2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/prefetch.h>
57 #include <linux/mm_inline.h>
58 #include <linux/migrate.h>
59 #include <linux/page-debug-flags.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
63 #include <asm/sections.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
69 static DEFINE_MUTEX(pcp_batch_high_lock
);
70 #define MIN_PERCPU_PAGELIST_FRACTION (8)
72 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
73 DEFINE_PER_CPU(int, numa_node
);
74 EXPORT_PER_CPU_SYMBOL(numa_node
);
77 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
79 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
80 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
81 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
82 * defined in <linux/topology.h>.
84 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
85 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
86 int _node_numa_mem_
[MAX_NUMNODES
];
90 * Array of node states.
92 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
93 [N_POSSIBLE
] = NODE_MASK_ALL
,
94 [N_ONLINE
] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY
] = { { [0] = 1UL } },
103 [N_CPU
] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states
);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock
);
111 unsigned long totalram_pages __read_mostly
;
112 unsigned long totalreserve_pages __read_mostly
;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly
;
121 int percpu_pagelist_fraction
;
122 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask
;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex
));
139 if (saved_gfp_mask
) {
140 gfp_allowed_mask
= saved_gfp_mask
;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex
));
148 WARN_ON(saved_gfp_mask
);
149 saved_gfp_mask
= gfp_allowed_mask
;
150 gfp_allowed_mask
&= ~GFP_IOFS
;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly
;
165 static void __free_pages_ok(struct page
*page
, unsigned int order
);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages
);
193 static char * const zone_names
[MAX_NR_ZONES
] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes
= 1024;
208 int user_min_free_kbytes
= -1;
210 static unsigned long __meminitdata nr_kernel_pages
;
211 static unsigned long __meminitdata nr_all_pages
;
212 static unsigned long __meminitdata dma_reserve
;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
217 static unsigned long __initdata required_kernelcore
;
218 static unsigned long __initdata required_movablecore
;
219 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone
);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
228 int nr_online_nodes __read_mostly
= 1;
229 EXPORT_SYMBOL(nr_node_ids
);
230 EXPORT_SYMBOL(nr_online_nodes
);
233 int page_group_by_mobility_disabled __read_mostly
;
235 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
237 if (unlikely(page_group_by_mobility_disabled
&&
238 migratetype
< MIGRATE_PCPTYPES
))
239 migratetype
= MIGRATE_UNMOVABLE
;
241 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
242 PB_migrate
, PB_migrate_end
);
245 bool oom_killer_disabled __read_mostly
;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
252 unsigned long pfn
= page_to_pfn(page
);
253 unsigned long sp
, start_pfn
;
256 seq
= zone_span_seqbegin(zone
);
257 start_pfn
= zone
->zone_start_pfn
;
258 sp
= zone
->spanned_pages
;
259 if (!zone_spans_pfn(zone
, pfn
))
261 } while (zone_span_seqretry(zone
, seq
));
264 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
265 pfn
, zone_to_nid(zone
), zone
->name
,
266 start_pfn
, start_pfn
+ sp
);
271 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
273 if (!pfn_valid_within(page_to_pfn(page
)))
275 if (zone
!= page_zone(page
))
281 * Temporary debugging check for pages not lying within a given zone.
283 static int bad_range(struct zone
*zone
, struct page
*page
)
285 if (page_outside_zone_boundaries(zone
, page
))
287 if (!page_is_consistent(zone
, page
))
293 static inline int bad_range(struct zone
*zone
, struct page
*page
)
299 static void bad_page(struct page
*page
, const char *reason
,
300 unsigned long bad_flags
)
302 static unsigned long resume
;
303 static unsigned long nr_shown
;
304 static unsigned long nr_unshown
;
306 /* Don't complain about poisoned pages */
307 if (PageHWPoison(page
)) {
308 page_mapcount_reset(page
); /* remove PageBuddy */
313 * Allow a burst of 60 reports, then keep quiet for that minute;
314 * or allow a steady drip of one report per second.
316 if (nr_shown
== 60) {
317 if (time_before(jiffies
, resume
)) {
323 "BUG: Bad page state: %lu messages suppressed\n",
330 resume
= jiffies
+ 60 * HZ
;
332 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
333 current
->comm
, page_to_pfn(page
));
334 dump_page_badflags(page
, reason
, bad_flags
);
339 /* Leave bad fields for debug, except PageBuddy could make trouble */
340 page_mapcount_reset(page
); /* remove PageBuddy */
341 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
345 * Higher-order pages are called "compound pages". They are structured thusly:
347 * The first PAGE_SIZE page is called the "head page".
349 * The remaining PAGE_SIZE pages are called "tail pages".
351 * All pages have PG_compound set. All tail pages have their ->first_page
352 * pointing at the head page.
354 * The first tail page's ->lru.next holds the address of the compound page's
355 * put_page() function. Its ->lru.prev holds the order of allocation.
356 * This usage means that zero-order pages may not be compound.
359 static void free_compound_page(struct page
*page
)
361 __free_pages_ok(page
, compound_order(page
));
364 void prep_compound_page(struct page
*page
, unsigned long order
)
367 int nr_pages
= 1 << order
;
369 set_compound_page_dtor(page
, free_compound_page
);
370 set_compound_order(page
, order
);
372 for (i
= 1; i
< nr_pages
; i
++) {
373 struct page
*p
= page
+ i
;
374 set_page_count(p
, 0);
375 p
->first_page
= page
;
376 /* Make sure p->first_page is always valid for PageTail() */
382 /* update __split_huge_page_refcount if you change this function */
383 static int destroy_compound_page(struct page
*page
, unsigned long order
)
386 int nr_pages
= 1 << order
;
389 if (unlikely(compound_order(page
) != order
)) {
390 bad_page(page
, "wrong compound order", 0);
394 __ClearPageHead(page
);
396 for (i
= 1; i
< nr_pages
; i
++) {
397 struct page
*p
= page
+ i
;
399 if (unlikely(!PageTail(p
))) {
400 bad_page(page
, "PageTail not set", 0);
402 } else if (unlikely(p
->first_page
!= page
)) {
403 bad_page(page
, "first_page not consistent", 0);
412 static inline void prep_zero_page(struct page
*page
, unsigned int order
,
418 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
419 * and __GFP_HIGHMEM from hard or soft interrupt context.
421 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
422 for (i
= 0; i
< (1 << order
); i
++)
423 clear_highpage(page
+ i
);
426 #ifdef CONFIG_DEBUG_PAGEALLOC
427 unsigned int _debug_guardpage_minorder
;
429 static int __init
debug_guardpage_minorder_setup(char *buf
)
433 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
434 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
437 _debug_guardpage_minorder
= res
;
438 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
441 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
443 static inline void set_page_guard_flag(struct page
*page
)
445 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
448 static inline void clear_page_guard_flag(struct page
*page
)
450 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
453 static inline void set_page_guard_flag(struct page
*page
) { }
454 static inline void clear_page_guard_flag(struct page
*page
) { }
457 static inline void set_page_order(struct page
*page
, unsigned int order
)
459 set_page_private(page
, order
);
460 __SetPageBuddy(page
);
463 static inline void rmv_page_order(struct page
*page
)
465 __ClearPageBuddy(page
);
466 set_page_private(page
, 0);
470 * Locate the struct page for both the matching buddy in our
471 * pair (buddy1) and the combined O(n+1) page they form (page).
473 * 1) Any buddy B1 will have an order O twin B2 which satisfies
474 * the following equation:
476 * For example, if the starting buddy (buddy2) is #8 its order
478 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
480 * 2) Any buddy B will have an order O+1 parent P which
481 * satisfies the following equation:
484 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
486 static inline unsigned long
487 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
489 return page_idx
^ (1 << order
);
493 * This function checks whether a page is free && is the buddy
494 * we can do coalesce a page and its buddy if
495 * (a) the buddy is not in a hole &&
496 * (b) the buddy is in the buddy system &&
497 * (c) a page and its buddy have the same order &&
498 * (d) a page and its buddy are in the same zone.
500 * For recording whether a page is in the buddy system, we set ->_mapcount
501 * PAGE_BUDDY_MAPCOUNT_VALUE.
502 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
503 * serialized by zone->lock.
505 * For recording page's order, we use page_private(page).
507 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
510 if (!pfn_valid_within(page_to_pfn(buddy
)))
513 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
514 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
516 if (page_zone_id(page
) != page_zone_id(buddy
))
522 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
523 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
526 * zone check is done late to avoid uselessly
527 * calculating zone/node ids for pages that could
530 if (page_zone_id(page
) != page_zone_id(buddy
))
539 * Freeing function for a buddy system allocator.
541 * The concept of a buddy system is to maintain direct-mapped table
542 * (containing bit values) for memory blocks of various "orders".
543 * The bottom level table contains the map for the smallest allocatable
544 * units of memory (here, pages), and each level above it describes
545 * pairs of units from the levels below, hence, "buddies".
546 * At a high level, all that happens here is marking the table entry
547 * at the bottom level available, and propagating the changes upward
548 * as necessary, plus some accounting needed to play nicely with other
549 * parts of the VM system.
550 * At each level, we keep a list of pages, which are heads of continuous
551 * free pages of length of (1 << order) and marked with _mapcount
552 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
554 * So when we are allocating or freeing one, we can derive the state of the
555 * other. That is, if we allocate a small block, and both were
556 * free, the remainder of the region must be split into blocks.
557 * If a block is freed, and its buddy is also free, then this
558 * triggers coalescing into a block of larger size.
563 static inline void __free_one_page(struct page
*page
,
565 struct zone
*zone
, unsigned int order
,
568 unsigned long page_idx
;
569 unsigned long combined_idx
;
570 unsigned long uninitialized_var(buddy_idx
);
573 VM_BUG_ON(!zone_is_initialized(zone
));
575 if (unlikely(PageCompound(page
)))
576 if (unlikely(destroy_compound_page(page
, order
)))
579 VM_BUG_ON(migratetype
== -1);
581 page_idx
= pfn
& ((1 << MAX_ORDER
) - 1);
583 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
584 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
586 while (order
< MAX_ORDER
-1) {
587 buddy_idx
= __find_buddy_index(page_idx
, order
);
588 buddy
= page
+ (buddy_idx
- page_idx
);
589 if (!page_is_buddy(page
, buddy
, order
))
592 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
593 * merge with it and move up one order.
595 if (page_is_guard(buddy
)) {
596 clear_page_guard_flag(buddy
);
597 set_page_private(page
, 0);
598 __mod_zone_freepage_state(zone
, 1 << order
,
601 list_del(&buddy
->lru
);
602 zone
->free_area
[order
].nr_free
--;
603 rmv_page_order(buddy
);
605 combined_idx
= buddy_idx
& page_idx
;
606 page
= page
+ (combined_idx
- page_idx
);
607 page_idx
= combined_idx
;
610 set_page_order(page
, order
);
613 * If this is not the largest possible page, check if the buddy
614 * of the next-highest order is free. If it is, it's possible
615 * that pages are being freed that will coalesce soon. In case,
616 * that is happening, add the free page to the tail of the list
617 * so it's less likely to be used soon and more likely to be merged
618 * as a higher order page
620 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
621 struct page
*higher_page
, *higher_buddy
;
622 combined_idx
= buddy_idx
& page_idx
;
623 higher_page
= page
+ (combined_idx
- page_idx
);
624 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
625 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
626 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
627 list_add_tail(&page
->lru
,
628 &zone
->free_area
[order
].free_list
[migratetype
]);
633 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
635 zone
->free_area
[order
].nr_free
++;
638 static inline int free_pages_check(struct page
*page
)
640 const char *bad_reason
= NULL
;
641 unsigned long bad_flags
= 0;
643 if (unlikely(page_mapcount(page
)))
644 bad_reason
= "nonzero mapcount";
645 if (unlikely(page
->mapping
!= NULL
))
646 bad_reason
= "non-NULL mapping";
647 if (unlikely(atomic_read(&page
->_count
) != 0))
648 bad_reason
= "nonzero _count";
649 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
650 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
651 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
653 if (unlikely(mem_cgroup_bad_page_check(page
)))
654 bad_reason
= "cgroup check failed";
655 if (unlikely(bad_reason
)) {
656 bad_page(page
, bad_reason
, bad_flags
);
659 page_cpupid_reset_last(page
);
660 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
661 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
666 * Frees a number of pages from the PCP lists
667 * Assumes all pages on list are in same zone, and of same order.
668 * count is the number of pages to free.
670 * If the zone was previously in an "all pages pinned" state then look to
671 * see if this freeing clears that state.
673 * And clear the zone's pages_scanned counter, to hold off the "all pages are
674 * pinned" detection logic.
676 static void free_pcppages_bulk(struct zone
*zone
, int count
,
677 struct per_cpu_pages
*pcp
)
682 unsigned long nr_scanned
;
684 spin_lock(&zone
->lock
);
685 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
687 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
691 struct list_head
*list
;
694 * Remove pages from lists in a round-robin fashion. A
695 * batch_free count is maintained that is incremented when an
696 * empty list is encountered. This is so more pages are freed
697 * off fuller lists instead of spinning excessively around empty
702 if (++migratetype
== MIGRATE_PCPTYPES
)
704 list
= &pcp
->lists
[migratetype
];
705 } while (list_empty(list
));
707 /* This is the only non-empty list. Free them all. */
708 if (batch_free
== MIGRATE_PCPTYPES
)
709 batch_free
= to_free
;
712 int mt
; /* migratetype of the to-be-freed page */
714 page
= list_entry(list
->prev
, struct page
, lru
);
715 /* must delete as __free_one_page list manipulates */
716 list_del(&page
->lru
);
717 mt
= get_freepage_migratetype(page
);
718 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
719 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
720 trace_mm_page_pcpu_drain(page
, 0, mt
);
721 if (likely(!is_migrate_isolate_page(page
))) {
722 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1);
723 if (is_migrate_cma(mt
))
724 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
, 1);
726 } while (--to_free
&& --batch_free
&& !list_empty(list
));
728 spin_unlock(&zone
->lock
);
731 static void free_one_page(struct zone
*zone
,
732 struct page
*page
, unsigned long pfn
,
736 unsigned long nr_scanned
;
737 spin_lock(&zone
->lock
);
738 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
740 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
742 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
743 if (unlikely(!is_migrate_isolate(migratetype
)))
744 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
745 spin_unlock(&zone
->lock
);
748 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
753 trace_mm_page_free(page
, order
);
754 kmemcheck_free_shadow(page
, order
);
757 page
->mapping
= NULL
;
758 for (i
= 0; i
< (1 << order
); i
++)
759 bad
+= free_pages_check(page
+ i
);
763 if (!PageHighMem(page
)) {
764 debug_check_no_locks_freed(page_address(page
),
766 debug_check_no_obj_freed(page_address(page
),
769 arch_free_page(page
, order
);
770 kernel_map_pages(page
, 1 << order
, 0);
775 static void __free_pages_ok(struct page
*page
, unsigned int order
)
779 unsigned long pfn
= page_to_pfn(page
);
781 if (!free_pages_prepare(page
, order
))
784 migratetype
= get_pfnblock_migratetype(page
, pfn
);
785 local_irq_save(flags
);
786 __count_vm_events(PGFREE
, 1 << order
);
787 set_freepage_migratetype(page
, migratetype
);
788 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
789 local_irq_restore(flags
);
792 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
794 unsigned int nr_pages
= 1 << order
;
795 struct page
*p
= page
;
799 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
801 __ClearPageReserved(p
);
802 set_page_count(p
, 0);
804 __ClearPageReserved(p
);
805 set_page_count(p
, 0);
807 page_zone(page
)->managed_pages
+= nr_pages
;
808 set_page_refcounted(page
);
809 __free_pages(page
, order
);
813 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
814 void __init
init_cma_reserved_pageblock(struct page
*page
)
816 unsigned i
= pageblock_nr_pages
;
817 struct page
*p
= page
;
820 __ClearPageReserved(p
);
821 set_page_count(p
, 0);
824 set_pageblock_migratetype(page
, MIGRATE_CMA
);
826 if (pageblock_order
>= MAX_ORDER
) {
827 i
= pageblock_nr_pages
;
830 set_page_refcounted(p
);
831 __free_pages(p
, MAX_ORDER
- 1);
832 p
+= MAX_ORDER_NR_PAGES
;
833 } while (i
-= MAX_ORDER_NR_PAGES
);
835 set_page_refcounted(page
);
836 __free_pages(page
, pageblock_order
);
839 adjust_managed_page_count(page
, pageblock_nr_pages
);
844 * The order of subdivision here is critical for the IO subsystem.
845 * Please do not alter this order without good reasons and regression
846 * testing. Specifically, as large blocks of memory are subdivided,
847 * the order in which smaller blocks are delivered depends on the order
848 * they're subdivided in this function. This is the primary factor
849 * influencing the order in which pages are delivered to the IO
850 * subsystem according to empirical testing, and this is also justified
851 * by considering the behavior of a buddy system containing a single
852 * large block of memory acted on by a series of small allocations.
853 * This behavior is a critical factor in sglist merging's success.
857 static inline void expand(struct zone
*zone
, struct page
*page
,
858 int low
, int high
, struct free_area
*area
,
861 unsigned long size
= 1 << high
;
867 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
869 #ifdef CONFIG_DEBUG_PAGEALLOC
870 if (high
< debug_guardpage_minorder()) {
872 * Mark as guard pages (or page), that will allow to
873 * merge back to allocator when buddy will be freed.
