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 if (unlikely(has_isolate_pageblock(zone
))) {
719 mt
= get_pageblock_migratetype(page
);
720 if (is_migrate_isolate(mt
))
723 __mod_zone_freepage_state(zone
, 1, mt
);
726 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
727 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
728 trace_mm_page_pcpu_drain(page
, 0, mt
);
729 } while (--to_free
&& --batch_free
&& !list_empty(list
));
731 spin_unlock(&zone
->lock
);
734 static void free_one_page(struct zone
*zone
,
735 struct page
*page
, unsigned long pfn
,
739 unsigned long nr_scanned
;
740 spin_lock(&zone
->lock
);
741 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
743 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
745 if (unlikely(has_isolate_pageblock(zone
) ||
746 is_migrate_isolate(migratetype
))) {
747 migratetype
= get_pfnblock_migratetype(page
, pfn
);
748 if (is_migrate_isolate(migratetype
))
751 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
754 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
755 spin_unlock(&zone
->lock
);
758 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
763 trace_mm_page_free(page
, order
);
764 kmemcheck_free_shadow(page
, order
);
767 page
->mapping
= NULL
;
768 for (i
= 0; i
< (1 << order
); i
++)
769 bad
+= free_pages_check(page
+ i
);
773 if (!PageHighMem(page
)) {
774 debug_check_no_locks_freed(page_address(page
),
776 debug_check_no_obj_freed(page_address(page
),
779 arch_free_page(page
, order
);
780 kernel_map_pages(page
, 1 << order
, 0);
785 static void __free_pages_ok(struct page
*page
, unsigned int order
)
789 unsigned long pfn
= page_to_pfn(page
);
791 if (!free_pages_prepare(page
, order
))
794 migratetype
= get_pfnblock_migratetype(page
, pfn
);
795 local_irq_save(flags
);
796 __count_vm_events(PGFREE
, 1 << order
);
797 set_freepage_migratetype(page
, migratetype
);
798 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
799 local_irq_restore(flags
);
802 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
804 unsigned int nr_pages
= 1 << order
;
805 struct page
*p
= page
;
809 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
811 __ClearPageReserved(p
);
812 set_page_count(p
, 0);
814 __ClearPageReserved(p
);
815 set_page_count(p
, 0);
817 page_zone(page
)->managed_pages
+= nr_pages
;
818 set_page_refcounted(page
);
819 __free_pages(page
, order
);
823 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
824 void __init
init_cma_reserved_pageblock(struct page
*page
)
826 unsigned i
= pageblock_nr_pages
;
827 struct page
*p
= page
;
830 __ClearPageReserved(p
);
831 set_page_count(p
, 0);
834 set_pageblock_migratetype(page
, MIGRATE_CMA
);
836 if (pageblock_order
>= MAX_ORDER
) {
837 i
= pageblock_nr_pages
;
840 set_page_refcounted(p
);
841 __free_pages(p
, MAX_ORDER
- 1);
842 p
+= MAX_ORDER_NR_PAGES
;
843 } while (i
-= MAX_ORDER_NR_PAGES
);
845 set_page_refcounted(page
);
846 __free_pages(page
, pageblock_order
);
849 adjust_managed_page_count(page
, pageblock_nr_pages
);
854 * The order of subdivision here is critical for the IO subsystem.
855 * Please do not alter this order without good reasons and regression
856 * testing. Specifically, as large blocks of memory are subdivided,
857 * the order in which smaller blocks are delivered depends on the order
858 * they're subdivided in this function. This is the primary factor
859 * influencing the order in which pages are delivered to the IO
860 * subsystem according to empirical testing, and this is also justified
861 * by considering the behavior of a buddy system containing a single
862 * large block of memory acted on by a series of small allocations.
863 * This behavior is a critical factor in sglist merging's success.
867 static inline void expand(struct zone
*zone
, struct page
*page
,
868 int low
, int high
, struct free_area
*area
,
871 unsigned long size
= 1 << high
;
877 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
879 #ifdef CONFIG_DEBUG_PAGEALLOC
880 if (high
< debug_guardpage_minorder()) {
882 * Mark as guard pages (or page), that will allow to
883 * merge back to allocator when buddy will be freed.
884 * Corresponding page table entries will not be touched,
885 * pages will stay not present in virtual address space
887 INIT_LIST_HEAD(&page
[size
].lru
);
888 set_page_guard_flag(&page
[size
]);
889 set_page_private(&page
[size
], high
);
890 /* Guard pages are not available for any usage */
891 __mod_zone_freepage_state(zone
, -(1 << high
),
896 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
898 set_page_order(&page
[size
], high
);
903 * This page is about to be returned from the page allocator
905 static inline int check_new_page(struct page
*page
)
907 const char *bad_reason
= NULL
;
908 unsigned long bad_flags
= 0;
910 if (unlikely(page_mapcount(page
)))
911 bad_reason
= "nonzero mapcount";
912 if (unlikely(page
->mapping
!= NULL
))
913 bad_reason
= "non-NULL mapping";
914 if (unlikely(atomic_read(&page
->_count
) != 0))
915 bad_reason
= "nonzero _count";
916 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
917 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
918 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
920 if (unlikely(mem_cgroup_bad_page_check(page
)))
921 bad_reason
= "cgroup check failed";
922 if (unlikely(bad_reason
)) {
923 bad_page(page
, bad_reason
, bad_flags
);
929 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
)
933 for (i
= 0; i
< (1 << order
); i
++) {
934 struct page
*p
= page
+ i
;
935 if (unlikely(check_new_page(p
)))
939 set_page_private(page
, 0);
940 set_page_refcounted(page
);
942 arch_alloc_page(page
, order
);
943 kernel_map_pages(page
, 1 << order
, 1);
945 if (gfp_flags
& __GFP_ZERO
)
946 prep_zero_page(page
, order
, gfp_flags
);
948 if (order
&& (gfp_flags
& __GFP_COMP
))
949 prep_compound_page(page
, order
);
955 * Go through the free lists for the given migratetype and remove
956 * the smallest available page from the freelists
959 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
962 unsigned int current_order
;
963 struct free_area
*area
;
966 /* Find a page of the appropriate size in the preferred list */
967 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
968 area
= &(zone
->free_area
[current_order
]);
969 if (list_empty(&area
->free_list
[migratetype
]))
972 page
= list_entry(area
->free_list
[migratetype
].next
,
974 list_del(&page
->lru
);
975 rmv_page_order(page
);
977 expand(zone
, page
, order
, current_order
, area
, migratetype
);
978 set_freepage_migratetype(page
, migratetype
);
987 * This array describes the order lists are fallen back to when
988 * the free lists for the desirable migrate type are depleted
990 static int fallbacks
[MIGRATE_TYPES
][4] = {
991 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
992 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
994 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
995 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
997 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
999 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
1000 #ifdef CONFIG_MEMORY_ISOLATION
1001 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
1006 * Move the free pages in a range to the free lists of the requested type.
1007 * Note that start_page and end_pages are not aligned on a pageblock
1008 * boundary. If alignment is required, use move_freepages_block()
1010 int move_freepages(struct zone
*zone
,
1011 struct page
*start_page
, struct page
*end_page
,
1015 unsigned long order
;
1016 int pages_moved
= 0;
1018 #ifndef CONFIG_HOLES_IN_ZONE
1020 * page_zone is not safe to call in this context when
1021 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1022 * anyway as we check zone boundaries in move_freepages_block().
1023 * Remove at a later date when no bug reports exist related to
1024 * grouping pages by mobility
1026 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1029 for (page
= start_page
; page
<= end_page
;) {
1030 /* Make sure we are not inadvertently changing nodes */
1031 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1033 if (!pfn_valid_within(page_to_pfn(page
))) {
1038 if (!PageBuddy(page
)) {
1043 order
= page_order(page
);
1044 list_move(&page
->lru
,
1045 &zone
->free_area
[order
].free_list
[migratetype
]);
1046 set_freepage_migratetype(page
, migratetype
);
1048 pages_moved
+= 1 << order
;
1054 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1057 unsigned long start_pfn
, end_pfn
;
1058 struct page
*start_page
, *end_page
;
1060 start_pfn
= page_to_pfn(page
);
1061 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1062 start_page
= pfn_to_page(start_pfn
);
1063 end_page
= start_page
+ pageblock_nr_pages
- 1;
1064 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1066 /* Do not cross zone boundaries */
1067 if (!zone_spans_pfn(zone
, start_pfn
))
1069 if (!zone_spans_pfn(zone
, end_pfn
))
1072 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1075 static void change_pageblock_range(struct page
*pageblock_page
,
1076 int start_order
, int migratetype
)
1078 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1080 while (nr_pageblocks
--) {
1081 set_pageblock_migratetype(pageblock_page
, migratetype
);
1082 pageblock_page
+= pageblock_nr_pages
;
1087 * If breaking a large block of pages, move all free pages to the preferred
1088 * allocation list. If falling back for a reclaimable kernel allocation, be
1089 * more aggressive about taking ownership of free pages.
1091 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1092 * nor move CMA pages to different free lists. We don't want unmovable pages
1093 * to be allocated from MIGRATE_CMA areas.
1095 * Returns the new migratetype of the pageblock (or the same old migratetype
1096 * if it was unchanged).
1098 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1099 int start_type
, int fallback_type
)
1101 int current_order
= page_order(page
);
1104 * When borrowing from MIGRATE_CMA, we need to release the excess
1105 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1106 * is set to CMA so it is returned to the correct freelist in case
1107 * the page ends up being not actually allocated from the pcp lists.
