 val
->totalram
= managed_pages
;
3134 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3135 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3136 #ifdef CONFIG_HIGHMEM
3137 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3138 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3144 val
->mem_unit
= PAGE_SIZE
;
3149 * Determine whether the node should be displayed or not, depending on whether
3150 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3152 bool skip_free_areas_node(unsigned int flags
, int nid
)
3155 unsigned int cpuset_mems_cookie
;
3157 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3161 cpuset_mems_cookie
= read_mems_allowed_begin();
3162 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3163 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3168 #define K(x) ((x) << (PAGE_SHIFT-10))
3170 static void show_migration_types(unsigned char type
)
3172 static const char types
[MIGRATE_TYPES
] = {
3173 [MIGRATE_UNMOVABLE
] = 'U',
3174 [MIGRATE_RECLAIMABLE
] = 'E',
3175 [MIGRATE_MOVABLE
] = 'M',
3176 [MIGRATE_RESERVE
] = 'R',
3178 [MIGRATE_CMA
] = 'C',
3180 #ifdef CONFIG_MEMORY_ISOLATION
3181 [MIGRATE_ISOLATE
] = 'I',
3184 char tmp
[MIGRATE_TYPES
+ 1];
3188 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3189 if (type
& (1 << i
))
3194 printk("(%s) ", tmp
);
3198 * Show free area list (used inside shift_scroll-lock stuff)
3199 * We also calculate the percentage fragmentation. We do this by counting the
3200 * memory on each free list with the exception of the first item on the list.
3201 * Suppresses nodes that are not allowed by current's cpuset if
3202 * SHOW_MEM_FILTER_NODES is passed.
3204 void show_free_areas(unsigned int filter
)
3209 for_each_populated_zone(zone
) {
3210 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3213 printk("%s per-cpu:\n", zone
->name
);
3215 for_each_online_cpu(cpu
) {
3216 struct per_cpu_pageset
*pageset
;
3218 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3220 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3221 cpu
, pageset
->pcp
.high
,
3222 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3226 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3227 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3229 " dirty:%lu writeback:%lu unstable:%lu\n"
3230 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3231 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3233 global_page_state(NR_ACTIVE_ANON
),
3234 global_page_state(NR_INACTIVE_ANON
),
3235 global_page_state(NR_ISOLATED_ANON
),
3236 global_page_state(NR_ACTIVE_FILE
),
3237 global_page_state(NR_INACTIVE_FILE
),
3238 global_page_state(NR_ISOLATED_FILE
),
3239 global_page_state(NR_UNEVICTABLE
),
3240 global_page_state(NR_FILE_DIRTY
),
3241 global_page_state(NR_WRITEBACK
),
3242 global_page_state(NR_UNSTABLE_NFS
),
3243 global_page_state(NR_FREE_PAGES
),
3244 global_page_state(NR_SLAB_RECLAIMABLE
),
3245 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3246 global_page_state(NR_FILE_MAPPED
),
3247 global_page_state(NR_SHMEM
),
3248 global_page_state(NR_PAGETABLE
),
3249 global_page_state(NR_BOUNCE
),
3250 global_page_state(NR_FREE_CMA_PAGES
));
3252 for_each_populated_zone(zone
) {
3255 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3263 " active_anon:%lukB"
3264 " inactive_anon:%lukB"
3265 " active_file:%lukB"
3266 " inactive_file:%lukB"
3267 " unevictable:%lukB"
3268 " isolated(anon):%lukB"
3269 " isolated(file):%lukB"
3277 " slab_reclaimable:%lukB"
3278 " slab_unreclaimable:%lukB"
3279 " kernel_stack:%lukB"
3284 " writeback_tmp:%lukB"
3285 " pages_scanned:%lu"
3286 " all_unreclaimable? %s"
3289 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3290 K(min_wmark_pages(zone
)),
3291 K(low_wmark_pages(zone
)),
3292 K(high_wmark_pages(zone
)),
3293 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3294 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3295 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3296 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3297 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3298 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3299 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3300 K(zone
->present_pages
),
3301 K(zone
->managed_pages
),
3302 K(zone_page_state(zone
, NR_MLOCK
)),
3303 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3304 K(zone_page_state(zone
, NR_WRITEBACK
)),
3305 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3306 K(zone_page_state(zone
, NR_SHMEM
)),
3307 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3308 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3309 zone_page_state(zone
, NR_KERNEL_STACK
) *
3311 K(zone_page_state(zone
, NR_PAGETABLE
)),
3312 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3313 K(zone_page_state(zone
, NR_BOUNCE
)),
3314 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3315 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3316 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3317 (!zone_reclaimable(zone
) ? "yes" : "no")
3319 printk("lowmem_reserve[]:");
3320 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3321 printk(" %ld", zone
->lowmem_reserve
[i
]);
3325 for_each_populated_zone(zone
) {
3326 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3327 unsigned char types
[MAX_ORDER
];
3329 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3332 printk("%s: ", zone
->name
);
3334 spin_lock_irqsave(&zone
->lock
, flags
);
3335 for (order
= 0; order
< MAX_ORDER
; order
++) {
3336 struct free_area
*area
= &zone
->free_area
[order
];
3339 nr
[order
] = area
->nr_free
;
3340 total
+= nr
[order
] << order
;
3343 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3344 if (!list_empty(&area
->free_list
[type
]))
3345 types
[order
] |= 1 << type
;
3348 spin_unlock_irqrestore(&zone
->lock
, flags
);
3349 for (order
= 0; order
< MAX_ORDER
; order
++) {
3350 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3352 show_migration_types(types
[order
]);
3354 printk("= %lukB\n", K(total
));
3357 hugetlb_show_meminfo();
3359 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3361 show_swap_cache_info();
3364 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3366 zoneref
->zone
= zone
;
3367 zoneref
->zone_idx
= zone_idx(zone
);
3371 * Builds allocation fallback zone lists.
3373 * Add all populated zones of a node to the zonelist.
3375 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3379 enum zone_type zone_type
= MAX_NR_ZONES
;
3383 zone
= pgdat
->node_zones
+ zone_type
;
3384 if (populated_zone(zone
)) {
3385 zoneref_set_zone(zone
,
3386 &zonelist
->_zonerefs
[nr_zones
++]);
3387 check_highest_zone(zone_type
);
3389 } while (zone_type
);
3397 * 0 = automatic detection of better ordering.
3398 * 1 = order by ([node] distance, -zonetype)
3399 * 2 = order by (-zonetype, [node] distance)
3401 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3402 * the same zonelist. So only NUMA can configure this param.
3404 #define ZONELIST_ORDER_DEFAULT 0
3405 #define ZONELIST_ORDER_NODE 1
3406 #define ZONELIST_ORDER_ZONE 2
3408 /* zonelist order in the kernel.
3409 * set_zonelist_order() will set this to NODE or ZONE.
3411 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3412 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3416 /* The value user specified ....changed by config */
3417 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3418 /* string for sysctl */
3419 #define NUMA_ZONELIST_ORDER_LEN 16
3420 char numa_zonelist_order
[16] = "default";
3423 * interface for configure zonelist ordering.
3424 * command line option "numa_zonelist_order"
3425 * = "[dD]efault - default, automatic configuration.
3426 * = "[nN]ode - order by node locality, then by zone within node
3427 * = "[zZ]one - order by zone, then by locality within zone
3430 static int __parse_numa_zonelist_order(char *s
)
3432 if (*s
== 'd' || *s
== 'D') {
3433 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3434 } else if (*s
== 'n' || *s
== 'N') {
3435 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3436 } else if (*s
== 'z' || *s
== 'Z') {
3437 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3440 "Ignoring invalid numa_zonelist_order value: "
3447 static __init
int setup_numa_zonelist_order(char *s
)
3454 ret
= __parse_numa_zonelist_order(s
);
3456 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3460 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3463 * sysctl handler for numa_zonelist_order
3465 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3466 void __user
*buffer
, size_t *length
,
3469 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3471 static DEFINE_MUTEX(zl_order_mutex
);
3473 mutex_lock(&zl_order_mutex
);
3475 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3479 strcpy(saved_string
, (char *)table
->data
);
3481 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3485 int oldval
= user_zonelist_order
;
3487 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3490 * bogus value. restore saved string
3492 strncpy((char *)table
->data
, saved_string
,
3493 NUMA_ZONELIST_ORDER_LEN
);
3494 user_zonelist_order
= oldval
;
3495 } else if (oldval
!= user_zonelist_order
) {
3496 mutex_lock(&zonelists_mutex
);
3497 build_all_zonelists(NULL
, NULL
);
3498 mutex_unlock(&zonelists_mutex
);
3502 mutex_unlock(&zl_order_mutex
);
3507 #define MAX_NODE_LOAD (nr_online_nodes)
3508 static int node_load
[MAX_NUMNODES
];
3511 * find_next_best_node - find the next node that should appear in a given node's fallback list
3512 * @node: node whose fallback list we're appending
3513 * @used_node_mask: nodemask_t of already used nodes
3515 * We use a number of factors to determine which is the next node that should
3516 * appear on a given node's fallback list. The node should not have appeared
3517 * already in @node's fallback list, and it should be the next closest node
3518 * according to the distance array (which contains arbitrary distance values
3519 * from each node to each node in the system), and should also prefer nodes
3520 * with no CPUs, since presumably they'll have very little allocation pressure
3521 * on them otherwise.
3522 * It returns -1 if no node is found.
3524 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3527 int min_val
= INT_MAX
;
3528 int best_node
= NUMA_NO_NODE
;
3529 const struct cpumask
*tmp
= cpumask_of_node(0);
3531 /* Use the local node if we haven't already */
3532 if (!node_isset(node
, *used_node_mask
)) {
3533 node_set(node
, *used_node_mask
);
3537 for_each_node_state(n
, N_MEMORY
) {
3539 /* Don't want a node to appear more than once */
3540 if (node_isset(n
, *used_node_mask
))
3543 /* Use the distance array to find the distance */
3544 val
= node_distance(node
, n
);
3546 /* Penalize nodes under us ("prefer the next node") */
3549 /* Give preference to headless and unused nodes */
3550 tmp
= cpumask_of_node(n
);
3551 if (!cpumask_empty(tmp
))
3552 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3554 /* Slight preference for less loaded node */
3555 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3556 val
+= node_load
[n
];
3558 if (val
< min_val
) {
3565 node_set(best_node
, *used_node_mask
);
3572 * Build zonelists ordered by node and zones within node.