874 * Corresponding page table entries will not be touched,
875 * pages will stay not present in virtual address space
877 INIT_LIST_HEAD(&page
[size
].lru
);
878 set_page_guard_flag(&page
[size
]);
879 set_page_private(&page
[size
], high
);
880 /* Guard pages are not available for any usage */
881 __mod_zone_freepage_state(zone
, -(1 << high
),
886 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
888 set_page_order(&page
[size
], high
);
893 * This page is about to be returned from the page allocator
895 static inline int check_new_page(struct page
*page
)
897 const char *bad_reason
= NULL
;
898 unsigned long bad_flags
= 0;
900 if (unlikely(page_mapcount(page
)))
901 bad_reason
= "nonzero mapcount";
902 if (unlikely(page
->mapping
!= NULL
))
903 bad_reason
= "non-NULL mapping";
904 if (unlikely(atomic_read(&page
->_count
) != 0))
905 bad_reason
= "nonzero _count";
906 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
907 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
908 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
910 if (unlikely(mem_cgroup_bad_page_check(page
)))
911 bad_reason
= "cgroup check failed";
912 if (unlikely(bad_reason
)) {
913 bad_page(page
, bad_reason
, bad_flags
);
919 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
)
923 for (i
= 0; i
< (1 << order
); i
++) {
924 struct page
*p
= page
+ i
;
925 if (unlikely(check_new_page(p
)))
929 set_page_private(page
, 0);
930 set_page_refcounted(page
);
932 arch_alloc_page(page
, order
);
933 kernel_map_pages(page
, 1 << order
, 1);
935 if (gfp_flags
& __GFP_ZERO
)
936 prep_zero_page(page
, order
, gfp_flags
);
938 if (order
&& (gfp_flags
& __GFP_COMP
))
939 prep_compound_page(page
, order
);
945 * Go through the free lists for the given migratetype and remove
946 * the smallest available page from the freelists
949 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
952 unsigned int current_order
;
953 struct free_area
*area
;
956 /* Find a page of the appropriate size in the preferred list */
957 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
958 area
= &(zone
->free_area
[current_order
]);
959 if (list_empty(&area
->free_list
[migratetype
]))
962 page
= list_entry(area
->free_list
[migratetype
].next
,
964 list_del(&page
->lru
);
965 rmv_page_order(page
);
967 expand(zone
, page
, order
, current_order
, area
, migratetype
);
968 set_freepage_migratetype(page
, migratetype
);
977 * This array describes the order lists are fallen back to when
978 * the free lists for the desirable migrate type are depleted
980 static int fallbacks
[MIGRATE_TYPES
][4] = {
981 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
982 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
984 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
985 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
987 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
989 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
990 #ifdef CONFIG_MEMORY_ISOLATION
991 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
996 * Move the free pages in a range to the free lists of the requested type.
997 * Note that start_page and end_pages are not aligned on a pageblock
998 * boundary. If alignment is required, use move_freepages_block()
1000 int move_freepages(struct zone
*zone
,
1001 struct page
*start_page
, struct page
*end_page
,
1005 unsigned long order
;
1006 int pages_moved
= 0;
1008 #ifndef CONFIG_HOLES_IN_ZONE
1010 * page_zone is not safe to call in this context when
1011 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1012 * anyway as we check zone boundaries in move_freepages_block().
1013 * Remove at a later date when no bug reports exist related to
1014 * grouping pages by mobility
1016 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1019 for (page
= start_page
; page
<= end_page
;) {
1020 /* Make sure we are not inadvertently changing nodes */
1021 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1023 if (!pfn_valid_within(page_to_pfn(page
))) {
1028 if (!PageBuddy(page
)) {
1033 order
= page_order(page
);
1034 list_move(&page
->lru
,
1035 &zone
->free_area
[order
].free_list
[migratetype
]);
1036 set_freepage_migratetype(page
, migratetype
);
1038 pages_moved
+= 1 << order
;
1044 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1047 unsigned long start_pfn
, end_pfn
;
1048 struct page
*start_page
, *end_page
;
1050 start_pfn
= page_to_pfn(page
);
1051 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1052 start_page
= pfn_to_page(start_pfn
);
1053 end_page
= start_page
+ pageblock_nr_pages
- 1;
1054 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1056 /* Do not cross zone boundaries */
1057 if (!zone_spans_pfn(zone
, start_pfn
))
1059 if (!zone_spans_pfn(zone
, end_pfn
))
1062 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1065 static void change_pageblock_range(struct page
*pageblock_page
,
1066 int start_order
, int migratetype
)
1068 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1070 while (nr_pageblocks
--) {
1071 set_pageblock_migratetype(pageblock_page
, migratetype
);
1072 pageblock_page
+= pageblock_nr_pages
;
1077 * If breaking a large block of pages, move all free pages to the preferred
1078 * allocation list. If falling back for a reclaimable kernel allocation, be
1079 * more aggressive about taking ownership of free pages.
1081 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1082 * nor move CMA pages to different free lists. We don't want unmovable pages
1083 * to be allocated from MIGRATE_CMA areas.
1085 * Returns the new migratetype of the pageblock (or the same old migratetype
1086 * if it was unchanged).
1088 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1089 int start_type
, int fallback_type
)
1091 int current_order
= page_order(page
);
1094 * When borrowing from MIGRATE_CMA, we need to release the excess
1095 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1096 * is set to CMA so it is returned to the correct freelist in case
1097 * the page ends up being not actually allocated from the pcp lists.
1099 if (is_migrate_cma(fallback_type
))
1100 return fallback_type
;
1102 /* Take ownership for orders >= pageblock_order */
1103 if (current_order
>= pageblock_order
) {
1104 change_pageblock_range(page
, current_order
, start_type
);
1108 if (current_order
>= pageblock_order
/ 2 ||
1109 start_type
== MIGRATE_RECLAIMABLE
||
1110 page_group_by_mobility_disabled
) {
1113 pages
= move_freepages_block(zone
, page
, start_type
);
1115 /* Claim the whole block if over half of it is free */
1116 if (pages
>= (1 << (pageblock_order
-1)) ||
1117 page_group_by_mobility_disabled
) {
1119 set_pageblock_migratetype(page
, start_type
);
1125 return fallback_type
;
1128 /* Remove an element from the buddy allocator from the fallback list */
1129 static inline struct page
*
1130 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1132 struct free_area
*area
;
1133 unsigned int current_order
;
1135 int migratetype
, new_type
, i
;
1137 /* Find the largest possible block of pages in the other list */
1138 for (current_order
= MAX_ORDER
-1;
1139 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1142 migratetype
= fallbacks
[start_migratetype
][i
];
1144 /* MIGRATE_RESERVE handled later if necessary */
1145 if (migratetype
== MIGRATE_RESERVE
)
1148 area
= &(zone
->free_area
[current_order
]);
1149 if (list_empty(&area
->free_list
[migratetype
]))
1152 page
= list_entry(area
->free_list
[migratetype
].next
,
1156 new_type
= try_to_steal_freepages(zone
, page
,
1160 /* Remove the page from the freelists */
1161 list_del(&page
->lru
);
1162 rmv_page_order(page
);
1164 expand(zone
, page
, order
, current_order
, area
,
1166 /* The freepage_migratetype may differ from pageblock's
1167 * migratetype depending on the decisions in
1168 * try_to_steal_freepages. This is OK as long as it does
1169 * not differ for MIGRATE_CMA type.
1171 set_freepage_migratetype(page
, new_type
);
1173 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1174 start_migratetype
, migratetype
, new_type
);
1184 * Do the hard work of removing an element from the buddy allocator.
1185 * Call me with the zone->lock already held.
1187 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1193 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1195 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1196 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1199 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1200 * is used because __rmqueue_smallest is an inline function
1201 * and we want just one call site
1204 migratetype
= MIGRATE_RESERVE
;
1209 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1214 * Obtain a specified number of elements from the buddy allocator, all under
1215 * a single hold of the lock, for efficiency. Add them to the supplied list.
1216 * Returns the number of new pages which were placed at *list.
1218 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1219 unsigned long count
, struct list_head
*list
,
1220 int migratetype
, bool cold
)
1224 spin_lock(&zone
->lock
);
1225 for (i
= 0; i
< count
; ++i
) {
1226 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1227 if (unlikely(page
== NULL
))
1231 * Split buddy pages returned by expand() are received here
1232 * in physical page order. The page is added to the callers and
1233 * list and the list head then moves forward. From the callers
1234 * perspective, the linked list is ordered by page number in
1235 * some conditions. This is useful for IO devices that can
1236 * merge IO requests if the physical pages are ordered
1240 list_add(&page
->lru
, list
);
1242 list_add_tail(&page
->lru
, list
);
1244 if (is_migrate_cma(get_freepage_migratetype(page
)))
1245 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1248 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1249 spin_unlock(&zone
->lock
);
1255 * Called from the vmstat counter updater to drain pagesets of this
1256 * currently executing processor on remote nodes after they have
1259 * Note that this function must be called with the thread pinned to
1260 * a single processor.
1262 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1264 unsigned long flags
;
1265 int to_drain
, batch
;
1267 local_irq_save(flags
);
1268 batch
= ACCESS_ONCE(pcp
->batch
);
1269 to_drain
= min(pcp
->count
, batch
);
1271 free_pcppages_bulk(zone
, to_drain
, pcp
);
1272 pcp
->count
-= to_drain
;
1274 local_irq_restore(flags
);
1279 * Drain pages of the indicated processor.
1281 * The processor must either be the current processor and the
1282 * thread pinned to the current processor or a processor that
1285 static void drain_pages(unsigned int cpu
)
1287 unsigned long flags
;
1290 for_each_populated_zone(zone
) {
1291 struct per_cpu_pageset
*pset
;
1292 struct per_cpu_pages
*pcp
;
1294 local_irq_save(flags
);
1295 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1299 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1302 local_irq_restore(flags
);
1307 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1309 void drain_local_pages(void *arg
)
1311 drain_pages(smp_processor_id());
1315 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1317 * Note that this code is protected against sending an IPI to an offline
1318 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1319 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1320 * nothing keeps CPUs from showing up after we populated the cpumask and
1321 * before the call to on_each_cpu_mask().
1323 void drain_all_pages(void)
1326 struct per_cpu_pageset
*pcp
;
1330 * Allocate in the BSS so we wont require allocation in
1331 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1333 static cpumask_t cpus_with_pcps
;
1336 * We don't care about racing with CPU hotplug event
1337 * as offline notification will cause the notified
1338 * cpu to drain that CPU pcps and on_each_cpu_mask
1339 * disables preemption as part of its processing
1341 for_each_online_cpu(cpu
) {
1342 bool has_pcps
= false;
1343 for_each_populated_zone(zone
) {
1344 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1345 if (pcp
->pcp
.count
) {
1351 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1353 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1355 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1358 #ifdef CONFIG_HIBERNATION
1360 void mark_free_pages(struct zone
*zone
)
1362 unsigned long pfn
, max_zone_pfn
;
1363 unsigned long flags
;
1364 unsigned int order
, t
;
1365 struct list_head
*curr
;
1367 if (zone_is_empty(zone
))
1370 spin_lock_irqsave(&zone
->lock
, flags
);
1372 max_zone_pfn
= zone_end_pfn(zone
);
1373 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1374 if (pfn_valid(pfn
)) {
1375 struct page
*page
= pfn_to_page(pfn
);
1377 if (!swsusp_page_is_forbidden(page
))
1378 swsusp_unset_page_free(page
);
1381 for_each_migratetype_order(order
, t
) {
1382 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1385 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1386 for (i
= 0; i
< (1UL << order
); i
++)
1387 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1390 spin_unlock_irqrestore(&zone
->lock
, flags
);
1392 #endif /* CONFIG_PM */
1395 * Free a 0-order page
1396 * cold == true ? free a cold page : free a hot page
1398 void free_hot_cold_page(struct page
*page
, bool cold
)
1400 struct zone
*zone
= page_zone(page
);
1401 struct per_cpu_pages
*pcp
;
1402 unsigned long flags
;
1403 unsigned long pfn
= page_to_pfn(page
);
1406 if (!free_pages_prepare(page
, 0))
1409 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1410 set_freepage_migratetype(page
, migratetype
);
1411 local_irq_save(flags
);
1412 __count_vm_event(PGFREE
);
1415 * We only track unmovable, reclaimable and movable on pcp lists.
1416 * Free ISOLATE pages back to the allocator because they are being
1417 * offlined but treat RESERVE as movable pages so we can get those
1418 * areas back if necessary. Otherwise, we may have to free
1419 * excessively into the page allocator
1421 if (migratetype
>= MIGRATE_PCPTYPES
) {
1422 if (unlikely(is_migrate_isolate(migratetype
))) {
1423 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1426 migratetype
= MIGRATE_MOVABLE
;
1429 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1431 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1433 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1435 if (pcp
->count
>= pcp
->high
) {
1436 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1437 free_pcppages_bulk(zone
, batch
, pcp
);
1438 pcp
->count
-= batch
;
1442 local_irq_restore(flags
);
1446 * Free a list of 0-order pages
1448 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1450 struct page
*page
, *next
;
1452 list_for_each_entry_safe(page
, next
, list
, lru
) {
1453 trace_mm_page_free_batched(page
, cold
);
1454 free_hot_cold_page(page
, cold
);
1459 * split_page takes a non-compound higher-order page, and splits it into
1460 * n (1<<order) sub-pages: page[0..n]
1461 * Each sub-page must be freed individually.
1463 * Note: this is probably too low level an operation for use in drivers.
1464 * Please consult with lkml before using this in your driver.
1466 void split_page(struct page
*page
, unsigned int order
)
1470 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1471 VM_BUG_ON_PAGE(!page_count(page
), page
);
1473 #ifdef CONFIG_KMEMCHECK
1475 * Split shadow pages too, because free(page[0]) would
1476 * otherwise free the whole shadow.
1478 if (kmemcheck_page_is_tracked(page
))
1479 split_page(virt_to_page(page
[0].shadow
), order
);
1482 for (i
= 1; i
< (1 << order
); i
++)
1483 set_page_refcounted(page
+ i
);
1485 EXPORT_SYMBOL_GPL(split_page
);
1487 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1489 unsigned long watermark
;
1493 BUG_ON(!PageBuddy(page
));
1495 zone
= page_zone(page
);
1496 mt
= get_pageblock_migratetype(page
);
1498 if (!is_migrate_isolate(mt
)) {
1499 /* Obey watermarks as if the page was being allocated */
1500 watermark
= low_wmark_pages(zone
) + (1 << order
);
1501 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1504 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1507 /* Remove page from free list */
1508 list_del(&page
->lru
);
1509 zone
->free_area
[order
].nr_free
--;
1510 rmv_page_order(page
);
1512 /* Set the pageblock if the isolated page is at least a pageblock */
1513 if (order
>= pageblock_order
- 1) {
1514 struct page
*endpage
= page
+ (1 << order
) - 1;
1515 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1516 int mt
= get_pageblock_migratetype(page
);
1517 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1518 set_pageblock_migratetype(page
,
1523 return 1UL << order
;
1527 * Similar to split_page except the page is already free. As this is only
1528 * being used for migration, the migratetype of the block also changes.
1529 * As this is called with interrupts disabled, the caller is responsible
1530 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1533 * Note: this is probably too low level an operation for use in drivers.
1534 * Please consult with lkml before using this in your driver.
1536 int split_free_page(struct page
*page
)
1541 order
= page_order(page
);
1543 nr_pages
= __isolate_free_page(page
, order
);
1547 /* Split into individual pages */
1548 set_page_refcounted(page
);
1549 split_page(page
, order
);
1554 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1555 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1559 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1560 struct zone
*zone
, unsigned int order
,
1561 gfp_t gfp_flags
, int migratetype
)
1563 unsigned long flags
;
1565 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1568 if (likely(order
== 0)) {
1569 struct per_cpu_pages
*pcp
;
1570 struct list_head
*list
;
1572 local_irq_save(flags
);
1573 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1574 list
= &pcp
->lists
[migratetype
];
1575 if (list_empty(list
)) {
1576 pcp
->count
+= rmqueue_bulk(zone
, 0,
1579 if (unlikely(list_empty(list
)))
1584 page
= list_entry(list
->prev
, struct page
, lru
);
1586 page
= list_entry(list
->next
, struct page
, lru
);
1588 list_del(&page
->lru
);
1591 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1593 * __GFP_NOFAIL is not to be used in new code.
1595 * All __GFP_NOFAIL callers should be fixed so that they
1596 * properly detect and handle allocation failures.
1598 * We most definitely don't want callers attempting to
1599 * allocate greater than order-1 page units with
1602 WARN_ON_ONCE(order
> 1);
1604 spin_lock_irqsave(&zone
->lock
, flags
);
1605 page
= __rmqueue(zone
, order
, migratetype
);
1606 spin_unlock(&zone
->lock
);
1609 __mod_zone_freepage_state(zone
, -(1 << order
),
1610 get_freepage_migratetype(page
));
1613 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1614 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1615 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1616 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1618 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1619 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1620 local_irq_restore(flags
);
1622 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1623 if (prep_new_page(page
, order
, gfp_flags
))
1628 local_irq_restore(flags
);
1632 #ifdef CONFIG_FAIL_PAGE_ALLOC
1635 struct fault_attr attr
;
1637 u32 ignore_gfp_highmem
;
1638 u32 ignore_gfp_wait
;
1640 } fail_page_alloc
= {
1641 .attr
= FAULT_ATTR_INITIALIZER
,
1642 .ignore_gfp_wait
= 1,
1643 .ignore_gfp_highmem
= 1,
1647 static int __init
setup_fail_page_alloc(char *str
)
1649 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1651 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1653 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1655 if (order
< fail_page_alloc
.min_order
)
1657 if (gfp_mask
& __GFP_NOFAIL
)
1659 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1661 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1664 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1667 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1669 static int __init
fail_page_alloc_debugfs(void)
1671 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1674 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1675 &fail_page_alloc
.attr
);
1677 return PTR_ERR(dir
);
1679 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1680 &fail_page_alloc
.ignore_gfp_wait
))
1682 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1683 &fail_page_alloc
.ignore_gfp_highmem
))
1685 if (!debugfs_create_u32("min-order", mode
, dir
,
1686 &fail_page_alloc
.min_order
))
1691 debugfs_remove_recursive(dir
);
1696 late_initcall(fail_page_alloc_debugfs
);
1698 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1700 #else /* CONFIG_FAIL_PAGE_ALLOC */
1702 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1707 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1710 * Return true if free pages are above 'mark'. This takes into account the order
1711 * of the allocation.