1109 if (is_migrate_cma(fallback_type
))
1110 return fallback_type
;
1112 /* Take ownership for orders >= pageblock_order */
1113 if (current_order
>= pageblock_order
) {
1114 change_pageblock_range(page
, current_order
, start_type
);
1118 if (current_order
>= pageblock_order
/ 2 ||
1119 start_type
== MIGRATE_RECLAIMABLE
||
1120 page_group_by_mobility_disabled
) {
1123 pages
= move_freepages_block(zone
, page
, start_type
);
1125 /* Claim the whole block if over half of it is free */
1126 if (pages
>= (1 << (pageblock_order
-1)) ||
1127 page_group_by_mobility_disabled
) {
1129 set_pageblock_migratetype(page
, start_type
);
1135 return fallback_type
;
1138 /* Remove an element from the buddy allocator from the fallback list */
1139 static inline struct page
*
1140 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1142 struct free_area
*area
;
1143 unsigned int current_order
;
1145 int migratetype
, new_type
, i
;
1147 /* Find the largest possible block of pages in the other list */
1148 for (current_order
= MAX_ORDER
-1;
1149 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1152 migratetype
= fallbacks
[start_migratetype
][i
];
1154 /* MIGRATE_RESERVE handled later if necessary */
1155 if (migratetype
== MIGRATE_RESERVE
)
1158 area
= &(zone
->free_area
[current_order
]);
1159 if (list_empty(&area
->free_list
[migratetype
]))
1162 page
= list_entry(area
->free_list
[migratetype
].next
,
1166 new_type
= try_to_steal_freepages(zone
, page
,
1170 /* Remove the page from the freelists */
1171 list_del(&page
->lru
);
1172 rmv_page_order(page
);
1174 expand(zone
, page
, order
, current_order
, area
,
1176 /* The freepage_migratetype may differ from pageblock's
1177 * migratetype depending on the decisions in
1178 * try_to_steal_freepages. This is OK as long as it does
1179 * not differ for MIGRATE_CMA type.
1181 set_freepage_migratetype(page
, new_type
);
1183 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1184 start_migratetype
, migratetype
, new_type
);
1194 * Do the hard work of removing an element from the buddy allocator.
1195 * Call me with the zone->lock already held.
1197 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1203 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1205 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1206 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1209 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1210 * is used because __rmqueue_smallest is an inline function
1211 * and we want just one call site
1214 migratetype
= MIGRATE_RESERVE
;
1219 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1224 * Obtain a specified number of elements from the buddy allocator, all under
1225 * a single hold of the lock, for efficiency. Add them to the supplied list.
1226 * Returns the number of new pages which were placed at *list.
1228 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1229 unsigned long count
, struct list_head
*list
,
1230 int migratetype
, bool cold
)
1234 spin_lock(&zone
->lock
);
1235 for (i
= 0; i
< count
; ++i
) {
1236 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1237 if (unlikely(page
== NULL
))
1241 * Split buddy pages returned by expand() are received here
1242 * in physical page order. The page is added to the callers and
1243 * list and the list head then moves forward. From the callers
1244 * perspective, the linked list is ordered by page number in
1245 * some conditions. This is useful for IO devices that can
1246 * merge IO requests if the physical pages are ordered
1250 list_add(&page
->lru
, list
);
1252 list_add_tail(&page
->lru
, list
);
1254 if (is_migrate_cma(get_freepage_migratetype(page
)))
1255 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1258 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1259 spin_unlock(&zone
->lock
);
1265 * Called from the vmstat counter updater to drain pagesets of this
1266 * currently executing processor on remote nodes after they have
1269 * Note that this function must be called with the thread pinned to
1270 * a single processor.
1272 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1274 unsigned long flags
;
1275 int to_drain
, batch
;
1277 local_irq_save(flags
);
1278 batch
= ACCESS_ONCE(pcp
->batch
);
1279 to_drain
= min(pcp
->count
, batch
);
1281 free_pcppages_bulk(zone
, to_drain
, pcp
);
1282 pcp
->count
-= to_drain
;
1284 local_irq_restore(flags
);
1289 * Drain pages of the indicated processor.
1291 * The processor must either be the current processor and the
1292 * thread pinned to the current processor or a processor that
1295 static void drain_pages(unsigned int cpu
)
1297 unsigned long flags
;
1300 for_each_populated_zone(zone
) {
1301 struct per_cpu_pageset
*pset
;
1302 struct per_cpu_pages
*pcp
;
1304 local_irq_save(flags
);
1305 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1309 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1312 local_irq_restore(flags
);
1317 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1319 void drain_local_pages(void *arg
)
1321 drain_pages(smp_processor_id());
1325 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1327 * Note that this code is protected against sending an IPI to an offline
1328 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1329 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1330 * nothing keeps CPUs from showing up after we populated the cpumask and
1331 * before the call to on_each_cpu_mask().
1333 void drain_all_pages(void)
1336 struct per_cpu_pageset
*pcp
;
1340 * Allocate in the BSS so we wont require allocation in
1341 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1343 static cpumask_t cpus_with_pcps
;
1346 * We don't care about racing with CPU hotplug event
1347 * as offline notification will cause the notified
1348 * cpu to drain that CPU pcps and on_each_cpu_mask
1349 * disables preemption as part of its processing
1351 for_each_online_cpu(cpu
) {
1352 bool has_pcps
= false;
1353 for_each_populated_zone(zone
) {
1354 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1355 if (pcp
->pcp
.count
) {
1361 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1363 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1365 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1368 #ifdef CONFIG_HIBERNATION
1370 void mark_free_pages(struct zone
*zone
)
1372 unsigned long pfn
, max_zone_pfn
;
1373 unsigned long flags
;
1374 unsigned int order
, t
;
1375 struct list_head
*curr
;
1377 if (zone_is_empty(zone
))
1380 spin_lock_irqsave(&zone
->lock
, flags
);
1382 max_zone_pfn
= zone_end_pfn(zone
);
1383 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1384 if (pfn_valid(pfn
)) {
1385 struct page
*page
= pfn_to_page(pfn
);
1387 if (!swsusp_page_is_forbidden(page
))
1388 swsusp_unset_page_free(page
);
1391 for_each_migratetype_order(order
, t
) {
1392 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1395 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1396 for (i
= 0; i
< (1UL << order
); i
++)
1397 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1400 spin_unlock_irqrestore(&zone
->lock
, flags
);
1402 #endif /* CONFIG_PM */
1405 * Free a 0-order page
1406 * cold == true ? free a cold page : free a hot page
1408 void free_hot_cold_page(struct page
*page
, bool cold
)
1410 struct zone
*zone
= page_zone(page
);
1411 struct per_cpu_pages
*pcp
;
1412 unsigned long flags
;
1413 unsigned long pfn
= page_to_pfn(page
);
1416 if (!free_pages_prepare(page
, 0))
1419 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1420 set_freepage_migratetype(page
, migratetype
);
1421 local_irq_save(flags
);
1422 __count_vm_event(PGFREE
);
1425 * We only track unmovable, reclaimable and movable on pcp lists.
1426 * Free ISOLATE pages back to the allocator because they are being
1427 * offlined but treat RESERVE as movable pages so we can get those
1428 * areas back if necessary. Otherwise, we may have to free
1429 * excessively into the page allocator
1431 if (migratetype
>= MIGRATE_PCPTYPES
) {
1432 if (unlikely(is_migrate_isolate(migratetype
))) {
1433 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1436 migratetype
= MIGRATE_MOVABLE
;
1439 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1441 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1443 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1445 if (pcp
->count
>= pcp
->high
) {
1446 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1447 free_pcppages_bulk(zone
, batch
, pcp
);
1448 pcp
->count
-= batch
;
1452 local_irq_restore(flags
);
1456 * Free a list of 0-order pages
1458 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1460 struct page
*page
, *next
;
1462 list_for_each_entry_safe(page
, next
, list
, lru
) {
1463 trace_mm_page_free_batched(page
, cold
);
1464 free_hot_cold_page(page
, cold
);
1469 * split_page takes a non-compound higher-order page, and splits it into
1470 * n (1<<order) sub-pages: page[0..n]
1471 * Each sub-page must be freed individually.
1473 * Note: this is probably too low level an operation for use in drivers.
1474 * Please consult with lkml before using this in your driver.
1476 void split_page(struct page
*page
, unsigned int order
)
1480 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1481 VM_BUG_ON_PAGE(!page_count(page
), page
);
1483 #ifdef CONFIG_KMEMCHECK
1485 * Split shadow pages too, because free(page[0]) would
1486 * otherwise free the whole shadow.
1488 if (kmemcheck_page_is_tracked(page
))
1489 split_page(virt_to_page(page
[0].shadow
), order
);
1492 for (i
= 1; i
< (1 << order
); i
++)
1493 set_page_refcounted(page
+ i
);
1495 EXPORT_SYMBOL_GPL(split_page
);
1497 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1499 unsigned long watermark
;
1503 BUG_ON(!PageBuddy(page
));
1505 zone
= page_zone(page
);
1506 mt
= get_pageblock_migratetype(page
);
1508 if (!is_migrate_isolate(mt
)) {
1509 /* Obey watermarks as if the page was being allocated */
1510 watermark
= low_wmark_pages(zone
) + (1 << order
);
1511 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1514 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1517 /* Remove page from free list */
1518 list_del(&page
->lru
);
1519 zone
->free_area
[order
].nr_free
--;
1520 rmv_page_order(page
);
1522 /* Set the pageblock if the isolated page is at least a pageblock */
1523 if (order
>= pageblock_order
- 1) {
1524 struct page
*endpage
= page
+ (1 << order
) - 1;
1525 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1526 int mt
= get_pageblock_migratetype(page
);
1527 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1528 set_pageblock_migratetype(page
,
1533 return 1UL << order
;
1537 * Similar to split_page except the page is already free. As this is only
1538 * being used for migration, the migratetype of the block also changes.
