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_ext.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_ext.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
62 #include <linux/page_owner.h>
64 #include <asm/sections.h>
65 #include <asm/tlbflush.h>
66 #include <asm/div64.h>
69 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
70 static DEFINE_MUTEX(pcp_batch_high_lock
);
71 #define MIN_PERCPU_PAGELIST_FRACTION (8)
73 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74 DEFINE_PER_CPU(int, numa_node
);
75 EXPORT_PER_CPU_SYMBOL(numa_node
);
78 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
85 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
86 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
87 int _node_numa_mem_
[MAX_NUMNODES
];
91 * Array of node states.
93 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
94 [N_POSSIBLE
] = NODE_MASK_ALL
,
95 [N_ONLINE
] = { { [0] = 1UL } },
97 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
99 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
101 #ifdef CONFIG_MOVABLE_NODE
102 [N_MEMORY
] = { { [0] = 1UL } },
104 [N_CPU
] = { { [0] = 1UL } },
107 EXPORT_SYMBOL(node_states
);
109 /* Protect totalram_pages and zone->managed_pages */
110 static DEFINE_SPINLOCK(managed_page_count_lock
);
112 unsigned long totalram_pages __read_mostly
;
113 unsigned long totalreserve_pages __read_mostly
;
114 unsigned long totalcma_pages __read_mostly
;
116 * When calculating the number of globally allowed dirty pages, there
117 * is a certain number of per-zone reserves that should not be
118 * considered dirtyable memory. This is the sum of those reserves
119 * over all existing zones that contribute dirtyable memory.
121 unsigned long dirty_balance_reserve __read_mostly
;
123 int percpu_pagelist_fraction
;
124 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
126 #ifdef CONFIG_PM_SLEEP
128 * The following functions are used by the suspend/hibernate code to temporarily
129 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
130 * while devices are suspended. To avoid races with the suspend/hibernate code,
131 * they should always be called with pm_mutex held (gfp_allowed_mask also should
132 * only be modified with pm_mutex held, unless the suspend/hibernate code is
133 * guaranteed not to run in parallel with that modification).
136 static gfp_t saved_gfp_mask
;
138 void pm_restore_gfp_mask(void)
140 WARN_ON(!mutex_is_locked(&pm_mutex
));
141 if (saved_gfp_mask
) {
142 gfp_allowed_mask
= saved_gfp_mask
;
147 void pm_restrict_gfp_mask(void)
149 WARN_ON(!mutex_is_locked(&pm_mutex
));
150 WARN_ON(saved_gfp_mask
);
151 saved_gfp_mask
= gfp_allowed_mask
;
152 gfp_allowed_mask
&= ~GFP_IOFS
;
155 bool pm_suspended_storage(void)
157 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
161 #endif /* CONFIG_PM_SLEEP */
163 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
164 int pageblock_order __read_mostly
;
167 static void __free_pages_ok(struct page
*page
, unsigned int order
);
170 * results with 256, 32 in the lowmem_reserve sysctl:
171 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
172 * 1G machine -> (16M dma, 784M normal, 224M high)
173 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
174 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
175 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
177 * TBD: should special case ZONE_DMA32 machines here - in those we normally
178 * don't need any ZONE_NORMAL reservation
180 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
181 #ifdef CONFIG_ZONE_DMA
184 #ifdef CONFIG_ZONE_DMA32
187 #ifdef CONFIG_HIGHMEM
193 EXPORT_SYMBOL(totalram_pages
);
195 static char * const zone_names
[MAX_NR_ZONES
] = {
196 #ifdef CONFIG_ZONE_DMA
199 #ifdef CONFIG_ZONE_DMA32
203 #ifdef CONFIG_HIGHMEM
209 int min_free_kbytes
= 1024;
210 int user_min_free_kbytes
= -1;
212 static unsigned long __meminitdata nr_kernel_pages
;
213 static unsigned long __meminitdata nr_all_pages
;
214 static unsigned long __meminitdata dma_reserve
;
216 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
217 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
218 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
219 static unsigned long __initdata required_kernelcore
;
220 static unsigned long __initdata required_movablecore
;
221 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
223 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
225 EXPORT_SYMBOL(movable_zone
);
226 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
229 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
230 int nr_online_nodes __read_mostly
= 1;
231 EXPORT_SYMBOL(nr_node_ids
);
232 EXPORT_SYMBOL(nr_online_nodes
);
235 int page_group_by_mobility_disabled __read_mostly
;
237 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
239 if (unlikely(page_group_by_mobility_disabled
&&
240 migratetype
< MIGRATE_PCPTYPES
))
241 migratetype
= MIGRATE_UNMOVABLE
;
243 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
244 PB_migrate
, PB_migrate_end
);
247 bool oom_killer_disabled __read_mostly
;
249 #ifdef CONFIG_DEBUG_VM
250 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
254 unsigned long pfn
= page_to_pfn(page
);
255 unsigned long sp
, start_pfn
;
258 seq
= zone_span_seqbegin(zone
);
259 start_pfn
= zone
->zone_start_pfn
;
260 sp
= zone
->spanned_pages
;
261 if (!zone_spans_pfn(zone
, pfn
))
263 } while (zone_span_seqretry(zone
, seq
));
266 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
267 pfn
, zone_to_nid(zone
), zone
->name
,
268 start_pfn
, start_pfn
+ sp
);
273 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
275 if (!pfn_valid_within(page_to_pfn(page
)))
277 if (zone
!= page_zone(page
))
283 * Temporary debugging check for pages not lying within a given zone.
285 static int bad_range(struct zone
*zone
, struct page
*page
)
287 if (page_outside_zone_boundaries(zone
, page
))
289 if (!page_is_consistent(zone
, page
))
295 static inline int bad_range(struct zone
*zone
, struct page
*page
)
301 static void bad_page(struct page
*page
, const char *reason
,
302 unsigned long bad_flags
)
304 static unsigned long resume
;
305 static unsigned long nr_shown
;
306 static unsigned long nr_unshown
;
308 /* Don't complain about poisoned pages */
309 if (PageHWPoison(page
)) {
310 page_mapcount_reset(page
); /* remove PageBuddy */
315 * Allow a burst of 60 reports, then keep quiet for that minute;
316 * or allow a steady drip of one report per second.
318 if (nr_shown
== 60) {
319 if (time_before(jiffies
, resume
)) {
325 "BUG: Bad page state: %lu messages suppressed\n",
332 resume
= jiffies
+ 60 * HZ
;
334 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
335 current
->comm
, page_to_pfn(page
));
336 dump_page_badflags(page
, reason
, bad_flags
);
341 /* Leave bad fields for debug, except PageBuddy could make trouble */
342 page_mapcount_reset(page
); /* remove PageBuddy */
343 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
347 * Higher-order pages are called "compound pages". They are structured thusly:
349 * The first PAGE_SIZE page is called the "head page".
351 * The remaining PAGE_SIZE pages are called "tail pages".
353 * All pages have PG_compound set. All tail pages have their ->first_page
354 * pointing at the head page.
356 * The first tail page's ->lru.next holds the address of the compound page's
357 * put_page() function. Its ->lru.prev holds the order of allocation.
358 * This usage means that zero-order pages may not be compound.
361 static void free_compound_page(struct page
*page
)
363 __free_pages_ok(page
, compound_order(page
));
366 void prep_compound_page(struct page
*page
, unsigned long order
)
369 int nr_pages
= 1 << order
;
371 set_compound_page_dtor(page
, free_compound_page
);
372 set_compound_order(page
, order
);
374 for (i
= 1; i
< nr_pages
; i
++) {
375 struct page
*p
= page
+ i
;
376 set_page_count(p
, 0);
377 p
->first_page
= page
;
378 /* Make sure p->first_page is always valid for PageTail() */
384 static inline void prep_zero_page(struct page
*page
, unsigned int order
,
390 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
391 * and __GFP_HIGHMEM from hard or soft interrupt context.
393 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
394 for (i
= 0; i
< (1 << order
); i
++)
395 clear_highpage(page
+ i
);
398 #ifdef CONFIG_DEBUG_PAGEALLOC
399 unsigned int _debug_guardpage_minorder
;
400 bool _debug_pagealloc_enabled __read_mostly
;
401 bool _debug_guardpage_enabled __read_mostly
;
403 static int __init
early_debug_pagealloc(char *buf
)
408 if (strcmp(buf
, "on") == 0)
409 _debug_pagealloc_enabled
= true;
413 early_param("debug_pagealloc", early_debug_pagealloc
);
415 static bool need_debug_guardpage(void)
417 /* If we don't use debug_pagealloc, we don't need guard page */
418 if (!debug_pagealloc_enabled())
424 static void init_debug_guardpage(void)
426 if (!debug_pagealloc_enabled())
429 _debug_guardpage_enabled
= true;
432 struct page_ext_operations debug_guardpage_ops
= {
433 .need
= need_debug_guardpage
,
434 .init
= init_debug_guardpage
,
437 static int __init
debug_guardpage_minorder_setup(char *buf
)
441 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
442 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
445 _debug_guardpage_minorder
= res
;
446 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
449 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
451 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
452 unsigned int order
, int migratetype
)
454 struct page_ext
*page_ext
;
456 if (!debug_guardpage_enabled())
459 page_ext
= lookup_page_ext(page
);
460 __set_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
462 INIT_LIST_HEAD(&page
->lru
);
463 set_page_private(page
, order
);
464 /* Guard pages are not available for any usage */
465 __mod_zone_freepage_state(zone
, -(1 << order
), migratetype
);
468 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
469 unsigned int order
, int migratetype
)
471 struct page_ext
*page_ext
;
473 if (!debug_guardpage_enabled())
476 page_ext
= lookup_page_ext(page
);
477 __clear_bit(PAGE_EXT_DEBUG_GUARD
, &page_ext
->flags
);
479 set_page_private(page
, 0);
480 if (!is_migrate_isolate(migratetype
))
481 __mod_zone_freepage_state(zone
, (1 << order
), migratetype
);
484 struct page_ext_operations debug_guardpage_ops
= { NULL
, };
485 static inline void set_page_guard(struct zone
*zone
, struct page
*page
,
486 unsigned int order
, int migratetype
) {}
487 static inline void clear_page_guard(struct zone
*zone
, struct page
*page
,
488 unsigned int order
, int migratetype
) {}
491 static inline void set_page_order(struct page
*page
, unsigned int order
)
493 set_page_private(page
, order
);
494 __SetPageBuddy(page
);
497 static inline void rmv_page_order(struct page
*page
)
499 __ClearPageBuddy(page
);
500 set_page_private(page
, 0);
504 * This function checks whether a page is free && is the buddy
505 * we can do coalesce a page and its buddy if
506 * (a) the buddy is not in a hole &&
507 * (b) the buddy is in the buddy system &&
508 * (c) a page and its buddy have the same order &&
509 * (d) a page and its buddy are in the same zone.
511 * For recording whether a page is in the buddy system, we set ->_mapcount
512 * PAGE_BUDDY_MAPCOUNT_VALUE.
513 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
514 * serialized by zone->lock.
516 * For recording page's order, we use page_private(page).
518 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
521 if (!pfn_valid_within(page_to_pfn(buddy
)))
524 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
525 if (page_zone_id(page
) != page_zone_id(buddy
))
528 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
533 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
535 * zone check is done late to avoid uselessly
536 * calculating zone/node ids for pages that could
539 if (page_zone_id(page
) != page_zone_id(buddy
))
542 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
550 * Freeing function for a buddy system allocator.
552 * The concept of a buddy system is to maintain direct-mapped table
553 * (containing bit values) for memory blocks of various "orders".
554 * The bottom level table contains the map for the smallest allocatable
555 * units of memory (here, pages), and each level above it describes
556 * pairs of units from the levels below, hence, "buddies".
557 * At a high level, all that happens here is marking the table entry
558 * at the bottom level available, and propagating the changes upward
559 * as necessary, plus some accounting needed to play nicely with other
560 * parts of the VM system.
561 * At each level, we keep a list of pages, which are heads of continuous
562 * free pages of length of (1 << order) and marked with _mapcount
563 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
565 * So when we are allocating or freeing one, we can derive the state of the
566 * other. That is, if we allocate a small block, and both were
567 * free, the remainder of the region must be split into blocks.
568 * If a block is freed, and its buddy is also free, then this
569 * triggers coalescing into a block of larger size.
574 static inline void __free_one_page(struct page
*page
,
576 struct zone
*zone
, unsigned int order
,
579 unsigned long page_idx
;
580 unsigned long combined_idx
;
581 unsigned long uninitialized_var(buddy_idx
);
583 int max_order
= MAX_ORDER
;
585 VM_BUG_ON(!zone_is_initialized(zone
));
586 VM_BUG_ON_PAGE(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
, page
);
588 VM_BUG_ON(migratetype
== -1);
589 if (is_migrate_isolate(migratetype
)) {
591 * We restrict max order of merging to prevent merge
592 * between freepages on isolate pageblock and normal
593 * pageblock. Without this, pageblock isolation
594 * could cause incorrect freepage accounting.
596 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
598 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
601 page_idx
= pfn
& ((1 << max_order
) - 1);
603 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
604 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
606 while (order
< max_order
- 1) {
607 buddy_idx
= __find_buddy_index(page_idx
, order
);
608 buddy
= page
+ (buddy_idx
- page_idx
);
609 if (!page_is_buddy(page
, buddy
, order
))
612 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
613 * merge with it and move up one order.
615 if (page_is_guard(buddy
)) {
616 clear_page_guard(zone
, buddy
, order
, migratetype
);
618 list_del(&buddy
->lru
);
619 zone
->free_area
[order
].nr_free
--;
620 rmv_page_order(buddy
);
622 combined_idx
= buddy_idx
& page_idx
;
623 page
= page
+ (combined_idx
- page_idx
);
624 page_idx
= combined_idx
;
627 set_page_order(page
, order
);
630 * If this is not the largest possible page, check if the buddy
631 * of the next-highest order is free. If it is, it's possible
632 * that pages are being freed that will coalesce soon. In case,
633 * that is happening, add the free page to the tail of the list
634 * so it's less likely to be used soon and more likely to be merged
635 * as a higher order page
637 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
638 struct page
*higher_page
, *higher_buddy
;
639 combined_idx
= buddy_idx
& page_idx
;
640 higher_page
= page
+ (combined_idx
- page_idx
);
641 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
642 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
643 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
644 list_add_tail(&page
->lru
,
645 &zone
->free_area
[order
].free_list
[migratetype
]);
650 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
652 zone
->free_area
[order
].nr_free
++;
655 static inline int free_pages_check(struct page
*page
)
657 const char *bad_reason
= NULL
;
658 unsigned long bad_flags
= 0;
660 if (unlikely(page_mapcount(page
)))
661 bad_reason
= "nonzero mapcount";
662 if (unlikely(page
->mapping
!= NULL
))
663 bad_reason
= "non-NULL mapping";
664 if (unlikely(atomic_read(&page
->_count
) != 0))
665 bad_reason
= "nonzero _count";
666 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
667 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
668 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
671 if (unlikely(page
->mem_cgroup
))
672 bad_reason
= "page still charged to cgroup";
674 if (unlikely(bad_reason
)) {
675 bad_page(page
, bad_reason
, bad_flags
);
678 page_cpupid_reset_last(page
);
679 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
680 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
685 * Frees a number of pages from the PCP lists
686 * Assumes all pages on list are in same zone, and of same order.
687 * count is the number of pages to free.
689 * If the zone was previously in an "all pages pinned" state then look to
690 * see if this freeing clears that state.
692 * And clear the zone's pages_scanned counter, to hold off the "all pages are
693 * pinned" detection logic.
695 static void free_pcppages_bulk(struct zone
*zone
, int count
,
696 struct per_cpu_pages
*pcp
)
701 unsigned long nr_scanned
;
703 spin_lock(&zone
->lock
);
704 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
706 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
710 struct list_head
*list
;
713 * Remove pages from lists in a round-robin fashion. A
714 * batch_free count is maintained that is incremented when an
715 * empty list is encountered. This is so more pages are freed
716 * off fuller lists instead of spinning excessively around empty
721 if (++migratetype
== MIGRATE_PCPTYPES
)
723 list
= &pcp
->lists
[migratetype
];
724 } while (list_empty(list
));
726 /* This is the only non-empty list. Free them all. */
727 if (batch_free
== MIGRATE_PCPTYPES
)
728 batch_free
= to_free
;
731 int mt
; /* migratetype of the to-be-freed page */
733 page
= list_entry(list
->prev
, struct page
, lru
);
734 /* must delete as __free_one_page list manipulates */
735 list_del(&page
->lru
);
736 mt
= get_freepage_migratetype(page
);
737 if (unlikely(has_isolate_pageblock(zone
)))
738 mt
= get_pageblock_migratetype(page
);
740 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
741 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
742 trace_mm_page_pcpu_drain(page
, 0, mt
);
743 } while (--to_free
&& --batch_free
&& !list_empty(list
));
745 spin_unlock(&zone
->lock
);
748 static void free_one_page(struct zone
*zone
,
749 struct page
*page
, unsigned long pfn
,
753 unsigned long nr_scanned
;
754 spin_lock(&zone
->lock
);
755 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
757 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
759 if (unlikely(has_isolate_pageblock(zone
) ||
760 is_migrate_isolate(migratetype
))) {
761 migratetype
= get_pfnblock_migratetype(page
, pfn
);
763 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
764 spin_unlock(&zone
->lock
);
767 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
772 VM_BUG_ON_PAGE(PageTail(page
), page
);
773 VM_BUG_ON_PAGE(PageHead(page
) && compound_order(page
) != order
, page
);
775 trace_mm_page_free(page
, order
);
776 kmemcheck_free_shadow(page
, order
);
779 page
->mapping
= NULL
;
780 for (i
= 0; i
< (1 << order
); i
++)
781 bad
+= free_pages_check(page
+ i
);
785 reset_page_owner(page
, order
);
787 if (!PageHighMem(page
)) {
788 debug_check_no_locks_freed(page_address(page
),
790 debug_check_no_obj_freed(page_address(page
),
793 arch_free_page(page
, order
);
794 kernel_map_pages(page
, 1 << order
, 0);
799 static void __free_pages_ok(struct page
*page
, unsigned int order
)
803 unsigned long pfn
= page_to_pfn(page
);
805 if (!free_pages_prepare(page
, order
))
808 migratetype
= get_pfnblock_migratetype(page
, pfn
);
809 local_irq_save(flags
);
810 __count_vm_events(PGFREE
, 1 << order
);
811 set_freepage_migratetype(page
, migratetype
);
812 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
813 local_irq_restore(flags
);
816 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
818 unsigned int nr_pages
= 1 << order
;
819 struct page
*p
= page
;
823 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
825 __ClearPageReserved(p
);
826 set_page_count(p
, 0);
828 __ClearPageReserved(p
);
829 set_page_count(p
, 0);
831 page_zone(page
)->managed_pages
+= nr_pages
;
832 set_page_refcounted(page
);
833 __free_pages(page
, order
);
837 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
838 void __init
init_cma_reserved_pageblock(struct page
*page
)
840 unsigned i
= pageblock_nr_pages
;
841 struct page
*p
= page
;
844 __ClearPageReserved(p
);
845 set_page_count(p
, 0);
848 set_pageblock_migratetype(page
, MIGRATE_CMA
);
850 if (pageblock_order
>= MAX_ORDER
) {
851 i
= pageblock_nr_pages
;
854 set_page_refcounted(p
);
855 __free_pages(p
, MAX_ORDER
- 1);
856 p
+= MAX_ORDER_NR_PAGES
;
857 } while (i
-= MAX_ORDER_NR_PAGES
);
859 set_page_refcounted(page
);
860 __free_pages(page
, pageblock_order
);
863 adjust_managed_page_count(page
, pageblock_nr_pages
);
868 * The order of subdivision here is critical for the IO subsystem.