1713 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1714 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1717 /* free_pages my go negative - that's OK */
1722 free_pages
-= (1 << order
) - 1;
1723 if (alloc_flags
& ALLOC_HIGH
)
1725 if (alloc_flags
& ALLOC_HARDER
)
1728 /* If allocation can't use CMA areas don't use free CMA pages */
1729 if (!(alloc_flags
& ALLOC_CMA
))
1730 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1733 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1735 for (o
= 0; o
< order
; o
++) {
1736 /* At the next order, this order's pages become unavailable */
1737 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1739 /* Require fewer higher order pages to be free */
1742 if (free_pages
<= min
)
1748 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1749 int classzone_idx
, int alloc_flags
)
1751 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1752 zone_page_state(z
, NR_FREE_PAGES
));
1755 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1756 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1758 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1760 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1761 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1763 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1769 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1770 * skip over zones that are not allowed by the cpuset, or that have
1771 * been recently (in last second) found to be nearly full. See further
1772 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1773 * that have to skip over a lot of full or unallowed zones.
1775 * If the zonelist cache is present in the passed zonelist, then
1776 * returns a pointer to the allowed node mask (either the current
1777 * tasks mems_allowed, or node_states[N_MEMORY].)
1779 * If the zonelist cache is not available for this zonelist, does
1780 * nothing and returns NULL.
1782 * If the fullzones BITMAP in the zonelist cache is stale (more than
1783 * a second since last zap'd) then we zap it out (clear its bits.)
1785 * We hold off even calling zlc_setup, until after we've checked the
1786 * first zone in the zonelist, on the theory that most allocations will
1787 * be satisfied from that first zone, so best to examine that zone as
1788 * quickly as we can.
1790 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1792 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1793 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1795 zlc
= zonelist
->zlcache_ptr
;
1799 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1800 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1801 zlc
->last_full_zap
= jiffies
;
1804 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1805 &cpuset_current_mems_allowed
:
1806 &node_states
[N_MEMORY
];
1807 return allowednodes
;
1811 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1812 * if it is worth looking at further for free memory:
1813 * 1) Check that the zone isn't thought to be full (doesn't have its
1814 * bit set in the zonelist_cache fullzones BITMAP).
1815 * 2) Check that the zones node (obtained from the zonelist_cache
1816 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1817 * Return true (non-zero) if zone is worth looking at further, or
1818 * else return false (zero) if it is not.
1820 * This check -ignores- the distinction between various watermarks,
1821 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1822 * found to be full for any variation of these watermarks, it will
1823 * be considered full for up to one second by all requests, unless
1824 * we are so low on memory on all allowed nodes that we are forced
1825 * into the second scan of the zonelist.
1827 * In the second scan we ignore this zonelist cache and exactly
1828 * apply the watermarks to all zones, even it is slower to do so.
1829 * We are low on memory in the second scan, and should leave no stone
1830 * unturned looking for a free page.
1832 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1833 nodemask_t
*allowednodes
)
1835 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1836 int i
; /* index of *z in zonelist zones */
1837 int n
; /* node that zone *z is on */
1839 zlc
= zonelist
->zlcache_ptr
;
1843 i
= z
- zonelist
->_zonerefs
;
1846 /* This zone is worth trying if it is allowed but not full */
1847 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1851 * Given 'z' scanning a zonelist, set the corresponding bit in
1852 * zlc->fullzones, so that subsequent attempts to allocate a page
1853 * from that zone don't waste time re-examining it.
1855 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1857 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1858 int i
; /* index of *z in zonelist zones */
1860 zlc
= zonelist
->zlcache_ptr
;
1864 i
= z
- zonelist
->_zonerefs
;
1866 set_bit(i
, zlc
->fullzones
);
1870 * clear all zones full, called after direct reclaim makes progress so that
1871 * a zone that was recently full is not skipped over for up to a second
1873 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1875 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1877 zlc
= zonelist
->zlcache_ptr
;
1881 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1884 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1886 return local_zone
->node
== zone
->node
;
1889 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1891 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1895 #else /* CONFIG_NUMA */
1897 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1902 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1903 nodemask_t
*allowednodes
)
1908 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1912 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1916 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1921 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1926 #endif /* CONFIG_NUMA */
1928 static void reset_alloc_batches(struct zone
*preferred_zone
)
1930 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
1933 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
1934 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
1935 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
1936 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1937 } while (zone
++ != preferred_zone
);
1941 * get_page_from_freelist goes through the zonelist trying to allocate
1944 static struct page
*
1945 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1946 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1947 struct zone
*preferred_zone
, int classzone_idx
, int migratetype
)
1950 struct page
*page
= NULL
;
1952 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1953 int zlc_active
= 0; /* set if using zonelist_cache */
1954 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1955 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
1956 (gfp_mask
& __GFP_WRITE
);
1957 int nr_fair_skipped
= 0;
1958 bool zonelist_rescan
;
1961 zonelist_rescan
= false;
1964 * Scan zonelist, looking for a zone with enough free.
1965 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1967 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1968 high_zoneidx
, nodemask
) {
1971 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1972 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1974 if (cpusets_enabled() &&
1975 (alloc_flags
& ALLOC_CPUSET
) &&
1976 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1979 * Distribute pages in proportion to the individual
1980 * zone size to ensure fair page aging. The zone a
1981 * page was allocated in should have no effect on the
1982 * time the page has in memory before being reclaimed.
1984 if (alloc_flags
& ALLOC_FAIR
) {
1985 if (!zone_local(preferred_zone
, zone
))
1987 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
1993 * When allocating a page cache page for writing, we
1994 * want to get it from a zone that is within its dirty
1995 * limit, such that no single zone holds more than its
1996 * proportional share of globally allowed dirty pages.
1997 * The dirty limits take into account the zone's
1998 * lowmem reserves and high watermark so that kswapd
1999 * should be able to balance it without having to
2000 * write pages from its LRU list.
2002 * This may look like it could increase pressure on
2003 * lower zones by failing allocations in higher zones
2004 * before they are full. But the pages that do spill
2005 * over are limited as the lower zones are protected
2006 * by this very same mechanism. It should not become
2007 * a practical burden to them.
2009 * XXX: For now, allow allocations to potentially
2010 * exceed the per-zone dirty limit in the slowpath
2011 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2012 * which is important when on a NUMA setup the allowed
2013 * zones are together not big enough to reach the
2014 * global limit. The proper fix for these situations
2015 * will require awareness of zones in the
2016 * dirty-throttling and the flusher threads.
2018 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2021 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2022 if (!zone_watermark_ok(zone
, order
, mark
,
2023 classzone_idx
, alloc_flags
)) {
2026 /* Checked here to keep the fast path fast */
2027 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2028 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2031 if (IS_ENABLED(CONFIG_NUMA
) &&
2032 !did_zlc_setup
&& nr_online_nodes
> 1) {
2034 * we do zlc_setup if there are multiple nodes
2035 * and before considering the first zone allowed
2038 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2043 if (zone_reclaim_mode
== 0 ||
2044 !zone_allows_reclaim(preferred_zone
, zone
))
2045 goto this_zone_full
;
2048 * As we may have just activated ZLC, check if the first
2049 * eligible zone has failed zone_reclaim recently.
2051 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2052 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2055 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2057 case ZONE_RECLAIM_NOSCAN
:
2060 case ZONE_RECLAIM_FULL
:
2061 /* scanned but unreclaimable */
2064 /* did we reclaim enough */
2065 if (zone_watermark_ok(zone
, order
, mark
,
2066 classzone_idx
, alloc_flags
))
2070 * Failed to reclaim enough to meet watermark.
2071 * Only mark the zone full if checking the min
2072 * watermark or if we failed to reclaim just
2073 * 1<<order pages or else the page allocator
2074 * fastpath will prematurely mark zones full
2075 * when the watermark is between the low and
2078 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2079 ret
== ZONE_RECLAIM_SOME
)
2080 goto this_zone_full
;
2087 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2088 gfp_mask
, migratetype
);
2092 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2093 zlc_mark_zone_full(zonelist
, z
);
2098 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2099 * necessary to allocate the page. The expectation is
2100 * that the caller is taking steps that will free more
2101 * memory. The caller should avoid the page being used
2102 * for !PFMEMALLOC purposes.
2104 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2109 * The first pass makes sure allocations are spread fairly within the
2110 * local node. However, the local node might have free pages left
2111 * after the fairness batches are exhausted, and remote zones haven't
2112 * even been considered yet. Try once more without fairness, and
2113 * include remote zones now, before entering the slowpath and waking
2114 * kswapd: prefer spilling to a remote zone over swapping locally.
2116 if (alloc_flags
& ALLOC_FAIR
) {
2117 alloc_flags
&= ~ALLOC_FAIR
;
2118 if (nr_fair_skipped
) {
2119 zonelist_rescan
= true;
2120 reset_alloc_batches(preferred_zone
);
2122 if (nr_online_nodes
> 1)
2123 zonelist_rescan
= true;
2126 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2127 /* Disable zlc cache for second zonelist scan */
2129 zonelist_rescan
= true;
2132 if (zonelist_rescan
)
2139 * Large machines with many possible nodes should not always dump per-node
2140 * meminfo in irq context.
2142 static inline bool should_suppress_show_mem(void)
2147 ret
= in_interrupt();
2152 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2153 DEFAULT_RATELIMIT_INTERVAL
,
2154 DEFAULT_RATELIMIT_BURST
);
2156 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2158 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2160 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2161 debug_guardpage_minorder() > 0)
2165 * This documents exceptions given to allocations in certain
2166 * contexts that are allowed to allocate outside current's set
2169 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2170 if (test_thread_flag(TIF_MEMDIE
) ||
2171 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2172 filter
&= ~SHOW_MEM_FILTER_NODES
;
2173 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2174 filter
&= ~SHOW_MEM_FILTER_NODES
;
2177 struct va_format vaf
;
2180 va_start(args
, fmt
);
2185 pr_warn("%pV", &vaf
);
2190 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2191 current
->comm
, order
, gfp_mask
);
2194 if (!should_suppress_show_mem())
2199 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2200 unsigned long did_some_progress
,
2201 unsigned long pages_reclaimed
)
2203 /* Do not loop if specifically requested */
2204 if (gfp_mask
& __GFP_NORETRY
)
2207 /* Always retry if specifically requested */
2208 if (gfp_mask
& __GFP_NOFAIL
)
2212 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2213 * making forward progress without invoking OOM. Suspend also disables
2214 * storage devices so kswapd will not help. Bail if we are suspending.
2216 if (!did_some_progress
&& pm_suspended_storage())
2220 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2221 * means __GFP_NOFAIL, but that may not be true in other
2224 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2228 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2229 * specified, then we retry until we no longer reclaim any pages
2230 * (above), or we've reclaimed an order of pages at least as
2231 * large as the allocation's order. In both cases, if the
2232 * allocation still fails, we stop retrying.
2234 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2240 static inline struct page
*
2241 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2242 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2243 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2244 int classzone_idx
, int migratetype
)
2248 /* Acquire the per-zone oom lock for each zone */
2249 if (!oom_zonelist_trylock(zonelist
, gfp_mask
)) {
2250 schedule_timeout_uninterruptible(1);
2255 * Go through the zonelist yet one more time, keep very high watermark
2256 * here, this is only to catch a parallel oom killing, we must fail if
2257 * we're still under heavy pressure.
2259 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2260 order
, zonelist
, high_zoneidx
,
2261 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2262 preferred_zone
, classzone_idx
, migratetype
);
2266 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2267 /* The OOM killer will not help higher order allocs */
2268 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2270 /* The OOM killer does not needlessly kill tasks for lowmem */
2271 if (high_zoneidx
< ZONE_NORMAL
)
2274 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2275 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2276 * The caller should handle page allocation failure by itself if
2277 * it specifies __GFP_THISNODE.
2278 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2280 if (gfp_mask
& __GFP_THISNODE
)
2283 /* Exhausted what can be done so it's blamo time */
2284 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2287 oom_zonelist_unlock(zonelist
, gfp_mask
);
2291 #ifdef CONFIG_COMPACTION
2292 /* Try memory compaction for high-order allocations before reclaim */
2293 static struct page
*
2294 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2295 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2296 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2297 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2298 int *contended_compaction
, bool *deferred_compaction
)
2300 struct zone
*last_compact_zone
= NULL
;
2301 unsigned long compact_result
;
2307 current
->flags
|= PF_MEMALLOC
;
2308 compact_result
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2310 contended_compaction
,
2311 &last_compact_zone
);
2312 current
->flags
&= ~PF_MEMALLOC
;
2314 switch (compact_result
) {
2315 case COMPACT_DEFERRED
:
2316 *deferred_compaction
= true;
2318 case COMPACT_SKIPPED
:
2325 * At least in one zone compaction wasn't deferred or skipped, so let's
2326 * count a compaction stall
2328 count_vm_event(COMPACTSTALL
);
2330 /* Page migration frees to the PCP lists but we want merging */
2331 drain_pages(get_cpu());
2334 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2335 order
, zonelist
, high_zoneidx
,
2336 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2337 preferred_zone
, classzone_idx
, migratetype
);
2340 struct zone
*zone
= page_zone(page
);
2342 zone
->compact_blockskip_flush
= false;
2343 compaction_defer_reset(zone
, order
, true);
2344 count_vm_event(COMPACTSUCCESS
);
2349 * last_compact_zone is where try_to_compact_pages thought allocation
2350 * should succeed, so it did not defer compaction. But here we know
2351 * that it didn't succeed, so we do the defer.
2353 if (last_compact_zone
&& mode
!= MIGRATE_ASYNC
)
2354 defer_compaction(last_compact_zone
, order
);
2357 * It's bad if compaction run occurs and fails. The most likely reason
2358 * is that pages exist, but not enough to satisfy watermarks.
2360 count_vm_event(COMPACTFAIL
);
2367 static inline struct page
*
2368 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2369 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2370 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2371 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2372 int *contended_compaction
, bool *deferred_compaction
)
2376 #endif /* CONFIG_COMPACTION */
2378 /* Perform direct synchronous page reclaim */
2380 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2381 nodemask_t
*nodemask
)
2383 struct reclaim_state reclaim_state
;
2388 /* We now go into synchronous reclaim */
2389 cpuset_memory_pressure_bump();
2390 current
->flags
|= PF_MEMALLOC
;
2391 lockdep_set_current_reclaim_state(gfp_mask
);
2392 reclaim_state
.reclaimed_slab
= 0;
2393 current
->reclaim_state
= &reclaim_state
;
2395 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2397 current
->reclaim_state
= NULL
;
2398 lockdep_clear_current_reclaim_state();
2399 current
->flags
&= ~PF_MEMALLOC
;
2406 /* The really slow allocator path where we enter direct reclaim */
2407 static inline struct page
*
2408 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2409 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2410 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2411 int classzone_idx
, int migratetype
, unsigned long *did_some_progress
)
2413 struct page
*page
= NULL
;
2414 bool drained
= false;
2416 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2418 if (unlikely(!(*did_some_progress
)))
2421 /* After successful reclaim, reconsider all zones for allocation */
2422 if (IS_ENABLED(CONFIG_NUMA
))
2423 zlc_clear_zones_full(zonelist
);
2426 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2427 zonelist
, high_zoneidx
,
2428 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2429 preferred_zone
, classzone_idx
,
2433 * If an allocation failed after direct reclaim, it could be because
2434 * pages are pinned on the per-cpu lists. Drain them and try again
2436 if (!page
&& !drained
) {
2446 * This is called in the allocator slow-path if the allocation request is of
2447 * sufficient urgency to ignore watermarks and take other desperate measures
2449 static inline struct page
*
2450 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2451 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2452 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2453 int classzone_idx
, int migratetype
)
2458 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2459 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2460 preferred_zone
, classzone_idx
, migratetype
);
2462 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2463 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2464 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2469 static void wake_all_kswapds(unsigned int order
,
2470 struct zonelist
*zonelist
,
2471 enum zone_type high_zoneidx
,
2472 struct zone
*preferred_zone
,
2473 nodemask_t
*nodemask
)
2478 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2479 high_zoneidx
, nodemask
)
2480 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2484 gfp_to_alloc_flags(gfp_t gfp_mask
)
2486 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2487 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2489 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2490 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2493 * The caller may dip into page reserves a bit more if the caller
2494 * cannot run direct reclaim, or if the caller has realtime scheduling
2495 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2496 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2498 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2502 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2503 * if it can't schedule.
2505 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2506 alloc_flags
|= ALLOC_HARDER
;
2508 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2509 * comment for __cpuset_node_allowed_softwall().
2511 alloc_flags
&= ~ALLOC_CPUSET
;
2512 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2513 alloc_flags
|= ALLOC_HARDER
;
2515 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2516 if (gfp_mask
& __GFP_MEMALLOC
)
2517 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2518 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2519 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2520 else if (!in_interrupt() &&
2521 ((current
->flags
& PF_MEMALLOC
) ||
2522 unlikely(test_thread_flag(TIF_MEMDIE
))))
2523 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2526 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2527 alloc_flags
|= ALLOC_CMA
;
2532 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2534 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2537 static inline struct page
*
2538 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2539 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2540 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2541 int classzone_idx
, int migratetype
)
2543 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2544 struct page
*page
= NULL
;
2546 unsigned long pages_reclaimed
= 0;
2547 unsigned long did_some_progress
;
2548 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2549 bool deferred_compaction
= false;
2550 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2553 * In the slowpath, we sanity check order to avoid ever trying to
2554 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2555 * be using allocators in order of preference for an area that is
2558 if (order
>= MAX_ORDER
) {
2559 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2564 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2565 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2566 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2567 * using a larger set of nodes after it has established that the
2568 * allowed per node queues are empty and that nodes are
2571 if (IS_ENABLED(CONFIG_NUMA
) &&
2572 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2576 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2577 wake_all_kswapds(order
, zonelist
, high_zoneidx
,
2578 preferred_zone
, nodemask
);
2581 * OK, we're below the kswapd watermark and have kicked background
2582 * reclaim. Now things get more complex, so set up alloc_flags according
2583 * to how we want to proceed.
2585 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2588 * Find the true preferred zone if the allocation is unconstrained by
2591 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
) {
2592 struct zoneref
*preferred_zoneref
;
2593 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2594 NULL
, &preferred_zone
);
2595 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2599 /* This is the last chance, in general, before the goto nopage. */
2600 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2601 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2602 preferred_zone
, classzone_idx
, migratetype
);
2606 /* Allocate without watermarks if the context allows */
2607 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2609 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2610 * the allocation is high priority and these type of
2611 * allocations are system rather than user orientated
2613 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2615 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2616 zonelist
, high_zoneidx
, nodemask
,
2617 preferred_zone
, classzone_idx
, migratetype
);
2623 /* Atomic allocations - we can't balance anything */
2626 * All existing users of the deprecated __GFP_NOFAIL are
2627 * blockable, so warn of any new users that actually allow this
2628 * type of allocation to fail.