1539 * As this is called with interrupts disabled, the caller is responsible
1540 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1543 * Note: this is probably too low level an operation for use in drivers.
1544 * Please consult with lkml before using this in your driver.
1546 int split_free_page(struct page
*page
)
1551 order
= page_order(page
);
1553 nr_pages
= __isolate_free_page(page
, order
);
1557 /* Split into individual pages */
1558 set_page_refcounted(page
);
1559 split_page(page
, order
);
1564 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1565 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1569 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1570 struct zone
*zone
, unsigned int order
,
1571 gfp_t gfp_flags
, int migratetype
)
1573 unsigned long flags
;
1575 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1578 if (likely(order
== 0)) {
1579 struct per_cpu_pages
*pcp
;
1580 struct list_head
*list
;
1582 local_irq_save(flags
);
1583 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1584 list
= &pcp
->lists
[migratetype
];
1585 if (list_empty(list
)) {
1586 pcp
->count
+= rmqueue_bulk(zone
, 0,
1589 if (unlikely(list_empty(list
)))
1594 page
= list_entry(list
->prev
, struct page
, lru
);
1596 page
= list_entry(list
->next
, struct page
, lru
);
1598 list_del(&page
->lru
);
1601 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1603 * __GFP_NOFAIL is not to be used in new code.
1605 * All __GFP_NOFAIL callers should be fixed so that they
1606 * properly detect and handle allocation failures.
1608 * We most definitely don't want callers attempting to
1609 * allocate greater than order-1 page units with
1612 WARN_ON_ONCE(order
> 1);
1614 spin_lock_irqsave(&zone
->lock
, flags
);
1615 page
= __rmqueue(zone
, order
, migratetype
);
1616 spin_unlock(&zone
->lock
);
1619 __mod_zone_freepage_state(zone
, -(1 << order
),
1620 get_freepage_migratetype(page
));
1623 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1624 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1625 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1626 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1628 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1629 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1630 local_irq_restore(flags
);
1632 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1633 if (prep_new_page(page
, order
, gfp_flags
))
1638 local_irq_restore(flags
);
1642 #ifdef CONFIG_FAIL_PAGE_ALLOC
1645 struct fault_attr attr
;
1647 u32 ignore_gfp_highmem
;
1648 u32 ignore_gfp_wait
;
1650 } fail_page_alloc
= {
1651 .attr
= FAULT_ATTR_INITIALIZER
,
1652 .ignore_gfp_wait
= 1,
1653 .ignore_gfp_highmem
= 1,
1657 static int __init
setup_fail_page_alloc(char *str
)
1659 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1661 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1663 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1665 if (order
< fail_page_alloc
.min_order
)
1667 if (gfp_mask
& __GFP_NOFAIL
)
1669 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1671 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1674 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1677 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1679 static int __init
fail_page_alloc_debugfs(void)
1681 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1684 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1685 &fail_page_alloc
.attr
);
1687 return PTR_ERR(dir
);
1689 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1690 &fail_page_alloc
.ignore_gfp_wait
))
1692 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1693 &fail_page_alloc
.ignore_gfp_highmem
))
1695 if (!debugfs_create_u32("min-order", mode
, dir
,
1696 &fail_page_alloc
.min_order
))
1701 debugfs_remove_recursive(dir
);
1706 late_initcall(fail_page_alloc_debugfs
);
1708 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1710 #else /* CONFIG_FAIL_PAGE_ALLOC */
1712 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1717 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1720 * Return true if free pages are above 'mark'. This takes into account the order
1721 * of the allocation.
1723 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1724 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1727 /* free_pages my go negative - that's OK */
1732 free_pages
-= (1 << order
) - 1;
1733 if (alloc_flags
& ALLOC_HIGH
)
1735 if (alloc_flags
& ALLOC_HARDER
)
1738 /* If allocation can't use CMA areas don't use free CMA pages */
1739 if (!(alloc_flags
& ALLOC_CMA
))
1740 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1743 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1745 for (o
= 0; o
< order
; o
++) {
1746 /* At the next order, this order's pages become unavailable */
1747 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1749 /* Require fewer higher order pages to be free */
1752 if (free_pages
<= min
)
1758 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1759 int classzone_idx
, int alloc_flags
)
1761 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1762 zone_page_state(z
, NR_FREE_PAGES
));
1765 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1766 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1768 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1770 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1771 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1773 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1779 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1780 * skip over zones that are not allowed by the cpuset, or that have
1781 * been recently (in last second) found to be nearly full. See further
1782 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1783 * that have to skip over a lot of full or unallowed zones.
1785 * If the zonelist cache is present in the passed zonelist, then
1786 * returns a pointer to the allowed node mask (either the current
1787 * tasks mems_allowed, or node_states[N_MEMORY].)
1789 * If the zonelist cache is not available for this zonelist, does
1790 * nothing and returns NULL.
1792 * If the fullzones BITMAP in the zonelist cache is stale (more than
1793 * a second since last zap'd) then we zap it out (clear its bits.)
1795 * We hold off even calling zlc_setup, until after we've checked the
1796 * first zone in the zonelist, on the theory that most allocations will
1797 * be satisfied from that first zone, so best to examine that zone as
1798 * quickly as we can.
1800 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1802 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1803 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1805 zlc
= zonelist
->zlcache_ptr
;
1809 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1810 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1811 zlc
->last_full_zap
= jiffies
;
1814 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1815 &cpuset_current_mems_allowed
:
1816 &node_states
[N_MEMORY
];
1817 return allowednodes
;
1821 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1822 * if it is worth looking at further for free memory:
1823 * 1) Check that the zone isn't thought to be full (doesn't have its
1824 * bit set in the zonelist_cache fullzones BITMAP).
1825 * 2) Check that the zones node (obtained from the zonelist_cache
1826 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1827 * Return true (non-zero) if zone is worth looking at further, or
1828 * else return false (zero) if it is not.
1830 * This check -ignores- the distinction between various watermarks,
1831 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1832 * found to be full for any variation of these watermarks, it will
1833 * be considered full for up to one second by all requests, unless
1834 * we are so low on memory on all allowed nodes that we are forced
1835 * into the second scan of the zonelist.
1837 * In the second scan we ignore this zonelist cache and exactly
1838 * apply the watermarks to all zones, even it is slower to do so.
1839 * We are low on memory in the second scan, and should leave no stone
1840 * unturned looking for a free page.
1842 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1843 nodemask_t
*allowednodes
)
1845 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1846 int i
; /* index of *z in zonelist zones */
1847 int n
; /* node that zone *z is on */
1849 zlc
= zonelist
->zlcache_ptr
;
1853 i
= z
- zonelist
->_zonerefs
;
1856 /* This zone is worth trying if it is allowed but not full */
1857 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1861 * Given 'z' scanning a zonelist, set the corresponding bit in
1862 * zlc->fullzones, so that subsequent attempts to allocate a page
1863 * from that zone don't waste time re-examining it.
1865 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1867 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1868 int i
; /* index of *z in zonelist zones */
1870 zlc
= zonelist
->zlcache_ptr
;
1874 i
= z
- zonelist
->_zonerefs
;
1876 set_bit(i
, zlc
->fullzones
);
1880 * clear all zones full, called after direct reclaim makes progress so that
1881 * a zone that was recently full is not skipped over for up to a second
1883 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1885 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1887 zlc
= zonelist
->zlcache_ptr
;
1891 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1894 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1896 return local_zone
->node
== zone
->node
;
1899 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1901 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1905 #else /* CONFIG_NUMA */
1907 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1912 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1913 nodemask_t
*allowednodes
)
1918 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1922 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1926 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1931 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1936 #endif /* CONFIG_NUMA */
1938 static void reset_alloc_batches(struct zone
*preferred_zone
)
1940 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
1943 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
1944 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
1945 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
1946 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1947 } while (zone
++ != preferred_zone
);
1951 * get_page_from_freelist goes through the zonelist trying to allocate
1954 static struct page
*
1955 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1956 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1957 struct zone
*preferred_zone
, int classzone_idx
, int migratetype
)
1960 struct page
*page
= NULL
;
1962 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1963 int zlc_active
= 0; /* set if using zonelist_cache */
1964 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1965 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
1966 (gfp_mask
& __GFP_WRITE
);
1967 int nr_fair_skipped
= 0;
1968 bool zonelist_rescan
;
1971 zonelist_rescan
= false;
1974 * Scan zonelist, looking for a zone with enough free.
1975 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1977 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1978 high_zoneidx
, nodemask
) {
1981 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1982 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1984 if (cpusets_enabled() &&
1985 (alloc_flags
& ALLOC_CPUSET
) &&
1986 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1989 * Distribute pages in proportion to the individual
1990 * zone size to ensure fair page aging. The zone a
1991 * page was allocated in should have no effect on the
1992 * time the page has in memory before being reclaimed.