3573 * This results in maximum locality--normal zone overflows into local
3574 * DMA zone, if any--but risks exhausting DMA zone.
3576 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3579 struct zonelist
*zonelist
;
3581 zonelist
= &pgdat
->node_zonelists
[0];
3582 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3584 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3585 zonelist
->_zonerefs
[j
].zone
= NULL
;
3586 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3590 * Build gfp_thisnode zonelists
3592 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3595 struct zonelist
*zonelist
;
3597 zonelist
= &pgdat
->node_zonelists
[1];
3598 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3599 zonelist
->_zonerefs
[j
].zone
= NULL
;
3600 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3604 * Build zonelists ordered by zone and nodes within zones.
3605 * This results in conserving DMA zone[s] until all Normal memory is
3606 * exhausted, but results in overflowing to remote node while memory
3607 * may still exist in local DMA zone.
3609 static int node_order
[MAX_NUMNODES
];
3611 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3614 int zone_type
; /* needs to be signed */
3616 struct zonelist
*zonelist
;
3618 zonelist
= &pgdat
->node_zonelists
[0];
3620 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3621 for (j
= 0; j
< nr_nodes
; j
++) {
3622 node
= node_order
[j
];
3623 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3624 if (populated_zone(z
)) {
3626 &zonelist
->_zonerefs
[pos
++]);
3627 check_highest_zone(zone_type
);
3631 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3632 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3635 #if defined(CONFIG_64BIT)
3637 * Devices that require DMA32/DMA are relatively rare and do not justify a
3638 * penalty to every machine in case the specialised case applies. Default
3639 * to Node-ordering on 64-bit NUMA machines
3641 static int default_zonelist_order(void)
3643 return ZONELIST_ORDER_NODE
;
3647 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3648 * by the kernel. If processes running on node 0 deplete the low memory zone
3649 * then reclaim will occur more frequency increasing stalls and potentially
3650 * be easier to OOM if a large percentage of the zone is under writeback or
3651 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3652 * Hence, default to zone ordering on 32-bit.
3654 static int default_zonelist_order(void)
3656 return ZONELIST_ORDER_ZONE
;
3658 #endif /* CONFIG_64BIT */
3660 static void set_zonelist_order(void)
3662 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3663 current_zonelist_order
= default_zonelist_order();
3665 current_zonelist_order
= user_zonelist_order
;
3668 static void build_zonelists(pg_data_t
*pgdat
)
3672 nodemask_t used_mask
;
3673 int local_node
, prev_node
;
3674 struct zonelist
*zonelist
;
3675 int order
= current_zonelist_order
;
3677 /* initialize zonelists */
3678 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3679 zonelist
= pgdat
->node_zonelists
+ i
;
3680 zonelist
->_zonerefs
[0].zone
= NULL
;
3681 zonelist
->_zonerefs
[0].zone_idx
= 0;
3684 /* NUMA-aware ordering of nodes */
3685 local_node
= pgdat
->node_id
;
3686 load
= nr_online_nodes
;
3687 prev_node
= local_node
;
3688 nodes_clear(used_mask
);
3690 memset(node_order
, 0, sizeof(node_order
));
3693 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3695 * We don't want to pressure a particular node.
3696 * So adding penalty to the first node in same
3697 * distance group to make it round-robin.
3699 if (node_distance(local_node
, node
) !=
3700 node_distance(local_node
, prev_node
))
3701 node_load
[node
] = load
;
3705 if (order
== ZONELIST_ORDER_NODE
)
3706 build_zonelists_in_node_order(pgdat
, node
);
3708 node_order
[j
++] = node
; /* remember order */
3711 if (order
== ZONELIST_ORDER_ZONE
) {
3712 /* calculate node order -- i.e., DMA last! */
3713 build_zonelists_in_zone_order(pgdat
, j
);
3716 build_thisnode_zonelists(pgdat
);
3719 /* Construct the zonelist performance cache - see further mmzone.h */
3720 static void build_zonelist_cache(pg_data_t
*pgdat
)
3722 struct zonelist
*zonelist
;
3723 struct zonelist_cache
*zlc
;
3726 zonelist
= &pgdat
->node_zonelists
[0];
3727 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3728 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3729 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3730 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3733 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3735 * Return node id of node used for "local" allocations.
3736 * I.e., first node id of first zone in arg node's generic zonelist.
3737 * Used for initializing percpu 'numa_mem', which is used primarily
3738 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3740 int local_memory_node(int node
)
3744 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3745 gfp_zone(GFP_KERNEL
),
3752 #else /* CONFIG_NUMA */
3754 static void set_zonelist_order(void)
3756 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3759 static void build_zonelists(pg_data_t
*pgdat
)
3761 int node
, local_node
;
3763 struct zonelist
*zonelist
;
3765 local_node
= pgdat
->node_id
;
3767 zonelist
= &pgdat
->node_zonelists
[0];
3768 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3771 * Now we build the zonelist so that it contains the zones
3772 * of all the other nodes.
3773 * We don't want to pressure a particular node, so when
3774 * building the zones for node N, we make sure that the
3775 * zones coming right after the local ones are those from
3776 * node N+1 (modulo N)
3778 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3779 if (!node_online(node
))
3781 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3783 for (node
= 0; node
< local_node
; node
++) {
3784 if (!node_online(node
))
3786 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3789 zonelist
->_zonerefs
[j
].zone
= NULL
;
3790 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3793 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3794 static void build_zonelist_cache(pg_data_t
*pgdat
)
3796 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3799 #endif /* CONFIG_NUMA */
3802 * Boot pageset table. One per cpu which is going to be used for all
3803 * zones and all nodes. The parameters will be set in such a way
3804 * that an item put on a list will immediately be handed over to
3805 * the buddy list. This is safe since pageset manipulation is done
3806 * with interrupts disabled.
3808 * The boot_pagesets must be kept even after bootup is complete for
3809 * unused processors and/or zones. They do play a role for bootstrapping
3810 * hotplugged processors.
3812 * zoneinfo_show() and maybe other functions do
3813 * not check if the processor is online before following the pageset pointer.
3814 * Other parts of the kernel may not check if the zone is available.
3816 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3817 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3818 static void setup_zone_pageset(struct zone
*zone
);
3821 * Global mutex to protect against size modification of zonelists
3822 * as well as to serialize pageset setup for the new populated zone.
3824 DEFINE_MUTEX(zonelists_mutex
);
3826 /* return values int ....just for stop_machine() */
3827 static int __build_all_zonelists(void *data
)
3831 pg_data_t
*self
= data
;
3834 memset(node_load
, 0, sizeof(node_load
));
3837 if (self
&& !node_online(self
->node_id
)) {
3838 build_zonelists(self
);
3839 build_zonelist_cache(self
);
3842 for_each_online_node(nid
) {
3843 pg_data_t
*pgdat
= NODE_DATA(nid
);
3845 build_zonelists(pgdat
);
3846 build_zonelist_cache(pgdat
);
3850 * Initialize the boot_pagesets that are going to be used
3851 * for bootstrapping processors. The real pagesets for
3852 * each zone will be allocated later when the per cpu
3853 * allocator is available.
3855 * boot_pagesets are used also for bootstrapping offline
3856 * cpus if the system is already booted because the pagesets
3857 * are needed to initialize allocators on a specific cpu too.
3858 * F.e. the percpu allocator needs the page allocator which
3859 * needs the percpu allocator in order to allocate its pagesets
3860 * (a chicken-egg dilemma).
3862 for_each_possible_cpu(cpu
) {
3863 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3865 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3867 * We now know the "local memory node" for each node--
3868 * i.e., the node of the first zone in the generic zonelist.
3869 * Set up numa_mem percpu variable for on-line cpus. During
3870 * boot, only the boot cpu should be on-line; we'll init the
3871 * secondary cpus' numa_mem as they come on-line. During
3872 * node/memory hotplug, we'll fixup all on-line cpus.
3874 if (cpu_online(cpu
))
3875 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3883 * Called with zonelists_mutex held always
3884 * unless system_state == SYSTEM_BOOTING.
3886 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3888 set_zonelist_order();
3890 if (system_state
== SYSTEM_BOOTING
) {
3891 __build_all_zonelists(NULL
);
3892 mminit_verify_zonelist();
3893 cpuset_init_current_mems_allowed();
3895 #ifdef CONFIG_MEMORY_HOTPLUG
3897 setup_zone_pageset(zone
);
3899 /* we have to stop all cpus to guarantee there is no user
3901 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3902 /* cpuset refresh routine should be here */
3904 vm_total_pages
= nr_free_pagecache_pages();
3906 * Disable grouping by mobility if the number of pages in the
3907 * system is too low to allow the mechanism to work. It would be
3908 * more accurate, but expensive to check per-zone. This check is
3909 * made on memory-hotadd so a system can start with mobility
3910 * disabled and enable it later
3912 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3913 page_group_by_mobility_disabled
= 1;
3915 page_group_by_mobility_disabled
= 0;
3917 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3918 "Total pages: %ld\n",
3920 zonelist_order_name
[current_zonelist_order
],
3921 page_group_by_mobility_disabled
? "off" : "on",
3924 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
3929 * Helper functions to size the waitqueue hash table.
3930 * Essentially these want to choose hash table sizes sufficiently
3931 * large so that collisions trying to wait on pages are rare.
3932 * But in fact, the number of active page waitqueues on typical
3933 * systems is ridiculously low, less than 200. So this is even
3934 * conservative, even though it seems large.
3936 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3937 * waitqueues, i.e. the size of the waitq table given the number of pages.
3939 #define PAGES_PER_WAITQUEUE 256
3941 #ifndef CONFIG_MEMORY_HOTPLUG
3942 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3944 unsigned long size
= 1;
3946 pages
/= PAGES_PER_WAITQUEUE
;
3948 while (size
< pages
)
3952 * Once we have dozens or even hundreds of threads sleeping
3953 * on IO we've got bigger problems than wait queue collision.
3954 * Limit the size of the wait table to a reasonable size.