2630 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2634 /* Avoid recursion of direct reclaim */
2635 if (current
->flags
& PF_MEMALLOC
)
2638 /* Avoid allocations with no watermarks from looping endlessly */
2639 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2643 * Try direct compaction. The first pass is asynchronous. Subsequent
2644 * attempts after direct reclaim are synchronous
2646 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2647 high_zoneidx
, nodemask
, alloc_flags
,
2649 classzone_idx
, migratetype
,
2650 migration_mode
, &contended_compaction
,
2651 &deferred_compaction
);
2655 /* Checks for THP-specific high-order allocations */
2656 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2658 * If compaction is deferred for high-order allocations, it is
2659 * because sync compaction recently failed. If this is the case
2660 * and the caller requested a THP allocation, we do not want
2661 * to heavily disrupt the system, so we fail the allocation
2662 * instead of entering direct reclaim.
2664 if (deferred_compaction
)
2668 * In all zones where compaction was attempted (and not
2669 * deferred or skipped), lock contention has been detected.
2670 * For THP allocation we do not want to disrupt the others
2671 * so we fallback to base pages instead.
2673 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2677 * If compaction was aborted due to need_resched(), we do not
2678 * want to further increase allocation latency, unless it is
2679 * khugepaged trying to collapse.
2681 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2682 && !(current
->flags
& PF_KTHREAD
))
2687 * It can become very expensive to allocate transparent hugepages at
2688 * fault, so use asynchronous memory compaction for THP unless it is
2689 * khugepaged trying to collapse.
2691 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2692 (current
->flags
& PF_KTHREAD
))
2693 migration_mode
= MIGRATE_SYNC_LIGHT
;
2695 /* Try direct reclaim and then allocating */
2696 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2697 zonelist
, high_zoneidx
,
2699 alloc_flags
, preferred_zone
,
2700 classzone_idx
, migratetype
,
2701 &did_some_progress
);
2706 * If we failed to make any progress reclaiming, then we are
2707 * running out of options and have to consider going OOM
2709 if (!did_some_progress
) {
2710 if (oom_gfp_allowed(gfp_mask
)) {
2711 if (oom_killer_disabled
)
2713 /* Coredumps can quickly deplete all memory reserves */
2714 if ((current
->flags
& PF_DUMPCORE
) &&
2715 !(gfp_mask
& __GFP_NOFAIL
))
2717 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2718 zonelist
, high_zoneidx
,
2719 nodemask
, preferred_zone
,
2720 classzone_idx
, migratetype
);
2724 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2726 * The oom killer is not called for high-order
2727 * allocations that may fail, so if no progress
2728 * is being made, there are no other options and
2729 * retrying is unlikely to help.
2731 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2734 * The oom killer is not called for lowmem
2735 * allocations to prevent needlessly killing
2738 if (high_zoneidx
< ZONE_NORMAL
)
2746 /* Check if we should retry the allocation */
2747 pages_reclaimed
+= did_some_progress
;
2748 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2750 /* Wait for some write requests to complete then retry */
2751 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2755 * High-order allocations do not necessarily loop after
2756 * direct reclaim and reclaim/compaction depends on compaction
2757 * being called after reclaim so call directly if necessary
2759 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2760 high_zoneidx
, nodemask
, alloc_flags
,
2762 classzone_idx
, migratetype
,
2763 migration_mode
, &contended_compaction
,
2764 &deferred_compaction
);
2770 warn_alloc_failed(gfp_mask
, order
, NULL
);
2773 if (kmemcheck_enabled
)
2774 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2780 * This is the 'heart' of the zoned buddy allocator.
2783 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2784 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2786 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2787 struct zone
*preferred_zone
;
2788 struct zoneref
*preferred_zoneref
;
2789 struct page
*page
= NULL
;
2790 int migratetype
= gfpflags_to_migratetype(gfp_mask
);
2791 unsigned int cpuset_mems_cookie
;
2792 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2795 gfp_mask
&= gfp_allowed_mask
;
2797 lockdep_trace_alloc(gfp_mask
);
2799 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2801 if (should_fail_alloc_page(gfp_mask
, order
))
2805 * Check the zones suitable for the gfp_mask contain at least one
2806 * valid zone. It's possible to have an empty zonelist as a result
2807 * of GFP_THISNODE and a memoryless node
2809 if (unlikely(!zonelist
->_zonerefs
->zone
))
2812 if (IS_ENABLED(CONFIG_CMA
) && migratetype
== MIGRATE_MOVABLE
)
2813 alloc_flags
|= ALLOC_CMA
;
2816 cpuset_mems_cookie
= read_mems_allowed_begin();
2818 /* The preferred zone is used for statistics later */
2819 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2820 nodemask
? : &cpuset_current_mems_allowed
,
2822 if (!preferred_zone
)
2824 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2826 /* First allocation attempt */
2827 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2828 zonelist
, high_zoneidx
, alloc_flags
,
2829 preferred_zone
, classzone_idx
, migratetype
);
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 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2837 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2838 zonelist
, high_zoneidx
, nodemask
,
2839 preferred_zone
, classzone_idx
, migratetype
);
2842 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2846 * When updating a task's mems_allowed, it is possible to race with
2847 * parallel threads in such a way that an allocation can fail while
2848 * the mask is being updated. If a page allocation is about to fail,
2849 * check if the cpuset changed during allocation and if so, retry.
2851 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2856 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2859 * Common helper functions.
2861 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2866 * __get_free_pages() returns a 32-bit address, which cannot represent
2869 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2871 page
= alloc_pages(gfp_mask
, order
);
2874 return (unsigned long) page_address(page
);
2876 EXPORT_SYMBOL(__get_free_pages
);
2878 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2880 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2882 EXPORT_SYMBOL(get_zeroed_page
);
2884 void __free_pages(struct page
*page
, unsigned int order
)
2886 if (put_page_testzero(page
)) {
2888 free_hot_cold_page(page
, false);
2890 __free_pages_ok(page
, order
);
2894 EXPORT_SYMBOL(__free_pages
);
2896 void free_pages(unsigned long addr
, unsigned int order
)
2899 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2900 __free_pages(virt_to_page((void *)addr
), order
);
2904 EXPORT_SYMBOL(free_pages
);
2907 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2908 * of the current memory cgroup.
2910 * It should be used when the caller would like to use kmalloc, but since the
2911 * allocation is large, it has to fall back to the page allocator.
2913 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2916 struct mem_cgroup
*memcg
= NULL
;
2918 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2920 page
= alloc_pages(gfp_mask
, order
);
2921 memcg_kmem_commit_charge(page
, memcg
, order
);
2925 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2928 struct mem_cgroup
*memcg
= NULL
;
2930 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2932 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2933 memcg_kmem_commit_charge(page
, memcg
, order
);
2938 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2941 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2943 memcg_kmem_uncharge_pages(page
, order
);
2944 __free_pages(page
, order
);
2947 void free_kmem_pages(unsigned long addr
, unsigned int order
)
2950 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2951 __free_kmem_pages(virt_to_page((void *)addr
), order
);
2955 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2958 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2959 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2961 split_page(virt_to_page((void *)addr
), order
);
2962 while (used
< alloc_end
) {
2967 return (void *)addr
;
2971 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2972 * @size: the number of bytes to allocate
2973 * @gfp_mask: GFP flags for the allocation
2975 * This function is similar to alloc_pages(), except that it allocates the
2976 * minimum number of pages to satisfy the request. alloc_pages() can only
2977 * allocate memory in power-of-two pages.
2979 * This function is also limited by MAX_ORDER.
2981 * Memory allocated by this function must be released by free_pages_exact().
2983 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2985 unsigned int order
= get_order(size
);
2988 addr
= __get_free_pages(gfp_mask
, order
);
2989 return make_alloc_exact(addr
, order
, size
);
2991 EXPORT_SYMBOL(alloc_pages_exact
);
2994 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2996 * @nid: the preferred node ID where memory should be allocated
2997 * @size: the number of bytes to allocate
2998 * @gfp_mask: GFP flags for the allocation
3000 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3002 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3005 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3007 unsigned order
= get_order(size
);
3008 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3011 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3015 * free_pages_exact - release memory allocated via alloc_pages_exact()
3016 * @virt: the value returned by alloc_pages_exact.
3017 * @size: size of allocation, same value as passed to alloc_pages_exact().
3019 * Release the memory allocated by a previous call to alloc_pages_exact.
3021 void free_pages_exact(void *virt
, size_t size
)
3023 unsigned long addr
= (unsigned long)virt
;
3024 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3026 while (addr
< end
) {
3031 EXPORT_SYMBOL(free_pages_exact
);
3034 * nr_free_zone_pages - count number of pages beyond high watermark
3035 * @offset: The zone index of the highest zone
3037 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3038 * high watermark within all zones at or below a given zone index. For each
3039 * zone, the number of pages is calculated as:
3040 * managed_pages - high_pages
3042 static unsigned long nr_free_zone_pages(int offset
)
3047 /* Just pick one node, since fallback list is circular */
3048 unsigned long sum
= 0;
3050 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3052 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3053 unsigned long size
= zone
->managed_pages
;
3054 unsigned long high
= high_wmark_pages(zone
);
3063 * nr_free_buffer_pages - count number of pages beyond high watermark
3065 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3066 * watermark within ZONE_DMA and ZONE_NORMAL.
3068 unsigned long nr_free_buffer_pages(void)
3070 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3072 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3075 * nr_free_pagecache_pages - count number of pages beyond high watermark
3077 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3078 * high watermark within all zones.
3080 unsigned long nr_free_pagecache_pages(void)
3082 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3085 static inline void show_node(struct zone
*zone
)
3087 if (IS_ENABLED(CONFIG_NUMA
))
3088 printk("Node %d ", zone_to_nid(zone
));
3091 void si_meminfo(struct sysinfo
*val
)
3093 val
->totalram
= totalram_pages
;
3094 val
->sharedram
= global_page_state(NR_SHMEM
);
3095 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3096 val
->bufferram
= nr_blockdev_pages();
3097 val
->totalhigh
= totalhigh_pages
;
3098 val
->freehigh
= nr_free_highpages();
3099 val
->mem_unit
= PAGE_SIZE
;
3102 EXPORT_SYMBOL(si_meminfo
);
3105 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3107 int zone_type
; /* needs to be signed */
3108 unsigned long managed_pages
= 0;
3109 pg_data_t
*pgdat
= NODE_DATA(nid
);
3111 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3112 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3113 val
->totalram
= managed_pages
;
3114 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3115 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3116 #ifdef CONFIG_HIGHMEM
3117 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3118 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3124 val
->mem_unit
= PAGE_SIZE
;
3129 * Determine whether the node should be displayed or not, depending on whether
3130 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3132 bool skip_free_areas_node(unsigned int flags
, int nid
)
3135 unsigned int cpuset_mems_cookie
;
3137 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3141 cpuset_mems_cookie
= read_mems_allowed_begin();
3142 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3143 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3148 #define K(x) ((x) << (PAGE_SHIFT-10))
3150 static void show_migration_types(unsigned char type
)
3152 static const char types
[MIGRATE_TYPES
] = {
3153 [MIGRATE_UNMOVABLE
] = 'U',
3154 [MIGRATE_RECLAIMABLE
] = 'E',
3155 [MIGRATE_MOVABLE
] = 'M',
3156 [MIGRATE_RESERVE
] = 'R',
3158 [MIGRATE_CMA
] = 'C',
3160 #ifdef CONFIG_MEMORY_ISOLATION
3161 [MIGRATE_ISOLATE
] = 'I',
3164 char tmp
[MIGRATE_TYPES
+ 1];
3168 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3169 if (type
& (1 << i
))
3174 printk("(%s) ", tmp
);
3178 * Show free area list (used inside shift_scroll-lock stuff)
3179 * We also calculate the percentage fragmentation. We do this by counting the
3180 * memory on each free list with the exception of the first item on the list.
3181 * Suppresses nodes that are not allowed by current's cpuset if
3182 * SHOW_MEM_FILTER_NODES is passed.
3184 void show_free_areas(unsigned int filter
)
3189 for_each_populated_zone(zone
) {
3190 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3193 printk("%s per-cpu:\n", zone
->name
);
3195 for_each_online_cpu(cpu
) {
3196 struct per_cpu_pageset
*pageset
;
3198 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3200 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3201 cpu
, pageset
->pcp
.high
,
3202 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3206 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3207 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3209 " dirty:%lu writeback:%lu unstable:%lu\n"
3210 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3211 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3213 global_page_state(NR_ACTIVE_ANON
),
3214 global_page_state(NR_INACTIVE_ANON
),
3215 global_page_state(NR_ISOLATED_ANON
),
3216 global_page_state(NR_ACTIVE_FILE
),
3217 global_page_state(NR_INACTIVE_FILE
),
3218 global_page_state(NR_ISOLATED_FILE
),
3219 global_page_state(NR_UNEVICTABLE
),
3220 global_page_state(NR_FILE_DIRTY
),
3221 global_page_state(NR_WRITEBACK
),
3222 global_page_state(NR_UNSTABLE_NFS
),
3223 global_page_state(NR_FREE_PAGES
),
3224 global_page_state(NR_SLAB_RECLAIMABLE
),
3225 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3226 global_page_state(NR_FILE_MAPPED
),
3227 global_page_state(NR_SHMEM
),
3228 global_page_state(NR_PAGETABLE
),
3229 global_page_state(NR_BOUNCE
),
3230 global_page_state(NR_FREE_CMA_PAGES
));
3232 for_each_populated_zone(zone
) {
3235 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3243 " active_anon:%lukB"
3244 " inactive_anon:%lukB"
3245 " active_file:%lukB"
3246 " inactive_file:%lukB"
3247 " unevictable:%lukB"
3248 " isolated(anon):%lukB"
3249 " isolated(file):%lukB"
3257 " slab_reclaimable:%lukB"
3258 " slab_unreclaimable:%lukB"
3259 " kernel_stack:%lukB"
3264 " writeback_tmp:%lukB"
3265 " pages_scanned:%lu"
3266 " all_unreclaimable? %s"
3269 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3270 K(min_wmark_pages(zone
)),
3271 K(low_wmark_pages(zone
)),
3272 K(high_wmark_pages(zone
)),
3273 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3274 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3275 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3276 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3277 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3278 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3279 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3280 K(zone
->present_pages
),
3281 K(zone
->managed_pages
),
3282 K(zone_page_state(zone
, NR_MLOCK
)),
3283 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3284 K(zone_page_state(zone
, NR_WRITEBACK
)),
3285 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3286 K(zone_page_state(zone
, NR_SHMEM
)),
3287 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3288 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3289 zone_page_state(zone
, NR_KERNEL_STACK
) *
3291 K(zone_page_state(zone
, NR_PAGETABLE
)),
3292 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3293 K(zone_page_state(zone
, NR_BOUNCE
)),
3294 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3295 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3296 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3297 (!zone_reclaimable(zone
) ? "yes" : "no")
3299 printk("lowmem_reserve[]:");
3300 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3301 printk(" %ld", zone
->lowmem_reserve
[i
]);
3305 for_each_populated_zone(zone
) {
3306 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3307 unsigned char types
[MAX_ORDER
];
3309 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3312 printk("%s: ", zone
->name
);
3314 spin_lock_irqsave(&zone
->lock
, flags
);
3315 for (order
= 0; order
< MAX_ORDER
; order
++) {
3316 struct free_area
*area
= &zone
->free_area
[order
];
3319 nr
[order
] = area
->nr_free
;
3320 total
+= nr
[order
] << order
;
3323 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3324 if (!list_empty(&area
->free_list
[type
]))
3325 types
[order
] |= 1 << type
;
3328 spin_unlock_irqrestore(&zone
->lock
, flags
);
3329 for (order
= 0; order
< MAX_ORDER
; order
++) {
3330 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3332 show_migration_types(types
[order
]);
3334 printk("= %lukB\n", K(total
));
3337 hugetlb_show_meminfo();
3339 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3341 show_swap_cache_info();
3344 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3346 zoneref
->zone
= zone
;
3347 zoneref
->zone_idx
= zone_idx(zone
);
3351 * Builds allocation fallback zone lists.
3353 * Add all populated zones of a node to the zonelist.
3355 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3359 enum zone_type zone_type
= MAX_NR_ZONES
;
3363 zone
= pgdat
->node_zones
+ zone_type
;
3364 if (populated_zone(zone
)) {
3365 zoneref_set_zone(zone
,
3366 &zonelist
->_zonerefs
[nr_zones
++]);
3367 check_highest_zone(zone_type
);
3369 } while (zone_type
);
3377 * 0 = automatic detection of better ordering.
3378 * 1 = order by ([node] distance, -zonetype)
3379 * 2 = order by (-zonetype, [node] distance)
3381 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3382 * the same zonelist. So only NUMA can configure this param.
3384 #define ZONELIST_ORDER_DEFAULT 0
3385 #define ZONELIST_ORDER_NODE 1
3386 #define ZONELIST_ORDER_ZONE 2
3388 /* zonelist order in the kernel.
3389 * set_zonelist_order() will set this to NODE or ZONE.
3391 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3392 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3396 /* The value user specified ....changed by config */
3397 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3398 /* string for sysctl */
3399 #define NUMA_ZONELIST_ORDER_LEN 16
3400 char numa_zonelist_order
[16] = "default";
3403 * interface for configure zonelist ordering.
3404 * command line option "numa_zonelist_order"
3405 * = "[dD]efault - default, automatic configuration.