869 * Please do not alter this order without good reasons and regression
870 * testing. Specifically, as large blocks of memory are subdivided,
871 * the order in which smaller blocks are delivered depends on the order
872 * they're subdivided in this function. This is the primary factor
873 * influencing the order in which pages are delivered to the IO
874 * subsystem according to empirical testing, and this is also justified
875 * by considering the behavior of a buddy system containing a single
876 * large block of memory acted on by a series of small allocations.
877 * This behavior is a critical factor in sglist merging's success.
881 static inline void expand(struct zone
*zone
, struct page
*page
,
882 int low
, int high
, struct free_area
*area
,
885 unsigned long size
= 1 << high
;
891 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
893 if (IS_ENABLED(CONFIG_DEBUG_PAGEALLOC
) &&
894 debug_guardpage_enabled() &&
895 high
< debug_guardpage_minorder()) {
897 * Mark as guard pages (or page), that will allow to
898 * merge back to allocator when buddy will be freed.
899 * Corresponding page table entries will not be touched,
900 * pages will stay not present in virtual address space
902 set_page_guard(zone
, &page
[size
], high
, migratetype
);
905 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
907 set_page_order(&page
[size
], high
);
912 * This page is about to be returned from the page allocator
914 static inline int check_new_page(struct page
*page
)
916 const char *bad_reason
= NULL
;
917 unsigned long bad_flags
= 0;
919 if (unlikely(page_mapcount(page
)))
920 bad_reason
= "nonzero mapcount";
921 if (unlikely(page
->mapping
!= NULL
))
922 bad_reason
= "non-NULL mapping";
923 if (unlikely(atomic_read(&page
->_count
) != 0))
924 bad_reason
= "nonzero _count";
925 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
926 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
927 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
930 if (unlikely(page
->mem_cgroup
))
931 bad_reason
= "page still charged to cgroup";
933 if (unlikely(bad_reason
)) {
934 bad_page(page
, bad_reason
, bad_flags
);
940 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
,
945 for (i
= 0; i
< (1 << order
); i
++) {
946 struct page
*p
= page
+ i
;
947 if (unlikely(check_new_page(p
)))
951 set_page_private(page
, 0);
952 set_page_refcounted(page
);
954 arch_alloc_page(page
, order
);
955 kernel_map_pages(page
, 1 << order
, 1);
957 if (gfp_flags
& __GFP_ZERO
)
958 prep_zero_page(page
, order
, gfp_flags
);
960 if (order
&& (gfp_flags
& __GFP_COMP
))
961 prep_compound_page(page
, order
);
963 set_page_owner(page
, order
, gfp_flags
);
966 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was necessary to
967 * allocate the page. The expectation is that the caller is taking
968 * steps that will free more memory. The caller should avoid the page
969 * being used for !PFMEMALLOC purposes.
971 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
977 * Go through the free lists for the given migratetype and remove
978 * the smallest available page from the freelists
981 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
984 unsigned int current_order
;
985 struct free_area
*area
;
988 /* Find a page of the appropriate size in the preferred list */
989 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
990 area
= &(zone
->free_area
[current_order
]);
991 if (list_empty(&area
->free_list
[migratetype
]))
994 page
= list_entry(area
->free_list
[migratetype
].next
,
996 list_del(&page
->lru
);
997 rmv_page_order(page
);
999 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1000 set_freepage_migratetype(page
, migratetype
);
1009 * This array describes the order lists are fallen back to when
1010 * the free lists for the desirable migrate type are depleted
1012 static int fallbacks
[MIGRATE_TYPES
][4] = {
1013 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1014 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
1016 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
1017 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
1019 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
1021 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
1022 #ifdef CONFIG_MEMORY_ISOLATION
1023 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
1028 * Move the free pages in a range to the free lists of the requested type.
1029 * Note that start_page and end_pages are not aligned on a pageblock
1030 * boundary. If alignment is required, use move_freepages_block()
1032 int move_freepages(struct zone
*zone
,
1033 struct page
*start_page
, struct page
*end_page
,
1037 unsigned long order
;
1038 int pages_moved
= 0;
1040 #ifndef CONFIG_HOLES_IN_ZONE
1042 * page_zone is not safe to call in this context when
1043 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1044 * anyway as we check zone boundaries in move_freepages_block().
1045 * Remove at a later date when no bug reports exist related to
1046 * grouping pages by mobility
1048 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1051 for (page
= start_page
; page
<= end_page
;) {
1052 /* Make sure we are not inadvertently changing nodes */
1053 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1055 if (!pfn_valid_within(page_to_pfn(page
))) {
1060 if (!PageBuddy(page
)) {
1065 order
= page_order(page
);
1066 list_move(&page
->lru
,
1067 &zone
->free_area
[order
].free_list
[migratetype
]);
1068 set_freepage_migratetype(page
, migratetype
);
1070 pages_moved
+= 1 << order
;
1076 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1079 unsigned long start_pfn
, end_pfn
;
1080 struct page
*start_page
, *end_page
;
1082 start_pfn
= page_to_pfn(page
);
1083 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1084 start_page
= pfn_to_page(start_pfn
);
1085 end_page
= start_page
+ pageblock_nr_pages
- 1;
1086 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1088 /* Do not cross zone boundaries */
1089 if (!zone_spans_pfn(zone
, start_pfn
))
1091 if (!zone_spans_pfn(zone
, end_pfn
))
1094 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1097 static void change_pageblock_range(struct page
*pageblock_page
,
1098 int start_order
, int migratetype
)
1100 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1102 while (nr_pageblocks
--) {
1103 set_pageblock_migratetype(pageblock_page
, migratetype
);
1104 pageblock_page
+= pageblock_nr_pages
;
1109 * If breaking a large block of pages, move all free pages to the preferred
1110 * allocation list. If falling back for a reclaimable kernel allocation, be
1111 * more aggressive about taking ownership of free pages.
1113 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1114 * nor move CMA pages to different free lists. We don't want unmovable pages
1115 * to be allocated from MIGRATE_CMA areas.
1117 * Returns the new migratetype of the pageblock (or the same old migratetype
1118 * if it was unchanged).
1120 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1121 int start_type
, int fallback_type
)
1123 int current_order
= page_order(page
);
1126 * When borrowing from MIGRATE_CMA, we need to release the excess
1127 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1128 * is set to CMA so it is returned to the correct freelist in case
1129 * the page ends up being not actually allocated from the pcp lists.
1131 if (is_migrate_cma(fallback_type
))
1132 return fallback_type
;
1134 /* Take ownership for orders >= pageblock_order */
1135 if (current_order
>= pageblock_order
) {
1136 change_pageblock_range(page
, current_order
, start_type
);
1140 if (current_order
>= pageblock_order
/ 2 ||
1141 start_type
== MIGRATE_RECLAIMABLE
||
1142 page_group_by_mobility_disabled
) {
1145 pages
= move_freepages_block(zone
, page
, start_type
);
1147 /* Claim the whole block if over half of it is free */
1148 if (pages
>= (1 << (pageblock_order
-1)) ||
1149 page_group_by_mobility_disabled
) {
1151 set_pageblock_migratetype(page
, start_type
);
1157 return fallback_type
;
1160 /* Remove an element from the buddy allocator from the fallback list */
1161 static inline struct page
*
1162 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1164 struct free_area
*area
;
1165 unsigned int current_order
;
1167 int migratetype
, new_type
, i
;
1169 /* Find the largest possible block of pages in the other list */
1170 for (current_order
= MAX_ORDER
-1;
1171 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1174 migratetype
= fallbacks
[start_migratetype
][i
];
1176 /* MIGRATE_RESERVE handled later if necessary */
1177 if (migratetype
== MIGRATE_RESERVE
)
1180 area
= &(zone
->free_area
[current_order
]);
1181 if (list_empty(&area
->free_list
[migratetype
]))
1184 page
= list_entry(area
->free_list
[migratetype
].next
,
1188 new_type
= try_to_steal_freepages(zone
, page
,
1192 /* Remove the page from the freelists */
1193 list_del(&page
->lru
);
1194 rmv_page_order(page
);
1196 expand(zone
, page
, order
, current_order
, area
,
1198 /* The freepage_migratetype may differ from pageblock's
1199 * migratetype depending on the decisions in
1200 * try_to_steal_freepages. This is OK as long as it does
1201 * not differ for MIGRATE_CMA type.
1203 set_freepage_migratetype(page
, new_type
);
1205 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1206 start_migratetype
, migratetype
, new_type
);
1216 * Do the hard work of removing an element from the buddy allocator.
1217 * Call me with the zone->lock already held.
1219 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1225 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1227 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1228 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1231 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1232 * is used because __rmqueue_smallest is an inline function
1233 * and we want just one call site
1236 migratetype
= MIGRATE_RESERVE
;
1241 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1246 * Obtain a specified number of elements from the buddy allocator, all under
1247 * a single hold of the lock, for efficiency. Add them to the supplied list.
1248 * Returns the number of new pages which were placed at *list.
1250 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1251 unsigned long count
, struct list_head
*list
,
1252 int migratetype
, bool cold
)
1256 spin_lock(&zone
->lock
);
1257 for (i
= 0; i
< count
; ++i
) {
1258 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1259 if (unlikely(page
== NULL
))
1263 * Split buddy pages returned by expand() are received here
1264 * in physical page order. The page is added to the callers and
1265 * list and the list head then moves forward. From the callers
1266 * perspective, the linked list is ordered by page number in
1267 * some conditions. This is useful for IO devices that can
1268 * merge IO requests if the physical pages are ordered
1272 list_add(&page
->lru
, list
);
1274 list_add_tail(&page
->lru
, list
);
1276 if (is_migrate_cma(get_freepage_migratetype(page
)))
1277 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1280 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1281 spin_unlock(&zone
->lock
);
1287 * Called from the vmstat counter updater to drain pagesets of this
1288 * currently executing processor on remote nodes after they have
1291 * Note that this function must be called with the thread pinned to
1292 * a single processor.
1294 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1296 unsigned long flags
;
1297 int to_drain
, batch
;
1299 local_irq_save(flags
);
1300 batch
= ACCESS_ONCE(pcp
->batch
);
1301 to_drain
= min(pcp
->count
, batch
);
1303 free_pcppages_bulk(zone
, to_drain
, pcp
);
1304 pcp
->count
-= to_drain
;
1306 local_irq_restore(flags
);
1311 * Drain pcplists of the indicated processor and zone.
1313 * The processor must either be the current processor and the
1314 * thread pinned to the current processor or a processor that
1317 static void drain_pages_zone(unsigned int cpu
, struct zone
*zone
)
1319 unsigned long flags
;
1320 struct per_cpu_pageset
*pset
;
1321 struct per_cpu_pages
*pcp
;
1323 local_irq_save(flags
);
1324 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1328 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1331 local_irq_restore(flags
);
1335 * Drain pcplists of all zones on the indicated processor.
1337 * The processor must either be the current processor and the
1338 * thread pinned to the current processor or a processor that
1341 static void drain_pages(unsigned int cpu
)
1345 for_each_populated_zone(zone
) {
1346 drain_pages_zone(cpu
, zone
);
1351 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1353 * The CPU has to be pinned. When zone parameter is non-NULL, spill just
1354 * the single zone's pages.
1356 void drain_local_pages(struct zone
*zone
)
1358 int cpu
= smp_processor_id();
1361 drain_pages_zone(cpu
, zone
);
1367 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1369 * When zone parameter is non-NULL, spill just the single zone's pages.
1371 * Note that this code is protected against sending an IPI to an offline
1372 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1373 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1374 * nothing keeps CPUs from showing up after we populated the cpumask and
1375 * before the call to on_each_cpu_mask().
1377 void drain_all_pages(struct zone
*zone
)
1382 * Allocate in the BSS so we wont require allocation in
1383 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1385 static cpumask_t cpus_with_pcps
;
1388 * We don't care about racing with CPU hotplug event
1389 * as offline notification will cause the notified
1390 * cpu to drain that CPU pcps and on_each_cpu_mask
1391 * disables preemption as part of its processing
1393 for_each_online_cpu(cpu
) {
1394 struct per_cpu_pageset
*pcp
;
1396 bool has_pcps
= false;
1399 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1403 for_each_populated_zone(z
) {
1404 pcp
= per_cpu_ptr(z
->pageset
, cpu
);
1405 if (pcp
->pcp
.count
) {
1413 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1415 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1417 on_each_cpu_mask(&cpus_with_pcps
, (smp_call_func_t
) drain_local_pages
,
1421 #ifdef CONFIG_HIBERNATION
1423 void mark_free_pages(struct zone
*zone
)
1425 unsigned long pfn
, max_zone_pfn
;
1426 unsigned long flags
;
1427 unsigned int order
, t
;
1428 struct list_head
*curr
;
1430 if (zone_is_empty(zone
))
1433 spin_lock_irqsave(&zone
->lock
, flags
);
1435 max_zone_pfn
= zone_end_pfn(zone
);
1436 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1437 if (pfn_valid(pfn
)) {
1438 struct page
*page
= pfn_to_page(pfn
);
1440 if (!swsusp_page_is_forbidden(page
))
1441 swsusp_unset_page_free(page
);
1444 for_each_migratetype_order(order
, t
) {
1445 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1448 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1449 for (i
= 0; i
< (1UL << order
); i
++)
1450 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1453 spin_unlock_irqrestore(&zone
->lock
, flags
);
1455 #endif /* CONFIG_PM */
1458 * Free a 0-order page
1459 * cold == true ? free a cold page : free a hot page
1461 void free_hot_cold_page(struct page
*page
, bool cold
)
1463 struct zone
*zone
= page_zone(page
);
1464 struct per_cpu_pages
*pcp
;
1465 unsigned long flags
;
1466 unsigned long pfn
= page_to_pfn(page
);
1469 if (!free_pages_prepare(page
, 0))
1472 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1473 set_freepage_migratetype(page
, migratetype
);
1474 local_irq_save(flags
);
1475 __count_vm_event(PGFREE
);
1478 * We only track unmovable, reclaimable and movable on pcp lists.
1479 * Free ISOLATE pages back to the allocator because they are being
1480 * offlined but treat RESERVE as movable pages so we can get those
1481 * areas back if necessary. Otherwise, we may have to free
1482 * excessively into the page allocator
1484 if (migratetype
>= MIGRATE_PCPTYPES
) {
1485 if (unlikely(is_migrate_isolate(migratetype
))) {
1486 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1489 migratetype
= MIGRATE_MOVABLE
;
1492 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1494 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1496 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1498 if (pcp
->count
>= pcp
->high
) {
1499 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1500 free_pcppages_bulk(zone
, batch
, pcp
);
1501 pcp
->count
-= batch
;
1505 local_irq_restore(flags
);
1509 * Free a list of 0-order pages
1511 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1513 struct page
*page
, *next
;
1515 list_for_each_entry_safe(page
, next
, list
, lru
) {
1516 trace_mm_page_free_batched(page
, cold
);
1517 free_hot_cold_page(page
, cold
);
1522 * split_page takes a non-compound higher-order page, and splits it into
1523 * n (1<<order) sub-pages: page[0..n]
1524 * Each sub-page must be freed individually.
1526 * Note: this is probably too low level an operation for use in drivers.
1527 * Please consult with lkml before using this in your driver.
1529 void split_page(struct page
*page
, unsigned int order
)
1533 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1534 VM_BUG_ON_PAGE(!page_count(page
), page
);
1536 #ifdef CONFIG_KMEMCHECK
1538 * Split shadow pages too, because free(page[0]) would
1539 * otherwise free the whole shadow.
1541 if (kmemcheck_page_is_tracked(page
))
1542 split_page(virt_to_page(page
[0].shadow
), order
);
1545 set_page_owner(page
, 0, 0);
1546 for (i
= 1; i
< (1 << order
); i
++) {
1547 set_page_refcounted(page
+ i
);
1548 set_page_owner(page
+ i
, 0, 0);
1551 EXPORT_SYMBOL_GPL(split_page
);
1553 int __isolate_free_page(struct page
*page
, unsigned int order
)
1555 unsigned long watermark
;
1559 BUG_ON(!PageBuddy(page
));
1561 zone
= page_zone(page
);
1562 mt
= get_pageblock_migratetype(page
);
1564 if (!is_migrate_isolate(mt
)) {
1565 /* Obey watermarks as if the page was being allocated */
1566 watermark
= low_wmark_pages(zone
) + (1 << order
);
1567 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1570 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1573 /* Remove page from free list */
1574 list_del(&page
->lru
);
1575 zone
->free_area
[order
].nr_free
--;
1576 rmv_page_order(page
);
1578 /* Set the pageblock if the isolated page is at least a pageblock */
1579 if (order
>= pageblock_order
- 1) {
1580 struct page
*endpage
= page
+ (1 << order
) - 1;
1581 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1582 int mt
= get_pageblock_migratetype(page
);
1583 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1584 set_pageblock_migratetype(page
,
1589 set_page_owner(page
, order
, 0);
1590 return 1UL << order
;
1594 * Similar to split_page except the page is already free. As this is only
1595 * being used for migration, the migratetype of the block also changes.
1596 * As this is called with interrupts disabled, the caller is responsible
1597 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1600 * Note: this is probably too low level an operation for use in drivers.
1601 * Please consult with lkml before using this in your driver.
1603 int split_free_page(struct page
*page
)
1608 order
= page_order(page
);
1610 nr_pages
= __isolate_free_page(page
, order
);
1614 /* Split into individual pages */
1615 set_page_refcounted(page
);
1616 split_page(page
, order
);
1621 * Allocate a page from the given zone. Use pcplists for order-0 allocations.
1624 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1625 struct zone
*zone
, unsigned int order
,
1626 gfp_t gfp_flags
, int migratetype
)
1628 unsigned long flags
;
1630 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1632 if (likely(order
== 0)) {
1633 struct per_cpu_pages
*pcp
;
1634 struct list_head
*list
;
1636 local_irq_save(flags
);
1637 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1638 list
= &pcp
->lists
[migratetype
];
1639 if (list_empty(list
)) {
1640 pcp
->count
+= rmqueue_bulk(zone
, 0,
1643 if (unlikely(list_empty(list
)))
1648 page
= list_entry(list
->prev
, struct page
, lru
);
1650 page
= list_entry(list
->next
, struct page
, lru
);
1652 list_del(&page
->lru
);
1655 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1657 * __GFP_NOFAIL is not to be used in new code.