3956 size
= min(size
, 4096UL);
3958 return max(size
, 4UL);
3962 * A zone's size might be changed by hot-add, so it is not possible to determine
3963 * a suitable size for its wait_table. So we use the maximum size now.
3965 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3967 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3968 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3969 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3971 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3972 * or more by the traditional way. (See above). It equals:
3974 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3975 * ia64(16K page size) : = ( 8G + 4M)byte.
3976 * powerpc (64K page size) : = (32G +16M)byte.
3978 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3985 * This is an integer logarithm so that shifts can be used later
3986 * to extract the more random high bits from the multiplicative
3987 * hash function before the remainder is taken.
3989 static inline unsigned long wait_table_bits(unsigned long size
)
3995 * Check if a pageblock contains reserved pages
3997 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
4001 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4002 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
4009 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4010 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4011 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4012 * higher will lead to a bigger reserve which will get freed as contiguous
4013 * blocks as reclaim kicks in
4015 static void setup_zone_migrate_reserve(struct zone
*zone
)
4017 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4019 unsigned long block_migratetype
;
4024 * Get the start pfn, end pfn and the number of blocks to reserve
4025 * We have to be careful to be aligned to pageblock_nr_pages to
4026 * make sure that we always check pfn_valid for the first page in
4029 start_pfn
= zone
->zone_start_pfn
;
4030 end_pfn
= zone_end_pfn(zone
);
4031 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4032 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4036 * Reserve blocks are generally in place to help high-order atomic
4037 * allocations that are short-lived. A min_free_kbytes value that
4038 * would result in more than 2 reserve blocks for atomic allocations
4039 * is assumed to be in place to help anti-fragmentation for the
4040 * future allocation of hugepages at runtime.
4042 reserve
= min(2, reserve
);
4043 old_reserve
= zone
->nr_migrate_reserve_block
;
4045 /* When memory hot-add, we almost always need to do nothing */
4046 if (reserve
== old_reserve
)
4048 zone
->nr_migrate_reserve_block
= reserve
;
4050 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4051 if (!pfn_valid(pfn
))
4053 page
= pfn_to_page(pfn
);
4055 /* Watch out for overlapping nodes */
4056 if (page_to_nid(page
) != zone_to_nid(zone
))
4059 block_migratetype
= get_pageblock_migratetype(page
);
4061 /* Only test what is necessary when the reserves are not met */
4064 * Blocks with reserved pages will never free, skip
4067 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4068 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4071 /* If this block is reserved, account for it */
4072 if (block_migratetype
== MIGRATE_RESERVE
) {
4077 /* Suitable for reserving if this block is movable */
4078 if (block_migratetype
== MIGRATE_MOVABLE
) {
4079 set_pageblock_migratetype(page
,
4081 move_freepages_block(zone
, page
,
4086 } else if (!old_reserve
) {
4088 * At boot time we don't need to scan the whole zone
4089 * for turning off MIGRATE_RESERVE.
4095 * If the reserve is met and this is a previous reserved block,
4098 if (block_migratetype
== MIGRATE_RESERVE
) {
4099 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4100 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4106 * Initially all pages are reserved - free ones are freed
4107 * up by free_all_bootmem() once the early boot process is
4108 * done. Non-atomic initialization, single-pass.
4110 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4111 unsigned long start_pfn
, enum memmap_context context
)
4114 unsigned long end_pfn
= start_pfn
+ size
;
4118 if (highest_memmap_pfn
< end_pfn
- 1)
4119 highest_memmap_pfn
= end_pfn
- 1;
4121 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4122 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4124 * There can be holes in boot-time mem_map[]s
4125 * handed to this function. They do not
4126 * exist on hotplugged memory.
4128 if (context
== MEMMAP_EARLY
) {
4129 if (!early_pfn_valid(pfn
))
4131 if (!early_pfn_in_nid(pfn
, nid
))
4134 page
= pfn_to_page(pfn
);
4135 set_page_links(page
, zone
, nid
, pfn
);
4136 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4137 init_page_count(page
);
4138 page_mapcount_reset(page
);
4139 page_cpupid_reset_last(page
);
4140 SetPageReserved(page
);
4142 * Mark the block movable so that blocks are reserved for
4143 * movable at startup. This will force kernel allocations
4144 * to reserve their blocks rather than leaking throughout
4145 * the address space during boot when many long-lived
4146 * kernel allocations are made. Later some blocks near
4147 * the start are marked MIGRATE_RESERVE by
4148 * setup_zone_migrate_reserve()
4150 * bitmap is created for zone's valid pfn range. but memmap
4151 * can be created for invalid pages (for alignment)
4152 * check here not to call set_pageblock_migratetype() against
4155 if ((z
->zone_start_pfn
<= pfn
)
4156 && (pfn
< zone_end_pfn(z
))
4157 && !(pfn
& (pageblock_nr_pages
- 1)))
4158 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4160 INIT_LIST_HEAD(&page
->lru
);
4161 #ifdef WANT_PAGE_VIRTUAL
4162 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4163 if (!is_highmem_idx(zone
))
4164 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4169 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4171 unsigned int order
, t
;
4172 for_each_migratetype_order(order
, t
) {
4173 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4174 zone
->free_area
[order
].nr_free
= 0;
4178 #ifndef __HAVE_ARCH_MEMMAP_INIT
4179 #define memmap_init(size, nid, zone, start_pfn) \
4180 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4183 static int zone_batchsize(struct zone
*zone
)
4189 * The per-cpu-pages pools are set to around 1000th of the
4190 * size of the zone. But no more than 1/2 of a meg.
4192 * OK, so we don't know how big the cache is. So guess.
4194 batch
= zone
->managed_pages
/ 1024;
4195 if (batch
* PAGE_SIZE
> 512 * 1024)
4196 batch
= (512 * 1024) / PAGE_SIZE
;
4197 batch
/= 4; /* We effectively *= 4 below */
4202 * Clamp the batch to a 2^n - 1 value. Having a power
4203 * of 2 value was found to be more likely to have
4204 * suboptimal cache aliasing properties in some cases.
4206 * For example if 2 tasks are alternately allocating
4207 * batches of pages, one task can end up with a lot
4208 * of pages of one half of the possible page colors
4209 * and the other with pages of the other colors.
4211 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4216 /* The deferral and batching of frees should be suppressed under NOMMU
4219 * The problem is that NOMMU needs to be able to allocate large chunks
4220 * of contiguous memory as there's no hardware page translation to
4221 * assemble apparent contiguous memory from discontiguous pages.
4223 * Queueing large contiguous runs of pages for batching, however,
4224 * causes the pages to actually be freed in smaller chunks. As there
4225 * can be a significant delay between the individual batches being
4226 * recycled, this leads to the once large chunks of space being
4227 * fragmented and becoming unavailable for high-order allocations.
4234 * pcp->high and pcp->batch values are related and dependent on one another:
4235 * ->batch must never be higher then ->high.
4236 * The following function updates them in a safe manner without read side
4239 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4240 * those fields changing asynchronously (acording the the above rule).
4242 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4243 * outside of boot time (or some other assurance that no concurrent updaters
4246 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4247 unsigned long batch
)
4249 /* start with a fail safe value for batch */
4253 /* Update high, then batch, in order */
4260 /* a companion to pageset_set_high() */
4261 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4263 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4266 static void pageset_init(struct per_cpu_pageset
*p
)
4268 struct per_cpu_pages
*pcp
;
4271 memset(p
, 0, sizeof(*p
));
4275 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4276 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4279 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4282 pageset_set_batch(p
, batch
);
4286 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4287 * to the value high for the pageset p.
4289 static void pageset_set_high(struct per_cpu_pageset
*p
,
4292 unsigned long batch
= max(1UL, high
/ 4);
4293 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4294 batch
= PAGE_SHIFT
* 8;
4296 pageset_update(&p
->pcp
, high
, batch
);
4299 static void pageset_set_high_and_batch(struct zone
*zone
,
4300 struct per_cpu_pageset
*pcp
)
4302 if (percpu_pagelist_fraction
)
4303 pageset_set_high(pcp
,
4304 (zone
->managed_pages
/
4305 percpu_pagelist_fraction
));
4307 pageset_set_batch(pcp
, zone_batchsize(zone
));
4310 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4312 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4315 pageset_set_high_and_batch(zone
, pcp
);
4318 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4321 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4322 for_each_possible_cpu(cpu
)
4323 zone_pageset_init(zone
, cpu
);
4327 * Allocate per cpu pagesets and initialize them.
4328 * Before this call only boot pagesets were available.
4330 void __init
setup_per_cpu_pageset(void)
4334 for_each_populated_zone(zone
)
4335 setup_zone_pageset(zone
);
4338 static noinline __init_refok
4339 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4345 * The per-page waitqueue mechanism uses hashed waitqueues
4348 zone
->wait_table_hash_nr_entries
=
4349 wait_table_hash_nr_entries(zone_size_pages
);
4350 zone
->wait_table_bits
=
4351 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4352 alloc_size
= zone
->wait_table_hash_nr_entries
4353 * sizeof(wait_queue_head_t
);
4355 if (!slab_is_available()) {
4356 zone
->wait_table
= (wait_queue_head_t
*)
4357 memblock_virt_alloc_node_nopanic(
4358 alloc_size
, zone
->zone_pgdat
->node_id
);
4361 * This case means that a zone whose size was 0 gets new memory
4362 * via memory hot-add.
4363 * But it may be the case that a new node was hot-added. In
4364 * this case vmalloc() will not be able to use this new node's
4365 * memory - this wait_table must be initialized to use this new
4366 * node itself as well.
4367 * To use this new node's memory, further consideration will be
4370 zone
->wait_table
= vmalloc(alloc_size
);
4372 if (!zone
->wait_table
)
4375 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4376 init_waitqueue_head(zone
->wait_table
+ i
);
4381 static __meminit
void zone_pcp_init(struct zone
*zone
)
4384 * per cpu subsystem is not up at this point. The following code
4385 * relies on the ability of the linker to provide the
4386 * offset of a (static) per cpu variable into the per cpu area.