3406 * = "[nN]ode - order by node locality, then by zone within node
3407 * = "[zZ]one - order by zone, then by locality within zone
3410 static int __parse_numa_zonelist_order(char *s
)
3412 if (*s
== 'd' || *s
== 'D') {
3413 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3414 } else if (*s
== 'n' || *s
== 'N') {
3415 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3416 } else if (*s
== 'z' || *s
== 'Z') {
3417 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3420 "Ignoring invalid numa_zonelist_order value: "
3427 static __init
int setup_numa_zonelist_order(char *s
)
3434 ret
= __parse_numa_zonelist_order(s
);
3436 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3440 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3443 * sysctl handler for numa_zonelist_order
3445 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3446 void __user
*buffer
, size_t *length
,
3449 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3451 static DEFINE_MUTEX(zl_order_mutex
);
3453 mutex_lock(&zl_order_mutex
);
3455 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3459 strcpy(saved_string
, (char *)table
->data
);
3461 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3465 int oldval
= user_zonelist_order
;
3467 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3470 * bogus value. restore saved string
3472 strncpy((char *)table
->data
, saved_string
,
3473 NUMA_ZONELIST_ORDER_LEN
);
3474 user_zonelist_order
= oldval
;
3475 } else if (oldval
!= user_zonelist_order
) {
3476 mutex_lock(&zonelists_mutex
);
3477 build_all_zonelists(NULL
, NULL
);
3478 mutex_unlock(&zonelists_mutex
);
3482 mutex_unlock(&zl_order_mutex
);
3487 #define MAX_NODE_LOAD (nr_online_nodes)
3488 static int node_load
[MAX_NUMNODES
];
3491 * find_next_best_node - find the next node that should appear in a given node's fallback list
3492 * @node: node whose fallback list we're appending
3493 * @used_node_mask: nodemask_t of already used nodes
3495 * We use a number of factors to determine which is the next node that should
3496 * appear on a given node's fallback list. The node should not have appeared
3497 * already in @node's fallback list, and it should be the next closest node
3498 * according to the distance array (which contains arbitrary distance values
3499 * from each node to each node in the system), and should also prefer nodes
3500 * with no CPUs, since presumably they'll have very little allocation pressure
3501 * on them otherwise.
3502 * It returns -1 if no node is found.
3504 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3507 int min_val
= INT_MAX
;
3508 int best_node
= NUMA_NO_NODE
;
3509 const struct cpumask
*tmp
= cpumask_of_node(0);
3511 /* Use the local node if we haven't already */
3512 if (!node_isset(node
, *used_node_mask
)) {
3513 node_set(node
, *used_node_mask
);
3517 for_each_node_state(n
, N_MEMORY
) {
3519 /* Don't want a node to appear more than once */
3520 if (node_isset(n
, *used_node_mask
))
3523 /* Use the distance array to find the distance */
3524 val
= node_distance(node
, n
);
3526 /* Penalize nodes under us ("prefer the next node") */
3529 /* Give preference to headless and unused nodes */
3530 tmp
= cpumask_of_node(n
);
3531 if (!cpumask_empty(tmp
))
3532 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3534 /* Slight preference for less loaded node */
3535 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3536 val
+= node_load
[n
];
3538 if (val
< min_val
) {
3545 node_set(best_node
, *used_node_mask
);
3552 * Build zonelists ordered by node and zones within node.
3553 * This results in maximum locality--normal zone overflows into local
3554 * DMA zone, if any--but risks exhausting DMA zone.
3556 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3559 struct zonelist
*zonelist
;
3561 zonelist
= &pgdat
->node_zonelists
[0];
3562 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3564 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3565 zonelist
->_zonerefs
[j
].zone
= NULL
;
3566 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3570 * Build gfp_thisnode zonelists
3572 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3575 struct zonelist
*zonelist
;
3577 zonelist
= &pgdat
->node_zonelists
[1];
3578 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3579 zonelist
->_zonerefs
[j
].zone
= NULL
;
3580 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3584 * Build zonelists ordered by zone and nodes within zones.
3585 * This results in conserving DMA zone[s] until all Normal memory is
3586 * exhausted, but results in overflowing to remote node while memory
3587 * may still exist in local DMA zone.
3589 static int node_order
[MAX_NUMNODES
];
3591 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3594 int zone_type
; /* needs to be signed */
3596 struct zonelist
*zonelist
;
3598 zonelist
= &pgdat
->node_zonelists
[0];
3600 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3601 for (j
= 0; j
< nr_nodes
; j
++) {
3602 node
= node_order
[j
];
3603 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3604 if (populated_zone(z
)) {
3606 &zonelist
->_zonerefs
[pos
++]);
3607 check_highest_zone(zone_type
);
3611 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3612 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3615 #if defined(CONFIG_64BIT)
3617 * Devices that require DMA32/DMA are relatively rare and do not justify a
3618 * penalty to every machine in case the specialised case applies. Default
3619 * to Node-ordering on 64-bit NUMA machines
3621 static int default_zonelist_order(void)
3623 return ZONELIST_ORDER_NODE
;
3627 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3628 * by the kernel. If processes running on node 0 deplete the low memory zone
3629 * then reclaim will occur more frequency increasing stalls and potentially
3630 * be easier to OOM if a large percentage of the zone is under writeback or
3631 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3632 * Hence, default to zone ordering on 32-bit.
3634 static int default_zonelist_order(void)
3636 return ZONELIST_ORDER_ZONE
;
3638 #endif /* CONFIG_64BIT */
3640 static void set_zonelist_order(void)
3642 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3643 current_zonelist_order
= default_zonelist_order();
3645 current_zonelist_order
= user_zonelist_order
;
3648 static void build_zonelists(pg_data_t
*pgdat
)
3652 nodemask_t used_mask
;
3653 int local_node
, prev_node
;
3654 struct zonelist
*zonelist
;
3655 int order
= current_zonelist_order
;
3657 /* initialize zonelists */
3658 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3659 zonelist
= pgdat
->node_zonelists
+ i
;
3660 zonelist
->_zonerefs
[0].zone
= NULL
;
3661 zonelist
->_zonerefs
[0].zone_idx
= 0;
3664 /* NUMA-aware ordering of nodes */
3665 local_node
= pgdat
->node_id
;
3666 load
= nr_online_nodes
;
3667 prev_node
= local_node
;
3668 nodes_clear(used_mask
);
3670 memset(node_order
, 0, sizeof(node_order
));
3673 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3675 * We don't want to pressure a particular node.
3676 * So adding penalty to the first node in same
3677 * distance group to make it round-robin.
3679 if (node_distance(local_node
, node
) !=
3680 node_distance(local_node
, prev_node
))
3681 node_load
[node
] = load
;
3685 if (order
== ZONELIST_ORDER_NODE
)
3686 build_zonelists_in_node_order(pgdat
, node
);
3688 node_order
[j
++] = node
; /* remember order */
3691 if (order
== ZONELIST_ORDER_ZONE
) {
3692 /* calculate node order -- i.e., DMA last! */
3693 build_zonelists_in_zone_order(pgdat
, j
);
3696 build_thisnode_zonelists(pgdat
);
3699 /* Construct the zonelist performance cache - see further mmzone.h */
3700 static void build_zonelist_cache(pg_data_t
*pgdat
)
3702 struct zonelist
*zonelist
;
3703 struct zonelist_cache
*zlc
;
3706 zonelist
= &pgdat
->node_zonelists
[0];
3707 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3708 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3709 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3710 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3713 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3715 * Return node id of node used for "local" allocations.
3716 * I.e., first node id of first zone in arg node's generic zonelist.
3717 * Used for initializing percpu 'numa_mem', which is used primarily
3718 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3720 int local_memory_node(int node
)
3724 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3725 gfp_zone(GFP_KERNEL
),
3732 #else /* CONFIG_NUMA */
3734 static void set_zonelist_order(void)
3736 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3739 static void build_zonelists(pg_data_t
*pgdat
)
3741 int node
, local_node
;
3743 struct zonelist
*zonelist
;
3745 local_node
= pgdat
->node_id
;
3747 zonelist
= &pgdat
->node_zonelists
[0];
3748 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3751 * Now we build the zonelist so that it contains the zones
3752 * of all the other nodes.
3753 * We don't want to pressure a particular node, so when
3754 * building the zones for node N, we make sure that the
3755 * zones coming right after the local ones are those from
3756 * node N+1 (modulo N)
3758 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3759 if (!node_online(node
))
3761 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3763 for (node
= 0; node
< local_node
; node
++) {
3764 if (!node_online(node
))
3766 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3769 zonelist
->_zonerefs
[j
].zone
= NULL
;
3770 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3773 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3774 static void build_zonelist_cache(pg_data_t
*pgdat
)
3776 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3779 #endif /* CONFIG_NUMA */
3782 * Boot pageset table. One per cpu which is going to be used for all
3783 * zones and all nodes. The parameters will be set in such a way
3784 * that an item put on a list will immediately be handed over to
3785 * the buddy list. This is safe since pageset manipulation is done
3786 * with interrupts disabled.
3788 * The boot_pagesets must be kept even after bootup is complete for
3789 * unused processors and/or zones. They do play a role for bootstrapping
3790 * hotplugged processors.
3792 * zoneinfo_show() and maybe other functions do
3793 * not check if the processor is online before following the pageset pointer.
3794 * Other parts of the kernel may not check if the zone is available.
3796 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3797 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3798 static void setup_zone_pageset(struct zone
*zone
);
3801 * Global mutex to protect against size modification of zonelists
3802 * as well as to serialize pageset setup for the new populated zone.
3804 DEFINE_MUTEX(zonelists_mutex
);
3806 /* return values int ....just for stop_machine() */
3807 static int __build_all_zonelists(void *data
)
3811 pg_data_t
*self
= data
;
3814 memset(node_load
, 0, sizeof(node_load
));
3817 if (self
&& !node_online(self
->node_id
)) {
3818 build_zonelists(self
);
3819 build_zonelist_cache(self
);
3822 for_each_online_node(nid
) {
3823 pg_data_t
*pgdat
= NODE_DATA(nid
);
3825 build_zonelists(pgdat
);
3826 build_zonelist_cache(pgdat
);
3830 * Initialize the boot_pagesets that are going to be used
3831 * for bootstrapping processors. The real pagesets for
3832 * each zone will be allocated later when the per cpu
3833 * allocator is available.
3835 * boot_pagesets are used also for bootstrapping offline
3836 * cpus if the system is already booted because the pagesets
3837 * are needed to initialize allocators on a specific cpu too.
3838 * F.e. the percpu allocator needs the page allocator which
3839 * needs the percpu allocator in order to allocate its pagesets
3840 * (a chicken-egg dilemma).
3842 for_each_possible_cpu(cpu
) {
3843 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3845 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3847 * We now know the "local memory node" for each node--
3848 * i.e., the node of the first zone in the generic zonelist.
3849 * Set up numa_mem percpu variable for on-line cpus. During
3850 * boot, only the boot cpu should be on-line; we'll init the
3851 * secondary cpus' numa_mem as they come on-line. During
3852 * node/memory hotplug, we'll fixup all on-line cpus.
3854 if (cpu_online(cpu
))
3855 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3863 * Called with zonelists_mutex held always
3864 * unless system_state == SYSTEM_BOOTING.
3866 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3868 set_zonelist_order();
3870 if (system_state
== SYSTEM_BOOTING
) {
3871 __build_all_zonelists(NULL
);
3872 mminit_verify_zonelist();
3873 cpuset_init_current_mems_allowed();
3875 #ifdef CONFIG_MEMORY_HOTPLUG
3877 setup_zone_pageset(zone
);
3879 /* we have to stop all cpus to guarantee there is no user
3881 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3882 /* cpuset refresh routine should be here */
3884 vm_total_pages
= nr_free_pagecache_pages();
3886 * Disable grouping by mobility if the number of pages in the
3887 * system is too low to allow the mechanism to work. It would be
3888 * more accurate, but expensive to check per-zone. This check is
3889 * made on memory-hotadd so a system can start with mobility
3890 * disabled and enable it later
3892 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3893 page_group_by_mobility_disabled
= 1;
3895 page_group_by_mobility_disabled
= 0;
3897 printk("Built %i zonelists in %s order, mobility grouping %s. "
3898 "Total pages: %ld\n",
3900 zonelist_order_name
[current_zonelist_order
],
3901 page_group_by_mobility_disabled
? "off" : "on",
3904 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3909 * Helper functions to size the waitqueue hash table.
3910 * Essentially these want to choose hash table sizes sufficiently
3911 * large so that collisions trying to wait on pages are rare.
3912 * But in fact, the number of active page waitqueues on typical
3913 * systems is ridiculously low, less than 200. So this is even
3914 * conservative, even though it seems large.
3916 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3917 * waitqueues, i.e. the size of the waitq table given the number of pages.
3919 #define PAGES_PER_WAITQUEUE 256
3921 #ifndef CONFIG_MEMORY_HOTPLUG
3922 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3924 unsigned long size
= 1;
3926 pages
/= PAGES_PER_WAITQUEUE
;
3928 while (size
< pages
)
3932 * Once we have dozens or even hundreds of threads sleeping
3933 * on IO we've got bigger problems than wait queue collision.
3934 * Limit the size of the wait table to a reasonable size.
3936 size
= min(size
, 4096UL);
3938 return max(size
, 4UL);
3942 * A zone's size might be changed by hot-add, so it is not possible to determine
3943 * a suitable size for its wait_table. So we use the maximum size now.
3945 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3947 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3948 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3949 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3951 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3952 * or more by the traditional way. (See above). It equals:
3954 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3955 * ia64(16K page size) : = ( 8G + 4M)byte.
3956 * powerpc (64K page size) : = (32G +16M)byte.
3958 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3965 * This is an integer logarithm so that shifts can be used later
3966 * to extract the more random high bits from the multiplicative
3967 * hash function before the remainder is taken.
3969 static inline unsigned long wait_table_bits(unsigned long size
)
3975 * Check if a pageblock contains reserved pages
3977 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3981 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3982 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3989 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3990 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3991 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3992 * higher will lead to a bigger reserve which will get freed as contiguous
3993 * blocks as reclaim kicks in
3995 static void setup_zone_migrate_reserve(struct zone
*zone
)
3997 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3999 unsigned long block_migratetype
;
4004 * Get the start pfn, end pfn and the number of blocks to reserve
4005 * We have to be careful to be aligned to pageblock_nr_pages to
4006 * make sure that we always check pfn_valid for the first page in
4009 start_pfn
= zone
->zone_start_pfn
;
4010 end_pfn
= zone_end_pfn(zone
);
4011 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4012 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4016 * Reserve blocks are generally in place to help high-order atomic
4017 * allocations that are short-lived. A min_free_kbytes value that
4018 * would result in more than 2 reserve blocks for atomic allocations
4019 * is assumed to be in place to help anti-fragmentation for the
4020 * future allocation of hugepages at runtime.
4022 reserve
= min(2, reserve
);
4023 old_reserve
= zone
->nr_migrate_reserve_block
;
4025 /* When memory hot-add, we almost always need to do nothing */
4026 if (reserve
== old_reserve
)
4028 zone
->nr_migrate_reserve_block
= reserve
;
4030 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4031 if (!pfn_valid(pfn
))
4033 page
= pfn_to_page(pfn
);
4035 /* Watch out for overlapping nodes */
4036 if (page_to_nid(page
) != zone_to_nid(zone
))
4039 block_migratetype
= get_pageblock_migratetype(page
);
4041 /* Only test what is necessary when the reserves are not met */
4044 * Blocks with reserved pages will never free, skip
4047 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4048 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4051 /* If this block is reserved, account for it */
4052 if (block_migratetype
== MIGRATE_RESERVE
) {
4057 /* Suitable for reserving if this block is movable */
4058 if (block_migratetype
== MIGRATE_MOVABLE
) {
4059 set_pageblock_migratetype(page
,
4061 move_freepages_block(zone
, page
,
4066 } else if (!old_reserve
) {
4068 * At boot time we don't need to scan the whole zone
4069 * for turning off MIGRATE_RESERVE.
4075 * If the reserve is met and this is a previous reserved block,
4078 if (block_migratetype
== MIGRATE_RESERVE
) {
4079 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4080 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4086 * Initially all pages are reserved - free ones are freed
4087 * up by free_all_bootmem() once the early boot process is
4088 * done. Non-atomic initialization, single-pass.
4090 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4091 unsigned long start_pfn
, enum memmap_context context
)
4094 unsigned long end_pfn
= start_pfn
+ size
;
4098 if (highest_memmap_pfn
< end_pfn
- 1)
4099 highest_memmap_pfn
= end_pfn
- 1;
4101 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4102 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4104 * There can be holes in boot-time mem_map[]s
4105 * handed to this function. They do not
4106 * exist on hotplugged memory.
4108 if (context
== MEMMAP_EARLY
) {
4109 if (!early_pfn_valid(pfn
))
4111 if (!early_pfn_in_nid(pfn
, nid
))
4114 page
= pfn_to_page(pfn
);
4115 set_page_links(page
, zone
, nid
, pfn
);
4116 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4117 init_page_count(page
);
4118 page_mapcount_reset(page
);
4119 page_cpupid_reset_last(page
);
4120 SetPageReserved(page
);
4122 * Mark the block movable so that blocks are reserved for
4123 * movable at startup. This will force kernel allocations
4124 * to reserve their blocks rather than leaking throughout
4125 * the address space during boot when many long-lived
4126 * kernel allocations are made. Later some blocks near
4127 * the start are marked MIGRATE_RESERVE by
4128 * setup_zone_migrate_reserve()
4130 * bitmap is created for zone's valid pfn range. but memmap
4131 * can be created for invalid pages (for alignment)
4132 * check here not to call set_pageblock_migratetype() against
4135 if ((z
->zone_start_pfn
<= pfn
)
4136 && (pfn
< zone_end_pfn(z
))
4137 && !(pfn
& (pageblock_nr_pages
- 1)))
4138 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4140 INIT_LIST_HEAD(&page
->lru
);
4141 #ifdef WANT_PAGE_VIRTUAL
4142 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4143 if (!is_highmem_idx(zone
))
4144 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4149 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4151 unsigned int order
, t
;
4152 for_each_migratetype_order(order
, t
) {
4153 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4154 zone
->free_area
[order
].nr_free
= 0;
4158 #ifndef __HAVE_ARCH_MEMMAP_INIT
4159 #define memmap_init(size, nid, zone, start_pfn) \
4160 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4163 static int zone_batchsize(struct zone
*zone
)
4169 * The per-cpu-pages pools are set to around 1000th of the
4170 * size of the zone. But no more than 1/2 of a meg.
4172 * OK, so we don't know how big the cache is. So guess.