1659 * All __GFP_NOFAIL callers should be fixed so that they
1660 * properly detect and handle allocation failures.
1662 * We most definitely don't want callers attempting to
1663 * allocate greater than order-1 page units with
1666 WARN_ON_ONCE(order
> 1);
1668 spin_lock_irqsave(&zone
->lock
, flags
);
1669 page
= __rmqueue(zone
, order
, migratetype
);
1670 spin_unlock(&zone
->lock
);
1673 __mod_zone_freepage_state(zone
, -(1 << order
),
1674 get_freepage_migratetype(page
));
1677 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1678 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1679 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1680 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1682 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1683 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1684 local_irq_restore(flags
);
1686 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1690 local_irq_restore(flags
);
1694 #ifdef CONFIG_FAIL_PAGE_ALLOC
1697 struct fault_attr attr
;
1699 u32 ignore_gfp_highmem
;
1700 u32 ignore_gfp_wait
;
1702 } fail_page_alloc
= {
1703 .attr
= FAULT_ATTR_INITIALIZER
,
1704 .ignore_gfp_wait
= 1,
1705 .ignore_gfp_highmem
= 1,
1709 static int __init
setup_fail_page_alloc(char *str
)
1711 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1713 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1715 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1717 if (order
< fail_page_alloc
.min_order
)
1719 if (gfp_mask
& __GFP_NOFAIL
)
1721 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1723 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1726 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1729 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1731 static int __init
fail_page_alloc_debugfs(void)
1733 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1736 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1737 &fail_page_alloc
.attr
);
1739 return PTR_ERR(dir
);
1741 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1742 &fail_page_alloc
.ignore_gfp_wait
))
1744 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1745 &fail_page_alloc
.ignore_gfp_highmem
))
1747 if (!debugfs_create_u32("min-order", mode
, dir
,
1748 &fail_page_alloc
.min_order
))
1753 debugfs_remove_recursive(dir
);
1758 late_initcall(fail_page_alloc_debugfs
);
1760 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1762 #else /* CONFIG_FAIL_PAGE_ALLOC */
1764 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1769 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1772 * Return true if free pages are above 'mark'. This takes into account the order
1773 * of the allocation.
1775 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1776 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1779 /* free_pages may go negative - that's OK */
1784 free_pages
-= (1 << order
) - 1;
1785 if (alloc_flags
& ALLOC_HIGH
)
1787 if (alloc_flags
& ALLOC_HARDER
)
1790 /* If allocation can't use CMA areas don't use free CMA pages */
1791 if (!(alloc_flags
& ALLOC_CMA
))
1792 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1795 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1797 for (o
= 0; o
< order
; o
++) {
1798 /* At the next order, this order's pages become unavailable */
1799 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1801 /* Require fewer higher order pages to be free */
1804 if (free_pages
<= min
)
1810 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1811 int classzone_idx
, int alloc_flags
)
1813 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1814 zone_page_state(z
, NR_FREE_PAGES
));
1817 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1818 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1820 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1822 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1823 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1825 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1831 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1832 * skip over zones that are not allowed by the cpuset, or that have
1833 * been recently (in last second) found to be nearly full. See further
1834 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1835 * that have to skip over a lot of full or unallowed zones.
1837 * If the zonelist cache is present in the passed zonelist, then
1838 * returns a pointer to the allowed node mask (either the current
1839 * tasks mems_allowed, or node_states[N_MEMORY].)
1841 * If the zonelist cache is not available for this zonelist, does
1842 * nothing and returns NULL.
1844 * If the fullzones BITMAP in the zonelist cache is stale (more than
1845 * a second since last zap'd) then we zap it out (clear its bits.)
1847 * We hold off even calling zlc_setup, until after we've checked the
1848 * first zone in the zonelist, on the theory that most allocations will
1849 * be satisfied from that first zone, so best to examine that zone as
1850 * quickly as we can.
1852 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1854 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1855 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1857 zlc
= zonelist
->zlcache_ptr
;
1861 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1862 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1863 zlc
->last_full_zap
= jiffies
;
1866 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1867 &cpuset_current_mems_allowed
:
1868 &node_states
[N_MEMORY
];
1869 return allowednodes
;
1873 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1874 * if it is worth looking at further for free memory:
1875 * 1) Check that the zone isn't thought to be full (doesn't have its
1876 * bit set in the zonelist_cache fullzones BITMAP).
1877 * 2) Check that the zones node (obtained from the zonelist_cache
1878 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1879 * Return true (non-zero) if zone is worth looking at further, or
1880 * else return false (zero) if it is not.
1882 * This check -ignores- the distinction between various watermarks,
1883 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1884 * found to be full for any variation of these watermarks, it will
1885 * be considered full for up to one second by all requests, unless
1886 * we are so low on memory on all allowed nodes that we are forced
1887 * into the second scan of the zonelist.
1889 * In the second scan we ignore this zonelist cache and exactly
1890 * apply the watermarks to all zones, even it is slower to do so.
1891 * We are low on memory in the second scan, and should leave no stone
1892 * unturned looking for a free page.
1894 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1895 nodemask_t
*allowednodes
)
1897 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1898 int i
; /* index of *z in zonelist zones */
1899 int n
; /* node that zone *z is on */
1901 zlc
= zonelist
->zlcache_ptr
;
1905 i
= z
- zonelist
->_zonerefs
;
1908 /* This zone is worth trying if it is allowed but not full */
1909 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1913 * Given 'z' scanning a zonelist, set the corresponding bit in
1914 * zlc->fullzones, so that subsequent attempts to allocate a page
1915 * from that zone don't waste time re-examining it.
1917 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1919 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1920 int i
; /* index of *z in zonelist zones */
1922 zlc
= zonelist
->zlcache_ptr
;
1926 i
= z
- zonelist
->_zonerefs
;
1928 set_bit(i
, zlc
->fullzones
);
1932 * clear all zones full, called after direct reclaim makes progress so that
1933 * a zone that was recently full is not skipped over for up to a second
1935 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1937 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1939 zlc
= zonelist
->zlcache_ptr
;
1943 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1946 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1948 return local_zone
->node
== zone
->node
;
1951 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1953 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1957 #else /* CONFIG_NUMA */
1959 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1964 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1965 nodemask_t
*allowednodes
)
1970 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1974 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1978 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1983 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1988 #endif /* CONFIG_NUMA */
1990 static void reset_alloc_batches(struct zone
*preferred_zone
)
1992 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
1995 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
1996 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
1997 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
1998 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1999 } while (zone
++ != preferred_zone
);
2003 * get_page_from_freelist goes through the zonelist trying to allocate
2006 static struct page
*
2007 get_page_from_freelist(gfp_t gfp_mask
, unsigned int order
, int alloc_flags
,
2008 const struct alloc_context
*ac
)
2010 struct zonelist
*zonelist
= ac
->zonelist
;
2012 struct page
*page
= NULL
;
2014 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
2015 int zlc_active
= 0; /* set if using zonelist_cache */
2016 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
2017 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
2018 (gfp_mask
& __GFP_WRITE
);
2019 int nr_fair_skipped
= 0;
2020 bool zonelist_rescan
;
2023 zonelist_rescan
= false;
2026 * Scan zonelist, looking for a zone with enough free.
2027 * See also __cpuset_node_allowed() comment in kernel/cpuset.c.
2029 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
, ac
->high_zoneidx
,
2033 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2034 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2036 if (cpusets_enabled() &&
2037 (alloc_flags
& ALLOC_CPUSET
) &&
2038 !cpuset_zone_allowed(zone
, gfp_mask
))
2041 * Distribute pages in proportion to the individual
2042 * zone size to ensure fair page aging. The zone a
2043 * page was allocated in should have no effect on the
2044 * time the page has in memory before being reclaimed.
2046 if (alloc_flags
& ALLOC_FAIR
) {
2047 if (!zone_local(ac
->preferred_zone
, zone
))
2049 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2055 * When allocating a page cache page for writing, we
2056 * want to get it from a zone that is within its dirty
2057 * limit, such that no single zone holds more than its
2058 * proportional share of globally allowed dirty pages.
2059 * The dirty limits take into account the zone's
2060 * lowmem reserves and high watermark so that kswapd
2061 * should be able to balance it without having to
2062 * write pages from its LRU list.
2064 * This may look like it could increase pressure on
2065 * lower zones by failing allocations in higher zones
2066 * before they are full. But the pages that do spill
2067 * over are limited as the lower zones are protected
2068 * by this very same mechanism. It should not become
2069 * a practical burden to them.
2071 * XXX: For now, allow allocations to potentially
2072 * exceed the per-zone dirty limit in the slowpath
2073 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2074 * which is important when on a NUMA setup the allowed
2075 * zones are together not big enough to reach the
2076 * global limit. The proper fix for these situations
2077 * will require awareness of zones in the
2078 * dirty-throttling and the flusher threads.
2080 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2083 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2084 if (!zone_watermark_ok(zone
, order
, mark
,
2085 ac
->classzone_idx
, alloc_flags
)) {
2088 /* Checked here to keep the fast path fast */
2089 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2090 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2093 if (IS_ENABLED(CONFIG_NUMA
) &&
2094 !did_zlc_setup
&& nr_online_nodes
> 1) {
2096 * we do zlc_setup if there are multiple nodes
2097 * and before considering the first zone allowed
2100 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2105 if (zone_reclaim_mode
== 0 ||
2106 !zone_allows_reclaim(ac
->preferred_zone
, zone
))
2107 goto this_zone_full
;
2110 * As we may have just activated ZLC, check if the first
2111 * eligible zone has failed zone_reclaim recently.
2113 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2114 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2117 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2119 case ZONE_RECLAIM_NOSCAN
:
2122 case ZONE_RECLAIM_FULL
:
2123 /* scanned but unreclaimable */
2126 /* did we reclaim enough */
2127 if (zone_watermark_ok(zone
, order
, mark
,
2128 ac
->classzone_idx
, alloc_flags
))
2132 * Failed to reclaim enough to meet watermark.
2133 * Only mark the zone full if checking the min
2134 * watermark or if we failed to reclaim just
2135 * 1<<order pages or else the page allocator
2136 * fastpath will prematurely mark zones full
2137 * when the watermark is between the low and
2140 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2141 ret
== ZONE_RECLAIM_SOME
)
2142 goto this_zone_full
;
2149 page
= buffered_rmqueue(ac
->preferred_zone
, zone
, order
,
2150 gfp_mask
, ac
->migratetype
);
2152 if (prep_new_page(page
, order
, gfp_mask
, alloc_flags
))
2157 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2158 zlc_mark_zone_full(zonelist
, z
);
2162 * The first pass makes sure allocations are spread fairly within the
2163 * local node. However, the local node might have free pages left
2164 * after the fairness batches are exhausted, and remote zones haven't
2165 * even been considered yet. Try once more without fairness, and
2166 * include remote zones now, before entering the slowpath and waking
2167 * kswapd: prefer spilling to a remote zone over swapping locally.
2169 if (alloc_flags
& ALLOC_FAIR
) {
2170 alloc_flags
&= ~ALLOC_FAIR
;
2171 if (nr_fair_skipped
) {
2172 zonelist_rescan
= true;
2173 reset_alloc_batches(ac
->preferred_zone
);
2175 if (nr_online_nodes
> 1)
2176 zonelist_rescan
= true;
2179 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2180 /* Disable zlc cache for second zonelist scan */
2182 zonelist_rescan
= true;
2185 if (zonelist_rescan
)
2192 * Large machines with many possible nodes should not always dump per-node
2193 * meminfo in irq context.
2195 static inline bool should_suppress_show_mem(void)
2200 ret
= in_interrupt();
2205 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2206 DEFAULT_RATELIMIT_INTERVAL
,
2207 DEFAULT_RATELIMIT_BURST
);
2209 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2211 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2213 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2214 debug_guardpage_minorder() > 0)
2218 * This documents exceptions given to allocations in certain
2219 * contexts that are allowed to allocate outside current's set
2222 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2223 if (test_thread_flag(TIF_MEMDIE
) ||
2224 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2225 filter
&= ~SHOW_MEM_FILTER_NODES
;
2226 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2227 filter
&= ~SHOW_MEM_FILTER_NODES
;
2230 struct va_format vaf
;
2233 va_start(args
, fmt
);
2238 pr_warn("%pV", &vaf
);
2243 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2244 current
->comm
, order
, gfp_mask
);
2247 if (!should_suppress_show_mem())
2252 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2253 unsigned long did_some_progress
,
2254 unsigned long pages_reclaimed
)
2256 /* Do not loop if specifically requested */
2257 if (gfp_mask
& __GFP_NORETRY
)
2260 /* Always retry if specifically requested */
2261 if (gfp_mask
& __GFP_NOFAIL
)
2265 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2266 * making forward progress without invoking OOM. Suspend also disables
2267 * storage devices so kswapd will not help. Bail if we are suspending.
2269 if (!did_some_progress
&& pm_suspended_storage())
2273 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2274 * means __GFP_NOFAIL, but that may not be true in other
2277 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2281 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2282 * specified, then we retry until we no longer reclaim any pages
2283 * (above), or we've reclaimed an order of pages at least as
2284 * large as the allocation's order. In both cases, if the
2285 * allocation still fails, we stop retrying.
2287 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2293 static inline struct page
*
2294 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2295 const struct alloc_context
*ac
, unsigned long *did_some_progress
)
2299 *did_some_progress
= 0;
2301 if (oom_killer_disabled
)
2305 * Acquire the per-zone oom lock for each zone. If that
2306 * fails, somebody else is making progress for us.
2308 if (!oom_zonelist_trylock(ac
->zonelist
, gfp_mask
)) {
2309 *did_some_progress
= 1;
2310 schedule_timeout_uninterruptible(1);
2315 * PM-freezer should be notified that there might be an OOM killer on
2316 * its way to kill and wake somebody up. This is too early and we might
2317 * end up not killing anything but false positives are acceptable.
2318 * See freeze_processes.
2323 * Go through the zonelist yet one more time, keep very high watermark
2324 * here, this is only to catch a parallel oom killing, we must fail if
2325 * we're still under heavy pressure.
2327 page
= get_page_from_freelist(gfp_mask
| __GFP_HARDWALL
, order
,
2328 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
, ac
);
2332 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2333 /* Coredumps can quickly deplete all memory reserves */
2334 if (current
->flags
& PF_DUMPCORE
)
2336 /* The OOM killer will not help higher order allocs */
2337 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2339 /* The OOM killer does not needlessly kill tasks for lowmem */
2340 if (ac
->high_zoneidx
< ZONE_NORMAL
)
2342 /* The OOM killer does not compensate for light reclaim */
2343 if (!(gfp_mask
& __GFP_FS
))
2346 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2347 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2348 * The caller should handle page allocation failure by itself if
2349 * it specifies __GFP_THISNODE.
2350 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2352 if (gfp_mask
& __GFP_THISNODE
)
2355 /* Exhausted what can be done so it's blamo time */
2356 out_of_memory(ac
->zonelist
, gfp_mask
, order
, ac
->nodemask
, false);
2357 *did_some_progress
= 1;
2359 oom_zonelist_unlock(ac
->zonelist
, gfp_mask
);
2363 #ifdef CONFIG_COMPACTION
2364 /* Try memory compaction for high-order allocations before reclaim */
2365 static struct page
*
2366 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2367 int alloc_flags
, const struct alloc_context
*ac
,
2368 enum migrate_mode mode
, int *contended_compaction
,
2369 bool *deferred_compaction
)
2371 unsigned long compact_result
;
2377 current
->flags
|= PF_MEMALLOC
;
2378 compact_result
= try_to_compact_pages(gfp_mask
, order
, alloc_flags
, ac
,
2379 mode
, contended_compaction
);
2380 current
->flags
&= ~PF_MEMALLOC
;
2382 switch (compact_result
) {
2383 case COMPACT_DEFERRED
:
2384 *deferred_compaction
= true;
2386 case COMPACT_SKIPPED
:
2393 * At least in one zone compaction wasn't deferred or skipped, so let's
2394 * count a compaction stall
2396 count_vm_event(COMPACTSTALL
);
2398 page
= get_page_from_freelist(gfp_mask
, order
,
2399 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2402 struct zone
*zone
= page_zone(page
);
2404 zone
->compact_blockskip_flush
= false;
2405 compaction_defer_reset(zone
, order
, true);
2406 count_vm_event(COMPACTSUCCESS
);
2411 * It's bad if compaction run occurs and fails. The most likely reason
2412 * is that pages exist, but not enough to satisfy watermarks.
2414 count_vm_event(COMPACTFAIL
);
2421 static inline struct page
*
2422 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2423 int alloc_flags
, const struct alloc_context
*ac
,
2424 enum migrate_mode mode
, int *contended_compaction
,
2425 bool *deferred_compaction
)
2429 #endif /* CONFIG_COMPACTION */
2431 /* Perform direct synchronous page reclaim */
2433 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
,
2434 const struct alloc_context
*ac
)
2436 struct reclaim_state reclaim_state
;
2441 /* We now go into synchronous reclaim */
2442 cpuset_memory_pressure_bump();
2443 current
->flags
|= PF_MEMALLOC
;
2444 lockdep_set_current_reclaim_state(gfp_mask
);
2445 reclaim_state
.reclaimed_slab
= 0;
2446 current
->reclaim_state
= &reclaim_state
;
2448 progress
= try_to_free_pages(ac
->zonelist
, order
, gfp_mask
,
2451 current
->reclaim_state
= NULL
;
2452 lockdep_clear_current_reclaim_state();
2453 current
->flags
&= ~PF_MEMALLOC
;
2460 /* The really slow allocator path where we enter direct reclaim */
2461 static inline struct page
*
2462 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2463 int alloc_flags
, const struct alloc_context
*ac
,
2464 unsigned long *did_some_progress
)
2466 struct page
*page
= NULL
;
2467 bool drained
= false;
2469 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, ac
);
2470 if (unlikely(!(*did_some_progress
)))
2473 /* After successful reclaim, reconsider all zones for allocation */
2474 if (IS_ENABLED(CONFIG_NUMA
))
2475 zlc_clear_zones_full(ac
->zonelist
);
2478 page
= get_page_from_freelist(gfp_mask
, order
,
2479 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2482 * If an allocation failed after direct reclaim, it could be because
2483 * pages are pinned on the per-cpu lists. Drain them and try again
2485 if (!page
&& !drained
) {
2486 drain_all_pages(NULL
);
2495 * This is called in the allocator slow-path if the allocation request is of
2496 * sufficient urgency to ignore watermarks and take other desperate measures
2498 static inline struct page
*
2499 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2500 const struct alloc_context
*ac
)
2505 page
= get_page_from_freelist(gfp_mask
, order
,
2506 ALLOC_NO_WATERMARKS
, ac
);
2508 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2509 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
,
2511 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2516 static void wake_all_kswapds(unsigned int order
, const struct alloc_context
*ac
)
2521 for_each_zone_zonelist_nodemask(zone
, z
, ac
->zonelist
,
2522 ac
->high_zoneidx
, ac
->nodemask
)
2523 wakeup_kswapd(zone
, order
, zone_idx(ac
->preferred_zone
));
2527 gfp_to_alloc_flags(gfp_t gfp_mask
)
2529 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2530 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2532 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2533 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2536 * The caller may dip into page reserves a bit more if the caller
2537 * cannot run direct reclaim, or if the caller has realtime scheduling
2538 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2539 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2541 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2545 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2546 * if it can't schedule.