1994 if (alloc_flags
& ALLOC_FAIR
) {
1995 if (!zone_local(preferred_zone
, zone
))
1997 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
2003 * When allocating a page cache page for writing, we
2004 * want to get it from a zone that is within its dirty
2005 * limit, such that no single zone holds more than its
2006 * proportional share of globally allowed dirty pages.
2007 * The dirty limits take into account the zone's
2008 * lowmem reserves and high watermark so that kswapd
2009 * should be able to balance it without having to
2010 * write pages from its LRU list.
2012 * This may look like it could increase pressure on
2013 * lower zones by failing allocations in higher zones
2014 * before they are full. But the pages that do spill
2015 * over are limited as the lower zones are protected
2016 * by this very same mechanism. It should not become
2017 * a practical burden to them.
2019 * XXX: For now, allow allocations to potentially
2020 * exceed the per-zone dirty limit in the slowpath
2021 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2022 * which is important when on a NUMA setup the allowed
2023 * zones are together not big enough to reach the
2024 * global limit. The proper fix for these situations
2025 * will require awareness of zones in the
2026 * dirty-throttling and the flusher threads.
2028 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2031 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2032 if (!zone_watermark_ok(zone
, order
, mark
,
2033 classzone_idx
, alloc_flags
)) {
2036 /* Checked here to keep the fast path fast */
2037 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2038 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2041 if (IS_ENABLED(CONFIG_NUMA
) &&
2042 !did_zlc_setup
&& nr_online_nodes
> 1) {
2044 * we do zlc_setup if there are multiple nodes
2045 * and before considering the first zone allowed
2048 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2053 if (zone_reclaim_mode
== 0 ||
2054 !zone_allows_reclaim(preferred_zone
, zone
))
2055 goto this_zone_full
;
2058 * As we may have just activated ZLC, check if the first
2059 * eligible zone has failed zone_reclaim recently.
2061 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2062 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2065 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2067 case ZONE_RECLAIM_NOSCAN
:
2070 case ZONE_RECLAIM_FULL
:
2071 /* scanned but unreclaimable */
2074 /* did we reclaim enough */
2075 if (zone_watermark_ok(zone
, order
, mark
,
2076 classzone_idx
, alloc_flags
))
2080 * Failed to reclaim enough to meet watermark.
2081 * Only mark the zone full if checking the min
2082 * watermark or if we failed to reclaim just
2083 * 1<<order pages or else the page allocator
2084 * fastpath will prematurely mark zones full
2085 * when the watermark is between the low and
2088 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2089 ret
== ZONE_RECLAIM_SOME
)
2090 goto this_zone_full
;
2097 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2098 gfp_mask
, migratetype
);
2102 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2103 zlc_mark_zone_full(zonelist
, z
);
2108 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2109 * necessary to allocate the page. The expectation is
2110 * that the caller is taking steps that will free more
2111 * memory. The caller should avoid the page being used
2112 * for !PFMEMALLOC purposes.
2114 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2119 * The first pass makes sure allocations are spread fairly within the
2120 * local node. However, the local node might have free pages left
2121 * after the fairness batches are exhausted, and remote zones haven't
2122 * even been considered yet. Try once more without fairness, and
2123 * include remote zones now, before entering the slowpath and waking
2124 * kswapd: prefer spilling to a remote zone over swapping locally.
2126 if (alloc_flags
& ALLOC_FAIR
) {
2127 alloc_flags
&= ~ALLOC_FAIR
;
2128 if (nr_fair_skipped
) {
2129 zonelist_rescan
= true;
2130 reset_alloc_batches(preferred_zone
);
2132 if (nr_online_nodes
> 1)
2133 zonelist_rescan
= true;
2136 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2137 /* Disable zlc cache for second zonelist scan */
2139 zonelist_rescan
= true;
2142 if (zonelist_rescan
)
2149 * Large machines with many possible nodes should not always dump per-node
2150 * meminfo in irq context.
2152 static inline bool should_suppress_show_mem(void)
2157 ret
= in_interrupt();
2162 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2163 DEFAULT_RATELIMIT_INTERVAL
,
2164 DEFAULT_RATELIMIT_BURST
);
2166 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2168 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2170 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2171 debug_guardpage_minorder() > 0)
2175 * This documents exceptions given to allocations in certain
2176 * contexts that are allowed to allocate outside current's set
2179 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2180 if (test_thread_flag(TIF_MEMDIE
) ||
2181 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2182 filter
&= ~SHOW_MEM_FILTER_NODES
;
2183 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2184 filter
&= ~SHOW_MEM_FILTER_NODES
;
2187 struct va_format vaf
;
2190 va_start(args
, fmt
);
2195 pr_warn("%pV", &vaf
);
2200 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2201 current
->comm
, order
, gfp_mask
);
2204 if (!should_suppress_show_mem())
2209 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2210 unsigned long did_some_progress
,
2211 unsigned long pages_reclaimed
)
2213 /* Do not loop if specifically requested */
2214 if (gfp_mask
& __GFP_NORETRY
)
2217 /* Always retry if specifically requested */
2218 if (gfp_mask
& __GFP_NOFAIL
)
2222 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2223 * making forward progress without invoking OOM. Suspend also disables
2224 * storage devices so kswapd will not help. Bail if we are suspending.
2226 if (!did_some_progress
&& pm_suspended_storage())
2230 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2231 * means __GFP_NOFAIL, but that may not be true in other
2234 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2238 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2239 * specified, then we retry until we no longer reclaim any pages
2240 * (above), or we've reclaimed an order of pages at least as
2241 * large as the allocation's order. In both cases, if the
2242 * allocation still fails, we stop retrying.
2244 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2250 static inline struct page
*
2251 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2252 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2253 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2254 int classzone_idx
, int migratetype
)
2258 /* Acquire the per-zone oom lock for each zone */
2259 if (!oom_zonelist_trylock(zonelist
, gfp_mask
)) {
2260 schedule_timeout_uninterruptible(1);
2265 * PM-freezer should be notified that there might be an OOM killer on
2266 * its way to kill and wake somebody up. This is too early and we might
2267 * end up not killing anything but false positives are acceptable.
2268 * See freeze_processes.
2273 * Go through the zonelist yet one more time, keep very high watermark
2274 * here, this is only to catch a parallel oom killing, we must fail if
2275 * we're still under heavy pressure.
2277 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2278 order
, zonelist
, high_zoneidx
,
2279 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2280 preferred_zone
, classzone_idx
, migratetype
);
2284 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2285 /* The OOM killer will not help higher order allocs */
2286 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2288 /* The OOM killer does not needlessly kill tasks for lowmem */
2289 if (high_zoneidx
< ZONE_NORMAL
)
2292 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2293 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2294 * The caller should handle page allocation failure by itself if
2295 * it specifies __GFP_THISNODE.
2296 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2298 if (gfp_mask
& __GFP_THISNODE
)
2301 /* Exhausted what can be done so it's blamo time */
2302 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2305 oom_zonelist_unlock(zonelist
, gfp_mask
);
2309 #ifdef CONFIG_COMPACTION
2310 /* Try memory compaction for high-order allocations before reclaim */
2311 static struct page
*
2312 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2313 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2314 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2315 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2316 int *contended_compaction
, bool *deferred_compaction
)
2318 struct zone
*last_compact_zone
= NULL
;
2319 unsigned long compact_result
;
2325 current
->flags
|= PF_MEMALLOC
;
2326 compact_result
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2328 contended_compaction
,
2329 &last_compact_zone
);
2330 current
->flags
&= ~PF_MEMALLOC
;
2332 switch (compact_result
) {
2333 case COMPACT_DEFERRED
:
2334 *deferred_compaction
= true;
2336 case COMPACT_SKIPPED
:
2343 * At least in one zone compaction wasn't deferred or skipped, so let's
2344 * count a compaction stall
2346 count_vm_event(COMPACTSTALL
);
2348 /* Page migration frees to the PCP lists but we want merging */
2349 drain_pages(get_cpu());
2352 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2353 order
, zonelist
, high_zoneidx
,
2354 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2355 preferred_zone
, classzone_idx
, migratetype
);
2358 struct zone
*zone
= page_zone(page
);
2360 zone
->compact_blockskip_flush
= false;
2361 compaction_defer_reset(zone
, order
, true);
2362 count_vm_event(COMPACTSUCCESS
);
2367 * last_compact_zone is where try_to_compact_pages thought allocation
2368 * should succeed, so it did not defer compaction. But here we know
2369 * that it didn't succeed, so we do the defer.
2371 if (last_compact_zone
&& mode
!= MIGRATE_ASYNC
)
2372 defer_compaction(last_compact_zone
, order
);
2375 * It's bad if compaction run occurs and fails. The most likely reason
2376 * is that pages exist, but not enough to satisfy watermarks.