4388 zone
->pageset
= &boot_pageset
;
4390 if (populated_zone(zone
))
4391 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4392 zone
->name
, zone
->present_pages
,
4393 zone_batchsize(zone
));
4396 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4397 unsigned long zone_start_pfn
,
4399 enum memmap_context context
)
4401 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4403 ret
= zone_wait_table_init(zone
, size
);
4406 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4408 zone
->zone_start_pfn
= zone_start_pfn
;
4410 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4411 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4413 (unsigned long)zone_idx(zone
),
4414 zone_start_pfn
, (zone_start_pfn
+ size
));
4416 zone_init_free_lists(zone
);
4421 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4422 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4424 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4426 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4428 unsigned long start_pfn
, end_pfn
;
4431 * NOTE: The following SMP-unsafe globals are only used early in boot
4432 * when the kernel is running single-threaded.
4434 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4435 static int __meminitdata last_nid
;
4437 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4440 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4442 last_start_pfn
= start_pfn
;
4443 last_end_pfn
= end_pfn
;
4449 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4451 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4455 nid
= __early_pfn_to_nid(pfn
);
4458 /* just returns 0 */
4462 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4463 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4467 nid
= __early_pfn_to_nid(pfn
);
4468 if (nid
>= 0 && nid
!= node
)
4475 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4476 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4477 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4479 * If an architecture guarantees that all ranges registered contain no holes
4480 * and may be freed, this this function may be used instead of calling
4481 * memblock_free_early_nid() manually.
4483 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4485 unsigned long start_pfn
, end_pfn
;
4488 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4489 start_pfn
= min(start_pfn
, max_low_pfn
);
4490 end_pfn
= min(end_pfn
, max_low_pfn
);
4492 if (start_pfn
< end_pfn
)
4493 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4494 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4500 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4501 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4503 * If an architecture guarantees that all ranges registered contain no holes and may
4504 * be freed, this function may be used instead of calling memory_present() manually.
4506 void __init
sparse_memory_present_with_active_regions(int nid
)
4508 unsigned long start_pfn
, end_pfn
;
4511 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4512 memory_present(this_nid
, start_pfn
, end_pfn
);
4516 * get_pfn_range_for_nid - Return the start and end page frames for a node
4517 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4518 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4519 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4521 * It returns the start and end page frame of a node based on information
4522 * provided by memblock_set_node(). If called for a node
4523 * with no available memory, a warning is printed and the start and end
4526 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4527 unsigned long *start_pfn
, unsigned long *end_pfn
)
4529 unsigned long this_start_pfn
, this_end_pfn
;
4535 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4536 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4537 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4540 if (*start_pfn
== -1UL)
4545 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4546 * assumption is made that zones within a node are ordered in monotonic
4547 * increasing memory addresses so that the "highest" populated zone is used
4549 static void __init
find_usable_zone_for_movable(void)
4552 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4553 if (zone_index
== ZONE_MOVABLE
)
4556 if (arch_zone_highest_possible_pfn
[zone_index
] >
4557 arch_zone_lowest_possible_pfn
[zone_index
])
4561 VM_BUG_ON(zone_index
== -1);
4562 movable_zone
= zone_index
;
4566 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4567 * because it is sized independent of architecture. Unlike the other zones,
4568 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4569 * in each node depending on the size of each node and how evenly kernelcore
4570 * is distributed. This helper function adjusts the zone ranges
4571 * provided by the architecture for a given node by using the end of the
4572 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4573 * zones within a node are in order of monotonic increases memory addresses
4575 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4576 unsigned long zone_type
,
4577 unsigned long node_start_pfn
,
4578 unsigned long node_end_pfn
,
4579 unsigned long *zone_start_pfn
,
4580 unsigned long *zone_end_pfn
)
4582 /* Only adjust if ZONE_MOVABLE is on this node */
4583 if (zone_movable_pfn
[nid
]) {
4584 /* Size ZONE_MOVABLE */
4585 if (zone_type
== ZONE_MOVABLE
) {
4586 *zone_start_pfn
= zone_movable_pfn
[nid
];
4587 *zone_end_pfn
= min(node_end_pfn
,
4588 arch_zone_highest_possible_pfn
[movable_zone
]);
4590 /* Adjust for ZONE_MOVABLE starting within this range */
4591 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4592 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4593 *zone_end_pfn
= zone_movable_pfn
[nid
];
4595 /* Check if this whole range is within ZONE_MOVABLE */
4596 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4597 *zone_start_pfn
= *zone_end_pfn
;
4602 * Return the number of pages a zone spans in a node, including holes
4603 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4605 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4606 unsigned long zone_type
,
4607 unsigned long node_start_pfn
,
4608 unsigned long node_end_pfn
,
4609 unsigned long *ignored
)
4611 unsigned long zone_start_pfn
, zone_end_pfn
;
4613 /* Get the start and end of the zone */
4614 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4615 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4616 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4617 node_start_pfn
, node_end_pfn
,
4618 &zone_start_pfn
, &zone_end_pfn
);
4620 /* Check that this node has pages within the zone's required range */
4621 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4624 /* Move the zone boundaries inside the node if necessary */
4625 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4626 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4628 /* Return the spanned pages */
4629 return zone_end_pfn
- zone_start_pfn
;
4633 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4634 * then all holes in the requested range will be accounted for.
4636 unsigned long __meminit
__absent_pages_in_range(int nid
,
4637 unsigned long range_start_pfn
,
4638 unsigned long range_end_pfn
)
4640 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4641 unsigned long start_pfn
, end_pfn
;
4644 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4645 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4646 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4647 nr_absent
-= end_pfn
- start_pfn
;
4653 * absent_pages_in_range - Return number of page frames in holes within a range
4654 * @start_pfn: The start PFN to start searching for holes
4655 * @end_pfn: The end PFN to stop searching for holes
4657 * It returns the number of pages frames in memory holes within a range.
4659 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4660 unsigned long end_pfn
)
4662 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4665 /* Return the number of page frames in holes in a zone on a node */
4666 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4667 unsigned long zone_type
,
4668 unsigned long node_start_pfn
,
4669 unsigned long node_end_pfn
,
4670 unsigned long *ignored
)
4672 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4673 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4674 unsigned long zone_start_pfn
, zone_end_pfn
;
4676 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4677 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4679 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4680 node_start_pfn
, node_end_pfn
,
4681 &zone_start_pfn
, &zone_end_pfn
);
4682 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4685 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4686 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4687 unsigned long zone_type
,
4688 unsigned long node_start_pfn
,
4689 unsigned long node_end_pfn
,
4690 unsigned long *zones_size
)
4692 return zones_size
[zone_type
];
4695 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4696 unsigned long zone_type
,
4697 unsigned long node_start_pfn
,
4698 unsigned long node_end_pfn
,
4699 unsigned long *zholes_size
)
4704 return zholes_size
[zone_type
];
4707 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4709 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4710 unsigned long node_start_pfn
,
4711 unsigned long node_end_pfn
,
4712 unsigned long *zones_size
,
4713 unsigned long *zholes_size
)
4715 unsigned long realtotalpages
, totalpages
= 0;
4718 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4719 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4723 pgdat
->node_spanned_pages
= totalpages
;
4725 realtotalpages
= totalpages
;
4726 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4728 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4729 node_start_pfn
, node_end_pfn
,
4731 pgdat
->node_present_pages
= realtotalpages
;
4732 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4736 #ifndef CONFIG_SPARSEMEM
4738 * Calculate the size of the zone->blockflags rounded to an unsigned long
4739 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4740 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4741 * round what is now in bits to nearest long in bits, then return it in
4744 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4746 unsigned long usemapsize
;
4748 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4749 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4750 usemapsize
= usemapsize
>> pageblock_order
;
4751 usemapsize
*= NR_PAGEBLOCK_BITS
;
4752 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4754 return usemapsize
/ 8;
4757 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4759 unsigned long zone_start_pfn
,
4760 unsigned long zonesize
)
4762 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4763 zone
->pageblock_flags
= NULL
;
4765 zone
->pageblock_flags
=
4766 memblock_virt_alloc_node_nopanic(usemapsize
,
4770 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4771 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4772 #endif /* CONFIG_SPARSEMEM */
4774 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4776 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4777 void __paginginit
set_pageblock_order(void)
4781 /* Check that pageblock_nr_pages has not already been setup */
4782 if (pageblock_order
)
4785 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4786 order
= HUGETLB_PAGE_ORDER
;
4788 order
= MAX_ORDER
- 1;
4791 * Assume the largest contiguous order of interest is a huge page.
4792 * This value may be variable depending on boot parameters on IA64 and
4795 pageblock_order
= order
;
4797 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4800 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4801 * is unused as pageblock_order is set at compile-time. See
4802 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4805 void __paginginit
set_pageblock_order(void)
4809 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4811 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4812 unsigned long present_pages
)
4814 unsigned long pages
= spanned_pages
;
4817 * Provide a more accurate estimation if there are holes within
4818 * the zone and SPARSEMEM is in use. If there are holes within the
4819 * zone, each populated memory region may cost us one or two extra
4820 * memmap pages due to alignment because memmap pages for each
4821 * populated regions may not naturally algined on page boundary.
4822 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4824 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4825 IS_ENABLED(CONFIG_SPARSEMEM
))
4826 pages
= present_pages
;
4828 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4832 * Set up the zone data structures:
4833 * - mark all pages reserved
4834 * - mark all memory queues empty
4835 * - clear the memory bitmaps
4837 * NOTE: pgdat should get zeroed by caller.