4174 batch
= zone
->managed_pages
/ 1024;
4175 if (batch
* PAGE_SIZE
> 512 * 1024)
4176 batch
= (512 * 1024) / PAGE_SIZE
;
4177 batch
/= 4; /* We effectively *= 4 below */
4182 * Clamp the batch to a 2^n - 1 value. Having a power
4183 * of 2 value was found to be more likely to have
4184 * suboptimal cache aliasing properties in some cases.
4186 * For example if 2 tasks are alternately allocating
4187 * batches of pages, one task can end up with a lot
4188 * of pages of one half of the possible page colors
4189 * and the other with pages of the other colors.
4191 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4196 /* The deferral and batching of frees should be suppressed under NOMMU
4199 * The problem is that NOMMU needs to be able to allocate large chunks
4200 * of contiguous memory as there's no hardware page translation to
4201 * assemble apparent contiguous memory from discontiguous pages.
4203 * Queueing large contiguous runs of pages for batching, however,
4204 * causes the pages to actually be freed in smaller chunks. As there
4205 * can be a significant delay between the individual batches being
4206 * recycled, this leads to the once large chunks of space being
4207 * fragmented and becoming unavailable for high-order allocations.
4214 * pcp->high and pcp->batch values are related and dependent on one another:
4215 * ->batch must never be higher then ->high.
4216 * The following function updates them in a safe manner without read side
4219 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4220 * those fields changing asynchronously (acording the the above rule).
4222 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4223 * outside of boot time (or some other assurance that no concurrent updaters
4226 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4227 unsigned long batch
)
4229 /* start with a fail safe value for batch */
4233 /* Update high, then batch, in order */
4240 /* a companion to pageset_set_high() */
4241 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4243 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4246 static void pageset_init(struct per_cpu_pageset
*p
)
4248 struct per_cpu_pages
*pcp
;
4251 memset(p
, 0, sizeof(*p
));
4255 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4256 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4259 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4262 pageset_set_batch(p
, batch
);
4266 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4267 * to the value high for the pageset p.
4269 static void pageset_set_high(struct per_cpu_pageset
*p
,
4272 unsigned long batch
= max(1UL, high
/ 4);
4273 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4274 batch
= PAGE_SHIFT
* 8;
4276 pageset_update(&p
->pcp
, high
, batch
);
4279 static void pageset_set_high_and_batch(struct zone
*zone
,
4280 struct per_cpu_pageset
*pcp
)
4282 if (percpu_pagelist_fraction
)
4283 pageset_set_high(pcp
,
4284 (zone
->managed_pages
/
4285 percpu_pagelist_fraction
));
4287 pageset_set_batch(pcp
, zone_batchsize(zone
));
4290 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4292 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4295 pageset_set_high_and_batch(zone
, pcp
);
4298 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4301 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4302 for_each_possible_cpu(cpu
)
4303 zone_pageset_init(zone
, cpu
);
4307 * Allocate per cpu pagesets and initialize them.
4308 * Before this call only boot pagesets were available.
4310 void __init
setup_per_cpu_pageset(void)
4314 for_each_populated_zone(zone
)
4315 setup_zone_pageset(zone
);
4318 static noinline __init_refok
4319 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4325 * The per-page waitqueue mechanism uses hashed waitqueues
4328 zone
->wait_table_hash_nr_entries
=
4329 wait_table_hash_nr_entries(zone_size_pages
);
4330 zone
->wait_table_bits
=
4331 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4332 alloc_size
= zone
->wait_table_hash_nr_entries
4333 * sizeof(wait_queue_head_t
);
4335 if (!slab_is_available()) {
4336 zone
->wait_table
= (wait_queue_head_t
*)
4337 memblock_virt_alloc_node_nopanic(
4338 alloc_size
, zone
->zone_pgdat
->node_id
);
4341 * This case means that a zone whose size was 0 gets new memory
4342 * via memory hot-add.
4343 * But it may be the case that a new node was hot-added. In
4344 * this case vmalloc() will not be able to use this new node's
4345 * memory - this wait_table must be initialized to use this new
4346 * node itself as well.
4347 * To use this new node's memory, further consideration will be
4350 zone
->wait_table
= vmalloc(alloc_size
);
4352 if (!zone
->wait_table
)
4355 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4356 init_waitqueue_head(zone
->wait_table
+ i
);
4361 static __meminit
void zone_pcp_init(struct zone
*zone
)
4364 * per cpu subsystem is not up at this point. The following code
4365 * relies on the ability of the linker to provide the
4366 * offset of a (static) per cpu variable into the per cpu area.
4368 zone
->pageset
= &boot_pageset
;
4370 if (populated_zone(zone
))
4371 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4372 zone
->name
, zone
->present_pages
,
4373 zone_batchsize(zone
));
4376 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4377 unsigned long zone_start_pfn
,
4379 enum memmap_context context
)
4381 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4383 ret
= zone_wait_table_init(zone
, size
);
4386 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4388 zone
->zone_start_pfn
= zone_start_pfn
;
4390 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4391 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4393 (unsigned long)zone_idx(zone
),
4394 zone_start_pfn
, (zone_start_pfn
+ size
));
4396 zone_init_free_lists(zone
);
4401 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4402 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4404 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4406 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4408 unsigned long start_pfn
, end_pfn
;
4411 * NOTE: The following SMP-unsafe globals are only used early in boot
4412 * when the kernel is running single-threaded.
4414 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4415 static int __meminitdata last_nid
;
4417 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4420 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4422 last_start_pfn
= start_pfn
;
4423 last_end_pfn
= end_pfn
;
4429 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4431 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4435 nid
= __early_pfn_to_nid(pfn
);
4438 /* just returns 0 */
4442 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4443 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4447 nid
= __early_pfn_to_nid(pfn
);
4448 if (nid
>= 0 && nid
!= node
)
4455 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4456 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4457 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4459 * If an architecture guarantees that all ranges registered contain no holes
4460 * and may be freed, this this function may be used instead of calling
4461 * memblock_free_early_nid() manually.
4463 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4465 unsigned long start_pfn
, end_pfn
;
4468 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4469 start_pfn
= min(start_pfn
, max_low_pfn
);
4470 end_pfn
= min(end_pfn
, max_low_pfn
);
4472 if (start_pfn
< end_pfn
)
4473 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4474 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4480 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4481 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4483 * If an architecture guarantees that all ranges registered contain no holes and may
4484 * be freed, this function may be used instead of calling memory_present() manually.
4486 void __init
sparse_memory_present_with_active_regions(int nid
)
4488 unsigned long start_pfn
, end_pfn
;
4491 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4492 memory_present(this_nid
, start_pfn
, end_pfn
);
4496 * get_pfn_range_for_nid - Return the start and end page frames for a node
4497 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4498 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4499 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4501 * It returns the start and end page frame of a node based on information
4502 * provided by memblock_set_node(). If called for a node
4503 * with no available memory, a warning is printed and the start and end
4506 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4507 unsigned long *start_pfn
, unsigned long *end_pfn
)
4509 unsigned long this_start_pfn
, this_end_pfn
;
4515 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4516 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4517 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4520 if (*start_pfn
== -1UL)
4525 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4526 * assumption is made that zones within a node are ordered in monotonic
4527 * increasing memory addresses so that the "highest" populated zone is used
4529 static void __init
find_usable_zone_for_movable(void)
4532 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4533 if (zone_index
== ZONE_MOVABLE
)
4536 if (arch_zone_highest_possible_pfn
[zone_index
] >
4537 arch_zone_lowest_possible_pfn
[zone_index
])
4541 VM_BUG_ON(zone_index
== -1);
4542 movable_zone
= zone_index
;
4546 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4547 * because it is sized independent of architecture. Unlike the other zones,
4548 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4549 * in each node depending on the size of each node and how evenly kernelcore
4550 * is distributed. This helper function adjusts the zone ranges
4551 * provided by the architecture for a given node by using the end of the
4552 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4553 * zones within a node are in order of monotonic increases memory addresses
4555 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4556 unsigned long zone_type
,
4557 unsigned long node_start_pfn
,
4558 unsigned long node_end_pfn
,
4559 unsigned long *zone_start_pfn
,
4560 unsigned long *zone_end_pfn
)
4562 /* Only adjust if ZONE_MOVABLE is on this node */
4563 if (zone_movable_pfn
[nid
]) {
4564 /* Size ZONE_MOVABLE */
4565 if (zone_type
== ZONE_MOVABLE
) {
4566 *zone_start_pfn
= zone_movable_pfn
[nid
];
4567 *zone_end_pfn
= min(node_end_pfn
,
4568 arch_zone_highest_possible_pfn
[movable_zone
]);
4570 /* Adjust for ZONE_MOVABLE starting within this range */
4571 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4572 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4573 *zone_end_pfn
= zone_movable_pfn
[nid
];
4575 /* Check if this whole range is within ZONE_MOVABLE */
4576 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4577 *zone_start_pfn
= *zone_end_pfn
;
4582 * Return the number of pages a zone spans in a node, including holes
4583 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4585 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4586 unsigned long zone_type
,
4587 unsigned long node_start_pfn
,
4588 unsigned long node_end_pfn
,
4589 unsigned long *ignored
)
4591 unsigned long zone_start_pfn
, zone_end_pfn
;
4593 /* Get the start and end of the zone */
4594 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4595 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4596 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4597 node_start_pfn
, node_end_pfn
,
4598 &zone_start_pfn
, &zone_end_pfn
);
4600 /* Check that this node has pages within the zone's required range */
4601 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4604 /* Move the zone boundaries inside the node if necessary */
4605 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4606 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4608 /* Return the spanned pages */
4609 return zone_end_pfn
- zone_start_pfn
;
4613 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4614 * then all holes in the requested range will be accounted for.
4616 unsigned long __meminit
__absent_pages_in_range(int nid
,
4617 unsigned long range_start_pfn
,
4618 unsigned long range_end_pfn
)
4620 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4621 unsigned long start_pfn
, end_pfn
;
4624 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4625 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4626 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4627 nr_absent
-= end_pfn
- start_pfn
;
4633 * absent_pages_in_range - Return number of page frames in holes within a range
4634 * @start_pfn: The start PFN to start searching for holes
4635 * @end_pfn: The end PFN to stop searching for holes
4637 * It returns the number of pages frames in memory holes within a range.
4639 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4640 unsigned long end_pfn
)
4642 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4645 /* Return the number of page frames in holes in a zone on a node */
4646 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4647 unsigned long zone_type
,
4648 unsigned long node_start_pfn
,
4649 unsigned long node_end_pfn
,
4650 unsigned long *ignored
)
4652 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4653 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4654 unsigned long zone_start_pfn
, zone_end_pfn
;
4656 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4657 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4659 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4660 node_start_pfn
, node_end_pfn
,
4661 &zone_start_pfn
, &zone_end_pfn
);
4662 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4665 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4666 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4667 unsigned long zone_type
,
4668 unsigned long node_start_pfn
,
4669 unsigned long node_end_pfn
,
4670 unsigned long *zones_size
)
4672 return zones_size
[zone_type
];
4675 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4676 unsigned long zone_type
,
4677 unsigned long node_start_pfn
,
4678 unsigned long node_end_pfn
,
4679 unsigned long *zholes_size
)
4684 return zholes_size
[zone_type
];
4687 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4689 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4690 unsigned long node_start_pfn
,
4691 unsigned long node_end_pfn
,
4692 unsigned long *zones_size
,
4693 unsigned long *zholes_size
)
4695 unsigned long realtotalpages
, totalpages
= 0;
4698 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4699 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4703 pgdat
->node_spanned_pages
= totalpages
;
4705 realtotalpages
= totalpages
;
4706 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4708 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4709 node_start_pfn
, node_end_pfn
,
4711 pgdat
->node_present_pages
= realtotalpages
;
4712 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4716 #ifndef CONFIG_SPARSEMEM
4718 * Calculate the size of the zone->blockflags rounded to an unsigned long
4719 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4720 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4721 * round what is now in bits to nearest long in bits, then return it in
4724 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4726 unsigned long usemapsize
;
4728 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4729 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4730 usemapsize
= usemapsize
>> pageblock_order
;
4731 usemapsize
*= NR_PAGEBLOCK_BITS
;
4732 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4734 return usemapsize
/ 8;
4737 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4739 unsigned long zone_start_pfn
,
4740 unsigned long zonesize
)
4742 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4743 zone
->pageblock_flags
= NULL
;
4745 zone
->pageblock_flags
=
4746 memblock_virt_alloc_node_nopanic(usemapsize
,
4750 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4751 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4752 #endif /* CONFIG_SPARSEMEM */
4754 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4756 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4757 void __paginginit
set_pageblock_order(void)
4761 /* Check that pageblock_nr_pages has not already been setup */
4762 if (pageblock_order
)
4765 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4766 order
= HUGETLB_PAGE_ORDER
;
4768 order
= MAX_ORDER
- 1;
4771 * Assume the largest contiguous order of interest is a huge page.
4772 * This value may be variable depending on boot parameters on IA64 and
4775 pageblock_order
= order
;
4777 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4780 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4781 * is unused as pageblock_order is set at compile-time. See
4782 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4785 void __paginginit
set_pageblock_order(void)
4789 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4791 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4792 unsigned long present_pages
)
4794 unsigned long pages
= spanned_pages
;
4797 * Provide a more accurate estimation if there are holes within
4798 * the zone and SPARSEMEM is in use. If there are holes within the
4799 * zone, each populated memory region may cost us one or two extra
4800 * memmap pages due to alignment because memmap pages for each
4801 * populated regions may not naturally algined on page boundary.
4802 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4804 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4805 IS_ENABLED(CONFIG_SPARSEMEM
))
4806 pages
= present_pages
;
4808 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4812 * Set up the zone data structures:
4813 * - mark all pages reserved
4814 * - mark all memory queues empty
4815 * - clear the memory bitmaps
4817 * NOTE: pgdat should get zeroed by caller.
4819 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4820 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4821 unsigned long *zones_size
, unsigned long *zholes_size
)
4824 int nid
= pgdat
->node_id
;
4825 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4828 pgdat_resize_init(pgdat
);
4829 #ifdef CONFIG_NUMA_BALANCING
4830 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4831 pgdat
->numabalancing_migrate_nr_pages
= 0;
4832 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4834 init_waitqueue_head(&pgdat
->kswapd_wait
);
4835 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4836 pgdat_page_cgroup_init(pgdat
);
4838 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4839 struct zone
*zone
= pgdat
->node_zones
+ j
;
4840 unsigned long size
, realsize
, freesize
, memmap_pages
;
4842 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4843 node_end_pfn
, zones_size
);
4844 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4850 * Adjust freesize so that it accounts for how much memory
4851 * is used by this zone for memmap. This affects the watermark
4852 * and per-cpu initialisations
4854 memmap_pages
= calc_memmap_size(size
, realsize
);
4855 if (freesize
>= memmap_pages
) {
4856 freesize
-= memmap_pages
;
4859 " %s zone: %lu pages used for memmap\n",
4860 zone_names
[j
], memmap_pages
);
4863 " %s zone: %lu pages exceeds freesize %lu\n",
4864 zone_names
[j
], memmap_pages
, freesize
);
4866 /* Account for reserved pages */
4867 if (j
== 0 && freesize
> dma_reserve
) {
4868 freesize
-= dma_reserve
;
4869 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4870 zone_names
[0], dma_reserve
);
4873 if (!is_highmem_idx(j
))
4874 nr_kernel_pages
+= freesize
;
4875 /* Charge for highmem memmap if there are enough kernel pages */
4876 else if (nr_kernel_pages
> memmap_pages
* 2)
4877 nr_kernel_pages
-= memmap_pages
;
4878 nr_all_pages
+= freesize
;
4880 zone
->spanned_pages
= size
;
4881 zone
->present_pages
= realsize
;
4883 * Set an approximate value for lowmem here, it will be adjusted
4884 * when the bootmem allocator frees pages into the buddy system.
4885 * And all highmem pages will be managed by the buddy system.
4887 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4890 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4892 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4894 zone
->name
= zone_names
[j
];
4895 spin_lock_init(&zone
->lock
);
4896 spin_lock_init(&zone
->lru_lock
);
4897 zone_seqlock_init(zone
);
4898 zone
->zone_pgdat
= pgdat
;
4899 zone_pcp_init(zone
);
4901 /* For bootup, initialized properly in watermark setup */
4902 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4904 lruvec_init(&zone
->lruvec
);
4908 set_pageblock_order();
4909 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4910 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4911 size
, MEMMAP_EARLY
);
4913 memmap_init(size
, nid
, j
, zone_start_pfn
);
4914 zone_start_pfn
+= size
;
4918 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4920 /* Skip empty nodes */
4921 if (!pgdat
->node_spanned_pages
)
4924 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4925 /* ia64 gets its own node_mem_map, before this, without bootmem */
4926 if (!pgdat
->node_mem_map
) {
4927 unsigned long size
, start
, end
;
4931 * The zone's endpoints aren't required to be MAX_ORDER
4932 * aligned but the node_mem_map endpoints must be in order
4933 * for the buddy allocator to function correctly.
4935 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4936 end
= pgdat_end_pfn(pgdat
);
4937 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4938 size
= (end
- start
) * sizeof(struct page
);
4939 map
= alloc_remap(pgdat
->node_id
, size
);
4941 map
= memblock_virt_alloc_node_nopanic(size
,
4943 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4945 #ifndef CONFIG_NEED_MULTIPLE_NODES
4947 * With no DISCONTIG, the global mem_map is just set as node 0's
4949 if (pgdat
== NODE_DATA(0)) {
4950 mem_map
= NODE_DATA(0)->node_mem_map
;
4951 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4952 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4953 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4954 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4957 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4960 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4961 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4963 pg_data_t
*pgdat
= NODE_DATA(nid
);
4964 unsigned long start_pfn
= 0;
4965 unsigned long end_pfn
= 0;
4967 /* pg_data_t should be reset to zero when it's allocated */
4968 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4970 pgdat
->node_id
= nid
;
4971 pgdat
->node_start_pfn
= node_start_pfn
;
4972 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4973 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4975 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
4976 zones_size
, zholes_size
);
4978 alloc_node_mem_map(pgdat
);
4979 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4980 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4981 nid
, (unsigned long)pgdat
,
4982 (unsigned long)pgdat
->node_mem_map
);
4985 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
4986 zones_size
, zholes_size
);
4989 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4991 #if MAX_NUMNODES > 1
4993 * Figure out the number of possible node ids.