2378 count_vm_event(COMPACTFAIL
);
2385 static inline struct page
*
2386 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2387 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2388 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2389 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2390 int *contended_compaction
, bool *deferred_compaction
)
2394 #endif /* CONFIG_COMPACTION */
2396 /* Perform direct synchronous page reclaim */
2398 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2399 nodemask_t
*nodemask
)
2401 struct reclaim_state reclaim_state
;
2406 /* We now go into synchronous reclaim */
2407 cpuset_memory_pressure_bump();
2408 current
->flags
|= PF_MEMALLOC
;
2409 lockdep_set_current_reclaim_state(gfp_mask
);
2410 reclaim_state
.reclaimed_slab
= 0;
2411 current
->reclaim_state
= &reclaim_state
;
2413 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2415 current
->reclaim_state
= NULL
;
2416 lockdep_clear_current_reclaim_state();
2417 current
->flags
&= ~PF_MEMALLOC
;
2424 /* The really slow allocator path where we enter direct reclaim */
2425 static inline struct page
*
2426 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2427 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2428 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2429 int classzone_idx
, int migratetype
, unsigned long *did_some_progress
)
2431 struct page
*page
= NULL
;
2432 bool drained
= false;
2434 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2436 if (unlikely(!(*did_some_progress
)))
2439 /* After successful reclaim, reconsider all zones for allocation */
2440 if (IS_ENABLED(CONFIG_NUMA
))
2441 zlc_clear_zones_full(zonelist
);
2444 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2445 zonelist
, high_zoneidx
,
2446 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2447 preferred_zone
, classzone_idx
,
2451 * If an allocation failed after direct reclaim, it could be because
2452 * pages are pinned on the per-cpu lists. Drain them and try again
2454 if (!page
&& !drained
) {
2464 * This is called in the allocator slow-path if the allocation request is of
2465 * sufficient urgency to ignore watermarks and take other desperate measures
2467 static inline struct page
*
2468 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2469 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2470 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2471 int classzone_idx
, int migratetype
)
2476 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2477 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2478 preferred_zone
, classzone_idx
, migratetype
);
2480 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2481 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2482 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2487 static void wake_all_kswapds(unsigned int order
,
2488 struct zonelist
*zonelist
,
2489 enum zone_type high_zoneidx
,
2490 struct zone
*preferred_zone
,
2491 nodemask_t
*nodemask
)
2496 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2497 high_zoneidx
, nodemask
)
2498 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2502 gfp_to_alloc_flags(gfp_t gfp_mask
)
2504 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2505 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2507 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2508 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2511 * The caller may dip into page reserves a bit more if the caller
2512 * cannot run direct reclaim, or if the caller has realtime scheduling
2513 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2514 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2516 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2520 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2521 * if it can't schedule.
2523 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2524 alloc_flags
|= ALLOC_HARDER
;
2526 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2527 * comment for __cpuset_node_allowed_softwall().
2529 alloc_flags
&= ~ALLOC_CPUSET
;
2530 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2531 alloc_flags
|= ALLOC_HARDER
;
2533 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2534 if (gfp_mask
& __GFP_MEMALLOC
)
2535 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2536 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2537 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2538 else if (!in_interrupt() &&
2539 ((current
->flags
& PF_MEMALLOC
) ||
2540 unlikely(test_thread_flag(TIF_MEMDIE
))))
2541 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2544 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2545 alloc_flags
|= ALLOC_CMA
;
2550 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2552 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2555 static inline struct page
*
2556 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2557 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2558 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2559 int classzone_idx
, int migratetype
)
2561 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2562 struct page
*page
= NULL
;
2564 unsigned long pages_reclaimed
= 0;
2565 unsigned long did_some_progress
;
2566 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2567 bool deferred_compaction
= false;
2568 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2571 * In the slowpath, we sanity check order to avoid ever trying to
2572 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2573 * be using allocators in order of preference for an area that is
2576 if (order
>= MAX_ORDER
) {
2577 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2582 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2583 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2584 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2585 * using a larger set of nodes after it has established that the
2586 * allowed per node queues are empty and that nodes are
2589 if (IS_ENABLED(CONFIG_NUMA
) &&
2590 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2594 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2595 wake_all_kswapds(order
, zonelist
, high_zoneidx
,
2596 preferred_zone
, nodemask
);
2599 * OK, we're below the kswapd watermark and have kicked background
2600 * reclaim. Now things get more complex, so set up alloc_flags according
2601 * to how we want to proceed.
2603 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2606 * Find the true preferred zone if the allocation is unconstrained by
2609 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
) {
2610 struct zoneref
*preferred_zoneref
;
2611 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2612 NULL
, &preferred_zone
);
2613 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2617 /* This is the last chance, in general, before the goto nopage. */
2618 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2619 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2620 preferred_zone
, classzone_idx
, migratetype
);
2624 /* Allocate without watermarks if the context allows */
2625 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2627 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2628 * the allocation is high priority and these type of
2629 * allocations are system rather than user orientated
2631 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2633 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2634 zonelist
, high_zoneidx
, nodemask
,
2635 preferred_zone
, classzone_idx
, migratetype
);
2641 /* Atomic allocations - we can't balance anything */
2644 * All existing users of the deprecated __GFP_NOFAIL are
2645 * blockable, so warn of any new users that actually allow this
2646 * type of allocation to fail.
2648 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2652 /* Avoid recursion of direct reclaim */
2653 if (current
->flags
& PF_MEMALLOC
)
2656 /* Avoid allocations with no watermarks from looping endlessly */
2657 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2661 * Try direct compaction. The first pass is asynchronous. Subsequent
2662 * attempts after direct reclaim are synchronous
2664 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2665 high_zoneidx
, nodemask
, alloc_flags
,
2667 classzone_idx
, migratetype
,
2668 migration_mode
, &contended_compaction
,
2669 &deferred_compaction
);
2673 /* Checks for THP-specific high-order allocations */
2674 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2676 * If compaction is deferred for high-order allocations, it is
2677 * because sync compaction recently failed. If this is the case
2678 * and the caller requested a THP allocation, we do not want
2679 * to heavily disrupt the system, so we fail the allocation
2680 * instead of entering direct reclaim.
2682 if (deferred_compaction
)
2686 * In all zones where compaction was attempted (and not
2687 * deferred or skipped), lock contention has been detected.
2688 * For THP allocation we do not want to disrupt the others
2689 * so we fallback to base pages instead.
2691 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2695 * If compaction was aborted due to need_resched(), we do not
2696 * want to further increase allocation latency, unless it is
2697 * khugepaged trying to collapse.
2699 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2700 && !(current
->flags
& PF_KTHREAD
))
2705 * It can become very expensive to allocate transparent hugepages at
2706 * fault, so use asynchronous memory compaction for THP unless it is
2707 * khugepaged trying to collapse.
2709 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2710 (current
->flags
& PF_KTHREAD
))
2711 migration_mode
= MIGRATE_SYNC_LIGHT
;
2713 /* Try direct reclaim and then allocating */
2714 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2715 zonelist
, high_zoneidx
,
2717 alloc_flags
, preferred_zone
,
2718 classzone_idx
, migratetype
,
2719 &did_some_progress
);
2724 * If we failed to make any progress reclaiming, then we are
2725 * running out of options and have to consider going OOM
2727 if (!did_some_progress
) {
2728 if (oom_gfp_allowed(gfp_mask
)) {
2729 if (oom_killer_disabled
)
2731 /* Coredumps can quickly deplete all memory reserves */
2732 if ((current
->flags
& PF_DUMPCORE
) &&
2733 !(gfp_mask
& __GFP_NOFAIL
))
2735 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2736 zonelist
, high_zoneidx
,
2737 nodemask
, preferred_zone
,
2738 classzone_idx
, migratetype
);
2742 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2744 * The oom killer is not called for high-order
2745 * allocations that may fail, so if no progress
2746 * is being made, there are no other options and
2747 * retrying is unlikely to help.
2749 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2752 * The oom killer is not called for lowmem
2753 * allocations to prevent needlessly killing
2756 if (high_zoneidx
< ZONE_NORMAL
)
2764 /* Check if we should retry the allocation */
2765 pages_reclaimed
+= did_some_progress
;
2766 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2768 /* Wait for some write requests to complete then retry */
2769 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2773 * High-order allocations do not necessarily loop after
2774 * direct reclaim and reclaim/compaction depends on compaction
2775 * being called after reclaim so call directly if necessary
2777 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2778 high_zoneidx
, nodemask
, alloc_flags
,
2780 classzone_idx
, migratetype
,
2781 migration_mode
, &contended_compaction
,
2782 &deferred_compaction
);
2788 warn_alloc_failed(gfp_mask
, order
, NULL
);
2791 if (kmemcheck_enabled
)
2792 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2798 * This is the 'heart' of the zoned buddy allocator.
2801 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2802 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2804 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2805 struct zone
*preferred_zone
;
2806 struct zoneref
*preferred_zoneref
;
2807 struct page
*page
= NULL
;
2808 int migratetype
= gfpflags_to_migratetype(gfp_mask
);
2809 unsigned int cpuset_mems_cookie
;
2810 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2813 gfp_mask
&= gfp_allowed_mask
;
2815 lockdep_trace_alloc(gfp_mask
);
2817 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2819 if (should_fail_alloc_page(gfp_mask
, order
))
2823 * Check the zones suitable for the gfp_mask contain at least one
2824 * valid zone. It's possible to have an empty zonelist as a result
2825 * of GFP_THISNODE and a memoryless node
2827 if (unlikely(!zonelist
->_zonerefs
->zone
))
2830 if (IS_ENABLED(CONFIG_CMA
) && migratetype
== MIGRATE_MOVABLE
)
2831 alloc_flags
|= ALLOC_CMA
;
2834 cpuset_mems_cookie
= read_mems_allowed_begin();
2836 /* The preferred zone is used for statistics later */
2837 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2838 nodemask
? : &cpuset_current_mems_allowed
,
2840 if (!preferred_zone
)
2842 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2844 /* First allocation attempt */
2845 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2846 zonelist
, high_zoneidx
, alloc_flags
,
2847 preferred_zone
, classzone_idx
, migratetype
);
2848 if (unlikely(!page
)) {
2850 * Runtime PM, block IO and its error handling path
2851 * can deadlock because I/O on the device might not
2854 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2855 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2856 zonelist
, high_zoneidx
, nodemask
,
2857 preferred_zone
, classzone_idx
, migratetype
);
2860 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2864 * When updating a task's mems_allowed, it is possible to race with
2865 * parallel threads in such a way that an allocation can fail while
2866 * the mask is being updated. If a page allocation is about to fail,
2867 * check if the cpuset changed during allocation and if so, retry.