2548 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2549 alloc_flags
|= ALLOC_HARDER
;
2551 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2552 * comment for __cpuset_node_allowed().
2554 alloc_flags
&= ~ALLOC_CPUSET
;
2555 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2556 alloc_flags
|= ALLOC_HARDER
;
2558 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2559 if (gfp_mask
& __GFP_MEMALLOC
)
2560 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2561 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2562 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2563 else if (!in_interrupt() &&
2564 ((current
->flags
& PF_MEMALLOC
) ||
2565 unlikely(test_thread_flag(TIF_MEMDIE
))))
2566 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2569 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2570 alloc_flags
|= ALLOC_CMA
;
2575 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2577 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2580 static inline struct page
*
2581 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2582 struct alloc_context
*ac
)
2584 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2585 struct page
*page
= NULL
;
2587 unsigned long pages_reclaimed
= 0;
2588 unsigned long did_some_progress
;
2589 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2590 bool deferred_compaction
= false;
2591 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2594 * In the slowpath, we sanity check order to avoid ever trying to
2595 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2596 * be using allocators in order of preference for an area that is
2599 if (order
>= MAX_ORDER
) {
2600 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2605 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2606 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2607 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2608 * using a larger set of nodes after it has established that the
2609 * allowed per node queues are empty and that nodes are
2612 if (IS_ENABLED(CONFIG_NUMA
) &&
2613 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2617 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2618 wake_all_kswapds(order
, ac
);
2621 * OK, we're below the kswapd watermark and have kicked background
2622 * reclaim. Now things get more complex, so set up alloc_flags according
2623 * to how we want to proceed.
2625 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2628 * Find the true preferred zone if the allocation is unconstrained by
2631 if (!(alloc_flags
& ALLOC_CPUSET
) && !ac
->nodemask
) {
2632 struct zoneref
*preferred_zoneref
;
2633 preferred_zoneref
= first_zones_zonelist(ac
->zonelist
,
2634 ac
->high_zoneidx
, NULL
, &ac
->preferred_zone
);
2635 ac
->classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2638 /* This is the last chance, in general, before the goto nopage. */
2639 page
= get_page_from_freelist(gfp_mask
, order
,
2640 alloc_flags
& ~ALLOC_NO_WATERMARKS
, ac
);
2644 /* Allocate without watermarks if the context allows */
2645 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2647 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2648 * the allocation is high priority and these type of
2649 * allocations are system rather than user orientated
2651 ac
->zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2653 page
= __alloc_pages_high_priority(gfp_mask
, order
, ac
);
2660 /* Atomic allocations - we can't balance anything */
2663 * All existing users of the deprecated __GFP_NOFAIL are
2664 * blockable, so warn of any new users that actually allow this
2665 * type of allocation to fail.
2667 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2671 /* Avoid recursion of direct reclaim */
2672 if (current
->flags
& PF_MEMALLOC
)
2675 /* Avoid allocations with no watermarks from looping endlessly */
2676 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2680 * Try direct compaction. The first pass is asynchronous. Subsequent
2681 * attempts after direct reclaim are synchronous
2683 page
= __alloc_pages_direct_compact(gfp_mask
, order
, alloc_flags
, ac
,
2685 &contended_compaction
,
2686 &deferred_compaction
);
2690 /* Checks for THP-specific high-order allocations */
2691 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2693 * If compaction is deferred for high-order allocations, it is
2694 * because sync compaction recently failed. If this is the case
2695 * and the caller requested a THP allocation, we do not want
2696 * to heavily disrupt the system, so we fail the allocation
2697 * instead of entering direct reclaim.
2699 if (deferred_compaction
)
2703 * In all zones where compaction was attempted (and not
2704 * deferred or skipped), lock contention has been detected.
2705 * For THP allocation we do not want to disrupt the others
2706 * so we fallback to base pages instead.
2708 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2712 * If compaction was aborted due to need_resched(), we do not
2713 * want to further increase allocation latency, unless it is
2714 * khugepaged trying to collapse.
2716 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2717 && !(current
->flags
& PF_KTHREAD
))
2722 * It can become very expensive to allocate transparent hugepages at
2723 * fault, so use asynchronous memory compaction for THP unless it is
2724 * khugepaged trying to collapse.
2726 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2727 (current
->flags
& PF_KTHREAD
))
2728 migration_mode
= MIGRATE_SYNC_LIGHT
;
2730 /* Try direct reclaim and then allocating */
2731 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
, alloc_flags
, ac
,
2732 &did_some_progress
);
2736 /* Check if we should retry the allocation */
2737 pages_reclaimed
+= did_some_progress
;
2738 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2741 * If we fail to make progress by freeing individual
2742 * pages, but the allocation wants us to keep going,
2743 * start OOM killing tasks.
2745 if (!did_some_progress
) {
2746 page
= __alloc_pages_may_oom(gfp_mask
, order
, ac
,
2747 &did_some_progress
);
2750 if (!did_some_progress
)
2753 /* Wait for some write requests to complete then retry */
2754 wait_iff_congested(ac
->preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2758 * High-order allocations do not necessarily loop after
2759 * direct reclaim and reclaim/compaction depends on compaction
2760 * being called after reclaim so call directly if necessary
2762 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2763 alloc_flags
, ac
, migration_mode
,
2764 &contended_compaction
,
2765 &deferred_compaction
);
2771 warn_alloc_failed(gfp_mask
, order
, NULL
);
2777 * This is the 'heart' of the zoned buddy allocator.
2780 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2781 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2783 struct zoneref
*preferred_zoneref
;
2784 struct page
*page
= NULL
;
2785 unsigned int cpuset_mems_cookie
;
2786 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2787 gfp_t alloc_mask
; /* The gfp_t that was actually used for allocation */
2788 struct alloc_context ac
= {
2789 .high_zoneidx
= gfp_zone(gfp_mask
),
2790 .nodemask
= nodemask
,
2791 .migratetype
= gfpflags_to_migratetype(gfp_mask
),
2794 gfp_mask
&= gfp_allowed_mask
;
2796 lockdep_trace_alloc(gfp_mask
);
2798 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2800 if (should_fail_alloc_page(gfp_mask
, order
))
2804 * Check the zones suitable for the gfp_mask contain at least one
2805 * valid zone. It's possible to have an empty zonelist as a result
2806 * of GFP_THISNODE and a memoryless node
2808 if (unlikely(!zonelist
->_zonerefs
->zone
))
2811 if (IS_ENABLED(CONFIG_CMA
) && ac
.migratetype
== MIGRATE_MOVABLE
)
2812 alloc_flags
|= ALLOC_CMA
;
2815 cpuset_mems_cookie
= read_mems_allowed_begin();
2817 /* We set it here, as __alloc_pages_slowpath might have changed it */
2818 ac
.zonelist
= zonelist
;
2819 /* The preferred zone is used for statistics later */
2820 preferred_zoneref
= first_zones_zonelist(ac
.zonelist
, ac
.high_zoneidx
,
2821 ac
.nodemask
? : &cpuset_current_mems_allowed
,
2822 &ac
.preferred_zone
);
2823 if (!ac
.preferred_zone
)
2825 ac
.classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2827 /* First allocation attempt */
2828 alloc_mask
= gfp_mask
|__GFP_HARDWALL
;
2829 page
= get_page_from_freelist(alloc_mask
, order
, alloc_flags
, &ac
);
2830 if (unlikely(!page
)) {
2832 * Runtime PM, block IO and its error handling path
2833 * can deadlock because I/O on the device might not
2836 alloc_mask
= memalloc_noio_flags(gfp_mask
);
2838 page
= __alloc_pages_slowpath(alloc_mask
, order
, &ac
);
2841 if (kmemcheck_enabled
&& page
)
2842 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2844 trace_mm_page_alloc(page
, order
, alloc_mask
, ac
.migratetype
);
2848 * When updating a task's mems_allowed, it is possible to race with
2849 * parallel threads in such a way that an allocation can fail while
2850 * the mask is being updated. If a page allocation is about to fail,
2851 * check if the cpuset changed during allocation and if so, retry.
2853 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2858 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2861 * Common helper functions.
2863 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2868 * __get_free_pages() returns a 32-bit address, which cannot represent
2871 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2873 page
= alloc_pages(gfp_mask
, order
);
2876 return (unsigned long) page_address(page
);
2878 EXPORT_SYMBOL(__get_free_pages
);
2880 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2882 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2884 EXPORT_SYMBOL(get_zeroed_page
);
2886 void __free_pages(struct page
*page
, unsigned int order
)
2888 if (put_page_testzero(page
)) {
2890 free_hot_cold_page(page
, false);
2892 __free_pages_ok(page
, order
);
2896 EXPORT_SYMBOL(__free_pages
);
2898 void free_pages(unsigned long addr
, unsigned int order
)
2901 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2902 __free_pages(virt_to_page((void *)addr
), order
);
2906 EXPORT_SYMBOL(free_pages
);
2909 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2910 * of the current memory cgroup.
2912 * It should be used when the caller would like to use kmalloc, but since the
2913 * allocation is large, it has to fall back to the page allocator.
2915 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2918 struct mem_cgroup
*memcg
= NULL
;
2920 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2922 page
= alloc_pages(gfp_mask
, order
);
2923 memcg_kmem_commit_charge(page
, memcg
, order
);
2927 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2930 struct mem_cgroup
*memcg
= NULL
;
2932 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2934 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2935 memcg_kmem_commit_charge(page
, memcg
, order
);
2940 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2943 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2945 memcg_kmem_uncharge_pages(page
, order
);
2946 __free_pages(page
, order
);
2949 void free_kmem_pages(unsigned long addr
, unsigned int order
)
2952 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2953 __free_kmem_pages(virt_to_page((void *)addr
), order
);
2957 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2960 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2961 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2963 split_page(virt_to_page((void *)addr
), order
);
2964 while (used
< alloc_end
) {
2969 return (void *)addr
;
2973 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2974 * @size: the number of bytes to allocate
2975 * @gfp_mask: GFP flags for the allocation
2977 * This function is similar to alloc_pages(), except that it allocates the
2978 * minimum number of pages to satisfy the request. alloc_pages() can only
2979 * allocate memory in power-of-two pages.
2981 * This function is also limited by MAX_ORDER.
2983 * Memory allocated by this function must be released by free_pages_exact().
2985 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2987 unsigned int order
= get_order(size
);
2990 addr
= __get_free_pages(gfp_mask
, order
);
2991 return make_alloc_exact(addr
, order
, size
);
2993 EXPORT_SYMBOL(alloc_pages_exact
);
2996 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2998 * @nid: the preferred node ID where memory should be allocated
2999 * @size: the number of bytes to allocate
3000 * @gfp_mask: GFP flags for the allocation
3002 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3004 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3007 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3009 unsigned order
= get_order(size
);
3010 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3013 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3017 * free_pages_exact - release memory allocated via alloc_pages_exact()
3018 * @virt: the value returned by alloc_pages_exact.
3019 * @size: size of allocation, same value as passed to alloc_pages_exact().
3021 * Release the memory allocated by a previous call to alloc_pages_exact.
3023 void free_pages_exact(void *virt
, size_t size
)
3025 unsigned long addr
= (unsigned long)virt
;
3026 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3028 while (addr
< end
) {
3033 EXPORT_SYMBOL(free_pages_exact
);
3036 * nr_free_zone_pages - count number of pages beyond high watermark
3037 * @offset: The zone index of the highest zone
3039 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3040 * high watermark within all zones at or below a given zone index. For each
3041 * zone, the number of pages is calculated as:
3042 * managed_pages - high_pages
3044 static unsigned long nr_free_zone_pages(int offset
)
3049 /* Just pick one node, since fallback list is circular */
3050 unsigned long sum
= 0;
3052 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3054 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3055 unsigned long size
= zone
->managed_pages
;
3056 unsigned long high
= high_wmark_pages(zone
);
3065 * nr_free_buffer_pages - count number of pages beyond high watermark
3067 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3068 * watermark within ZONE_DMA and ZONE_NORMAL.
3070 unsigned long nr_free_buffer_pages(void)
3072 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3074 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3077 * nr_free_pagecache_pages - count number of pages beyond high watermark
3079 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3080 * high watermark within all zones.
3082 unsigned long nr_free_pagecache_pages(void)
3084 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3087 static inline void show_node(struct zone
*zone
)
3089 if (IS_ENABLED(CONFIG_NUMA
))
3090 printk("Node %d ", zone_to_nid(zone
));
3093 void si_meminfo(struct sysinfo
*val
)
3095 val
->totalram
= totalram_pages
;
3096 val
->sharedram
= global_page_state(NR_SHMEM
);
3097 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3098 val
->bufferram
= nr_blockdev_pages();
3099 val
->totalhigh
= totalhigh_pages
;
3100 val
->freehigh
= nr_free_highpages();
3101 val
->mem_unit
= PAGE_SIZE
;
3104 EXPORT_SYMBOL(si_meminfo
);
3107 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3109 int zone_type
; /* needs to be signed */
3110 unsigned long managed_pages
= 0;
3111 pg_data_t
*pgdat
= NODE_DATA(nid
);
3113 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3114 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3115 val
->totalram
= managed_pages
;
3116 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3117 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3118 #ifdef CONFIG_HIGHMEM
3119 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3120 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3126 val
->mem_unit
= PAGE_SIZE
;
3131 * Determine whether the node should be displayed or not, depending on whether
3132 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3134 bool skip_free_areas_node(unsigned int flags
, int nid
)
3137 unsigned int cpuset_mems_cookie
;
3139 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3143 cpuset_mems_cookie
= read_mems_allowed_begin();
3144 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3145 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3150 #define K(x) ((x) << (PAGE_SHIFT-10))
3152 static void show_migration_types(unsigned char type
)
3154 static const char types
[MIGRATE_TYPES
] = {
3155 [MIGRATE_UNMOVABLE
] = 'U',
3156 [MIGRATE_RECLAIMABLE
] = 'E',
3157 [MIGRATE_MOVABLE
] = 'M',
3158 [MIGRATE_RESERVE
] = 'R',
3160 [MIGRATE_CMA
] = 'C',
3162 #ifdef CONFIG_MEMORY_ISOLATION
3163 [MIGRATE_ISOLATE
] = 'I',
3166 char tmp
[MIGRATE_TYPES
+ 1];
3170 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3171 if (type
& (1 << i
))
3176 printk("(%s) ", tmp
);
3180 * Show free area list (used inside shift_scroll-lock stuff)
3181 * We also calculate the percentage fragmentation. We do this by counting the
3182 * memory on each free list with the exception of the first item on the list.
3183 * Suppresses nodes that are not allowed by current's cpuset if
3184 * SHOW_MEM_FILTER_NODES is passed.
3186 void show_free_areas(unsigned int filter
)
3191 for_each_populated_zone(zone
) {
3192 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3195 printk("%s per-cpu:\n", zone
->name
);
3197 for_each_online_cpu(cpu
) {
3198 struct per_cpu_pageset
*pageset
;
3200 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3202 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3203 cpu
, pageset
->pcp
.high
,
3204 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3208 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3209 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3211 " dirty:%lu writeback:%lu unstable:%lu\n"
3212 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3213 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3215 global_page_state(NR_ACTIVE_ANON
),
3216 global_page_state(NR_INACTIVE_ANON
),
3217 global_page_state(NR_ISOLATED_ANON
),
3218 global_page_state(NR_ACTIVE_FILE
),
3219 global_page_state(NR_INACTIVE_FILE
),
3220 global_page_state(NR_ISOLATED_FILE
),
3221 global_page_state(NR_UNEVICTABLE
),
3222 global_page_state(NR_FILE_DIRTY
),
3223 global_page_state(NR_WRITEBACK
),
3224 global_page_state(NR_UNSTABLE_NFS
),
3225 global_page_state(NR_FREE_PAGES
),
3226 global_page_state(NR_SLAB_RECLAIMABLE
),
3227 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3228 global_page_state(NR_FILE_MAPPED
),
3229 global_page_state(NR_SHMEM
),
3230 global_page_state(NR_PAGETABLE
),
3231 global_page_state(NR_BOUNCE
),
3232 global_page_state(NR_FREE_CMA_PAGES
));
3234 for_each_populated_zone(zone
) {
3237 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3245 " active_anon:%lukB"
3246 " inactive_anon:%lukB"
3247 " active_file:%lukB"
3248 " inactive_file:%lukB"
3249 " unevictable:%lukB"
3250 " isolated(anon):%lukB"
3251 " isolated(file):%lukB"
3259 " slab_reclaimable:%lukB"
3260 " slab_unreclaimable:%lukB"
3261 " kernel_stack:%lukB"
3266 " writeback_tmp:%lukB"
3267 " pages_scanned:%lu"
3268 " all_unreclaimable? %s"
3271 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3272 K(min_wmark_pages(zone
)),
3273 K(low_wmark_pages(zone
)),
3274 K(high_wmark_pages(zone
)),
3275 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3276 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3277 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3278 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3279 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3280 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3281 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3282 K(zone
->present_pages
),
3283 K(zone
->managed_pages
),
3284 K(zone_page_state(zone
, NR_MLOCK
)),
3285 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3286 K(zone_page_state(zone
, NR_WRITEBACK
)),
3287 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3288 K(zone_page_state(zone
, NR_SHMEM
)),
3289 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3290 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3291 zone_page_state(zone
, NR_KERNEL_STACK
) *
3293 K(zone_page_state(zone
, NR_PAGETABLE
)),
3294 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3295 K(zone_page_state(zone
, NR_BOUNCE
)),
3296 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3297 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3298 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3299 (!zone_reclaimable(zone
) ? "yes" : "no")
3301 printk("lowmem_reserve[]:");
3302 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3303 printk(" %ld", zone
->lowmem_reserve
[i
]);
3307 for_each_populated_zone(zone
) {
3308 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3309 unsigned char types
[MAX_ORDER
];
3311 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3314 printk("%s: ", zone
->name
);
3316 spin_lock_irqsave(&zone
->lock
, flags
);
3317 for (order
= 0; order
< MAX_ORDER
; order
++) {
3318 struct free_area
*area
= &zone
->free_area
[order
];
3321 nr
[order
] = area
->nr_free
;
3322 total
+= nr
[order
] << order
;
3325 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3326 if (!list_empty(&area
->free_list
[type
]))
3327 types
[order
] |= 1 << type
;
3330 spin_unlock_irqrestore(&zone
->lock
, flags
);
3331 for (order
= 0; order
< MAX_ORDER
; order
++) {
3332 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3334 show_migration_types(types
[order
]);
3336 printk("= %lukB\n", K(total
));
3339 hugetlb_show_meminfo();
3341 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3343 show_swap_cache_info();
3346 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3348 zoneref
->zone
= zone
;
3349 zoneref
->zone_idx
= zone_idx(zone
);
3353 * Builds allocation fallback zone lists.