4839 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4840 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4841 unsigned long *zones_size
, unsigned long *zholes_size
)
4844 int nid
= pgdat
->node_id
;
4845 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4848 pgdat_resize_init(pgdat
);
4849 #ifdef CONFIG_NUMA_BALANCING
4850 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4851 pgdat
->numabalancing_migrate_nr_pages
= 0;
4852 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4854 init_waitqueue_head(&pgdat
->kswapd_wait
);
4855 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4856 pgdat_page_cgroup_init(pgdat
);
4858 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4859 struct zone
*zone
= pgdat
->node_zones
+ j
;
4860 unsigned long size
, realsize
, freesize
, memmap_pages
;
4862 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4863 node_end_pfn
, zones_size
);
4864 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4870 * Adjust freesize so that it accounts for how much memory
4871 * is used by this zone for memmap. This affects the watermark
4872 * and per-cpu initialisations
4874 memmap_pages
= calc_memmap_size(size
, realsize
);
4875 if (freesize
>= memmap_pages
) {
4876 freesize
-= memmap_pages
;
4879 " %s zone: %lu pages used for memmap\n",
4880 zone_names
[j
], memmap_pages
);
4883 " %s zone: %lu pages exceeds freesize %lu\n",
4884 zone_names
[j
], memmap_pages
, freesize
);
4886 /* Account for reserved pages */
4887 if (j
== 0 && freesize
> dma_reserve
) {
4888 freesize
-= dma_reserve
;
4889 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4890 zone_names
[0], dma_reserve
);
4893 if (!is_highmem_idx(j
))
4894 nr_kernel_pages
+= freesize
;
4895 /* Charge for highmem memmap if there are enough kernel pages */
4896 else if (nr_kernel_pages
> memmap_pages
* 2)
4897 nr_kernel_pages
-= memmap_pages
;
4898 nr_all_pages
+= freesize
;
4900 zone
->spanned_pages
= size
;
4901 zone
->present_pages
= realsize
;
4903 * Set an approximate value for lowmem here, it will be adjusted
4904 * when the bootmem allocator frees pages into the buddy system.
4905 * And all highmem pages will be managed by the buddy system.
4907 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4910 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4912 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4914 zone
->name
= zone_names
[j
];
4915 spin_lock_init(&zone
->lock
);
4916 spin_lock_init(&zone
->lru_lock
);
4917 zone_seqlock_init(zone
);
4918 zone
->zone_pgdat
= pgdat
;
4919 zone_pcp_init(zone
);
4921 /* For bootup, initialized properly in watermark setup */
4922 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4924 lruvec_init(&zone
->lruvec
);
4928 set_pageblock_order();
4929 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4930 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4931 size
, MEMMAP_EARLY
);
4933 memmap_init(size
, nid
, j
, zone_start_pfn
);
4934 zone_start_pfn
+= size
;
4938 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4940 /* Skip empty nodes */
4941 if (!pgdat
->node_spanned_pages
)
4944 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4945 /* ia64 gets its own node_mem_map, before this, without bootmem */
4946 if (!pgdat
->node_mem_map
) {
4947 unsigned long size
, start
, end
;
4951 * The zone's endpoints aren't required to be MAX_ORDER
4952 * aligned but the node_mem_map endpoints must be in order
4953 * for the buddy allocator to function correctly.
4955 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4956 end
= pgdat_end_pfn(pgdat
);
4957 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4958 size
= (end
- start
) * sizeof(struct page
);
4959 map
= alloc_remap(pgdat
->node_id
, size
);
4961 map
= memblock_virt_alloc_node_nopanic(size
,
4963 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4965 #ifndef CONFIG_NEED_MULTIPLE_NODES
4967 * With no DISCONTIG, the global mem_map is just set as node 0's
4969 if (pgdat
== NODE_DATA(0)) {
4970 mem_map
= NODE_DATA(0)->node_mem_map
;
4971 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4972 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4973 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4974 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4977 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4980 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4981 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4983 pg_data_t
*pgdat
= NODE_DATA(nid
);
4984 unsigned long start_pfn
= 0;
4985 unsigned long end_pfn
= 0;
4987 /* pg_data_t should be reset to zero when it's allocated */
4988 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4990 pgdat
->node_id
= nid
;
4991 pgdat
->node_start_pfn
= node_start_pfn
;
4992 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4993 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4994 printk(KERN_INFO
"Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid
,
4995 (u64
) start_pfn
<< PAGE_SHIFT
, (u64
) (end_pfn
<< PAGE_SHIFT
) - 1);
4997 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
4998 zones_size
, zholes_size
);
5000 alloc_node_mem_map(pgdat
);
5001 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5002 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5003 nid
, (unsigned long)pgdat
,
5004 (unsigned long)pgdat
->node_mem_map
);
5007 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
5008 zones_size
, zholes_size
);
5011 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5013 #if MAX_NUMNODES > 1
5015 * Figure out the number of possible node ids.
5017 void __init
setup_nr_node_ids(void)
5020 unsigned int highest
= 0;
5022 for_each_node_mask(node
, node_possible_map
)
5024 nr_node_ids
= highest
+ 1;
5029 * node_map_pfn_alignment - determine the maximum internode alignment
5031 * This function should be called after node map is populated and sorted.
5032 * It calculates the maximum power of two alignment which can distinguish
5035 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5036 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5037 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5038 * shifted, 1GiB is enough and this function will indicate so.
5040 * This is used to test whether pfn -> nid mapping of the chosen memory
5041 * model has fine enough granularity to avoid incorrect mapping for the
5042 * populated node map.
5044 * Returns the determined alignment in pfn's. 0 if there is no alignment
5045 * requirement (single node).
5047 unsigned long __init
node_map_pfn_alignment(void)
5049 unsigned long accl_mask
= 0, last_end
= 0;
5050 unsigned long start
, end
, mask
;
5054 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5055 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5062 * Start with a mask granular enough to pin-point to the
5063 * start pfn and tick off bits one-by-one until it becomes
5064 * too coarse to separate the current node from the last.
5066 mask
= ~((1 << __ffs(start
)) - 1);
5067 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5070 /* accumulate all internode masks */
5074 /* convert mask to number of pages */
5075 return ~accl_mask
+ 1;
5078 /* Find the lowest pfn for a node */
5079 static unsigned long __init
find_min_pfn_for_node(int nid
)
5081 unsigned long min_pfn
= ULONG_MAX
;
5082 unsigned long start_pfn
;
5085 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5086 min_pfn
= min(min_pfn
, start_pfn
);
5088 if (min_pfn
== ULONG_MAX
) {
5090 "Could not find start_pfn for node %d\n", nid
);
5098 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5100 * It returns the minimum PFN based on information provided via
5101 * memblock_set_node().
5103 unsigned long __init
find_min_pfn_with_active_regions(void)
5105 return find_min_pfn_for_node(MAX_NUMNODES
);
5109 * early_calculate_totalpages()
5110 * Sum pages in active regions for movable zone.
5111 * Populate N_MEMORY for calculating usable_nodes.
5113 static unsigned long __init
early_calculate_totalpages(void)
5115 unsigned long totalpages
= 0;
5116 unsigned long start_pfn
, end_pfn
;
5119 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5120 unsigned long pages
= end_pfn
- start_pfn
;
5122 totalpages
+= pages
;
5124 node_set_state(nid
, N_MEMORY
);
5130 * Find the PFN the Movable zone begins in each node. Kernel memory
5131 * is spread evenly between nodes as long as the nodes have enough
5132 * memory. When they don't, some nodes will have more kernelcore than
5135 static void __init
find_zone_movable_pfns_for_nodes(void)
5138 unsigned long usable_startpfn
;
5139 unsigned long kernelcore_node
, kernelcore_remaining
;
5140 /* save the state before borrow the nodemask */
5141 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5142 unsigned long totalpages
= early_calculate_totalpages();
5143 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5144 struct memblock_region
*r
;
5146 /* Need to find movable_zone earlier when movable_node is specified. */
5147 find_usable_zone_for_movable();
5150 * If movable_node is specified, ignore kernelcore and movablecore
5153 if (movable_node_is_enabled()) {
5154 for_each_memblock(memory
, r
) {
5155 if (!memblock_is_hotpluggable(r
))
5160 usable_startpfn
= PFN_DOWN(r
->base
);
5161 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5162 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5170 * If movablecore=nn[KMG] was specified, calculate what size of
5171 * kernelcore that corresponds so that memory usable for
5172 * any allocation type is evenly spread. If both kernelcore
5173 * and movablecore are specified, then the value of kernelcore
5174 * will be used for required_kernelcore if it's greater than
5175 * what movablecore would have allowed.
5177 if (required_movablecore
) {
5178 unsigned long corepages
;
5181 * Round-up so that ZONE_MOVABLE is at least as large as what
5182 * was requested by the user
5184 required_movablecore
=
5185 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5186 corepages
= totalpages
- required_movablecore
;
5188 required_kernelcore
= max(required_kernelcore
, corepages
);
5191 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5192 if (!required_kernelcore
)
5195 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5196 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5199 /* Spread kernelcore memory as evenly as possible throughout nodes */
5200 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5201 for_each_node_state(nid
, N_MEMORY
) {
5202 unsigned long start_pfn
, end_pfn
;
5205 * Recalculate kernelcore_node if the division per node
5206 * now exceeds what is necessary to satisfy the requested
5207 * amount of memory for the kernel
5209 if (required_kernelcore
< kernelcore_node
)
5210 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5213 * As the map is walked, we track how much memory is usable
5214 * by the kernel using kernelcore_remaining. When it is
5215 * 0, the rest of the node is usable by ZONE_MOVABLE
5217 kernelcore_remaining
= kernelcore_node
;
5219 /* Go through each range of PFNs within this node */
5220 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5221 unsigned long size_pages
;
5223 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5224 if (start_pfn
>= end_pfn
)
5227 /* Account for what is only usable for kernelcore */
5228 if (start_pfn
< usable_startpfn
) {
5229 unsigned long kernel_pages
;
5230 kernel_pages
= min(end_pfn
, usable_startpfn
)
5233 kernelcore_remaining
-= min(kernel_pages
,
5234 kernelcore_remaining
);
5235 required_kernelcore
-= min(kernel_pages
,
5236 required_kernelcore
);
5238 /* Continue if range is now fully accounted */
5239 if (end_pfn
<= usable_startpfn
) {
5242 * Push zone_movable_pfn to the end so
5243 * that if we have to rebalance
5244 * kernelcore across nodes, we will
5245 * not double account here
5247 zone_movable_pfn
[nid
] = end_pfn
;
5250 start_pfn
= usable_startpfn
;
5254 * The usable PFN range for ZONE_MOVABLE is from
5255 * start_pfn->end_pfn. Calculate size_pages as the
5256 * number of pages used as kernelcore
5258 size_pages
= end_pfn
- start_pfn
;
5259 if (size_pages
> kernelcore_remaining
)
5260 size_pages
= kernelcore_remaining
;
5261 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5264 * Some kernelcore has been met, update counts and
5265 * break if the kernelcore for this node has been
5268 required_kernelcore
-= min(required_kernelcore
,
5270 kernelcore_remaining
-= size_pages
;
5271 if (!kernelcore_remaining
)
5277 * If there is still required_kernelcore, we do another pass with one
5278 * less node in the count. This will push zone_movable_pfn[nid] further
5279 * along on the nodes that still have memory until kernelcore is
5283 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5287 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5288 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5289 zone_movable_pfn
[nid
] =
5290 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5293 /* restore the node_state */
5294 node_states
[N_MEMORY
] = saved_node_state
;
5297 /* Any regular or high memory on that node ? */
5298 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5300 enum zone_type zone_type
;
5302 if (N_MEMORY
== N_NORMAL_MEMORY
)
5305 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5306 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5307 if (populated_zone(zone
)) {
5308 node_set_state(nid
, N_HIGH_MEMORY
);
5309 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5310 zone_type
<= ZONE_NORMAL
)
5311 node_set_state(nid
, N_NORMAL_MEMORY
);
5318 * free_area_init_nodes - Initialise all pg_data_t and zone data
5319 * @max_zone_pfn: an array of max PFNs for each zone
5321 * This will call free_area_init_node() for each active node in the system.