2869 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2874 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2877 * Common helper functions.
2879 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2884 * __get_free_pages() returns a 32-bit address, which cannot represent
2887 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2889 page
= alloc_pages(gfp_mask
, order
);
2892 return (unsigned long) page_address(page
);
2894 EXPORT_SYMBOL(__get_free_pages
);
2896 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2898 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2900 EXPORT_SYMBOL(get_zeroed_page
);
2902 void __free_pages(struct page
*page
, unsigned int order
)
2904 if (put_page_testzero(page
)) {
2906 free_hot_cold_page(page
, false);
2908 __free_pages_ok(page
, order
);
2912 EXPORT_SYMBOL(__free_pages
);
2914 void free_pages(unsigned long addr
, unsigned int order
)
2917 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2918 __free_pages(virt_to_page((void *)addr
), order
);
2922 EXPORT_SYMBOL(free_pages
);
2925 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2926 * of the current memory cgroup.
2928 * It should be used when the caller would like to use kmalloc, but since the
2929 * allocation is large, it has to fall back to the page allocator.
2931 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2934 struct mem_cgroup
*memcg
= NULL
;
2936 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2938 page
= alloc_pages(gfp_mask
, order
);
2939 memcg_kmem_commit_charge(page
, memcg
, order
);
2943 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2946 struct mem_cgroup
*memcg
= NULL
;
2948 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2950 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2951 memcg_kmem_commit_charge(page
, memcg
, order
);
2956 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2959 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2961 memcg_kmem_uncharge_pages(page
, order
);
2962 __free_pages(page
, order
);
2965 void free_kmem_pages(unsigned long addr
, unsigned int order
)
2968 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2969 __free_kmem_pages(virt_to_page((void *)addr
), order
);
2973 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2976 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2977 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2979 split_page(virt_to_page((void *)addr
), order
);
2980 while (used
< alloc_end
) {
2985 return (void *)addr
;
2989 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2990 * @size: the number of bytes to allocate
2991 * @gfp_mask: GFP flags for the allocation
2993 * This function is similar to alloc_pages(), except that it allocates the
2994 * minimum number of pages to satisfy the request. alloc_pages() can only
2995 * allocate memory in power-of-two pages.
2997 * This function is also limited by MAX_ORDER.
2999 * Memory allocated by this function must be released by free_pages_exact().
3001 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
3003 unsigned int order
= get_order(size
);
3006 addr
= __get_free_pages(gfp_mask
, order
);
3007 return make_alloc_exact(addr
, order
, size
);
3009 EXPORT_SYMBOL(alloc_pages_exact
);
3012 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3014 * @nid: the preferred node ID where memory should be allocated
3015 * @size: the number of bytes to allocate
3016 * @gfp_mask: GFP flags for the allocation
3018 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3020 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3023 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3025 unsigned order
= get_order(size
);
3026 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3029 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3033 * free_pages_exact - release memory allocated via alloc_pages_exact()
3034 * @virt: the value returned by alloc_pages_exact.
3035 * @size: size of allocation, same value as passed to alloc_pages_exact().
3037 * Release the memory allocated by a previous call to alloc_pages_exact.
3039 void free_pages_exact(void *virt
, size_t size
)
3041 unsigned long addr
= (unsigned long)virt
;
3042 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3044 while (addr
< end
) {
3049 EXPORT_SYMBOL(free_pages_exact
);
3052 * nr_free_zone_pages - count number of pages beyond high watermark
3053 * @offset: The zone index of the highest zone
3055 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3056 * high watermark within all zones at or below a given zone index. For each
3057 * zone, the number of pages is calculated as:
3058 * managed_pages - high_pages
3060 static unsigned long nr_free_zone_pages(int offset
)
3065 /* Just pick one node, since fallback list is circular */
3066 unsigned long sum
= 0;
3068 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3070 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3071 unsigned long size
= zone
->managed_pages
;
3072 unsigned long high
= high_wmark_pages(zone
);
3081 * nr_free_buffer_pages - count number of pages beyond high watermark
3083 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3084 * watermark within ZONE_DMA and ZONE_NORMAL.
3086 unsigned long nr_free_buffer_pages(void)
3088 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3090 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3093 * nr_free_pagecache_pages - count number of pages beyond high watermark
3095 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3096 * high watermark within all zones.
3098 unsigned long nr_free_pagecache_pages(void)
3100 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3103 static inline void show_node(struct zone
*zone
)
3105 if (IS_ENABLED(CONFIG_NUMA
))
3106 printk("Node %d ", zone_to_nid(zone
));
3109 void si_meminfo(struct sysinfo
*val
)
3111 val
->totalram
= totalram_pages
;
3112 val
->sharedram
= global_page_state(NR_SHMEM
);
3113 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3114 val
->bufferram
= nr_blockdev_pages();
3115 val
->totalhigh
= totalhigh_pages
;
3116 val
->freehigh
= nr_free_highpages();
3117 val
->mem_unit
= PAGE_SIZE
;
3120 EXPORT_SYMBOL(si_meminfo
);
3123 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3125 int zone_type
; /* needs to be signed */
3126 unsigned long managed_pages
= 0;
3127 pg_data_t
*pgdat
= NODE_DATA(nid
);
3129 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3130 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3131 val
->totalram
= managed_pages
;
3132 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3133 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3134 #ifdef CONFIG_HIGHMEM
3135 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3136 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3142 val
->mem_unit
= PAGE_SIZE
;
3147 * Determine whether the node should be displayed or not, depending on whether
3148 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3150 bool skip_free_areas_node(unsigned int flags
, int nid
)
3153 unsigned int cpuset_mems_cookie
;
3155 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3159 cpuset_mems_cookie
= read_mems_allowed_begin();
3160 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3161 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3166 #define K(x) ((x) << (PAGE_SHIFT-10))
3168 static void show_migration_types(unsigned char type
)
3170 static const char types
[MIGRATE_TYPES
] = {
3171 [MIGRATE_UNMOVABLE
] = 'U',
3172 [MIGRATE_RECLAIMABLE
] = 'E',
3173 [MIGRATE_MOVABLE
] = 'M',
3174 [MIGRATE_RESERVE
] = 'R',
3176 [MIGRATE_CMA
] = 'C',
3178 #ifdef CONFIG_MEMORY_ISOLATION
3179 [MIGRATE_ISOLATE
] = 'I',
3182 char tmp
[MIGRATE_TYPES
+ 1];
3186 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3187 if (type
& (1 << i
))
3192 printk("(%s) ", tmp
);
3196 * Show free area list (used inside shift_scroll-lock stuff)
3197 * We also calculate the percentage fragmentation. We do this by counting the
3198 * memory on each free list with the exception of the first item on the list.
3199 * Suppresses nodes that are not allowed by current's cpuset if
3200 * SHOW_MEM_FILTER_NODES is passed.
3202 void show_free_areas(unsigned int filter
)
3207 for_each_populated_zone(zone
) {
3208 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3211 printk("%s per-cpu:\n", zone
->name
);
3213 for_each_online_cpu(cpu
) {
3214 struct per_cpu_pageset
*pageset
;
3216 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3218 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3219 cpu
, pageset
->pcp
.high
,
3220 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3224 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3225 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3227 " dirty:%lu writeback:%lu unstable:%lu\n"
3228 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3229 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3231 global_page_state(NR_ACTIVE_ANON
),
3232 global_page_state(NR_INACTIVE_ANON
),
3233 global_page_state(NR_ISOLATED_ANON
),
3234 global_page_state(NR_ACTIVE_FILE
),
3235 global_page_state(NR_INACTIVE_FILE
),
3236 global_page_state(NR_ISOLATED_FILE
),
3237 global_page_state(NR_UNEVICTABLE
),
3238 global_page_state(NR_FILE_DIRTY
),
3239 global_page_state(NR_WRITEBACK
),
3240 global_page_state(NR_UNSTABLE_NFS
),
3241 global_page_state(NR_FREE_PAGES
),
3242 global_page_state(NR_SLAB_RECLAIMABLE
),
3243 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3244 global_page_state(NR_FILE_MAPPED
),
3245 global_page_state(NR_SHMEM
),
3246 global_page_state(NR_PAGETABLE
),
3247 global_page_state(NR_BOUNCE
),
3248 global_page_state(NR_FREE_CMA_PAGES
));
3250 for_each_populated_zone(zone
) {
3253 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3261 " active_anon:%lukB"
3262 " inactive_anon:%lukB"
3263 " active_file:%lukB"
3264 " inactive_file:%lukB"
3265 " unevictable:%lukB"
3266 " isolated(anon):%lukB"
3267 " isolated(file):%lukB"
3275 " slab_reclaimable:%lukB"
3276 " slab_unreclaimable:%lukB"
3277 " kernel_stack:%lukB"
3282 " writeback_tmp:%lukB"
3283 " pages_scanned:%lu"
3284 " all_unreclaimable? %s"
3287 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3288 K(min_wmark_pages(zone
)),
3289 K(low_wmark_pages(zone
)),
3290 K(high_wmark_pages(zone
)),
3291 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3292 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3293 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3294 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3295 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3296 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3297 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3298 K(zone
->present_pages
),
3299 K(zone
->managed_pages
),
3300 K(zone_page_state(zone
, NR_MLOCK
)),
3301 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3302 K(zone_page_state(zone
, NR_WRITEBACK
)),
3303 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3304 K(zone_page_state(zone
, NR_SHMEM
)),
3305 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3306 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3307 zone_page_state(zone
, NR_KERNEL_STACK
) *
3309 K(zone_page_state(zone
, NR_PAGETABLE
)),
3310 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3311 K(zone_page_state(zone
, NR_BOUNCE
)),
3312 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3313 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3314 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3315 (!zone_reclaimable(zone
) ? "yes" : "no")
3317 printk("lowmem_reserve[]:");
3318 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3319 printk(" %ld", zone
->lowmem_reserve
[i
]);
3323 for_each_populated_zone(zone
) {
3324 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3325 unsigned char types
[MAX_ORDER
];
3327 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3330 printk("%s: ", zone
->name
);
3332 spin_lock_irqsave(&zone
->lock
, flags
);
3333 for (order
= 0; order
< MAX_ORDER
; order
++) {
3334 struct free_area
*area
= &zone
->free_area
[order
];
3337 nr
[order
] = area
->nr_free
;
3338 total
+= nr
[order
] << order
;
3341 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3342 if (!list_empty(&area
->free_list
[type
]))
3343 types
[order
] |= 1 << type
;
3346 spin_unlock_irqrestore(&zone
->lock
, flags
);
3347 for (order
= 0; order
< MAX_ORDER
; order
++) {
3348 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3350 show_migration_types(types
[order
]);
3352 printk("= %lukB\n", K(total
));
3355 hugetlb_show_meminfo();
3357 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3359 show_swap_cache_info();
3362 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3364 zoneref
->zone
= zone
;
3365 zoneref
->zone_idx
= zone_idx(zone
);
3369 * Builds allocation fallback zone lists.