3355 * Add all populated zones of a node to the zonelist.
3357 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3361 enum zone_type zone_type
= MAX_NR_ZONES
;
3365 zone
= pgdat
->node_zones
+ zone_type
;
3366 if (populated_zone(zone
)) {
3367 zoneref_set_zone(zone
,
3368 &zonelist
->_zonerefs
[nr_zones
++]);
3369 check_highest_zone(zone_type
);
3371 } while (zone_type
);
3379 * 0 = automatic detection of better ordering.
3380 * 1 = order by ([node] distance, -zonetype)
3381 * 2 = order by (-zonetype, [node] distance)
3383 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3384 * the same zonelist. So only NUMA can configure this param.
3386 #define ZONELIST_ORDER_DEFAULT 0
3387 #define ZONELIST_ORDER_NODE 1
3388 #define ZONELIST_ORDER_ZONE 2
3390 /* zonelist order in the kernel.
3391 * set_zonelist_order() will set this to NODE or ZONE.
3393 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3394 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3398 /* The value user specified ....changed by config */
3399 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3400 /* string for sysctl */
3401 #define NUMA_ZONELIST_ORDER_LEN 16
3402 char numa_zonelist_order
[16] = "default";
3405 * interface for configure zonelist ordering.
3406 * command line option "numa_zonelist_order"
3407 * = "[dD]efault - default, automatic configuration.
3408 * = "[nN]ode - order by node locality, then by zone within node
3409 * = "[zZ]one - order by zone, then by locality within zone
3412 static int __parse_numa_zonelist_order(char *s
)
3414 if (*s
== 'd' || *s
== 'D') {
3415 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3416 } else if (*s
== 'n' || *s
== 'N') {
3417 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3418 } else if (*s
== 'z' || *s
== 'Z') {
3419 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3422 "Ignoring invalid numa_zonelist_order value: "
3429 static __init
int setup_numa_zonelist_order(char *s
)
3436 ret
= __parse_numa_zonelist_order(s
);
3438 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3442 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3445 * sysctl handler for numa_zonelist_order
3447 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3448 void __user
*buffer
, size_t *length
,
3451 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3453 static DEFINE_MUTEX(zl_order_mutex
);
3455 mutex_lock(&zl_order_mutex
);
3457 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3461 strcpy(saved_string
, (char *)table
->data
);
3463 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3467 int oldval
= user_zonelist_order
;
3469 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3472 * bogus value. restore saved string
3474 strncpy((char *)table
->data
, saved_string
,
3475 NUMA_ZONELIST_ORDER_LEN
);
3476 user_zonelist_order
= oldval
;
3477 } else if (oldval
!= user_zonelist_order
) {
3478 mutex_lock(&zonelists_mutex
);
3479 build_all_zonelists(NULL
, NULL
);
3480 mutex_unlock(&zonelists_mutex
);
3484 mutex_unlock(&zl_order_mutex
);
3489 #define MAX_NODE_LOAD (nr_online_nodes)
3490 static int node_load
[MAX_NUMNODES
];
3493 * find_next_best_node - find the next node that should appear in a given node's fallback list
3494 * @node: node whose fallback list we're appending
3495 * @used_node_mask: nodemask_t of already used nodes
3497 * We use a number of factors to determine which is the next node that should
3498 * appear on a given node's fallback list. The node should not have appeared
3499 * already in @node's fallback list, and it should be the next closest node
3500 * according to the distance array (which contains arbitrary distance values
3501 * from each node to each node in the system), and should also prefer nodes
3502 * with no CPUs, since presumably they'll have very little allocation pressure
3503 * on them otherwise.
3504 * It returns -1 if no node is found.
3506 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3509 int min_val
= INT_MAX
;
3510 int best_node
= NUMA_NO_NODE
;
3511 const struct cpumask
*tmp
= cpumask_of_node(0);
3513 /* Use the local node if we haven't already */
3514 if (!node_isset(node
, *used_node_mask
)) {
3515 node_set(node
, *used_node_mask
);
3519 for_each_node_state(n
, N_MEMORY
) {
3521 /* Don't want a node to appear more than once */
3522 if (node_isset(n
, *used_node_mask
))
3525 /* Use the distance array to find the distance */
3526 val
= node_distance(node
, n
);
3528 /* Penalize nodes under us ("prefer the next node") */
3531 /* Give preference to headless and unused nodes */
3532 tmp
= cpumask_of_node(n
);
3533 if (!cpumask_empty(tmp
))
3534 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3536 /* Slight preference for less loaded node */
3537 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3538 val
+= node_load
[n
];
3540 if (val
< min_val
) {
3547 node_set(best_node
, *used_node_mask
);
3554 * Build zonelists ordered by node and zones within node.
3555 * This results in maximum locality--normal zone overflows into local
3556 * DMA zone, if any--but risks exhausting DMA zone.
3558 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3561 struct zonelist
*zonelist
;
3563 zonelist
= &pgdat
->node_zonelists
[0];
3564 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3566 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3567 zonelist
->_zonerefs
[j
].zone
= NULL
;
3568 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3572 * Build gfp_thisnode zonelists
3574 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3577 struct zonelist
*zonelist
;
3579 zonelist
= &pgdat
->node_zonelists
[1];
3580 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3581 zonelist
->_zonerefs
[j
].zone
= NULL
;
3582 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3586 * Build zonelists ordered by zone and nodes within zones.
3587 * This results in conserving DMA zone[s] until all Normal memory is
3588 * exhausted, but results in overflowing to remote node while memory
3589 * may still exist in local DMA zone.
3591 static int node_order
[MAX_NUMNODES
];
3593 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3596 int zone_type
; /* needs to be signed */
3598 struct zonelist
*zonelist
;
3600 zonelist
= &pgdat
->node_zonelists
[0];
3602 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3603 for (j
= 0; j
< nr_nodes
; j
++) {
3604 node
= node_order
[j
];
3605 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3606 if (populated_zone(z
)) {
3608 &zonelist
->_zonerefs
[pos
++]);
3609 check_highest_zone(zone_type
);
3613 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3614 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3617 #if defined(CONFIG_64BIT)
3619 * Devices that require DMA32/DMA are relatively rare and do not justify a
3620 * penalty to every machine in case the specialised case applies. Default
3621 * to Node-ordering on 64-bit NUMA machines
3623 static int default_zonelist_order(void)
3625 return ZONELIST_ORDER_NODE
;
3629 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3630 * by the kernel. If processes running on node 0 deplete the low memory zone
3631 * then reclaim will occur more frequency increasing stalls and potentially
3632 * be easier to OOM if a large percentage of the zone is under writeback or
3633 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3634 * Hence, default to zone ordering on 32-bit.
3636 static int default_zonelist_order(void)
3638 return ZONELIST_ORDER_ZONE
;
3640 #endif /* CONFIG_64BIT */
3642 static void set_zonelist_order(void)
3644 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3645 current_zonelist_order
= default_zonelist_order();
3647 current_zonelist_order
= user_zonelist_order
;
3650 static void build_zonelists(pg_data_t
*pgdat
)
3654 nodemask_t used_mask
;
3655 int local_node
, prev_node
;
3656 struct zonelist
*zonelist
;
3657 int order
= current_zonelist_order
;
3659 /* initialize zonelists */
3660 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3661 zonelist
= pgdat
->node_zonelists
+ i
;
3662 zonelist
->_zonerefs
[0].zone
= NULL
;
3663 zonelist
->_zonerefs
[0].zone_idx
= 0;
3666 /* NUMA-aware ordering of nodes */
3667 local_node
= pgdat
->node_id
;
3668 load
= nr_online_nodes
;
3669 prev_node
= local_node
;
3670 nodes_clear(used_mask
);
3672 memset(node_order
, 0, sizeof(node_order
));
3675 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3677 * We don't want to pressure a particular node.
3678 * So adding penalty to the first node in same
3679 * distance group to make it round-robin.
3681 if (node_distance(local_node
, node
) !=
3682 node_distance(local_node
, prev_node
))
3683 node_load
[node
] = load
;
3687 if (order
== ZONELIST_ORDER_NODE
)
3688 build_zonelists_in_node_order(pgdat
, node
);
3690 node_order
[j
++] = node
; /* remember order */
3693 if (order
== ZONELIST_ORDER_ZONE
) {
3694 /* calculate node order -- i.e., DMA last! */
3695 build_zonelists_in_zone_order(pgdat
, j
);
3698 build_thisnode_zonelists(pgdat
);
3701 /* Construct the zonelist performance cache - see further mmzone.h */
3702 static void build_zonelist_cache(pg_data_t
*pgdat
)
3704 struct zonelist
*zonelist
;
3705 struct zonelist_cache
*zlc
;
3708 zonelist
= &pgdat
->node_zonelists
[0];
3709 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3710 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3711 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3712 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3715 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3717 * Return node id of node used for "local" allocations.
3718 * I.e., first node id of first zone in arg node's generic zonelist.
3719 * Used for initializing percpu 'numa_mem', which is used primarily
3720 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3722 int local_memory_node(int node
)
3726 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3727 gfp_zone(GFP_KERNEL
),
3734 #else /* CONFIG_NUMA */
3736 static void set_zonelist_order(void)
3738 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3741 static void build_zonelists(pg_data_t
*pgdat
)
3743 int node
, local_node
;
3745 struct zonelist
*zonelist
;
3747 local_node
= pgdat
->node_id
;
3749 zonelist
= &pgdat
->node_zonelists
[0];
3750 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3753 * Now we build the zonelist so that it contains the zones
3754 * of all the other nodes.
3755 * We don't want to pressure a particular node, so when
3756 * building the zones for node N, we make sure that the
3757 * zones coming right after the local ones are those from
3758 * node N+1 (modulo N)
3760 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3761 if (!node_online(node
))
3763 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3765 for (node
= 0; node
< local_node
; node
++) {
3766 if (!node_online(node
))
3768 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3771 zonelist
->_zonerefs
[j
].zone
= NULL
;
3772 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3775 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3776 static void build_zonelist_cache(pg_data_t
*pgdat
)
3778 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3781 #endif /* CONFIG_NUMA */
3784 * Boot pageset table. One per cpu which is going to be used for all
3785 * zones and all nodes. The parameters will be set in such a way
3786 * that an item put on a list will immediately be handed over to
3787 * the buddy list. This is safe since pageset manipulation is done
3788 * with interrupts disabled.
3790 * The boot_pagesets must be kept even after bootup is complete for
3791 * unused processors and/or zones. They do play a role for bootstrapping
3792 * hotplugged processors.
3794 * zoneinfo_show() and maybe other functions do
3795 * not check if the processor is online before following the pageset pointer.
3796 * Other parts of the kernel may not check if the zone is available.
3798 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3799 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3800 static void setup_zone_pageset(struct zone
*zone
);
3803 * Global mutex to protect against size modification of zonelists
3804 * as well as to serialize pageset setup for the new populated zone.
3806 DEFINE_MUTEX(zonelists_mutex
);
3808 /* return values int ....just for stop_machine() */
3809 static int __build_all_zonelists(void *data
)
3813 pg_data_t
*self
= data
;
3816 memset(node_load
, 0, sizeof(node_load
));
3819 if (self
&& !node_online(self
->node_id
)) {
3820 build_zonelists(self
);
3821 build_zonelist_cache(self
);
3824 for_each_online_node(nid
) {
3825 pg_data_t
*pgdat
= NODE_DATA(nid
);
3827 build_zonelists(pgdat
);
3828 build_zonelist_cache(pgdat
);
3832 * Initialize the boot_pagesets that are going to be used
3833 * for bootstrapping processors. The real pagesets for
3834 * each zone will be allocated later when the per cpu
3835 * allocator is available.
3837 * boot_pagesets are used also for bootstrapping offline
3838 * cpus if the system is already booted because the pagesets
3839 * are needed to initialize allocators on a specific cpu too.
3840 * F.e. the percpu allocator needs the page allocator which
3841 * needs the percpu allocator in order to allocate its pagesets
3842 * (a chicken-egg dilemma).
3844 for_each_possible_cpu(cpu
) {
3845 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3847 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3849 * We now know the "local memory node" for each node--
3850 * i.e., the node of the first zone in the generic zonelist.
3851 * Set up numa_mem percpu variable for on-line cpus. During
3852 * boot, only the boot cpu should be on-line; we'll init the
3853 * secondary cpus' numa_mem as they come on-line. During
3854 * node/memory hotplug, we'll fixup all on-line cpus.
3856 if (cpu_online(cpu
))
3857 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3865 * Called with zonelists_mutex held always
3866 * unless system_state == SYSTEM_BOOTING.
3868 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3870 set_zonelist_order();
3872 if (system_state
== SYSTEM_BOOTING
) {
3873 __build_all_zonelists(NULL
);
3874 mminit_verify_zonelist();
3875 cpuset_init_current_mems_allowed();
3877 #ifdef CONFIG_MEMORY_HOTPLUG
3879 setup_zone_pageset(zone
);
3881 /* we have to stop all cpus to guarantee there is no user
3883 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3884 /* cpuset refresh routine should be here */
3886 vm_total_pages
= nr_free_pagecache_pages();
3888 * Disable grouping by mobility if the number of pages in the
3889 * system is too low to allow the mechanism to work. It would be
3890 * more accurate, but expensive to check per-zone. This check is
3891 * made on memory-hotadd so a system can start with mobility
3892 * disabled and enable it later
3894 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3895 page_group_by_mobility_disabled
= 1;
3897 page_group_by_mobility_disabled
= 0;
3899 pr_info("Built %i zonelists in %s order, mobility grouping %s. "
3900 "Total pages: %ld\n",
3902 zonelist_order_name
[current_zonelist_order
],
3903 page_group_by_mobility_disabled
? "off" : "on",
3906 pr_info("Policy zone: %s\n", zone_names
[policy_zone
]);
3911 * Helper functions to size the waitqueue hash table.
3912 * Essentially these want to choose hash table sizes sufficiently
3913 * large so that collisions trying to wait on pages are rare.
3914 * But in fact, the number of active page waitqueues on typical
3915 * systems is ridiculously low, less than 200. So this is even
3916 * conservative, even though it seems large.
3918 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3919 * waitqueues, i.e. the size of the waitq table given the number of pages.
3921 #define PAGES_PER_WAITQUEUE 256
3923 #ifndef CONFIG_MEMORY_HOTPLUG
3924 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3926 unsigned long size
= 1;
3928 pages
/= PAGES_PER_WAITQUEUE
;
3930 while (size
< pages
)
3934 * Once we have dozens or even hundreds of threads sleeping
3935 * on IO we've got bigger problems than wait queue collision.
3936 * Limit the size of the wait table to a reasonable size.
3938 size
= min(size
, 4096UL);
3940 return max(size
, 4UL);
3944 * A zone's size might be changed by hot-add, so it is not possible to determine
3945 * a suitable size for its wait_table. So we use the maximum size now.
3947 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3949 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3950 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3951 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3953 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3954 * or more by the traditional way. (See above). It equals:
3956 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3957 * ia64(16K page size) : = ( 8G + 4M)byte.
3958 * powerpc (64K page size) : = (32G +16M)byte.
3960 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3967 * This is an integer logarithm so that shifts can be used later
3968 * to extract the more random high bits from the multiplicative
3969 * hash function before the remainder is taken.
3971 static inline unsigned long wait_table_bits(unsigned long size
)
3977 * Check if a pageblock contains reserved pages
3979 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3983 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3984 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3991 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3992 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3993 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3994 * higher will lead to a bigger reserve which will get freed as contiguous
3995 * blocks as reclaim kicks in
3997 static void setup_zone_migrate_reserve(struct zone
*zone
)
3999 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4001 unsigned long block_migratetype
;
4006 * Get the start pfn, end pfn and the number of blocks to reserve
4007 * We have to be careful to be aligned to pageblock_nr_pages to
4008 * make sure that we always check pfn_valid for the first page in
4011 start_pfn
= zone
->zone_start_pfn
;
4012 end_pfn
= zone_end_pfn(zone
);
4013 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4014 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4018 * Reserve blocks are generally in place to help high-order atomic
4019 * allocations that are short-lived. A min_free_kbytes value that
4020 * would result in more than 2 reserve blocks for atomic allocations
4021 * is assumed to be in place to help anti-fragmentation for the
4022 * future allocation of hugepages at runtime.
4024 reserve
= min(2, reserve
);
4025 old_reserve
= zone
->nr_migrate_reserve_block
;
4027 /* When memory hot-add, we almost always need to do nothing */
4028 if (reserve
== old_reserve
)
4030 zone
->nr_migrate_reserve_block
= reserve
;
4032 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4033 if (!pfn_valid(pfn
))
4035 page
= pfn_to_page(pfn
);
4037 /* Watch out for overlapping nodes */
4038 if (page_to_nid(page
) != zone_to_nid(zone
))
4041 block_migratetype
= get_pageblock_migratetype(page
);
4043 /* Only test what is necessary when the reserves are not met */
4046 * Blocks with reserved pages will never free, skip
4049 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4050 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4053 /* If this block is reserved, account for it */
4054 if (block_migratetype
== MIGRATE_RESERVE
) {
4059 /* Suitable for reserving if this block is movable */
4060 if (block_migratetype
== MIGRATE_MOVABLE
) {
4061 set_pageblock_migratetype(page
,
4063 move_freepages_block(zone
, page
,
4068 } else if (!old_reserve
) {
4070 * At boot time we don't need to scan the whole zone
4071 * for turning off MIGRATE_RESERVE.
4077 * If the reserve is met and this is a previous reserved block,
4080 if (block_migratetype
== MIGRATE_RESERVE
) {
4081 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4082 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4088 * Initially all pages are reserved - free ones are freed
4089 * up by free_all_bootmem() once the early boot process is
4090 * done. Non-atomic initialization, single-pass.
4092 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4093 unsigned long start_pfn
, enum memmap_context context
)
4096 unsigned long end_pfn
= start_pfn
+ size
;
4100 if (highest_memmap_pfn
< end_pfn
- 1)
4101 highest_memmap_pfn
= end_pfn
- 1;
4103 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4104 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4106 * There can be holes in boot-time mem_map[]s
4107 * handed to this function. They do not
4108 * exist on hotplugged memory.