5322 * Using the page ranges provided by memblock_set_node(), the size of each
5323 * zone in each node and their holes is calculated. If the maximum PFN
5324 * between two adjacent zones match, it is assumed that the zone is empty.
5325 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5326 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5327 * starts where the previous one ended. For example, ZONE_DMA32 starts
5328 * at arch_max_dma_pfn.
5330 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5332 unsigned long start_pfn
, end_pfn
;
5335 /* Record where the zone boundaries are */
5336 memset(arch_zone_lowest_possible_pfn
, 0,
5337 sizeof(arch_zone_lowest_possible_pfn
));
5338 memset(arch_zone_highest_possible_pfn
, 0,
5339 sizeof(arch_zone_highest_possible_pfn
));
5340 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5341 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5342 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5343 if (i
== ZONE_MOVABLE
)
5345 arch_zone_lowest_possible_pfn
[i
] =
5346 arch_zone_highest_possible_pfn
[i
-1];
5347 arch_zone_highest_possible_pfn
[i
] =
5348 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5350 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5351 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5353 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5354 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5355 find_zone_movable_pfns_for_nodes();
5357 /* Print out the zone ranges */
5358 pr_info("Zone ranges:\n");
5359 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5360 if (i
== ZONE_MOVABLE
)
5362 pr_info(" %-8s ", zone_names
[i
]);
5363 if (arch_zone_lowest_possible_pfn
[i
] ==
5364 arch_zone_highest_possible_pfn
[i
])
5367 pr_cont("[mem %0#10lx-%0#10lx]\n",
5368 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5369 (arch_zone_highest_possible_pfn
[i
]
5370 << PAGE_SHIFT
) - 1);
5373 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5374 pr_info("Movable zone start for each node\n");
5375 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5376 if (zone_movable_pfn
[i
])
5377 pr_info(" Node %d: %#010lx\n", i
,
5378 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5381 /* Print out the early node map */
5382 pr_info("Early memory node ranges\n");
5383 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5384 pr_info(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5385 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5387 /* Initialise every node */
5388 mminit_verify_pageflags_layout();
5389 setup_nr_node_ids();
5390 for_each_online_node(nid
) {
5391 pg_data_t
*pgdat
= NODE_DATA(nid
);
5392 free_area_init_node(nid
, NULL
,
5393 find_min_pfn_for_node(nid
), NULL
);
5395 /* Any memory on that node */
5396 if (pgdat
->node_present_pages
)
5397 node_set_state(nid
, N_MEMORY
);
5398 check_for_memory(pgdat
, nid
);
5402 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5404 unsigned long long coremem
;
5408 coremem
= memparse(p
, &p
);
5409 *core
= coremem
>> PAGE_SHIFT
;
5411 /* Paranoid check that UL is enough for the coremem value */
5412 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5418 * kernelcore=size sets the amount of memory for use for allocations that
5419 * cannot be reclaimed or migrated.
5421 static int __init
cmdline_parse_kernelcore(char *p
)
5423 return cmdline_parse_core(p
, &required_kernelcore
);
5427 * movablecore=size sets the amount of memory for use for allocations that
5428 * can be reclaimed or migrated.
5430 static int __init
cmdline_parse_movablecore(char *p
)
5432 return cmdline_parse_core(p
, &required_movablecore
);
5435 early_param("kernelcore", cmdline_parse_kernelcore
);
5436 early_param("movablecore", cmdline_parse_movablecore
);
5438 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5440 void adjust_managed_page_count(struct page
*page
, long count
)
5442 spin_lock(&managed_page_count_lock
);
5443 page_zone(page
)->managed_pages
+= count
;
5444 totalram_pages
+= count
;
5445 #ifdef CONFIG_HIGHMEM
5446 if (PageHighMem(page
))
5447 totalhigh_pages
+= count
;
5449 spin_unlock(&managed_page_count_lock
);
5451 EXPORT_SYMBOL(adjust_managed_page_count
);
5453 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5456 unsigned long pages
= 0;
5458 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5459 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5460 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5461 if ((unsigned int)poison
<= 0xFF)
5462 memset(pos
, poison
, PAGE_SIZE
);
5463 free_reserved_page(virt_to_page(pos
));
5467 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5468 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5472 EXPORT_SYMBOL(free_reserved_area
);
5474 #ifdef CONFIG_HIGHMEM
5475 void free_highmem_page(struct page
*page
)
5477 __free_reserved_page(page
);
5479 page_zone(page
)->managed_pages
++;
5485 void __init
mem_init_print_info(const char *str
)
5487 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5488 unsigned long init_code_size
, init_data_size
;
5490 physpages
= get_num_physpages();
5491 codesize
= _etext
- _stext
;
5492 datasize
= _edata
- _sdata
;
5493 rosize
= __end_rodata
- __start_rodata
;
5494 bss_size
= __bss_stop
- __bss_start
;
5495 init_data_size
= __init_end
- __init_begin
;
5496 init_code_size
= _einittext
- _sinittext
;
5499 * Detect special cases and adjust section sizes accordingly:
5500 * 1) .init.* may be embedded into .data sections
5501 * 2) .init.text.* may be out of [__init_begin, __init_end],
5502 * please refer to arch/tile/kernel/vmlinux.lds.S.
5503 * 3) .rodata.* may be embedded into .text or .data sections.
5505 #define adj_init_size(start, end, size, pos, adj) \
5507 if (start <= pos && pos < end && size > adj) \
5511 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5512 _sinittext
, init_code_size
);
5513 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5514 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5515 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5516 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5518 #undef adj_init_size
5520 pr_info("Memory: %luK/%luK available "
5521 "(%luK kernel code, %luK rwdata, %luK rodata, "
5522 "%luK init, %luK bss, %luK reserved"
5523 #ifdef CONFIG_HIGHMEM
5527 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5528 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5529 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5530 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5531 #ifdef CONFIG_HIGHMEM
5532 totalhigh_pages
<< (PAGE_SHIFT
-10),
5534 str
? ", " : "", str
? str
: "");
5538 * set_dma_reserve - set the specified number of pages reserved in the first zone
5539 * @new_dma_reserve: The number of pages to mark reserved
5541 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5542 * In the DMA zone, a significant percentage may be consumed by kernel image
5543 * and other unfreeable allocations which can skew the watermarks badly. This
5544 * function may optionally be used to account for unfreeable pages in the
5545 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5546 * smaller per-cpu batchsize.
5548 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5550 dma_reserve
= new_dma_reserve
;
5553 void __init
free_area_init(unsigned long *zones_size
)
5555 free_area_init_node(0, zones_size
,
5556 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5559 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5560 unsigned long action
, void *hcpu
)
5562 int cpu
= (unsigned long)hcpu
;
5564 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5565 lru_add_drain_cpu(cpu
);
5569 * Spill the event counters of the dead processor
5570 * into the current processors event counters.
5571 * This artificially elevates the count of the current
5574 vm_events_fold_cpu(cpu
);
5577 * Zero the differential counters of the dead processor
5578 * so that the vm statistics are consistent.
5580 * This is only okay since the processor is dead and cannot
5581 * race with what we are doing.
5583 cpu_vm_stats_fold(cpu
);
5588 void __init
page_alloc_init(void)
5590 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5594 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5595 * or min_free_kbytes changes.
5597 static void calculate_totalreserve_pages(void)
5599 struct pglist_data
*pgdat
;
5600 unsigned long reserve_pages
= 0;
5601 enum zone_type i
, j
;
5603 for_each_online_pgdat(pgdat
) {
5604 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5605 struct zone
*zone
= pgdat
->node_zones
+ i
;
5608 /* Find valid and maximum lowmem_reserve in the zone */
5609 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5610 if (zone
->lowmem_reserve
[j
] > max
)
5611 max
= zone
->lowmem_reserve
[j
];
5614 /* we treat the high watermark as reserved pages. */
5615 max
+= high_wmark_pages(zone
);
5617 if (max
> zone
->managed_pages
)
5618 max
= zone
->managed_pages
;
5619 reserve_pages
+= max
;
5621 * Lowmem reserves are not available to
5622 * GFP_HIGHUSER page cache allocations and
5623 * kswapd tries to balance zones to their high
5624 * watermark. As a result, neither should be
5625 * regarded as dirtyable memory, to prevent a
5626 * situation where reclaim has to clean pages
5627 * in order to balance the zones.
5629 zone
->dirty_balance_reserve
= max
;
5632 dirty_balance_reserve
= reserve_pages
;
5633 totalreserve_pages
= reserve_pages
;
5637 * setup_per_zone_lowmem_reserve - called whenever
5638 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5639 * has a correct pages reserved value, so an adequate number of
5640 * pages are left in the zone after a successful __alloc_pages().