3371 * Add all populated zones of a node to the zonelist.
3373 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3377 enum zone_type zone_type
= MAX_NR_ZONES
;
3381 zone
= pgdat
->node_zones
+ zone_type
;
3382 if (populated_zone(zone
)) {
3383 zoneref_set_zone(zone
,
3384 &zonelist
->_zonerefs
[nr_zones
++]);
3385 check_highest_zone(zone_type
);
3387 } while (zone_type
);
3395 * 0 = automatic detection of better ordering.
3396 * 1 = order by ([node] distance, -zonetype)
3397 * 2 = order by (-zonetype, [node] distance)
3399 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3400 * the same zonelist. So only NUMA can configure this param.
3402 #define ZONELIST_ORDER_DEFAULT 0
3403 #define ZONELIST_ORDER_NODE 1
3404 #define ZONELIST_ORDER_ZONE 2
3406 /* zonelist order in the kernel.
3407 * set_zonelist_order() will set this to NODE or ZONE.
3409 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3410 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3414 /* The value user specified ....changed by config */
3415 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3416 /* string for sysctl */
3417 #define NUMA_ZONELIST_ORDER_LEN 16
3418 char numa_zonelist_order
[16] = "default";
3421 * interface for configure zonelist ordering.
3422 * command line option "numa_zonelist_order"
3423 * = "[dD]efault - default, automatic configuration.
3424 * = "[nN]ode - order by node locality, then by zone within node
3425 * = "[zZ]one - order by zone, then by locality within zone
3428 static int __parse_numa_zonelist_order(char *s
)
3430 if (*s
== 'd' || *s
== 'D') {
3431 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3432 } else if (*s
== 'n' || *s
== 'N') {
3433 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3434 } else if (*s
== 'z' || *s
== 'Z') {
3435 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3438 "Ignoring invalid numa_zonelist_order value: "
3445 static __init
int setup_numa_zonelist_order(char *s
)
3452 ret
= __parse_numa_zonelist_order(s
);
3454 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3458 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3461 * sysctl handler for numa_zonelist_order
3463 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3464 void __user
*buffer
, size_t *length
,
3467 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3469 static DEFINE_MUTEX(zl_order_mutex
);
3471 mutex_lock(&zl_order_mutex
);
3473 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3477 strcpy(saved_string
, (char *)table
->data
);
3479 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3483 int oldval
= user_zonelist_order
;
3485 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3488 * bogus value. restore saved string
3490 strncpy((char *)table
->data
, saved_string
,
3491 NUMA_ZONELIST_ORDER_LEN
);
3492 user_zonelist_order
= oldval
;
3493 } else if (oldval
!= user_zonelist_order
) {
3494 mutex_lock(&zonelists_mutex
);
3495 build_all_zonelists(NULL
, NULL
);
3496 mutex_unlock(&zonelists_mutex
);
3500 mutex_unlock(&zl_order_mutex
);
3505 #define MAX_NODE_LOAD (nr_online_nodes)
3506 static int node_load
[MAX_NUMNODES
];
3509 * find_next_best_node - find the next node that should appear in a given node's fallback list
3510 * @node: node whose fallback list we're appending
3511 * @used_node_mask: nodemask_t of already used nodes
3513 * We use a number of factors to determine which is the next node that should
3514 * appear on a given node's fallback list. The node should not have appeared
3515 * already in @node's fallback list, and it should be the next closest node
3516 * according to the distance array (which contains arbitrary distance values
3517 * from each node to each node in the system), and should also prefer nodes
3518 * with no CPUs, since presumably they'll have very little allocation pressure
3519 * on them otherwise.
3520 * It returns -1 if no node is found.
3522 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3525 int min_val
= INT_MAX
;
3526 int best_node
= NUMA_NO_NODE
;
3527 const struct cpumask
*tmp
= cpumask_of_node(0);
3529 /* Use the local node if we haven't already */
3530 if (!node_isset(node
, *used_node_mask
)) {
3531 node_set(node
, *used_node_mask
);
3535 for_each_node_state(n
, N_MEMORY
) {
3537 /* Don't want a node to appear more than once */
3538 if (node_isset(n
, *used_node_mask
))
3541 /* Use the distance array to find the distance */
3542 val
= node_distance(node
, n
);
3544 /* Penalize nodes under us ("prefer the next node") */
3547 /* Give preference to headless and unused nodes */
3548 tmp
= cpumask_of_node(n
);
3549 if (!cpumask_empty(tmp
))
3550 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3552 /* Slight preference for less loaded node */
3553 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3554 val
+= node_load
[n
];
3556 if (val
< min_val
) {
3563 node_set(best_node
, *used_node_mask
);
3570 * Build zonelists ordered by node and zones within node.
3571 * This results in maximum locality--normal zone overflows into local
3572 * DMA zone, if any--but risks exhausting DMA zone.
3574 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3577 struct zonelist
*zonelist
;
3579 zonelist
= &pgdat
->node_zonelists
[0];
3580 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3582 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3583 zonelist
->_zonerefs
[j
].zone
= NULL
;
3584 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3588 * Build gfp_thisnode zonelists
3590 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3593 struct zonelist
*zonelist
;
3595 zonelist
= &pgdat
->node_zonelists
[1];
3596 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3597 zonelist
->_zonerefs
[j
].zone
= NULL
;
3598 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3602 * Build zonelists ordered by zone and nodes within zones.
3603 * This results in conserving DMA zone[s] until all Normal memory is
3604 * exhausted, but results in overflowing to remote node while memory
3605 * may still exist in local DMA zone.
3607 static int node_order
[MAX_NUMNODES
];
3609 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3612 int zone_type
; /* needs to be signed */
3614 struct zonelist
*zonelist
;
3616 zonelist
= &pgdat
->node_zonelists
[0];
3618 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3619 for (j
= 0; j
< nr_nodes
; j
++) {
3620 node
= node_order
[j
];
3621 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3622 if (populated_zone(z
)) {
3624 &zonelist
->_zonerefs
[pos
++]);
3625 check_highest_zone(zone_type
);
3629 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3630 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3633 #if defined(CONFIG_64BIT)
3635 * Devices that require DMA32/DMA are relatively rare and do not justify a
3636 * penalty to every machine in case the specialised case applies. Default
3637 * to Node-ordering on 64-bit NUMA machines
3639 static int default_zonelist_order(void)
3641 return ZONELIST_ORDER_NODE
;
3645 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3646 * by the kernel. If processes running on node 0 deplete the low memory zone
3647 * then reclaim will occur more frequency increasing stalls and potentially
3648 * be easier to OOM if a large percentage of the zone is under writeback or
3649 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3650 * Hence, default to zone ordering on 32-bit.
3652 static int default_zonelist_order(void)
3654 return ZONELIST_ORDER_ZONE
;
3656 #endif /* CONFIG_64BIT */
3658 static void set_zonelist_order(void)
3660 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3661 current_zonelist_order
= default_zonelist_order();
3663 current_zonelist_order
= user_zonelist_order
;
3666 static void build_zonelists(pg_data_t
*pgdat
)
3670 nodemask_t used_mask
;
3671 int local_node
, prev_node
;
3672 struct zonelist
*zonelist
;
3673 int order
= current_zonelist_order
;
3675 /* initialize zonelists */
3676 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3677 zonelist
= pgdat
->node_zonelists
+ i
;
3678 zonelist
->_zonerefs
[0].zone
= NULL
;
3679 zonelist
->_zonerefs
[0].zone_idx
= 0;
3682 /* NUMA-aware ordering of nodes */
3683 local_node
= pgdat
->node_id
;
3684 load
= nr_online_nodes
;
3685 prev_node
= local_node
;
3686 nodes_clear(used_mask
);
3688 memset(node_order
, 0, sizeof(node_order
));
3691 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3693 * We don't want to pressure a particular node.