4995 void __init
setup_nr_node_ids(void)
4998 unsigned int highest
= 0;
5000 for_each_node_mask(node
, node_possible_map
)
5002 nr_node_ids
= highest
+ 1;
5007 * node_map_pfn_alignment - determine the maximum internode alignment
5009 * This function should be called after node map is populated and sorted.
5010 * It calculates the maximum power of two alignment which can distinguish
5013 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5014 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5015 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5016 * shifted, 1GiB is enough and this function will indicate so.
5018 * This is used to test whether pfn -> nid mapping of the chosen memory
5019 * model has fine enough granularity to avoid incorrect mapping for the
5020 * populated node map.
5022 * Returns the determined alignment in pfn's. 0 if there is no alignment
5023 * requirement (single node).
5025 unsigned long __init
node_map_pfn_alignment(void)
5027 unsigned long accl_mask
= 0, last_end
= 0;
5028 unsigned long start
, end
, mask
;
5032 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5033 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5040 * Start with a mask granular enough to pin-point to the
5041 * start pfn and tick off bits one-by-one until it becomes
5042 * too coarse to separate the current node from the last.
5044 mask
= ~((1 << __ffs(start
)) - 1);
5045 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5048 /* accumulate all internode masks */
5052 /* convert mask to number of pages */
5053 return ~accl_mask
+ 1;
5056 /* Find the lowest pfn for a node */
5057 static unsigned long __init
find_min_pfn_for_node(int nid
)
5059 unsigned long min_pfn
= ULONG_MAX
;
5060 unsigned long start_pfn
;
5063 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5064 min_pfn
= min(min_pfn
, start_pfn
);
5066 if (min_pfn
== ULONG_MAX
) {
5068 "Could not find start_pfn for node %d\n", nid
);
5076 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5078 * It returns the minimum PFN based on information provided via
5079 * memblock_set_node().
5081 unsigned long __init
find_min_pfn_with_active_regions(void)
5083 return find_min_pfn_for_node(MAX_NUMNODES
);
5087 * early_calculate_totalpages()
5088 * Sum pages in active regions for movable zone.
5089 * Populate N_MEMORY for calculating usable_nodes.
5091 static unsigned long __init
early_calculate_totalpages(void)
5093 unsigned long totalpages
= 0;
5094 unsigned long start_pfn
, end_pfn
;
5097 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5098 unsigned long pages
= end_pfn
- start_pfn
;
5100 totalpages
+= pages
;
5102 node_set_state(nid
, N_MEMORY
);
5108 * Find the PFN the Movable zone begins in each node. Kernel memory
5109 * is spread evenly between nodes as long as the nodes have enough
5110 * memory. When they don't, some nodes will have more kernelcore than
5113 static void __init
find_zone_movable_pfns_for_nodes(void)
5116 unsigned long usable_startpfn
;
5117 unsigned long kernelcore_node
, kernelcore_remaining
;
5118 /* save the state before borrow the nodemask */
5119 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5120 unsigned long totalpages
= early_calculate_totalpages();
5121 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5122 struct memblock_region
*r
;
5124 /* Need to find movable_zone earlier when movable_node is specified. */
5125 find_usable_zone_for_movable();
5128 * If movable_node is specified, ignore kernelcore and movablecore
5131 if (movable_node_is_enabled()) {
5132 for_each_memblock(memory
, r
) {
5133 if (!memblock_is_hotpluggable(r
))
5138 usable_startpfn
= PFN_DOWN(r
->base
);
5139 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5140 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5148 * If movablecore=nn[KMG] was specified, calculate what size of
5149 * kernelcore that corresponds so that memory usable for
5150 * any allocation type is evenly spread. If both kernelcore
5151 * and movablecore are specified, then the value of kernelcore
5152 * will be used for required_kernelcore if it's greater than
5153 * what movablecore would have allowed.
5155 if (required_movablecore
) {
5156 unsigned long corepages
;
5159 * Round-up so that ZONE_MOVABLE is at least as large as what
5160 * was requested by the user
5162 required_movablecore
=
5163 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5164 corepages
= totalpages
- required_movablecore
;
5166 required_kernelcore
= max(required_kernelcore
, corepages
);
5169 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5170 if (!required_kernelcore
)
5173 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5174 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5177 /* Spread kernelcore memory as evenly as possible throughout nodes */
5178 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5179 for_each_node_state(nid
, N_MEMORY
) {
5180 unsigned long start_pfn
, end_pfn
;
5183 * Recalculate kernelcore_node if the division per node
5184 * now exceeds what is necessary to satisfy the requested
5185 * amount of memory for the kernel
5187 if (required_kernelcore
< kernelcore_node
)
5188 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5191 * As the map is walked, we track how much memory is usable
5192 * by the kernel using kernelcore_remaining. When it is
5193 * 0, the rest of the node is usable by ZONE_MOVABLE
5195 kernelcore_remaining
= kernelcore_node
;
5197 /* Go through each range of PFNs within this node */
5198 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5199 unsigned long size_pages
;
5201 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5202 if (start_pfn
>= end_pfn
)
5205 /* Account for what is only usable for kernelcore */
5206 if (start_pfn
< usable_startpfn
) {
5207 unsigned long kernel_pages
;
5208 kernel_pages
= min(end_pfn
, usable_startpfn
)
5211 kernelcore_remaining
-= min(kernel_pages
,
5212 kernelcore_remaining
);
5213 required_kernelcore
-= min(kernel_pages
,
5214 required_kernelcore
);
5216 /* Continue if range is now fully accounted */
5217 if (end_pfn
<= usable_startpfn
) {
5220 * Push zone_movable_pfn to the end so
5221 * that if we have to rebalance
5222 * kernelcore across nodes, we will
5223 * not double account here
5225 zone_movable_pfn
[nid
] = end_pfn
;
5228 start_pfn
= usable_startpfn
;
5232 * The usable PFN range for ZONE_MOVABLE is from
5233 * start_pfn->end_pfn. Calculate size_pages as the
5234 * number of pages used as kernelcore
5236 size_pages
= end_pfn
- start_pfn
;
5237 if (size_pages
> kernelcore_remaining
)
5238 size_pages
= kernelcore_remaining
;
5239 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5242 * Some kernelcore has been met, update counts and
5243 * break if the kernelcore for this node has been
5246 required_kernelcore
-= min(required_kernelcore
,
5248 kernelcore_remaining
-= size_pages
;
5249 if (!kernelcore_remaining
)
5255 * If there is still required_kernelcore, we do another pass with one
5256 * less node in the count. This will push zone_movable_pfn[nid] further
5257 * along on the nodes that still have memory until kernelcore is
5261 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5265 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5266 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5267 zone_movable_pfn
[nid
] =
5268 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5271 /* restore the node_state */
5272 node_states
[N_MEMORY
] = saved_node_state
;
5275 /* Any regular or high memory on that node ? */
5276 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5278 enum zone_type zone_type
;
5280 if (N_MEMORY
== N_NORMAL_MEMORY
)
5283 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5284 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5285 if (populated_zone(zone
)) {
5286 node_set_state(nid
, N_HIGH_MEMORY
);
5287 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5288 zone_type
<= ZONE_NORMAL
)
5289 node_set_state(nid
, N_NORMAL_MEMORY
);
5296 * free_area_init_nodes - Initialise all pg_data_t and zone data
5297 * @max_zone_pfn: an array of max PFNs for each zone
5299 * This will call free_area_init_node() for each active node in the system.
5300 * Using the page ranges provided by memblock_set_node(), the size of each
5301 * zone in each node and their holes is calculated. If the maximum PFN
5302 * between two adjacent zones match, it is assumed that the zone is empty.
5303 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5304 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5305 * starts where the previous one ended. For example, ZONE_DMA32 starts
5306 * at arch_max_dma_pfn.
5308 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5310 unsigned long start_pfn
, end_pfn
;
5313 /* Record where the zone boundaries are */
5314 memset(arch_zone_lowest_possible_pfn
, 0,
5315 sizeof(arch_zone_lowest_possible_pfn
));
5316 memset(arch_zone_highest_possible_pfn
, 0,
5317 sizeof(arch_zone_highest_possible_pfn
));
5318 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5319 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5320 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5321 if (i
== ZONE_MOVABLE
)
5323 arch_zone_lowest_possible_pfn
[i
] =
5324 arch_zone_highest_possible_pfn
[i
-1];
5325 arch_zone_highest_possible_pfn
[i
] =
5326 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5328 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5329 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5331 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5332 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5333 find_zone_movable_pfns_for_nodes();
5335 /* Print out the zone ranges */
5336 printk("Zone ranges:\n");
5337 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5338 if (i
== ZONE_MOVABLE
)
5340 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5341 if (arch_zone_lowest_possible_pfn
[i
] ==
5342 arch_zone_highest_possible_pfn
[i
])
5343 printk(KERN_CONT
"empty\n");
5345 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5346 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5347 (arch_zone_highest_possible_pfn
[i
]
5348 << PAGE_SHIFT
) - 1);
5351 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5352 printk("Movable zone start for each node\n");
5353 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5354 if (zone_movable_pfn
[i
])
5355 printk(" Node %d: %#010lx\n", i
,
5356 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5359 /* Print out the early node map */
5360 printk("Early memory node ranges\n");
5361 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5362 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5363 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5365 /* Initialise every node */
5366 mminit_verify_pageflags_layout();
5367 setup_nr_node_ids();
5368 for_each_online_node(nid
) {
5369 pg_data_t
*pgdat
= NODE_DATA(nid
);
5370 free_area_init_node(nid
, NULL
,
5371 find_min_pfn_for_node(nid
), NULL
);
5373 /* Any memory on that node */
5374 if (pgdat
->node_present_pages
)
5375 node_set_state(nid
, N_MEMORY
);
5376 check_for_memory(pgdat
, nid
);
5380 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5382 unsigned long long coremem
;
5386 coremem
= memparse(p
, &p
);
5387 *core
= coremem
>> PAGE_SHIFT
;
5389 /* Paranoid check that UL is enough for the coremem value */
5390 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5396 * kernelcore=size sets the amount of memory for use for allocations that
5397 * cannot be reclaimed or migrated.
5399 static int __init
cmdline_parse_kernelcore(char *p
)
5401 return cmdline_parse_core(p
, &required_kernelcore
);
5405 * movablecore=size sets the amount of memory for use for allocations that
5406 * can be reclaimed or migrated.
5408 static int __init
cmdline_parse_movablecore(char *p
)
5410 return cmdline_parse_core(p
, &required_movablecore
);
5413 early_param("kernelcore", cmdline_parse_kernelcore
);
5414 early_param("movablecore", cmdline_parse_movablecore
);
5416 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5418 void adjust_managed_page_count(struct page
*page
, long count
)
5420 spin_lock(&managed_page_count_lock
);
5421 page_zone(page
)->managed_pages
+= count
;
5422 totalram_pages
+= count
;
5423 #ifdef CONFIG_HIGHMEM
5424 if (PageHighMem(page
))
5425 totalhigh_pages
+= count
;
5427 spin_unlock(&managed_page_count_lock
);
5429 EXPORT_SYMBOL(adjust_managed_page_count
);
5431 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5434 unsigned long pages
= 0;
5436 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5437 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5438 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5439 if ((unsigned int)poison
<= 0xFF)
5440 memset(pos
, poison
, PAGE_SIZE
);
5441 free_reserved_page(virt_to_page(pos
));
5445 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5446 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5450 EXPORT_SYMBOL(free_reserved_area
);
5452 #ifdef CONFIG_HIGHMEM
5453 void free_highmem_page(struct page
*page
)
5455 __free_reserved_page(page
);
5457 page_zone(page
)->managed_pages
++;
5463 void __init
mem_init_print_info(const char *str
)
5465 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5466 unsigned long init_code_size
, init_data_size
;
5468 physpages
= get_num_physpages();
5469 codesize
= _etext
- _stext
;
5470 datasize
= _edata
- _sdata
;
5471 rosize
= __end_rodata
- __start_rodata
;
5472 bss_size
= __bss_stop
- __bss_start
;
5473 init_data_size
= __init_end
- __init_begin
;
5474 init_code_size
= _einittext
- _sinittext
;
5477 * Detect special cases and adjust section sizes accordingly:
5478 * 1) .init.* may be embedded into .data sections
5479 * 2) .init.text.* may be out of [__init_begin, __init_end],
5480 * please refer to arch/tile/kernel/vmlinux.lds.S.
5481 * 3) .rodata.* may be embedded into .text or .data sections.
5483 #define adj_init_size(start, end, size, pos, adj) \
5485 if (start <= pos && pos < end && size > adj) \
5489 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5490 _sinittext
, init_code_size
);
5491 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5492 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5493 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5494 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5496 #undef adj_init_size
5498 printk("Memory: %luK/%luK available "
5499 "(%luK kernel code, %luK rwdata, %luK rodata, "
5500 "%luK init, %luK bss, %luK reserved"
5501 #ifdef CONFIG_HIGHMEM
5505 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5506 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5507 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5508 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5509 #ifdef CONFIG_HIGHMEM
5510 totalhigh_pages
<< (PAGE_SHIFT
-10),
5512 str
? ", " : "", str
? str
: "");
5516 * set_dma_reserve - set the specified number of pages reserved in the first zone
5517 * @new_dma_reserve: The number of pages to mark reserved
5519 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5520 * In the DMA zone, a significant percentage may be consumed by kernel image
5521 * and other unfreeable allocations which can skew the watermarks badly. This
5522 * function may optionally be used to account for unfreeable pages in the
5523 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5524 * smaller per-cpu batchsize.
5526 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5528 dma_reserve
= new_dma_reserve
;
5531 void __init
free_area_init(unsigned long *zones_size
)
5533 free_area_init_node(0, zones_size
,
5534 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5537 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5538 unsigned long action
, void *hcpu
)
5540 int cpu
= (unsigned long)hcpu
;
5542 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5543 lru_add_drain_cpu(cpu
);
5547 * Spill the event counters of the dead processor
5548 * into the current processors event counters.
5549 * This artificially elevates the count of the current
5552 vm_events_fold_cpu(cpu
);
5555 * Zero the differential counters of the dead processor
5556 * so that the vm statistics are consistent.
5558 * This is only okay since the processor is dead and cannot
5559 * race with what we are doing.
5561 cpu_vm_stats_fold(cpu
);
5566 void __init
page_alloc_init(void)
5568 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5572 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5573 * or min_free_kbytes changes.
5575 static void calculate_totalreserve_pages(void)
5577 struct pglist_data
*pgdat
;
5578 unsigned long reserve_pages
= 0;
5579 enum zone_type i
, j
;
5581 for_each_online_pgdat(pgdat
) {
5582 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5583 struct zone
*zone
= pgdat
->node_zones
+ i
;
5586 /* Find valid and maximum lowmem_reserve in the zone */
5587 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5588 if (zone
->lowmem_reserve
[j
] > max
)
5589 max
= zone
->lowmem_reserve
[j
];
5592 /* we treat the high watermark as reserved pages. */
5593 max
+= high_wmark_pages(zone
);
5595 if (max
> zone
->managed_pages
)
5596 max
= zone
->managed_pages
;
5597 reserve_pages
+= max
;
5599 * Lowmem reserves are not available to
5600 * GFP_HIGHUSER page cache allocations and
5601 * kswapd tries to balance zones to their high
5602 * watermark. As a result, neither should be
5603 * regarded as dirtyable memory, to prevent a
5604 * situation where reclaim has to clean pages
5605 * in order to balance the zones.
5607 zone
->dirty_balance_reserve
= max
;
5610 dirty_balance_reserve
= reserve_pages
;
5611 totalreserve_pages
= reserve_pages
;
5615 * setup_per_zone_lowmem_reserve - called whenever
5616 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5617 * has a correct pages reserved value, so an adequate number of
5618 * pages are left in the zone after a successful __alloc_pages().
5620 static void setup_per_zone_lowmem_reserve(void)
5622 struct pglist_data
*pgdat
;
5623 enum zone_type j
, idx
;
5625 for_each_online_pgdat(pgdat
) {
5626 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5627 struct zone
*zone
= pgdat
->node_zones
+ j
;
5628 unsigned long managed_pages
= zone
->managed_pages
;
5630 zone
->lowmem_reserve
[j
] = 0;
5634 struct zone
*lower_zone
;
5638 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5639 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5641 lower_zone
= pgdat
->node_zones
+ idx
;
5642 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5643 sysctl_lowmem_reserve_ratio
[idx
];
5644 managed_pages
+= lower_zone
->managed_pages
;
5649 /* update totalreserve_pages */
5650 calculate_totalreserve_pages();
5653 static void __setup_per_zone_wmarks(void)
5655 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5656 unsigned long lowmem_pages
= 0;
5658 unsigned long flags
;
5660 /* Calculate total number of !ZONE_HIGHMEM pages */
5661 for_each_zone(zone
) {
5662 if (!is_highmem(zone
))
5663 lowmem_pages
+= zone
->managed_pages
;
5666 for_each_zone(zone
) {
5669 spin_lock_irqsave(&zone
->lock
, flags
);
5670 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5671 do_div(tmp
, lowmem_pages
);
5672 if (is_highmem(zone
)) {
5674 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5675 * need highmem pages, so cap pages_min to a small
5678 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5679 * deltas controls asynch page reclaim, and so should
5680 * not be capped for highmem.
5682 unsigned long min_pages
;
5684 min_pages
= zone
->managed_pages
/ 1024;
5685 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5686 zone
->watermark
[WMARK_MIN
] = min_pages
;
5689 * If it's a lowmem zone, reserve a number of pages
5690 * proportionate to the zone's size.
5692 zone
->watermark
[WMARK_MIN
] = tmp
;
5695 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5696 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5698 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5699 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5700 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5702 setup_zone_migrate_reserve(zone
);
5703 spin_unlock_irqrestore(&zone
->lock
, flags
);
5706 /* update totalreserve_pages */
5707 calculate_totalreserve_pages();
5711 * setup_per_zone_wmarks - called when min_free_kbytes changes
5712 * or when memory is hot-{added|removed}
5714 * Ensures that the watermark[min,low,high] values for each zone are set
5715 * correctly with respect to min_free_kbytes.