5642 static void setup_per_zone_lowmem_reserve(void)
5644 struct pglist_data
*pgdat
;
5645 enum zone_type j
, idx
;
5647 for_each_online_pgdat(pgdat
) {
5648 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5649 struct zone
*zone
= pgdat
->node_zones
+ j
;
5650 unsigned long managed_pages
= zone
->managed_pages
;
5652 zone
->lowmem_reserve
[j
] = 0;
5656 struct zone
*lower_zone
;
5660 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5661 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5663 lower_zone
= pgdat
->node_zones
+ idx
;
5664 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5665 sysctl_lowmem_reserve_ratio
[idx
];
5666 managed_pages
+= lower_zone
->managed_pages
;
5671 /* update totalreserve_pages */
5672 calculate_totalreserve_pages();
5675 static void __setup_per_zone_wmarks(void)
5677 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5678 unsigned long lowmem_pages
= 0;
5680 unsigned long flags
;
5682 /* Calculate total number of !ZONE_HIGHMEM pages */
5683 for_each_zone(zone
) {
5684 if (!is_highmem(zone
))
5685 lowmem_pages
+= zone
->managed_pages
;
5688 for_each_zone(zone
) {
5691 spin_lock_irqsave(&zone
->lock
, flags
);
5692 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5693 do_div(tmp
, lowmem_pages
);
5694 if (is_highmem(zone
)) {
5696 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5697 * need highmem pages, so cap pages_min to a small
5700 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5701 * deltas controls asynch page reclaim, and so should
5702 * not be capped for highmem.
5704 unsigned long min_pages
;
5706 min_pages
= zone
->managed_pages
/ 1024;
5707 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5708 zone
->watermark
[WMARK_MIN
] = min_pages
;
5711 * If it's a lowmem zone, reserve a number of pages
5712 * proportionate to the zone's size.
5714 zone
->watermark
[WMARK_MIN
] = tmp
;
5717 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5718 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5720 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5721 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5722 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5724 setup_zone_migrate_reserve(zone
);
5725 spin_unlock_irqrestore(&zone
->lock
, flags
);
5728 /* update totalreserve_pages */
5729 calculate_totalreserve_pages();
5733 * setup_per_zone_wmarks - called when min_free_kbytes changes
5734 * or when memory is hot-{added|removed}
5736 * Ensures that the watermark[min,low,high] values for each zone are set
5737 * correctly with respect to min_free_kbytes.
5739 void setup_per_zone_wmarks(void)
5741 mutex_lock(&zonelists_mutex
);
5742 __setup_per_zone_wmarks();
5743 mutex_unlock(&zonelists_mutex
);
5747 * The inactive anon list should be small enough that the VM never has to
5748 * do too much work, but large enough that each inactive page has a chance
5749 * to be referenced again before it is swapped out.
5751 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5752 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5753 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5754 * the anonymous pages are kept on the inactive list.
5757 * memory ratio inactive anon
5758 * -------------------------------------
5767 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5769 unsigned int gb
, ratio
;
5771 /* Zone size in gigabytes */
5772 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5774 ratio
= int_sqrt(10 * gb
);
5778 zone
->inactive_ratio
= ratio
;
5781 static void __meminit
setup_per_zone_inactive_ratio(void)
5786 calculate_zone_inactive_ratio(zone
);
5790 * Initialise min_free_kbytes.
5792 * For small machines we want it small (128k min). For large machines
5793 * we want it large (64MB max). But it is not linear, because network
5794 * bandwidth does not increase linearly with machine size. We use
5796 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5797 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5813 int __meminit
init_per_zone_wmark_min(void)
5815 unsigned long lowmem_kbytes
;
5816 int new_min_free_kbytes
;
5818 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5819 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5821 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5822 min_free_kbytes
= new_min_free_kbytes
;
5823 if (min_free_kbytes
< 128)
5824 min_free_kbytes
= 128;
5825 if (min_free_kbytes
> 65536)
5826 min_free_kbytes
= 65536;
5828 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5829 new_min_free_kbytes
, user_min_free_kbytes
);
5831 setup_per_zone_wmarks();
5832 refresh_zone_stat_thresholds();
5833 setup_per_zone_lowmem_reserve();
5834 setup_per_zone_inactive_ratio();
5837 module_init(init_per_zone_wmark_min
)
5840 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5841 * that we can call two helper functions whenever min_free_kbytes
5844 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5845 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5849 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5854 user_min_free_kbytes
= min_free_kbytes
;
5855 setup_per_zone_wmarks();
5861 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5862 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5867 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5872 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5873 sysctl_min_unmapped_ratio
) / 100;
5877 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5878 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5883 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5888 zone
->min_slab_pages
= (zone
->managed_pages
*
5889 sysctl_min_slab_ratio
) / 100;
5895 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5896 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5897 * whenever sysctl_lowmem_reserve_ratio changes.
5899 * The reserve ratio obviously has absolutely no relation with the
5900 * minimum watermarks. The lowmem reserve ratio can only make sense
5901 * if in function of the boot time zone sizes.
5903 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5904 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5906 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5907 setup_per_zone_lowmem_reserve();
5912 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5913 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5914 * pagelist can have before it gets flushed back to buddy allocator.
5916 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5917 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5920 int old_percpu_pagelist_fraction
;
5923 mutex_lock(&pcp_batch_high_lock
);
5924 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5926 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5927 if (!write
|| ret
< 0)
5930 /* Sanity checking to avoid pcp imbalance */
5931 if (percpu_pagelist_fraction
&&
5932 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5933 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5939 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
5942 for_each_populated_zone(zone
) {
5945 for_each_possible_cpu(cpu
)
5946 pageset_set_high_and_batch(zone
,
5947 per_cpu_ptr(zone
->pageset
, cpu
));
5950 mutex_unlock(&pcp_batch_high_lock
);
5954 int hashdist
= HASHDIST_DEFAULT
;
5957 static int __init
set_hashdist(char *str
)
5961 hashdist
= simple_strtoul(str
, &str
, 0);
5964 __setup("hashdist=", set_hashdist
);
5968 * allocate a large system hash table from bootmem
5969 * - it is assumed that the hash table must contain an exact power-of-2
5970 * quantity of entries
5971 * - limit is the number of hash buckets, not the total allocation size
5973 void *__init
alloc_large_system_hash(const char *tablename
,
5974 unsigned long bucketsize
,
5975 unsigned long numentries
,
5978 unsigned int *_hash_shift
,
5979 unsigned int *_hash_mask
,
5980 unsigned long low_limit
,
5981 unsigned long high_limit
)
5983 unsigned long long max
= high_limit
;
5984 unsigned long log2qty
, size
;
5987 /* allow the kernel cmdline to have a say */
5989 /* round applicable memory size up to nearest megabyte */
5990 numentries
= nr_kernel_pages
;
5992 /* It isn't necessary when PAGE_SIZE >= 1MB */
5993 if (PAGE_SHIFT
< 20)
5994 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
5996 /* limit to 1 bucket per 2^scale bytes of low memory */
5997 if (scale
> PAGE_SHIFT
)
5998 numentries
>>= (scale
- PAGE_SHIFT
);
6000 numentries
<<= (PAGE_SHIFT
- scale
);
6002 /* Make sure we've got at least a 0-order allocation.. */
6003 if (unlikely(flags
& HASH_SMALL
)) {
6004 /* Makes no sense without HASH_EARLY */
6005 WARN_ON(!(flags
& HASH_EARLY
));
6006 if (!(numentries
>> *_hash_shift
)) {
6007 numentries
= 1UL << *_hash_shift
;
6008 BUG_ON(!numentries
);
6010 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6011 numentries
= PAGE_SIZE
/ bucketsize
;
6013 numentries
= roundup_pow_of_two(numentries
);
6015 /* limit allocation size to 1/16 total memory by default */
6017 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6018 do_div(max
, bucketsize
);
6020 max
= min(max
, 0x80000000ULL
);
6022 if (numentries
< low_limit
)
6023 numentries
= low_limit
;
6024 if (numentries
> max
)
6027 log2qty
= ilog2(numentries
);
6030 size
= bucketsize
<< log2qty
;
6031 if (flags
& HASH_EARLY
)
6032 table
= memblock_virt_alloc_nopanic(size
, 0);
6034 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6037 * If bucketsize is not a power-of-two, we may free
6038 * some pages at the end of hash table which
6039 * alloc_pages_exact() automatically does
6041 if (get_order(size
) < MAX_ORDER
) {
6042 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6043 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6046 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6049 panic("Failed to allocate %s hash table\n", tablename
);
6051 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6054 ilog2(size
) - PAGE_SHIFT
,
6058 *_hash_shift
= log2qty
;
6060 *_hash_mask
= (1 << log2qty
) - 1;
6065 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6066 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6069 #ifdef CONFIG_SPARSEMEM
6070 return __pfn_to_section(pfn
)->pageblock_flags
;
6072 return zone
->pageblock_flags
;
6073 #endif /* CONFIG_SPARSEMEM */
6076 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6078 #ifdef CONFIG_SPARSEMEM
6079 pfn
&= (PAGES_PER_SECTION
-1);
6080 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6082 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6083 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6084 #endif /* CONFIG_SPARSEMEM */
6088 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6089 * @page: The page within the block of interest
6090 * @pfn: The target page frame number
6091 * @end_bitidx: The last bit of interest to retrieve
6092 * @mask: mask of bits that the caller is interested in
6094 * Return: pageblock_bits flags
6096 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6097 unsigned long end_bitidx
,
6101 unsigned long *bitmap
;
6102 unsigned long bitidx
, word_bitidx
;
6105 zone
= page_zone(page
);
6106 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6107 bitidx
= pfn_to_bitidx(zone
, pfn
);
6108 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6109 bitidx
&= (BITS_PER_LONG
-1);
6111 word
= bitmap
[word_bitidx
];
6112 bitidx
+= end_bitidx
;
6113 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6117 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6118 * @page: The page within the block of interest
6119 * @flags: The flags to set
6120 * @pfn: The target page frame number
6121 * @end_bitidx: The last bit of interest
6122 * @mask: mask of bits that the caller is interested in
6124 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6126 unsigned long end_bitidx
,
6130 unsigned long *bitmap
;
6131 unsigned long bitidx
, word_bitidx
;
6132 unsigned long old_word
, word
;
6134 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6136 zone
= page_zone(page
);
6137 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6138 bitidx
= pfn_to_bitidx(zone
, pfn
);
6139 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6140 bitidx
&= (BITS_PER_LONG
-1);
6142 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6144 bitidx
+= end_bitidx
;
6145 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6146 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6148 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6150 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6151 if (word
== old_word
)
6158 * This function checks whether pageblock includes unmovable pages or not.