4110 if (context
== MEMMAP_EARLY
) {
4111 if (!early_pfn_valid(pfn
))
4113 if (!early_pfn_in_nid(pfn
, nid
))
4116 page
= pfn_to_page(pfn
);
4117 set_page_links(page
, zone
, nid
, pfn
);
4118 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4119 init_page_count(page
);
4120 page_mapcount_reset(page
);
4121 page_cpupid_reset_last(page
);
4122 SetPageReserved(page
);
4124 * Mark the block movable so that blocks are reserved for
4125 * movable at startup. This will force kernel allocations
4126 * to reserve their blocks rather than leaking throughout
4127 * the address space during boot when many long-lived
4128 * kernel allocations are made. Later some blocks near
4129 * the start are marked MIGRATE_RESERVE by
4130 * setup_zone_migrate_reserve()
4132 * bitmap is created for zone's valid pfn range. but memmap
4133 * can be created for invalid pages (for alignment)
4134 * check here not to call set_pageblock_migratetype() against
4137 if ((z
->zone_start_pfn
<= pfn
)
4138 && (pfn
< zone_end_pfn(z
))
4139 && !(pfn
& (pageblock_nr_pages
- 1)))
4140 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4142 INIT_LIST_HEAD(&page
->lru
);
4143 #ifdef WANT_PAGE_VIRTUAL
4144 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4145 if (!is_highmem_idx(zone
))
4146 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4151 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4153 unsigned int order
, t
;
4154 for_each_migratetype_order(order
, t
) {
4155 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4156 zone
->free_area
[order
].nr_free
= 0;
4160 #ifndef __HAVE_ARCH_MEMMAP_INIT
4161 #define memmap_init(size, nid, zone, start_pfn) \
4162 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4165 static int zone_batchsize(struct zone
*zone
)
4171 * The per-cpu-pages pools are set to around 1000th of the
4172 * size of the zone. But no more than 1/2 of a meg.
4174 * OK, so we don't know how big the cache is. So guess.
4176 batch
= zone
->managed_pages
/ 1024;
4177 if (batch
* PAGE_SIZE
> 512 * 1024)
4178 batch
= (512 * 1024) / PAGE_SIZE
;
4179 batch
/= 4; /* We effectively *= 4 below */
4184 * Clamp the batch to a 2^n - 1 value. Having a power
4185 * of 2 value was found to be more likely to have
4186 * suboptimal cache aliasing properties in some cases.
4188 * For example if 2 tasks are alternately allocating
4189 * batches of pages, one task can end up with a lot
4190 * of pages of one half of the possible page colors
4191 * and the other with pages of the other colors.
4193 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4198 /* The deferral and batching of frees should be suppressed under NOMMU
4201 * The problem is that NOMMU needs to be able to allocate large chunks
4202 * of contiguous memory as there's no hardware page translation to
4203 * assemble apparent contiguous memory from discontiguous pages.
4205 * Queueing large contiguous runs of pages for batching, however,
4206 * causes the pages to actually be freed in smaller chunks. As there
4207 * can be a significant delay between the individual batches being
4208 * recycled, this leads to the once large chunks of space being
4209 * fragmented and becoming unavailable for high-order allocations.
4216 * pcp->high and pcp->batch values are related and dependent on one another:
4217 * ->batch must never be higher then ->high.
4218 * The following function updates them in a safe manner without read side
4221 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4222 * those fields changing asynchronously (acording the the above rule).
4224 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4225 * outside of boot time (or some other assurance that no concurrent updaters
4228 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4229 unsigned long batch
)
4231 /* start with a fail safe value for batch */
4235 /* Update high, then batch, in order */
4242 /* a companion to pageset_set_high() */
4243 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4245 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4248 static void pageset_init(struct per_cpu_pageset
*p
)
4250 struct per_cpu_pages
*pcp
;
4253 memset(p
, 0, sizeof(*p
));
4257 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4258 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4261 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4264 pageset_set_batch(p
, batch
);
4268 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4269 * to the value high for the pageset p.
4271 static void pageset_set_high(struct per_cpu_pageset
*p
,
4274 unsigned long batch
= max(1UL, high
/ 4);
4275 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4276 batch
= PAGE_SHIFT
* 8;
4278 pageset_update(&p
->pcp
, high
, batch
);
4281 static void pageset_set_high_and_batch(struct zone
*zone
,
4282 struct per_cpu_pageset
*pcp
)
4284 if (percpu_pagelist_fraction
)
4285 pageset_set_high(pcp
,
4286 (zone
->managed_pages
/
4287 percpu_pagelist_fraction
));
4289 pageset_set_batch(pcp
, zone_batchsize(zone
));
4292 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4294 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4297 pageset_set_high_and_batch(zone
, pcp
);
4300 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4303 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4304 for_each_possible_cpu(cpu
)
4305 zone_pageset_init(zone
, cpu
);
4309 * Allocate per cpu pagesets and initialize them.
4310 * Before this call only boot pagesets were available.
4312 void __init
setup_per_cpu_pageset(void)
4316 for_each_populated_zone(zone
)
4317 setup_zone_pageset(zone
);
4320 static noinline __init_refok
4321 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4327 * The per-page waitqueue mechanism uses hashed waitqueues
4330 zone
->wait_table_hash_nr_entries
=
4331 wait_table_hash_nr_entries(zone_size_pages
);
4332 zone
->wait_table_bits
=
4333 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4334 alloc_size
= zone
->wait_table_hash_nr_entries
4335 * sizeof(wait_queue_head_t
);
4337 if (!slab_is_available()) {
4338 zone
->wait_table
= (wait_queue_head_t
*)
4339 memblock_virt_alloc_node_nopanic(
4340 alloc_size
, zone
->zone_pgdat
->node_id
);
4343 * This case means that a zone whose size was 0 gets new memory
4344 * via memory hot-add.
4345 * But it may be the case that a new node was hot-added. In
4346 * this case vmalloc() will not be able to use this new node's
4347 * memory - this wait_table must be initialized to use this new
4348 * node itself as well.
4349 * To use this new node's memory, further consideration will be
4352 zone
->wait_table
= vmalloc(alloc_size
);
4354 if (!zone
->wait_table
)
4357 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4358 init_waitqueue_head(zone
->wait_table
+ i
);
4363 static __meminit
void zone_pcp_init(struct zone
*zone
)
4366 * per cpu subsystem is not up at this point. The following code
4367 * relies on the ability of the linker to provide the
4368 * offset of a (static) per cpu variable into the per cpu area.
4370 zone
->pageset
= &boot_pageset
;
4372 if (populated_zone(zone
))
4373 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4374 zone
->name
, zone
->present_pages
,
4375 zone_batchsize(zone
));
4378 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4379 unsigned long zone_start_pfn
,
4381 enum memmap_context context
)
4383 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4385 ret
= zone_wait_table_init(zone
, size
);
4388 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4390 zone
->zone_start_pfn
= zone_start_pfn
;
4392 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4393 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4395 (unsigned long)zone_idx(zone
),
4396 zone_start_pfn
, (zone_start_pfn
+ size
));
4398 zone_init_free_lists(zone
);
4403 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4404 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4406 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4408 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4410 unsigned long start_pfn
, end_pfn
;
4413 * NOTE: The following SMP-unsafe globals are only used early in boot
4414 * when the kernel is running single-threaded.
4416 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4417 static int __meminitdata last_nid
;
4419 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4422 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4424 last_start_pfn
= start_pfn
;
4425 last_end_pfn
= end_pfn
;
4431 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4433 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4437 nid
= __early_pfn_to_nid(pfn
);
4440 /* just returns 0 */
4444 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4445 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4449 nid
= __early_pfn_to_nid(pfn
);
4450 if (nid
>= 0 && nid
!= node
)
4457 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4458 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4459 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4461 * If an architecture guarantees that all ranges registered contain no holes
4462 * and may be freed, this this function may be used instead of calling
4463 * memblock_free_early_nid() manually.
4465 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4467 unsigned long start_pfn
, end_pfn
;
4470 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4471 start_pfn
= min(start_pfn
, max_low_pfn
);
4472 end_pfn
= min(end_pfn
, max_low_pfn
);
4474 if (start_pfn
< end_pfn
)
4475 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4476 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4482 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4483 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4485 * If an architecture guarantees that all ranges registered contain no holes and may
4486 * be freed, this function may be used instead of calling memory_present() manually.
4488 void __init
sparse_memory_present_with_active_regions(int nid
)
4490 unsigned long start_pfn
, end_pfn
;
4493 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4494 memory_present(this_nid
, start_pfn
, end_pfn
);
4498 * get_pfn_range_for_nid - Return the start and end page frames for a node
4499 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4500 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4501 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4503 * It returns the start and end page frame of a node based on information
4504 * provided by memblock_set_node(). If called for a node
4505 * with no available memory, a warning is printed and the start and end
4508 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4509 unsigned long *start_pfn
, unsigned long *end_pfn
)
4511 unsigned long this_start_pfn
, this_end_pfn
;
4517 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4518 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4519 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4522 if (*start_pfn
== -1UL)
4527 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4528 * assumption is made that zones within a node are ordered in monotonic
4529 * increasing memory addresses so that the "highest" populated zone is used
4531 static void __init
find_usable_zone_for_movable(void)
4534 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4535 if (zone_index
== ZONE_MOVABLE
)
4538 if (arch_zone_highest_possible_pfn
[zone_index
] >
4539 arch_zone_lowest_possible_pfn
[zone_index
])
4543 VM_BUG_ON(zone_index
== -1);
4544 movable_zone
= zone_index
;
4548 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4549 * because it is sized independent of architecture. Unlike the other zones,
4550 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4551 * in each node depending on the size of each node and how evenly kernelcore
4552 * is distributed. This helper function adjusts the zone ranges
4553 * provided by the architecture for a given node by using the end of the
4554 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4555 * zones within a node are in order of monotonic increases memory addresses
4557 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4558 unsigned long zone_type
,
4559 unsigned long node_start_pfn
,
4560 unsigned long node_end_pfn
,
4561 unsigned long *zone_start_pfn
,
4562 unsigned long *zone_end_pfn
)
4564 /* Only adjust if ZONE_MOVABLE is on this node */
4565 if (zone_movable_pfn
[nid
]) {
4566 /* Size ZONE_MOVABLE */
4567 if (zone_type
== ZONE_MOVABLE
) {
4568 *zone_start_pfn
= zone_movable_pfn
[nid
];
4569 *zone_end_pfn
= min(node_end_pfn
,
4570 arch_zone_highest_possible_pfn
[movable_zone
]);
4572 /* Adjust for ZONE_MOVABLE starting within this range */
4573 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4574 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4575 *zone_end_pfn
= zone_movable_pfn
[nid
];
4577 /* Check if this whole range is within ZONE_MOVABLE */
4578 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4579 *zone_start_pfn
= *zone_end_pfn
;
4584 * Return the number of pages a zone spans in a node, including holes
4585 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4587 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4588 unsigned long zone_type
,
4589 unsigned long node_start_pfn
,
4590 unsigned long node_end_pfn
,
4591 unsigned long *ignored
)
4593 unsigned long zone_start_pfn
, zone_end_pfn
;
4595 /* Get the start and end of the zone */
4596 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4597 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4598 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4599 node_start_pfn
, node_end_pfn
,
4600 &zone_start_pfn
, &zone_end_pfn
);
4602 /* Check that this node has pages within the zone's required range */
4603 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4606 /* Move the zone boundaries inside the node if necessary */
4607 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4608 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4610 /* Return the spanned pages */
4611 return zone_end_pfn
- zone_start_pfn
;
4615 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4616 * then all holes in the requested range will be accounted for.
4618 unsigned long __meminit
__absent_pages_in_range(int nid
,
4619 unsigned long range_start_pfn
,
4620 unsigned long range_end_pfn
)
4622 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4623 unsigned long start_pfn
, end_pfn
;
4626 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4627 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4628 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4629 nr_absent
-= end_pfn
- start_pfn
;
4635 * absent_pages_in_range - Return number of page frames in holes within a range
4636 * @start_pfn: The start PFN to start searching for holes
4637 * @end_pfn: The end PFN to stop searching for holes
4639 * It returns the number of pages frames in memory holes within a range.
4641 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4642 unsigned long end_pfn
)
4644 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4647 /* Return the number of page frames in holes in a zone on a node */
4648 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4649 unsigned long zone_type
,
4650 unsigned long node_start_pfn
,
4651 unsigned long node_end_pfn
,
4652 unsigned long *ignored
)
4654 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4655 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4656 unsigned long zone_start_pfn
, zone_end_pfn
;
4658 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4659 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4661 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4662 node_start_pfn
, node_end_pfn
,
4663 &zone_start_pfn
, &zone_end_pfn
);
4664 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4667 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4668 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4669 unsigned long zone_type
,
4670 unsigned long node_start_pfn
,
4671 unsigned long node_end_pfn
,
4672 unsigned long *zones_size
)
4674 return zones_size
[zone_type
];
4677 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4678 unsigned long zone_type
,
4679 unsigned long node_start_pfn
,
4680 unsigned long node_end_pfn
,
4681 unsigned long *zholes_size
)
4686 return zholes_size
[zone_type
];
4689 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4691 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4692 unsigned long node_start_pfn
,
4693 unsigned long node_end_pfn
,
4694 unsigned long *zones_size
,
4695 unsigned long *zholes_size
)
4697 unsigned long realtotalpages
, totalpages
= 0;
4700 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4701 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4705 pgdat
->node_spanned_pages
= totalpages
;
4707 realtotalpages
= totalpages
;
4708 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4710 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4711 node_start_pfn
, node_end_pfn
,
4713 pgdat
->node_present_pages
= realtotalpages
;
4714 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4718 #ifndef CONFIG_SPARSEMEM
4720 * Calculate the size of the zone->blockflags rounded to an unsigned long
4721 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4722 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4723 * round what is now in bits to nearest long in bits, then return it in
4726 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4728 unsigned long usemapsize
;
4730 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4731 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4732 usemapsize
= usemapsize
>> pageblock_order
;
4733 usemapsize
*= NR_PAGEBLOCK_BITS
;
4734 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4736 return usemapsize
/ 8;
4739 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4741 unsigned long zone_start_pfn
,
4742 unsigned long zonesize
)
4744 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4745 zone
->pageblock_flags
= NULL
;
4747 zone
->pageblock_flags
=
4748 memblock_virt_alloc_node_nopanic(usemapsize
,
4752 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4753 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4754 #endif /* CONFIG_SPARSEMEM */
4756 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4758 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4759 void __paginginit
set_pageblock_order(void)
4763 /* Check that pageblock_nr_pages has not already been setup */
4764 if (pageblock_order
)
4767 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4768 order
= HUGETLB_PAGE_ORDER
;
4770 order
= MAX_ORDER
- 1;
4773 * Assume the largest contiguous order of interest is a huge page.
4774 * This value may be variable depending on boot parameters on IA64 and
4777 pageblock_order
= order
;
4779 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4782 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4783 * is unused as pageblock_order is set at compile-time. See
4784 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4787 void __paginginit
set_pageblock_order(void)
4791 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4793 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4794 unsigned long present_pages
)
4796 unsigned long pages
= spanned_pages
;
4799 * Provide a more accurate estimation if there are holes within
4800 * the zone and SPARSEMEM is in use. If there are holes within the
4801 * zone, each populated memory region may cost us one or two extra
4802 * memmap pages due to alignment because memmap pages for each
4803 * populated regions may not naturally algined on page boundary.
4804 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4806 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4807 IS_ENABLED(CONFIG_SPARSEMEM
))
4808 pages
= present_pages
;
4810 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4814 * Set up the zone data structures:
4815 * - mark all pages reserved
4816 * - mark all memory queues empty
4817 * - clear the memory bitmaps
4819 * NOTE: pgdat should get zeroed by caller.
4821 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4822 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4823 unsigned long *zones_size
, unsigned long *zholes_size
)
4826 int nid
= pgdat
->node_id
;
4827 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4830 pgdat_resize_init(pgdat
);
4831 #ifdef CONFIG_NUMA_BALANCING
4832 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4833 pgdat
->numabalancing_migrate_nr_pages
= 0;
4834 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4836 init_waitqueue_head(&pgdat
->kswapd_wait
);
4837 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4838 pgdat_page_ext_init(pgdat
);
4840 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4841 struct zone
*zone
= pgdat
->node_zones
+ j
;
4842 unsigned long size
, realsize
, freesize
, memmap_pages
;
4844 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4845 node_end_pfn
, zones_size
);
4846 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4852 * Adjust freesize so that it accounts for how much memory
4853 * is used by this zone for memmap. This affects the watermark
4854 * and per-cpu initialisations
4856 memmap_pages
= calc_memmap_size(size
, realsize
);
4857 if (!is_highmem_idx(j
)) {
4858 if (freesize
>= memmap_pages
) {
4859 freesize
-= memmap_pages
;
4862 " %s zone: %lu pages used for memmap\n",
4863 zone_names
[j
], memmap_pages
);
4866 " %s zone: %lu pages exceeds freesize %lu\n",
4867 zone_names
[j
], memmap_pages
, freesize
);
4870 /* Account for reserved pages */
4871 if (j
== 0 && freesize
> dma_reserve
) {
4872 freesize
-= dma_reserve
;
4873 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4874 zone_names
[0], dma_reserve
);
4877 if (!is_highmem_idx(j
))
4878 nr_kernel_pages
+= freesize
;
4879 /* Charge for highmem memmap if there are enough kernel pages */
4880 else if (nr_kernel_pages
> memmap_pages
* 2)
4881 nr_kernel_pages
-= memmap_pages
;
4882 nr_all_pages
+= freesize
;
4884 zone
->spanned_pages
= size
;
4885 zone
->present_pages
= realsize
;
4887 * Set an approximate value for lowmem here, it will be adjusted
4888 * when the bootmem allocator frees pages into the buddy system.
4889 * And all highmem pages will be managed by the buddy system.
4891 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4894 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4896 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4898 zone
->name
= zone_names
[j
];
4899 spin_lock_init(&zone
->lock
);
4900 spin_lock_init(&zone
->lru_lock
);
4901 zone_seqlock_init(zone
);
4902 zone
->zone_pgdat
= pgdat
;
4903 zone_pcp_init(zone
);
4905 /* For bootup, initialized properly in watermark setup */
4906 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4908 lruvec_init(&zone
->lruvec
);
4912 set_pageblock_order();
4913 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4914 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4915 size
, MEMMAP_EARLY
);
4917 memmap_init(size
, nid
, j
, zone_start_pfn
);
4918 zone_start_pfn
+= size
;
4922 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4924 /* Skip empty nodes */
4925 if (!pgdat
->node_spanned_pages
)
4928 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4929 /* ia64 gets its own node_mem_map, before this, without bootmem */
4930 if (!pgdat
->node_mem_map
) {
4931 unsigned long size
, start
, end
;
4935 * The zone's endpoints aren't required to be MAX_ORDER
4936 * aligned but the node_mem_map endpoints must be in order
4937 * for the buddy allocator to function correctly.