5717 void setup_per_zone_wmarks(void)
5719 mutex_lock(&zonelists_mutex
);
5720 __setup_per_zone_wmarks();
5721 mutex_unlock(&zonelists_mutex
);
5725 * The inactive anon list should be small enough that the VM never has to
5726 * do too much work, but large enough that each inactive page has a chance
5727 * to be referenced again before it is swapped out.
5729 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5730 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5731 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5732 * the anonymous pages are kept on the inactive list.
5735 * memory ratio inactive anon
5736 * -------------------------------------
5745 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5747 unsigned int gb
, ratio
;
5749 /* Zone size in gigabytes */
5750 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5752 ratio
= int_sqrt(10 * gb
);
5756 zone
->inactive_ratio
= ratio
;
5759 static void __meminit
setup_per_zone_inactive_ratio(void)
5764 calculate_zone_inactive_ratio(zone
);
5768 * Initialise min_free_kbytes.
5770 * For small machines we want it small (128k min). For large machines
5771 * we want it large (64MB max). But it is not linear, because network
5772 * bandwidth does not increase linearly with machine size. We use
5774 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5775 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5791 int __meminit
init_per_zone_wmark_min(void)
5793 unsigned long lowmem_kbytes
;
5794 int new_min_free_kbytes
;
5796 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5797 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5799 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5800 min_free_kbytes
= new_min_free_kbytes
;
5801 if (min_free_kbytes
< 128)
5802 min_free_kbytes
= 128;
5803 if (min_free_kbytes
> 65536)
5804 min_free_kbytes
= 65536;
5806 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5807 new_min_free_kbytes
, user_min_free_kbytes
);
5809 setup_per_zone_wmarks();
5810 refresh_zone_stat_thresholds();
5811 setup_per_zone_lowmem_reserve();
5812 setup_per_zone_inactive_ratio();
5815 module_init(init_per_zone_wmark_min
)
5818 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5819 * that we can call two helper functions whenever min_free_kbytes
5822 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5823 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5827 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5832 user_min_free_kbytes
= min_free_kbytes
;
5833 setup_per_zone_wmarks();
5839 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5840 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5845 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5850 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5851 sysctl_min_unmapped_ratio
) / 100;
5855 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5856 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5861 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5866 zone
->min_slab_pages
= (zone
->managed_pages
*
5867 sysctl_min_slab_ratio
) / 100;
5873 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5874 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5875 * whenever sysctl_lowmem_reserve_ratio changes.
5877 * The reserve ratio obviously has absolutely no relation with the
5878 * minimum watermarks. The lowmem reserve ratio can only make sense
5879 * if in function of the boot time zone sizes.
5881 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5882 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5884 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5885 setup_per_zone_lowmem_reserve();
5890 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5891 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5892 * pagelist can have before it gets flushed back to buddy allocator.
5894 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5895 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5898 int old_percpu_pagelist_fraction
;
5901 mutex_lock(&pcp_batch_high_lock
);
5902 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5904 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5905 if (!write
|| ret
< 0)
5908 /* Sanity checking to avoid pcp imbalance */
5909 if (percpu_pagelist_fraction
&&
5910 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5911 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5917 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
5920 for_each_populated_zone(zone
) {
5923 for_each_possible_cpu(cpu
)
5924 pageset_set_high_and_batch(zone
,
5925 per_cpu_ptr(zone
->pageset
, cpu
));
5928 mutex_unlock(&pcp_batch_high_lock
);
5932 int hashdist
= HASHDIST_DEFAULT
;
5935 static int __init
set_hashdist(char *str
)
5939 hashdist
= simple_strtoul(str
, &str
, 0);
5942 __setup("hashdist=", set_hashdist
);
5946 * allocate a large system hash table from bootmem
5947 * - it is assumed that the hash table must contain an exact power-of-2
5948 * quantity of entries
5949 * - limit is the number of hash buckets, not the total allocation size
5951 void *__init
alloc_large_system_hash(const char *tablename
,
5952 unsigned long bucketsize
,
5953 unsigned long numentries
,
5956 unsigned int *_hash_shift
,
5957 unsigned int *_hash_mask
,
5958 unsigned long low_limit
,
5959 unsigned long high_limit
)
5961 unsigned long long max
= high_limit
;
5962 unsigned long log2qty
, size
;
5965 /* allow the kernel cmdline to have a say */
5967 /* round applicable memory size up to nearest megabyte */
5968 numentries
= nr_kernel_pages
;
5970 /* It isn't necessary when PAGE_SIZE >= 1MB */
5971 if (PAGE_SHIFT
< 20)
5972 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
5974 /* limit to 1 bucket per 2^scale bytes of low memory */
5975 if (scale
> PAGE_SHIFT
)
5976 numentries
>>= (scale
- PAGE_SHIFT
);
5978 numentries
<<= (PAGE_SHIFT
- scale
);
5980 /* Make sure we've got at least a 0-order allocation.. */
5981 if (unlikely(flags
& HASH_SMALL
)) {
5982 /* Makes no sense without HASH_EARLY */
5983 WARN_ON(!(flags
& HASH_EARLY
));
5984 if (!(numentries
>> *_hash_shift
)) {
5985 numentries
= 1UL << *_hash_shift
;
5986 BUG_ON(!numentries
);
5988 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5989 numentries
= PAGE_SIZE
/ bucketsize
;
5991 numentries
= roundup_pow_of_two(numentries
);
5993 /* limit allocation size to 1/16 total memory by default */
5995 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5996 do_div(max
, bucketsize
);
5998 max
= min(max
, 0x80000000ULL
);
6000 if (numentries
< low_limit
)
6001 numentries
= low_limit
;
6002 if (numentries
> max
)
6005 log2qty
= ilog2(numentries
);
6008 size
= bucketsize
<< log2qty
;
6009 if (flags
& HASH_EARLY
)
6010 table
= memblock_virt_alloc_nopanic(size
, 0);
6012 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6015 * If bucketsize is not a power-of-two, we may free
6016 * some pages at the end of hash table which
6017 * alloc_pages_exact() automatically does
6019 if (get_order(size
) < MAX_ORDER
) {
6020 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6021 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6024 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6027 panic("Failed to allocate %s hash table\n", tablename
);
6029 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6032 ilog2(size
) - PAGE_SHIFT
,
6036 *_hash_shift
= log2qty
;
6038 *_hash_mask
= (1 << log2qty
) - 1;
6043 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6044 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6047 #ifdef CONFIG_SPARSEMEM
6048 return __pfn_to_section(pfn
)->pageblock_flags
;
6050 return zone
->pageblock_flags
;
6051 #endif /* CONFIG_SPARSEMEM */
6054 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6056 #ifdef CONFIG_SPARSEMEM
6057 pfn
&= (PAGES_PER_SECTION
-1);
6058 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6060 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6061 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6062 #endif /* CONFIG_SPARSEMEM */
6066 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6067 * @page: The page within the block of interest
6068 * @pfn: The target page frame number
6069 * @end_bitidx: The last bit of interest to retrieve
6070 * @mask: mask of bits that the caller is interested in
6072 * Return: pageblock_bits flags
6074 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6075 unsigned long end_bitidx
,
6079 unsigned long *bitmap
;
6080 unsigned long bitidx
, word_bitidx
;
6083 zone
= page_zone(page
);
6084 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6085 bitidx
= pfn_to_bitidx(zone
, pfn
);
6086 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6087 bitidx
&= (BITS_PER_LONG
-1);
6089 word
= bitmap
[word_bitidx
];
6090 bitidx
+= end_bitidx
;
6091 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6095 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6096 * @page: The page within the block of interest
6097 * @flags: The flags to set
6098 * @pfn: The target page frame number
6099 * @end_bitidx: The last bit of interest
6100 * @mask: mask of bits that the caller is interested in
6102 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6104 unsigned long end_bitidx
,
6108 unsigned long *bitmap
;
6109 unsigned long bitidx
, word_bitidx
;
6110 unsigned long old_word
, word
;
6112 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6114 zone
= page_zone(page
);
6115 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6116 bitidx
= pfn_to_bitidx(zone
, pfn
);
6117 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6118 bitidx
&= (BITS_PER_LONG
-1);
6120 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6122 bitidx
+= end_bitidx
;
6123 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6124 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6126 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6128 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6129 if (word
== old_word
)
6136 * This function checks whether pageblock includes unmovable pages or not.
6137 * If @count is not zero, it is okay to include less @count unmovable pages
6139 * PageLRU check without isolation or lru_lock could race so that
6140 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6141 * expect this function should be exact.
6143 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6144 bool skip_hwpoisoned_pages
)
6146 unsigned long pfn
, iter
, found
;
6150 * For avoiding noise data, lru_add_drain_all() should be called
6151 * If ZONE_MOVABLE, the zone never contains unmovable pages
6153 if (zone_idx(zone
) == ZONE_MOVABLE
)
6155 mt
= get_pageblock_migratetype(page
);
6156 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6159 pfn
= page_to_pfn(page
);
6160 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6161 unsigned long check
= pfn
+ iter
;
6163 if (!pfn_valid_within(check
))
6166 page
= pfn_to_page(check
);
6169 * Hugepages are not in LRU lists, but they're movable.
6170 * We need not scan over tail pages bacause we don't
6171 * handle each tail page individually in migration.
6173 if (PageHuge(page
)) {
6174 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6179 * We can't use page_count without pin a page
6180 * because another CPU can free compound page.
6181 * This check already skips compound tails of THP
6182 * because their page->_count is zero at all time.
6184 if (!atomic_read(&page
->_count
)) {
6185 if (PageBuddy(page
))
6186 iter
+= (1 << page_order(page
)) - 1;
6191 * The HWPoisoned page may be not in buddy system, and
6192 * page_count() is not 0.
6194 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6200 * If there are RECLAIMABLE pages, we need to check it.
6201 * But now, memory offline itself doesn't call shrink_slab()
6202 * and it still to be fixed.
6205 * If the page is not RAM, page_count()should be 0.
6206 * we don't need more check. This is an _used_ not-movable page.
6208 * The problematic thing here is PG_reserved pages. PG_reserved
6209 * is set to both of a memory hole page and a _used_ kernel
6218 bool is_pageblock_removable_nolock(struct page
*page
)
6224 * We have to be careful here because we are iterating over memory
6225 * sections which are not zone aware so we might end up outside of
6226 * the zone but still within the section.
6227 * We have to take care about the node as well. If the node is offline
6228 * its NODE_DATA will be NULL - see page_zone.
6230 if (!node_online(page_to_nid(page
)))
6233 zone
= page_zone(page
);
6234 pfn
= page_to_pfn(page
);
6235 if (!zone_spans_pfn(zone
, pfn
))
6238 return !has_unmovable_pages(zone
, page
, 0, true);
6243 static unsigned long pfn_max_align_down(unsigned long pfn
)
6245 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6246 pageblock_nr_pages
) - 1);
6249 static unsigned long pfn_max_align_up(unsigned long pfn
)
6251 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6252 pageblock_nr_pages
));
6255 /* [start, end) must belong to a single zone. */
6256 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6257 unsigned long start
, unsigned long end
)
6259 /* This function is based on compact_zone() from compaction.c. */
6260 unsigned long nr_reclaimed
;
6261 unsigned long pfn
= start
;
6262 unsigned int tries
= 0;
6267 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6268 if (fatal_signal_pending(current
)) {
6273 if (list_empty(&cc
->migratepages
)) {
6274 cc
->nr_migratepages
= 0;
6275 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6281 } else if (++tries
== 5) {
6282 ret
= ret
< 0 ? ret
: -EBUSY
;
6286 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6288 cc
->nr_migratepages
-= nr_reclaimed
;
6290 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6291 NULL
, 0, cc
->mode
, MR_CMA
);
6294 putback_movable_pages(&cc
->migratepages
);
6301 * alloc_contig_range() -- tries to allocate given range of pages
6302 * @start: start PFN to allocate
6303 * @end: one-past-the-last PFN to allocate
6304 * @migratetype: migratetype of the underlaying pageblocks (either
6305 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6306 * in range must have the same migratetype and it must
6307 * be either of the two.
6309 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6310 * aligned, however it's the caller's responsibility to guarantee that
6311 * we are the only thread that changes migrate type of pageblocks the
6314 * The PFN range must belong to a single zone.
6316 * Returns zero on success or negative error code. On success all
6317 * pages which PFN is in [start, end) are allocated for the caller and
6318 * need to be freed with free_contig_range().
6320 int alloc_contig_range(unsigned long start
, unsigned long end
,
6321 unsigned migratetype
)
6323 unsigned long outer_start
, outer_end
;
6326 struct compact_control cc
= {
6327 .nr_migratepages
= 0,
6329 .zone
= page_zone(pfn_to_page(start
)),
6330 .mode
= MIGRATE_SYNC
,
6331 .ignore_skip_hint
= true,
6333 INIT_LIST_HEAD(&cc
.migratepages
);
6336 * What we do here is we mark all pageblocks in range as
6337 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6338 * have different sizes, and due to the way page allocator
6339 * work, we align the range to biggest of the two pages so
6340 * that page allocator won't try to merge buddies from
6341 * different pageblocks and change MIGRATE_ISOLATE to some
6342 * other migration type.
6344 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6345 * migrate the pages from an unaligned range (ie. pages that
6346 * we are interested in). This will put all the pages in
6347 * range back to page allocator as MIGRATE_ISOLATE.
6349 * When this is done, we take the pages in range from page
6350 * allocator removing them from the buddy system. This way
6351 * page allocator will never consider using them.
6353 * This lets us mark the pageblocks back as
6354 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6355 * aligned range but not in the unaligned, original range are
6356 * put back to page allocator so that buddy can use them.
6359 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6360 pfn_max_align_up(end
), migratetype
,
6365 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6370 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6371 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6372 * more, all pages in [start, end) are free in page allocator.
6373 * What we are going to do is to allocate all pages from
6374 * [start, end) (that is remove them from page allocator).
6376 * The only problem is that pages at the beginning and at the
6377 * end of interesting range may be not aligned with pages that
6378 * page allocator holds, ie. they can be part of higher order
6379 * pages. Because of this, we reserve the bigger range and
6380 * once this is done free the pages we are not interested in.
6382 * We don't have to hold zone->lock here because the pages are
6383 * isolated thus they won't get removed from buddy.
6386 lru_add_drain_all();
6390 outer_start
= start
;
6391 while (!PageBuddy(pfn_to_page(outer_start
))) {
6392 if (++order
>= MAX_ORDER
) {
6396 outer_start
&= ~0UL << order
;
6399 /* Make sure the range is really isolated. */
6400 if (test_pages_isolated(outer_start
, end
, false)) {
6401 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6408 /* Grab isolated pages from freelists. */
6409 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6415 /* Free head and tail (if any) */
6416 if (start
!= outer_start
)
6417 free_contig_range(outer_start
, start
- outer_start
);
6418 if (end
!= outer_end
)
6419 free_contig_range(end
, outer_end
- end
);
6422 undo_isolate_page_range(pfn_max_align_down(start
),
6423 pfn_max_align_up(end
), migratetype
);
6427 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6429 unsigned int count
= 0;
6431 for (; nr_pages
--; pfn
++) {
6432 struct page
*page
= pfn_to_page(pfn
);
6434 count
+= page_count(page
) != 1;
6437 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6441 #ifdef CONFIG_MEMORY_HOTPLUG
6443 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6444 * page high values need to be recalulated.
6446 void __meminit
zone_pcp_update(struct zone
*zone
)
6449 mutex_lock(&pcp_batch_high_lock
);
6450 for_each_possible_cpu(cpu
)
6451 pageset_set_high_and_batch(zone
,
6452 per_cpu_ptr(zone
->pageset
, cpu
));
6453 mutex_unlock(&pcp_batch_high_lock
);
6457 void zone_pcp_reset(struct zone
*zone
)
6459 unsigned long flags
;
6461 struct per_cpu_pageset
*pset
;
6463 /* avoid races with drain_pages() */
6464 local_irq_save(flags
);
6465 if (zone
->pageset
!= &boot_pageset
) {
6466 for_each_online_cpu(cpu
) {
6467 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6468 drain_zonestat(zone
, pset
);
6470 free_percpu(zone
->pageset
);
6471 zone
->pageset
= &boot_pageset
;
6473 local_irq_restore(flags
);
6476 #ifdef CONFIG_MEMORY_HOTREMOVE
6478 * All pages in the range must be isolated before calling this.
6481 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6485 unsigned int order
, i
;
6487 unsigned long flags
;
6488 /* find the first valid pfn */
6489 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6494 zone
= page_zone(pfn_to_page(pfn
));
6495 spin_lock_irqsave(&zone
->lock
, flags
);
6497 while (pfn
< end_pfn
) {
6498 if (!pfn_valid(pfn
)) {
6502 page
= pfn_to_page(pfn
);
6504 * The HWPoisoned page may be not in buddy system, and
6505 * page_count() is not 0.
6507 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6509 SetPageReserved(page
);
6513 BUG_ON(page_count(page
));
6514 BUG_ON(!PageBuddy(page
));
6515 order
= page_order(page
);
6516 #ifdef CONFIG_DEBUG_VM
6517 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6518 pfn
, 1 << order
, end_pfn
);
6520 list_del(&page
->lru
);
6521 rmv_page_order(page
);
6522 zone
->free_area
[order
].nr_free
--;
6523 for (i
= 0; i
< (1 << order
); i
++)
6524 SetPageReserved((page
+i
));
6525 pfn
+= (1 << order
);
6527 spin_unlock_irqrestore(&zone
->lock
, flags
);
6531 #ifdef CONFIG_MEMORY_FAILURE
6532 bool is_free_buddy_page(struct page
*page
)
6534 struct zone
*zone
= page_zone(page
);
6535 unsigned long pfn
= page_to_pfn(page
);
6536 unsigned long flags
;
6539 spin_lock_irqsave(&zone
->lock
, flags
);
6540 for (order
= 0; order
< MAX_ORDER
; order
++) {
6541 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6543 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6546 spin_unlock_irqrestore(&zone
->lock
, flags
);
6548 return order
< MAX_ORDER
;