6159 * If @count is not zero, it is okay to include less @count unmovable pages
6161 * PageLRU check without isolation or lru_lock could race so that
6162 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6163 * expect this function should be exact.
6165 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6166 bool skip_hwpoisoned_pages
)
6168 unsigned long pfn
, iter
, found
;
6172 * For avoiding noise data, lru_add_drain_all() should be called
6173 * If ZONE_MOVABLE, the zone never contains unmovable pages
6175 if (zone_idx(zone
) == ZONE_MOVABLE
)
6177 mt
= get_pageblock_migratetype(page
);
6178 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6181 pfn
= page_to_pfn(page
);
6182 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6183 unsigned long check
= pfn
+ iter
;
6185 if (!pfn_valid_within(check
))
6188 page
= pfn_to_page(check
);
6191 * Hugepages are not in LRU lists, but they're movable.
6192 * We need not scan over tail pages bacause we don't
6193 * handle each tail page individually in migration.
6195 if (PageHuge(page
)) {
6196 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6201 * We can't use page_count without pin a page
6202 * because another CPU can free compound page.
6203 * This check already skips compound tails of THP
6204 * because their page->_count is zero at all time.
6206 if (!atomic_read(&page
->_count
)) {
6207 if (PageBuddy(page
))
6208 iter
+= (1 << page_order(page
)) - 1;
6213 * The HWPoisoned page may be not in buddy system, and
6214 * page_count() is not 0.
6216 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6222 * If there are RECLAIMABLE pages, we need to check it.
6223 * But now, memory offline itself doesn't call shrink_slab()
6224 * and it still to be fixed.
6227 * If the page is not RAM, page_count()should be 0.
6228 * we don't need more check. This is an _used_ not-movable page.
6230 * The problematic thing here is PG_reserved pages. PG_reserved
6231 * is set to both of a memory hole page and a _used_ kernel
6240 bool is_pageblock_removable_nolock(struct page
*page
)
6246 * We have to be careful here because we are iterating over memory
6247 * sections which are not zone aware so we might end up outside of
6248 * the zone but still within the section.
6249 * We have to take care about the node as well. If the node is offline
6250 * its NODE_DATA will be NULL - see page_zone.
6252 if (!node_online(page_to_nid(page
)))
6255 zone
= page_zone(page
);
6256 pfn
= page_to_pfn(page
);
6257 if (!zone_spans_pfn(zone
, pfn
))
6260 return !has_unmovable_pages(zone
, page
, 0, true);
6265 static unsigned long pfn_max_align_down(unsigned long pfn
)
6267 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6268 pageblock_nr_pages
) - 1);
6271 static unsigned long pfn_max_align_up(unsigned long pfn
)
6273 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6274 pageblock_nr_pages
));
6277 /* [start, end) must belong to a single zone. */
6278 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6279 unsigned long start
, unsigned long end
)
6281 /* This function is based on compact_zone() from compaction.c. */
6282 unsigned long nr_reclaimed
;
6283 unsigned long pfn
= start
;
6284 unsigned int tries
= 0;
6289 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6290 if (fatal_signal_pending(current
)) {
6295 if (list_empty(&cc
->migratepages
)) {
6296 cc
->nr_migratepages
= 0;
6297 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6303 } else if (++tries
== 5) {
6304 ret
= ret
< 0 ? ret
: -EBUSY
;
6308 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6310 cc
->nr_migratepages
-= nr_reclaimed
;
6312 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6313 NULL
, 0, cc
->mode
, MR_CMA
);
6316 putback_movable_pages(&cc
->migratepages
);
6323 * alloc_contig_range() -- tries to allocate given range of pages
6324 * @start: start PFN to allocate
6325 * @end: one-past-the-last PFN to allocate
6326 * @migratetype: migratetype of the underlaying pageblocks (either
6327 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6328 * in range must have the same migratetype and it must
6329 * be either of the two.
6331 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6332 * aligned, however it's the caller's responsibility to guarantee that
6333 * we are the only thread that changes migrate type of pageblocks the
6336 * The PFN range must belong to a single zone.
6338 * Returns zero on success or negative error code. On success all
6339 * pages which PFN is in [start, end) are allocated for the caller and
6340 * need to be freed with free_contig_range().
6342 int alloc_contig_range(unsigned long start
, unsigned long end
,
6343 unsigned migratetype
)
6345 unsigned long outer_start
, outer_end
;
6348 struct compact_control cc
= {
6349 .nr_migratepages
= 0,
6351 .zone
= page_zone(pfn_to_page(start
)),
6352 .mode
= MIGRATE_SYNC
,
6353 .ignore_skip_hint
= true,
6355 INIT_LIST_HEAD(&cc
.migratepages
);
6358 * What we do here is we mark all pageblocks in range as
6359 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6360 * have different sizes, and due to the way page allocator
6361 * work, we align the range to biggest of the two pages so
6362 * that page allocator won't try to merge buddies from
6363 * different pageblocks and change MIGRATE_ISOLATE to some
6364 * other migration type.
6366 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6367 * migrate the pages from an unaligned range (ie. pages that
6368 * we are interested in). This will put all the pages in
6369 * range back to page allocator as MIGRATE_ISOLATE.
6371 * When this is done, we take the pages in range from page
6372 * allocator removing them from the buddy system. This way
6373 * page allocator will never consider using them.
6375 * This lets us mark the pageblocks back as
6376 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6377 * aligned range but not in the unaligned, original range are
6378 * put back to page allocator so that buddy can use them.
6381 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6382 pfn_max_align_up(end
), migratetype
,
6387 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6392 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6393 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6394 * more, all pages in [start, end) are free in page allocator.
6395 * What we are going to do is to allocate all pages from
6396 * [start, end) (that is remove them from page allocator).
6398 * The only problem is that pages at the beginning and at the
6399 * end of interesting range may be not aligned with pages that
6400 * page allocator holds, ie. they can be part of higher order
6401 * pages. Because of this, we reserve the bigger range and
6402 * once this is done free the pages we are not interested in.
6404 * We don't have to hold zone->lock here because the pages are
6405 * isolated thus they won't get removed from buddy.
6408 lru_add_drain_all();
6409 drain_all_pages(cc
.zone
);
6412 outer_start
= start
;
6413 while (!PageBuddy(pfn_to_page(outer_start
))) {
6414 if (++order
>= MAX_ORDER
) {
6418 outer_start
&= ~0UL << order
;
6421 /* Make sure the range is really isolated. */
6422 if (test_pages_isolated(outer_start
, end
, false)) {
6423 pr_info("%s: [%lx, %lx) PFNs busy\n",
6424 __func__
, outer_start
, end
);
6429 /* Grab isolated pages from freelists. */
6430 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6436 /* Free head and tail (if any) */
6437 if (start
!= outer_start
)
6438 free_contig_range(outer_start
, start
- outer_start
);
6439 if (end
!= outer_end
)
6440 free_contig_range(end
, outer_end
- end
);
6443 undo_isolate_page_range(pfn_max_align_down(start
),
6444 pfn_max_align_up(end
), migratetype
);
6448 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6450 unsigned int count
= 0;
6452 for (; nr_pages
--; pfn
++) {
6453 struct page
*page
= pfn_to_page(pfn
);
6455 count
+= page_count(page
) != 1;
6458 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6462 #ifdef CONFIG_MEMORY_HOTPLUG
6464 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6465 * page high values need to be recalulated.
6467 void __meminit
zone_pcp_update(struct zone
*zone
)
6470 mutex_lock(&pcp_batch_high_lock
);
6471 for_each_possible_cpu(cpu
)
6472 pageset_set_high_and_batch(zone
,
6473 per_cpu_ptr(zone
->pageset
, cpu
));
6474 mutex_unlock(&pcp_batch_high_lock
);
6478 void zone_pcp_reset(struct zone
*zone
)
6480 unsigned long flags
;
6482 struct per_cpu_pageset
*pset
;
6484 /* avoid races with drain_pages() */
6485 local_irq_save(flags
);
6486 if (zone
->pageset
!= &boot_pageset
) {
6487 for_each_online_cpu(cpu
) {
6488 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6489 drain_zonestat(zone
, pset
);
6491 free_percpu(zone
->pageset
);
6492 zone
->pageset
= &boot_pageset
;
6494 local_irq_restore(flags
);
6497 #ifdef CONFIG_MEMORY_HOTREMOVE
6499 * All pages in the range must be isolated before calling this.
6502 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6506 unsigned int order
, i
;
6508 unsigned long flags
;
6509 /* find the first valid pfn */
6510 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6515 zone
= page_zone(pfn_to_page(pfn
));
6516 spin_lock_irqsave(&zone
->lock
, flags
);
6518 while (pfn
< end_pfn
) {
6519 if (!pfn_valid(pfn
)) {
6523 page
= pfn_to_page(pfn
);
6525 * The HWPoisoned page may be not in buddy system, and
6526 * page_count() is not 0.
6528 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6530 SetPageReserved(page
);
6534 BUG_ON(page_count(page
));
6535 BUG_ON(!PageBuddy(page
));
6536 order
= page_order(page
);
6537 #ifdef CONFIG_DEBUG_VM
6538 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6539 pfn
, 1 << order
, end_pfn
);
6541 list_del(&page
->lru
);
6542 rmv_page_order(page
);
6543 zone
->free_area
[order
].nr_free
--;
6544 for (i
= 0; i
< (1 << order
); i
++)
6545 SetPageReserved((page
+i
));
6546 pfn
+= (1 << order
);
6548 spin_unlock_irqrestore(&zone
->lock
, flags
);
6552 #ifdef CONFIG_MEMORY_FAILURE
6553 bool is_free_buddy_page(struct page
*page
)
6555 struct zone
*zone
= page_zone(page
);
6556 unsigned long pfn
= page_to_pfn(page
);
6557 unsigned long flags
;
6560 spin_lock_irqsave(&zone
->lock
, flags
);
6561 for (order
= 0; order
< MAX_ORDER
; order
++) {
6562 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6564 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6567 spin_unlock_irqrestore(&zone
->lock
, flags
);
6569 return order
< MAX_ORDER
;