4939 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4940 end
= pgdat_end_pfn(pgdat
);
4941 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4942 size
= (end
- start
) * sizeof(struct page
);
4943 map
= alloc_remap(pgdat
->node_id
, size
);
4945 map
= memblock_virt_alloc_node_nopanic(size
,
4947 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4949 #ifndef CONFIG_NEED_MULTIPLE_NODES
4951 * With no DISCONTIG, the global mem_map is just set as node 0's
4953 if (pgdat
== NODE_DATA(0)) {
4954 mem_map
= NODE_DATA(0)->node_mem_map
;
4955 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4956 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4957 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4958 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4961 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4964 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4965 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4967 pg_data_t
*pgdat
= NODE_DATA(nid
);
4968 unsigned long start_pfn
= 0;
4969 unsigned long end_pfn
= 0;
4971 /* pg_data_t should be reset to zero when it's allocated */
4972 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4974 pgdat
->node_id
= nid
;
4975 pgdat
->node_start_pfn
= node_start_pfn
;
4976 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4977 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4978 printk(KERN_INFO
"Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid
,
4979 (u64
) start_pfn
<< PAGE_SHIFT
, (u64
) (end_pfn
<< PAGE_SHIFT
) - 1);
4981 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
4982 zones_size
, zholes_size
);
4984 alloc_node_mem_map(pgdat
);
4985 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4986 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4987 nid
, (unsigned long)pgdat
,
4988 (unsigned long)pgdat
->node_mem_map
);
4991 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
4992 zones_size
, zholes_size
);
4995 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4997 #if MAX_NUMNODES > 1
4999 * Figure out the number of possible node ids.
5001 void __init
setup_nr_node_ids(void)
5004 unsigned int highest
= 0;
5006 for_each_node_mask(node
, node_possible_map
)
5008 nr_node_ids
= highest
+ 1;
5013 * node_map_pfn_alignment - determine the maximum internode alignment
5015 * This function should be called after node map is populated and sorted.
5016 * It calculates the maximum power of two alignment which can distinguish
5019 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5020 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5021 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5022 * shifted, 1GiB is enough and this function will indicate so.
5024 * This is used to test whether pfn -> nid mapping of the chosen memory
5025 * model has fine enough granularity to avoid incorrect mapping for the
5026 * populated node map.
5028 * Returns the determined alignment in pfn's. 0 if there is no alignment
5029 * requirement (single node).
5031 unsigned long __init
node_map_pfn_alignment(void)
5033 unsigned long accl_mask
= 0, last_end
= 0;
5034 unsigned long start
, end
, mask
;
5038 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5039 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5046 * Start with a mask granular enough to pin-point to the
5047 * start pfn and tick off bits one-by-one until it becomes
5048 * too coarse to separate the current node from the last.
5050 mask
= ~((1 << __ffs(start
)) - 1);
5051 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5054 /* accumulate all internode masks */
5058 /* convert mask to number of pages */
5059 return ~accl_mask
+ 1;
5062 /* Find the lowest pfn for a node */
5063 static unsigned long __init
find_min_pfn_for_node(int nid
)
5065 unsigned long min_pfn
= ULONG_MAX
;
5066 unsigned long start_pfn
;
5069 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5070 min_pfn
= min(min_pfn
, start_pfn
);
5072 if (min_pfn
== ULONG_MAX
) {
5074 "Could not find start_pfn for node %d\n", nid
);
5082 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5084 * It returns the minimum PFN based on information provided via
5085 * memblock_set_node().
5087 unsigned long __init
find_min_pfn_with_active_regions(void)
5089 return find_min_pfn_for_node(MAX_NUMNODES
);
5093 * early_calculate_totalpages()
5094 * Sum pages in active regions for movable zone.
5095 * Populate N_MEMORY for calculating usable_nodes.
5097 static unsigned long __init
early_calculate_totalpages(void)
5099 unsigned long totalpages
= 0;
5100 unsigned long start_pfn
, end_pfn
;
5103 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5104 unsigned long pages
= end_pfn
- start_pfn
;
5106 totalpages
+= pages
;
5108 node_set_state(nid
, N_MEMORY
);
5114 * Find the PFN the Movable zone begins in each node. Kernel memory
5115 * is spread evenly between nodes as long as the nodes have enough
5116 * memory. When they don't, some nodes will have more kernelcore than
5119 static void __init
find_zone_movable_pfns_for_nodes(void)
5122 unsigned long usable_startpfn
;
5123 unsigned long kernelcore_node
, kernelcore_remaining
;
5124 /* save the state before borrow the nodemask */
5125 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5126 unsigned long totalpages
= early_calculate_totalpages();
5127 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5128 struct memblock_region
*r
;
5130 /* Need to find movable_zone earlier when movable_node is specified. */
5131 find_usable_zone_for_movable();
5134 * If movable_node is specified, ignore kernelcore and movablecore
5137 if (movable_node_is_enabled()) {
5138 for_each_memblock(memory
, r
) {
5139 if (!memblock_is_hotpluggable(r
))
5144 usable_startpfn
= PFN_DOWN(r
->base
);
5145 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5146 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5154 * If movablecore=nn[KMG] was specified, calculate what size of
5155 * kernelcore that corresponds so that memory usable for
5156 * any allocation type is evenly spread. If both kernelcore
5157 * and movablecore are specified, then the value of kernelcore
5158 * will be used for required_kernelcore if it's greater than
5159 * what movablecore would have allowed.
5161 if (required_movablecore
) {
5162 unsigned long corepages
;
5165 * Round-up so that ZONE_MOVABLE is at least as large as what
5166 * was requested by the user
5168 required_movablecore
=
5169 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5170 corepages
= totalpages
- required_movablecore
;
5172 required_kernelcore
= max(required_kernelcore
, corepages
);
5175 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5176 if (!required_kernelcore
)
5179 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5180 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5183 /* Spread kernelcore memory as evenly as possible throughout nodes */
5184 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5185 for_each_node_state(nid
, N_MEMORY
) {
5186 unsigned long start_pfn
, end_pfn
;
5189 * Recalculate kernelcore_node if the division per node
5190 * now exceeds what is necessary to satisfy the requested
5191 * amount of memory for the kernel
5193 if (required_kernelcore
< kernelcore_node
)
5194 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5197 * As the map is walked, we track how much memory is usable
5198 * by the kernel using kernelcore_remaining. When it is
5199 * 0, the rest of the node is usable by ZONE_MOVABLE
5201 kernelcore_remaining
= kernelcore_node
;
5203 /* Go through each range of PFNs within this node */
5204 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5205 unsigned long size_pages
;
5207 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5208 if (start_pfn
>= end_pfn
)
5211 /* Account for what is only usable for kernelcore */
5212 if (start_pfn
< usable_startpfn
) {
5213 unsigned long kernel_pages
;
5214 kernel_pages
= min(end_pfn
, usable_startpfn
)
5217 kernelcore_remaining
-= min(kernel_pages
,
5218 kernelcore_remaining
);
5219 required_kernelcore
-= min(kernel_pages
,
5220 required_kernelcore
);
5222 /* Continue if range is now fully accounted */
5223 if (end_pfn
<= usable_startpfn
) {
5226 * Push zone_movable_pfn to the end so
5227 * that if we have to rebalance
5228 * kernelcore across nodes, we will
5229 * not double account here
5231 zone_movable_pfn
[nid
] = end_pfn
;
5234 start_pfn
= usable_startpfn
;
5238 * The usable PFN range for ZONE_MOVABLE is from
5239 * start_pfn->end_pfn. Calculate size_pages as the
5240 * number of pages used as kernelcore
5242 size_pages
= end_pfn
- start_pfn
;
5243 if (size_pages
> kernelcore_remaining
)
5244 size_pages
= kernelcore_remaining
;
5245 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5248 * Some kernelcore has been met, update counts and
5249 * break if the kernelcore for this node has been
5252 required_kernelcore
-= min(required_kernelcore
,
5254 kernelcore_remaining
-= size_pages
;
5255 if (!kernelcore_remaining
)
5261 * If there is still required_kernelcore, we do another pass with one
5262 * less node in the count. This will push zone_movable_pfn[nid] further
5263 * along on the nodes that still have memory until kernelcore is
5267 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5271 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5272 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5273 zone_movable_pfn
[nid
] =
5274 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5277 /* restore the node_state */
5278 node_states
[N_MEMORY
] = saved_node_state
;
5281 /* Any regular or high memory on that node ? */
5282 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5284 enum zone_type zone_type
;
5286 if (N_MEMORY
== N_NORMAL_MEMORY
)
5289 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5290 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5291 if (populated_zone(zone
)) {
5292 node_set_state(nid
, N_HIGH_MEMORY
);
5293 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5294 zone_type
<= ZONE_NORMAL
)
5295 node_set_state(nid
, N_NORMAL_MEMORY
);
5302 * free_area_init_nodes - Initialise all pg_data_t and zone data
5303 * @max_zone_pfn: an array of max PFNs for each zone
5305 * This will call free_area_init_node() for each active node in the system.
5306 * Using the page ranges provided by memblock_set_node(), the size of each
5307 * zone in each node and their holes is calculated. If the maximum PFN
5308 * between two adjacent zones match, it is assumed that the zone is empty.
5309 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5310 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5311 * starts where the previous one ended. For example, ZONE_DMA32 starts
5312 * at arch_max_dma_pfn.
5314 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5316 unsigned long start_pfn
, end_pfn
;
5319 /* Record where the zone boundaries are */
5320 memset(arch_zone_lowest_possible_pfn
, 0,
5321 sizeof(arch_zone_lowest_possible_pfn
));
5322 memset(arch_zone_highest_possible_pfn
, 0,
5323 sizeof(arch_zone_highest_possible_pfn
));
5324 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5325 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5326 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5327 if (i
== ZONE_MOVABLE
)
5329 arch_zone_lowest_possible_pfn
[i
] =
5330 arch_zone_highest_possible_pfn
[i
-1];
5331 arch_zone_highest_possible_pfn
[i
] =
5332 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5334 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5335 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5337 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5338 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5339 find_zone_movable_pfns_for_nodes();
5341 /* Print out the zone ranges */
5342 pr_info("Zone ranges:\n");
5343 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5344 if (i
== ZONE_MOVABLE
)
5346 pr_info(" %-8s ", zone_names
[i
]);
5347 if (arch_zone_lowest_possible_pfn
[i
] ==
5348 arch_zone_highest_possible_pfn
[i
])
5351 pr_cont("[mem %0#10lx-%0#10lx]\n",
5352 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5353 (arch_zone_highest_possible_pfn
[i
]
5354 << PAGE_SHIFT
) - 1);
5357 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5358 pr_info("Movable zone start for each node\n");
5359 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5360 if (zone_movable_pfn
[i
])
5361 pr_info(" Node %d: %#010lx\n", i
,
5362 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5365 /* Print out the early node map */
5366 pr_info("Early memory node ranges\n");
5367 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5368 pr_info(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5369 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5371 /* Initialise every node */
5372 mminit_verify_pageflags_layout();
5373 setup_nr_node_ids();
5374 for_each_online_node(nid
) {
5375 pg_data_t
*pgdat
= NODE_DATA(nid
);
5376 free_area_init_node(nid
, NULL
,
5377 find_min_pfn_for_node(nid
), NULL
);
5379 /* Any memory on that node */
5380 if (pgdat
->node_present_pages
)
5381 node_set_state(nid
, N_MEMORY
);
5382 check_for_memory(pgdat
, nid
);
5386 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5388 unsigned long long coremem
;
5392 coremem
= memparse(p
, &p
);
5393 *core
= coremem
>> PAGE_SHIFT
;
5395 /* Paranoid check that UL is enough for the coremem value */
5396 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5402 * kernelcore=size sets the amount of memory for use for allocations that
5403 * cannot be reclaimed or migrated.
5405 static int __init
cmdline_parse_kernelcore(char *p
)
5407 return cmdline_parse_core(p
, &required_kernelcore
);
5411 * movablecore=size sets the amount of memory for use for allocations that
5412 * can be reclaimed or migrated.
5414 static int __init
cmdline_parse_movablecore(char *p
)
5416 return cmdline_parse_core(p
, &required_movablecore
);
5419 early_param("kernelcore", cmdline_parse_kernelcore
);
5420 early_param("movablecore", cmdline_parse_movablecore
);
5422 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5424 void adjust_managed_page_count(struct page
*page
, long count
)
5426 spin_lock(&managed_page_count_lock
);
5427 page_zone(page
)->managed_pages
+= count
;
5428 totalram_pages
+= count
;
5429 #ifdef CONFIG_HIGHMEM
5430 if (PageHighMem(page
))
5431 totalhigh_pages
+= count
;
5433 spin_unlock(&managed_page_count_lock
);
5435 EXPORT_SYMBOL(adjust_managed_page_count
);
5437 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5440 unsigned long pages
= 0;
5442 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5443 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5444 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5445 if ((unsigned int)poison
<= 0xFF)
5446 memset(pos
, poison
, PAGE_SIZE
);
5447 free_reserved_page(virt_to_page(pos
));
5451 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5452 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5456 EXPORT_SYMBOL(free_reserved_area
);
5458 #ifdef CONFIG_HIGHMEM
5459 void free_highmem_page(struct page
*page
)
5461 __free_reserved_page(page
);
5463 page_zone(page
)->managed_pages
++;
5469 void __init
mem_init_print_info(const char *str
)
5471 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5472 unsigned long init_code_size
, init_data_size
;
5474 physpages
= get_num_physpages();
5475 codesize
= _etext
- _stext
;
5476 datasize
= _edata
- _sdata
;
5477 rosize
= __end_rodata
- __start_rodata
;
5478 bss_size
= __bss_stop
- __bss_start
;
5479 init_data_size
= __init_end
- __init_begin
;
5480 init_code_size
= _einittext
- _sinittext
;
5483 * Detect special cases and adjust section sizes accordingly:
5484 * 1) .init.* may be embedded into .data sections
5485 * 2) .init.text.* may be out of [__init_begin, __init_end],
5486 * please refer to arch/tile/kernel/vmlinux.lds.S.
5487 * 3) .rodata.* may be embedded into .text or .data sections.
5489 #define adj_init_size(start, end, size, pos, adj) \
5491 if (start <= pos && pos < end && size > adj) \
5495 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5496 _sinittext
, init_code_size
);
5497 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5498 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5499 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5500 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5502 #undef adj_init_size
5504 pr_info("Memory: %luK/%luK available "
5505 "(%luK kernel code, %luK rwdata, %luK rodata, "
5506 "%luK init, %luK bss, %luK reserved, %luK cma-reserved"
5507 #ifdef CONFIG_HIGHMEM
5511 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5512 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5513 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5514 (physpages
- totalram_pages
- totalcma_pages
) << (PAGE_SHIFT
-10),
5515 totalcma_pages
<< (PAGE_SHIFT
-10),
5516 #ifdef CONFIG_HIGHMEM
5517 totalhigh_pages
<< (PAGE_SHIFT
-10),
5519 str
? ", " : "", str
? str
: "");
5523 * set_dma_reserve - set the specified number of pages reserved in the first zone
5524 * @new_dma_reserve: The number of pages to mark reserved
5526 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5527 * In the DMA zone, a significant percentage may be consumed by kernel image
5528 * and other unfreeable allocations which can skew the watermarks badly. This
5529 * function may optionally be used to account for unfreeable pages in the
5530 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5531 * smaller per-cpu batchsize.
5533 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5535 dma_reserve
= new_dma_reserve
;
5538 void __init
free_area_init(unsigned long *zones_size
)
5540 free_area_init_node(0, zones_size
,
5541 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5544 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5545 unsigned long action
, void *hcpu
)
5547 int cpu
= (unsigned long)hcpu
;
5549 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5550 lru_add_drain_cpu(cpu
);
5554 * Spill the event counters of the dead processor
5555 * into the current processors event counters.
5556 * This artificially elevates the count of the current
5559 vm_events_fold_cpu(cpu
);
5562 * Zero the differential counters of the dead processor
5563 * so that the vm statistics are consistent.
5565 * This is only okay since the processor is dead and cannot
5566 * race with what we are doing.
5568 cpu_vm_stats_fold(cpu
);
5573 void __init
page_alloc_init(void)
5575 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5579 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5580 * or min_free_kbytes changes.
5582 static void calculate_totalreserve_pages(void)
5584 struct pglist_data
*pgdat
;
5585 unsigned long reserve_pages
= 0;
5586 enum zone_type i
, j
;
5588 for_each_online_pgdat(pgdat
) {
5589 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5590 struct zone
*zone
= pgdat
->node_zones
+ i
;
5593 /* Find valid and maximum lowmem_reserve in the zone */
5594 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5595 if (zone
->lowmem_reserve
[j
] > max
)
5596 max
= zone
->lowmem_reserve
[j
];
5599 /* we treat the high watermark as reserved pages. */
5600 max
+= high_wmark_pages(zone
);
5602 if (max
> zone
->managed_pages
)
5603 max
= zone
->managed_pages
;
5604 reserve_pages
+= max
;
5606 * Lowmem reserves are not available to
5607 * GFP_HIGHUSER page cache allocations and
5608 * kswapd tries to balance zones to their high
5609 * watermark. As a result, neither should be
5610 * regarded as dirtyable memory, to prevent a
5611 * situation where reclaim has to clean pages
5612 * in order to balance the zones.
5614 zone
->dirty_balance_reserve
= max
;
5617 dirty_balance_reserve
= reserve_pages
;
5618 totalreserve_pages
= reserve_pages
;
5622 * setup_per_zone_lowmem_reserve - called whenever
5623 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5624 * has a correct pages reserved value, so an adequate number of
5625 * pages are left in the zone after a successful __alloc_pages().
5627 static void setup_per_zone_lowmem_reserve(void)
5629 struct pglist_data
*pgdat
;
5630 enum zone_type j
, idx
;
5632 for_each_online_pgdat(pgdat
) {
5633 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5634 struct zone
*zone
= pgdat
->node_zones
+ j
;
5635 unsigned long managed_pages
= zone
->managed_pages
;
5637 zone
->lowmem_reserve
[j
] = 0;
5641 struct zone
*lower_zone
;
5645 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5646 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5648 lower_zone
= pgdat
->node_zones
+ idx
;
5649 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5650 sysctl_lowmem_reserve_ratio
[idx
];
5651 managed_pages
+= lower_zone
->managed_pages
;
5656 /* update totalreserve_pages */
5657 calculate_totalreserve_pages();
5660 static void __setup_per_zone_wmarks(void)
5662 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5663 unsigned long lowmem_pages
= 0;
5665 unsigned long flags
;
5667 /* Calculate total number of !ZONE_HIGHMEM pages */
5668 for_each_zone(zone
) {
5669 if (!is_highmem(zone
))
5670 lowmem_pages
+= zone
->managed_pages
;
5673 for_each_zone(zone
) {
5676 spin_lock_irqsave(&zone
->lock
, flags
);
5677 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5678 do_div(tmp
, lowmem_pages
);
5679 if (is_highmem(zone
)) {
5681 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5682 * need highmem pages, so cap pages_min to a small
5685 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5686 * deltas controls asynch page reclaim, and so should
5687 * not be capped for highmem.
5689 unsigned long min_pages
;
5691 min_pages
= zone
->managed_pages
/ 1024;
5692 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5693 zone
->watermark
[WMARK_MIN
] = min_pages
;
5696 * If it's a lowmem zone, reserve a number of pages
5697 * proportionate to the zone's size.
5699 zone
->watermark
[WMARK_MIN
] = tmp
;
5702 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5703 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5705 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5706 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5707 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5709 setup_zone_migrate_reserve(zone
);
5710 spin_unlock_irqrestore(&zone
->lock
, flags
);
5713 /* update totalreserve_pages */
5714 calculate_totalreserve_pages();
5718 * setup_per_zone_wmarks - called when min_free_kbytes changes
5719 * or when memory is hot-{added|removed}
5721 * Ensures that the watermark[min,low,high] values for each zone are set
5722 * correctly with respect to min_free_kbytes.
5724 void setup_per_zone_wmarks(void)
5726 mutex_lock(&zonelists_mutex
);
5727 __setup_per_zone_wmarks();
5728 mutex_unlock(&zonelists_mutex
);
5732 * The inactive anon list should be small enough that the VM never has to
5733 * do too much work, but large enough that each inactive page has a chance
5734 * to be referenced again before it is swapped out.
5736 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5737 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5738 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5739 * the anonymous pages are kept on the inactive list.
5742 * memory ratio inactive anon
5743 * -------------------------------------
5752 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5754 unsigned int gb
, ratio
;
5756 /* Zone size in gigabytes */
5757 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5759 ratio
= int_sqrt(10 * gb
);
5763 zone
->inactive_ratio
= ratio
;
5766 static void __meminit
setup_per_zone_inactive_ratio(void)
5771 calculate_zone_inactive_ratio(zone
);
5775 * Initialise min_free_kbytes.
5777 * For small machines we want it small (128k min). For large machines
5778 * we want it large (64MB max). But it is not linear, because network
5779 * bandwidth does not increase linearly with machine size. We use
5781 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5782 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5798 int __meminit
init_per_zone_wmark_min(void)
5800 unsigned long lowmem_kbytes
;
5801 int new_min_free_kbytes
;
5803 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5804 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5806 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5807 min_free_kbytes
= new_min_free_kbytes
;
5808 if (min_free_kbytes
< 128)
5809 min_free_kbytes
= 128;
5810 if (min_free_kbytes
> 65536)
5811 min_free_kbytes
= 65536;
5813 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5814 new_min_free_kbytes
, user_min_free_kbytes
);
5816 setup_per_zone_wmarks();
5817 refresh_zone_stat_thresholds();
5818 setup_per_zone_lowmem_reserve();
5819 setup_per_zone_inactive_ratio();
5822 module_init(init_per_zone_wmark_min
)
5825 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5826 * that we can call two helper functions whenever min_free_kbytes
5829 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5830 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5834 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5839 user_min_free_kbytes
= min_free_kbytes
;
5840 setup_per_zone_wmarks();
5846 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5847 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5852 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5857 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5858 sysctl_min_unmapped_ratio
) / 100;
5862 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5863 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5868 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5873 zone
->min_slab_pages
= (zone
->managed_pages
*
5874 sysctl_min_slab_ratio
) / 100;
5880 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5881 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5882 * whenever sysctl_lowmem_reserve_ratio changes.
5884 * The reserve ratio obviously has absolutely no relation with the
5885 * minimum watermarks. The lowmem reserve ratio can only make sense
5886 * if in function of the boot time zone sizes.
5888 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5889 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5891 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5892 setup_per_zone_lowmem_reserve();
5897 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5898 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5899 * pagelist can have before it gets flushed back to buddy allocator.
5901 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5902 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5905 int old_percpu_pagelist_fraction
;
5908 mutex_lock(&pcp_batch_high_lock
);
5909 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5911 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5912 if (!write
|| ret
< 0)
5915 /* Sanity checking to avoid pcp imbalance */
5916 if (percpu_pagelist_fraction
&&
5917 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5918 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5924 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
5927 for_each_populated_zone(zone
) {
5930 for_each_possible_cpu(cpu
)
5931 pageset_set_high_and_batch(zone
,
5932 per_cpu_ptr(zone
->pageset
, cpu
));
5935 mutex_unlock(&pcp_batch_high_lock
);
5939 int hashdist
= HASHDIST_DEFAULT
;
5942 static int __init
set_hashdist(char *str
)
5946 hashdist
= simple_strtoul(str
, &str
, 0);
5949 __setup("hashdist=", set_hashdist
);
5953 * allocate a large system hash table from bootmem
5954 * - it is assumed that the hash table must contain an exact power-of-2
5955 * quantity of entries
5956 * - limit is the number of hash buckets, not the total allocation size
5958 void *__init
alloc_large_system_hash(const char *tablename
,
5959 unsigned long bucketsize
,
5960 unsigned long numentries
,
5963 unsigned int *_hash_shift
,
5964 unsigned int *_hash_mask
,
5965 unsigned long low_limit
,
5966 unsigned long high_limit
)
5968 unsigned long long max
= high_limit
;
5969 unsigned long log2qty
, size
;
5972 /* allow the kernel cmdline to have a say */
5974 /* round applicable memory size up to nearest megabyte */
5975 numentries
= nr_kernel_pages
;
5977 /* It isn't necessary when PAGE_SIZE >= 1MB */
5978 if (PAGE_SHIFT
< 20)
5979 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
5981 /* limit to 1 bucket per 2^scale bytes of low memory */
5982 if (scale
> PAGE_SHIFT
)
5983 numentries
>>= (scale
- PAGE_SHIFT
);
5985 numentries
<<= (PAGE_SHIFT
- scale
);
5987 /* Make sure we've got at least a 0-order allocation.. */
5988 if (unlikely(flags
& HASH_SMALL
)) {
5989 /* Makes no sense without HASH_EARLY */
5990 WARN_ON(!(flags
& HASH_EARLY
));
5991 if (!(numentries
>> *_hash_shift
)) {
5992 numentries
= 1UL << *_hash_shift
;
5993 BUG_ON(!numentries
);
5995 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5996 numentries
= PAGE_SIZE
/ bucketsize
;
5998 numentries
= roundup_pow_of_two(numentries
);
6000 /* limit allocation size to 1/16 total memory by default */
6002 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6003 do_div(max
, bucketsize
);
6005 max
= min(max
, 0x80000000ULL
);
6007 if (numentries
< low_limit
)
6008 numentries
= low_limit
;
6009 if (numentries
> max
)
6012 log2qty
= ilog2(numentries
);
6015 size
= bucketsize
<< log2qty
;
6016 if (flags
& HASH_EARLY
)
6017 table
= memblock_virt_alloc_nopanic(size
, 0);
6019 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6022 * If bucketsize is not a power-of-two, we may free
6023 * some pages at the end of hash table which
6024 * alloc_pages_exact() automatically does
6026 if (get_order(size
) < MAX_ORDER
) {
6027 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6028 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6031 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6034 panic("Failed to allocate %s hash table\n", tablename
);
6036 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6039 ilog2(size
) - PAGE_SHIFT
,
6043 *_hash_shift
= log2qty
;
6045 *_hash_mask
= (1 << log2qty
) - 1;
6050 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6051 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6054 #ifdef CONFIG_SPARSEMEM
6055 return __pfn_to_section(pfn
)->pageblock_flags
;
6057 return zone
->pageblock_flags
;
6058 #endif /* CONFIG_SPARSEMEM */
6061 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6063 #ifdef CONFIG_SPARSEMEM
6064 pfn
&= (PAGES_PER_SECTION
-1);
6065 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6067 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6068 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6069 #endif /* CONFIG_SPARSEMEM */
6073 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6074 * @page: The page within the block of interest
6075 * @pfn: The target page frame number
6076 * @end_bitidx: The last bit of interest to retrieve
6077 * @mask: mask of bits that the caller is interested in
6079 * Return: pageblock_bits flags
6081 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6082 unsigned long end_bitidx
,
6086 unsigned long *bitmap
;
6087 unsigned long bitidx
, word_bitidx
;
6090 zone
= page_zone(page
);
6091 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6092 bitidx
= pfn_to_bitidx(zone
, pfn
);
6093 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6094 bitidx
&= (BITS_PER_LONG
-1);
6096 word
= bitmap
[word_bitidx
];
6097 bitidx
+= end_bitidx
;
6098 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6102 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6103 * @page: The page within the block of interest
6104 * @flags: The flags to set
6105 * @pfn: The target page frame number
6106 * @end_bitidx: The last bit of interest
6107 * @mask: mask of bits that the caller is interested in
6109 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6111 unsigned long end_bitidx
,
6115 unsigned long *bitmap
;
6116 unsigned long bitidx
, word_bitidx
;
6117 unsigned long old_word
, word
;
6119 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6121 zone
= page_zone(page
);
6122 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6123 bitidx
= pfn_to_bitidx(zone
, pfn
);
6124 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6125 bitidx
&= (BITS_PER_LONG
-1);
6127 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6129 bitidx
+= end_bitidx
;
6130 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6131 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6133 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6135 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6136 if (word
== old_word
)
6143 * This function checks whether pageblock includes unmovable pages or not.
6144 * If @count is not zero, it is okay to include less @count unmovable pages
6146 * PageLRU check without isolation or lru_lock could race so that
6147 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6148 * expect this function should be exact.
6150 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6151 bool skip_hwpoisoned_pages
)
6153 unsigned long pfn
, iter
, found
;
6157 * For avoiding noise data, lru_add_drain_all() should be called
6158 * If ZONE_MOVABLE, the zone never contains unmovable pages
6160 if (zone_idx(zone
) == ZONE_MOVABLE
)
6162 mt
= get_pageblock_migratetype(page
);
6163 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6166 pfn
= page_to_pfn(page
);
6167 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6168 unsigned long check
= pfn
+ iter
;
6170 if (!pfn_valid_within(check
))
6173 page
= pfn_to_page(check
);
6176 * Hugepages are not in LRU lists, but they're movable.
6177 * We need not scan over tail pages bacause we don't
6178 * handle each tail page individually in migration.
6180 if (PageHuge(page
)) {
6181 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6186 * We can't use page_count without pin a page
6187 * because another CPU can free compound page.
6188 * This check already skips compound tails of THP
6189 * because their page->_count is zero at all time.
6191 if (!atomic_read(&page
->_count
)) {
6192 if (PageBuddy(page
))
6193 iter
+= (1 << page_order(page
)) - 1;
6198 * The HWPoisoned page may be not in buddy system, and
6199 * page_count() is not 0.
6201 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6207 * If there are RECLAIMABLE pages, we need to check
6208 * it. But now, memory offline itself doesn't call
6209 * shrink_node_slabs() and it still to be fixed.
6212 * If the page is not RAM, page_count()should be 0.
6213 * we don't need more check. This is an _used_ not-movable page.
6215 * The problematic thing here is PG_reserved pages. PG_reserved
6216 * is set to both of a memory hole page and a _used_ kernel
6225 bool is_pageblock_removable_nolock(struct page
*page
)
6231 * We have to be careful here because we are iterating over memory
6232 * sections which are not zone aware so we might end up outside of
6233 * the zone but still within the section.
6234 * We have to take care about the node as well. If the node is offline
6235 * its NODE_DATA will be NULL - see page_zone.
6237 if (!node_online(page_to_nid(page
)))
6240 zone
= page_zone(page
);
6241 pfn
= page_to_pfn(page
);
6242 if (!zone_spans_pfn(zone
, pfn
))
6245 return !has_unmovable_pages(zone
, page
, 0, true);
6250 static unsigned long pfn_max_align_down(unsigned long pfn
)
6252 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6253 pageblock_nr_pages
) - 1);
6256 static unsigned long pfn_max_align_up(unsigned long pfn
)
6258 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6259 pageblock_nr_pages
));
6262 /* [start, end) must belong to a single zone. */
6263 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6264 unsigned long start
, unsigned long end
)
6266 /* This function is based on compact_zone() from compaction.c. */
6267 unsigned long nr_reclaimed
;
6268 unsigned long pfn
= start
;
6269 unsigned int tries
= 0;
6274 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6275 if (fatal_signal_pending(current
)) {
6280 if (list_empty(&cc
->migratepages
)) {
6281 cc
->nr_migratepages
= 0;
6282 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6288 } else if (++tries
== 5) {
6289 ret
= ret
< 0 ? ret
: -EBUSY
;
6293 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6295 cc
->nr_migratepages
-= nr_reclaimed
;
6297 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6298 NULL
, 0, cc
->mode
, MR_CMA
);
6301 putback_movable_pages(&cc
->migratepages
);
6308 * alloc_contig_range() -- tries to allocate given range of pages
6309 * @start: start PFN to allocate
6310 * @end: one-past-the-last PFN to allocate
6311 * @migratetype: migratetype of the underlaying pageblocks (either
6312 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6313 * in range must have the same migratetype and it must
6314 * be either of the two.
6316 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6317 * aligned, however it's the caller's responsibility to guarantee that
6318 * we are the only thread that changes migrate type of pageblocks the
6321 * The PFN range must belong to a single zone.
6323 * Returns zero on success or negative error code. On success all
6324 * pages which PFN is in [start, end) are allocated for the caller and
6325 * need to be freed with free_contig_range().
6327 int alloc_contig_range(unsigned long start
, unsigned long end
,
6328 unsigned migratetype
)
6330 unsigned long outer_start
, outer_end
;
6333 struct compact_control cc
= {
6334 .nr_migratepages
= 0,
6336 .zone
= page_zone(pfn_to_page(start
)),
6337 .mode
= MIGRATE_SYNC
,
6338 .ignore_skip_hint
= true,
6340 INIT_LIST_HEAD(&cc
.migratepages
);
6343 * What we do here is we mark all pageblocks in range as
6344 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6345 * have different sizes, and due to the way page allocator
6346 * work, we align the range to biggest of the two pages so
6347 * that page allocator won't try to merge buddies from
6348 * different pageblocks and change MIGRATE_ISOLATE to some
6349 * other migration type.
6351 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6352 * migrate the pages from an unaligned range (ie. pages that
6353 * we are interested in). This will put all the pages in
6354 * range back to page allocator as MIGRATE_ISOLATE.
6356 * When this is done, we take the pages in range from page
6357 * allocator removing them from the buddy system. This way
6358 * page allocator will never consider using them.
6360 * This lets us mark the pageblocks back as
6361 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6362 * aligned range but not in the unaligned, original range are
6363 * put back to page allocator so that buddy can use them.
6366 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6367 pfn_max_align_up(end
), migratetype
,
6372 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6377 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6378 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6379 * more, all pages in [start, end) are free in page allocator.
6380 * What we are going to do is to allocate all pages from
6381 * [start, end) (that is remove them from page allocator).
6383 * The only problem is that pages at the beginning and at the
6384 * end of interesting range may be not aligned with pages that
6385 * page allocator holds, ie. they can be part of higher order
6386 * pages. Because of this, we reserve the bigger range and
6387 * once this is done free the pages we are not interested in.
6389 * We don't have to hold zone->lock here because the pages are
6390 * isolated thus they won't get removed from buddy.
6393 lru_add_drain_all();
6394 drain_all_pages(cc
.zone
);
6397 outer_start
= start
;
6398 while (!PageBuddy(pfn_to_page(outer_start
))) {
6399 if (++order
>= MAX_ORDER
) {
6403 outer_start
&= ~0UL << order
;
6406 /* Make sure the range is really isolated. */
6407 if (test_pages_isolated(outer_start
, end
, false)) {
6408 pr_info("%s: [%lx, %lx) PFNs busy\n",
6409 __func__
, outer_start
, end
);
6414 /* Grab isolated pages from freelists. */
6415 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6421 /* Free head and tail (if any) */
6422 if (start
!= outer_start
)
6423 free_contig_range(outer_start
, start
- outer_start
);
6424 if (end
!= outer_end
)
6425 free_contig_range(end
, outer_end
- end
);
6428 undo_isolate_page_range(pfn_max_align_down(start
),
6429 pfn_max_align_up(end
), migratetype
);
6433 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6435 unsigned int count
= 0;
6437 for (; nr_pages
--; pfn
++) {
6438 struct page
*page
= pfn_to_page(pfn
);
6440 count
+= page_count(page
) != 1;
6443 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6447 #ifdef CONFIG_MEMORY_HOTPLUG
6449 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6450 * page high values need to be recalulated.
6452 void __meminit
zone_pcp_update(struct zone
*zone
)
6455 mutex_lock(&pcp_batch_high_lock
);
6456 for_each_possible_cpu(cpu
)
6457 pageset_set_high_and_batch(zone
,
6458 per_cpu_ptr(zone
->pageset
, cpu
));
6459 mutex_unlock(&pcp_batch_high_lock
);
6463 void zone_pcp_reset(struct zone
*zone
)
6465 unsigned long flags
;
6467 struct per_cpu_pageset
*pset
;
6469 /* avoid races with drain_pages() */
6470 local_irq_save(flags
);
6471 if (zone
->pageset
!= &boot_pageset
) {
6472 for_each_online_cpu(cpu
) {
6473 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6474 drain_zonestat(zone
, pset
);
6476 free_percpu(zone
->pageset
);
6477 zone
->pageset
= &boot_pageset
;
6479 local_irq_restore(flags
);
6482 #ifdef CONFIG_MEMORY_HOTREMOVE
6484 * All pages in the range must be isolated before calling this.
6487 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6491 unsigned int order
, i
;
6493 unsigned long flags
;
6494 /* find the first valid pfn */
6495 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6500 zone
= page_zone(pfn_to_page(pfn
));
6501 spin_lock_irqsave(&zone
->lock
, flags
);
6503 while (pfn
< end_pfn
) {
6504 if (!pfn_valid(pfn
)) {
6508 page
= pfn_to_page(pfn
);
6510 * The HWPoisoned page may be not in buddy system, and
6511 * page_count() is not 0.
6513 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6515 SetPageReserved(page
);
6519 BUG_ON(page_count(page
));
6520 BUG_ON(!PageBuddy(page
));
6521 order
= page_order(page
);
6522 #ifdef CONFIG_DEBUG_VM
6523 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6524 pfn
, 1 << order
, end_pfn
);
6526 list_del(&page
->lru
);
6527 rmv_page_order(page
);
6528 zone
->free_area
[order
].nr_free
--;
6529 for (i
= 0; i
< (1 << order
); i
++)
6530 SetPageReserved((page
+i
));
6531 pfn
+= (1 << order
);
6533 spin_unlock_irqrestore(&zone
->lock
, flags
);
6537 #ifdef CONFIG_MEMORY_FAILURE
6538 bool is_free_buddy_page(struct page
*page
)
6540 struct zone
*zone
= page_zone(page
);
6541 unsigned long pfn
= page_to_pfn(page
);
6542 unsigned long flags
;
6545 spin_lock_irqsave(&zone
->lock
, flags
);
6546 for (order
= 0; order
< MAX_ORDER
; order
++) {
6547 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6549 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6552 spin_unlock_irqrestore(&zone
->lock
, flags
);
6554 return order
< MAX_ORDER
;