3694 * So adding penalty to the first node in same
3695 * distance group to make it round-robin.
3697 if (node_distance(local_node
, node
) !=
3698 node_distance(local_node
, prev_node
))
3699 node_load
[node
] = load
;
3703 if (order
== ZONELIST_ORDER_NODE
)
3704 build_zonelists_in_node_order(pgdat
, node
);
3706 node_order
[j
++] = node
; /* remember order */
3709 if (order
== ZONELIST_ORDER_ZONE
) {
3710 /* calculate node order -- i.e., DMA last! */
3711 build_zonelists_in_zone_order(pgdat
, j
);
3714 build_thisnode_zonelists(pgdat
);
3717 /* Construct the zonelist performance cache - see further mmzone.h */
3718 static void build_zonelist_cache(pg_data_t
*pgdat
)
3720 struct zonelist
*zonelist
;
3721 struct zonelist_cache
*zlc
;
3724 zonelist
= &pgdat
->node_zonelists
[0];
3725 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3726 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3727 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3728 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3731 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3733 * Return node id of node used for "local" allocations.
3734 * I.e., first node id of first zone in arg node's generic zonelist.
3735 * Used for initializing percpu 'numa_mem', which is used primarily
3736 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3738 int local_memory_node(int node
)
3742 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3743 gfp_zone(GFP_KERNEL
),
3750 #else /* CONFIG_NUMA */
3752 static void set_zonelist_order(void)
3754 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3757 static void build_zonelists(pg_data_t
*pgdat
)
3759 int node
, local_node
;
3761 struct zonelist
*zonelist
;
3763 local_node
= pgdat
->node_id
;
3765 zonelist
= &pgdat
->node_zonelists
[0];
3766 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3769 * Now we build the zonelist so that it contains the zones
3770 * of all the other nodes.
3771 * We don't want to pressure a particular node, so when
3772 * building the zones for node N, we make sure that the
3773 * zones coming right after the local ones are those from
3774 * node N+1 (modulo N)
3776 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3777 if (!node_online(node
))
3779 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3781 for (node
= 0; node
< local_node
; node
++) {
3782 if (!node_online(node
))
3784 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3787 zonelist
->_zonerefs
[j
].zone
= NULL
;
3788 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3791 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3792 static void build_zonelist_cache(pg_data_t
*pgdat
)
3794 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3797 #endif /* CONFIG_NUMA */
3800 * Boot pageset table. One per cpu which is going to be used for all
3801 * zones and all nodes. The parameters will be set in such a way
3802 * that an item put on a list will immediately be handed over to
3803 * the buddy list. This is safe since pageset manipulation is done
3804 * with interrupts disabled.
3806 * The boot_pagesets must be kept even after bootup is complete for
3807 * unused processors and/or zones. They do play a role for bootstrapping
3808 * hotplugged processors.
3810 * zoneinfo_show() and maybe other functions do
3811 * not check if the processor is online before following the pageset pointer.
3812 * Other parts of the kernel may not check if the zone is available.
3814 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3815 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3816 static void setup_zone_pageset(struct zone
*zone
);
3819 * Global mutex to protect against size modification of zonelists
3820 * as well as to serialize pageset setup for the new populated zone.
3822 DEFINE_MUTEX(zonelists_mutex
);
3824 /* return values int ....just for stop_machine() */
3825 static int __build_all_zonelists(void *data
)
3829 pg_data_t
*self
= data
;
3832 memset(node_load
, 0, sizeof(node_load
));
3835 if (self
&& !node_online(self
->node_id
)) {
3836 build_zonelists(self
);
3837 build_zonelist_cache(self
);
3840 for_each_online_node(nid
) {
3841 pg_data_t
*pgdat
= NODE_DATA(nid
);
3843 build_zonelists(pgdat
);
3844 build_zonelist_cache(pgdat
);
3848 * Initialize the boot_pagesets that are going to be used
3849 * for bootstrapping processors. The real pagesets for
3850 * each zone will be allocated later when the per cpu
3851 * allocator is available.
3853 * boot_pagesets are used also for bootstrapping offline
3854 * cpus if the system is already booted because the pagesets
3855 * are needed to initialize allocators on a specific cpu too.
3856 * F.e. the percpu allocator needs the page allocator which
3857 * needs the percpu allocator in order to allocate its pagesets
3858 * (a chicken-egg dilemma).
3860 for_each_possible_cpu(cpu
) {
3861 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3863 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3865 * We now know the "local memory node" for each node--
3866 * i.e., the node of the first zone in the generic zonelist.
3867 * Set up numa_mem percpu variable for on-line cpus. During
3868 * boot, only the boot cpu should be on-line; we'll init the
3869 * secondary cpus' numa_mem as they come on-line. During
3870 * node/memory hotplug, we'll fixup all on-line cpus.
3872 if (cpu_online(cpu
))
3873 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3881 * Called with zonelists_mutex held always
3882 * unless system_state == SYSTEM_BOOTING.
3884 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3886 set_zonelist_order();
3888 if (system_state
== SYSTEM_BOOTING
) {
3889 __build_all_zonelists(NULL
);
3890 mminit_verify_zonelist();
3891 cpuset_init_current_mems_allowed();
3893 #ifdef CONFIG_MEMORY_HOTPLUG
3895 setup_zone_pageset(zone
);
3897 /* we have to stop all cpus to guarantee there is no user
3899 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3900 /* cpuset refresh routine should be here */
3902 vm_total_pages
= nr_free_pagecache_pages();
3904 * Disable grouping by mobility if the number of pages in the
3905 * system is too low to allow the mechanism to work. It would be
3906 * more accurate, but expensive to check per-zone. This check is
3907 * made on memory-hotadd so a system can start with mobility
3908 * disabled and enable it later
3910 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3911 page_group_by_mobility_disabled
= 1;
3913 page_group_by_mobility_disabled
= 0;
3915 printk("Built %i zonelists in %s order, mobility grouping %s. "
3916 "Total pages: %ld\n",
3918 zonelist_order_name
[current_zonelist_order
],
3919 page_group_by_mobility_disabled
? "off" : "on",
3922 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3927 * Helper functions to size the waitqueue hash table.
3928 * Essentially these want to choose hash table sizes sufficiently
3929 * large so that collisions trying to wait on pages are rare.
3930 * But in fact, the number of active page waitqueues on typical
3931 * systems is ridiculously low, less than 200. So this is even
3932 * conservative, even though it seems large.
3934 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3935 * waitqueues, i.e. the size of the waitq table given the number of pages.
3937 #define PAGES_PER_WAITQUEUE 256
3939 #ifndef CONFIG_MEMORY_HOTPLUG
3940 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3942 unsigned long size
= 1;
3944 pages
/= PAGES_PER_WAITQUEUE
;
3946 while (size
< pages
)
3950 * Once we have dozens or even hundreds of threads sleeping
3951 * on IO we've got bigger problems than wait queue collision.
3952 * Limit the size of the wait table to a reasonable size.
3954 size
= min(size
, 4096UL);
3956 return max(size
, 4UL);
3960 * A zone's size might be changed by hot-add, so it is not possible to determine
3961 * a suitable size for its wait_table. So we use the maximum size now.
3963 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3965 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3966 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3967 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3969 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3970 * or more by the traditional way. (See above). It equals:
3972 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3973 * ia64(16K page size) : = ( 8G + 4M)byte.
3974 * powerpc (64K page size) : = (32G +16M)byte.
3976 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3983 * This is an integer logarithm so that shifts can be used later
3984 * to extract the more random high bits from the multiplicative
3985 * hash function before the remainder is taken.
3987 static inline unsigned long wait_table_bits(unsigned long size
)
3993 * Check if a pageblock contains reserved pages
3995 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3999 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4000 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
4007 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4008 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4009 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4010 * higher will lead to a bigger reserve which will get freed as contiguous
4011 * blocks as reclaim kicks in
4013 static void setup_zone_migrate_reserve(struct zone
*zone
)
4015 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4017 unsigned long block_migratetype
;
4022 * Get the start pfn, end pfn and the number of blocks to reserve
4023 * We have to be careful to be aligned to pageblock_nr_pages to
4024 * make sure that we always check pfn_valid for the first page in
4027 start_pfn
= zone
->zone_start_pfn
;
4028 end_pfn
= zone_end_pfn(zone
);
4029 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4030 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4034 * Reserve blocks are generally in place to help high-order atomic
4035 * allocations that are short-lived. A min_free_kbytes value that
4036 * would result in more than 2 reserve blocks for atomic allocations
4037 * is assumed to be in place to help anti-fragmentation for the
4038 * future allocation of hugepages at runtime.
4040 reserve
= min(2, reserve
);
4041 old_reserve
= zone
->nr_migrate_reserve_block
;
4043 /* When memory hot-add, we almost always need to do nothing */
4044 if (reserve
== old_reserve
)
4046 zone
->nr_migrate_reserve_block
= reserve
;
4048 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4049 if (!pfn_valid(pfn
))
4051 page
= pfn_to_page(pfn
);
4053 /* Watch out for overlapping nodes */
4054 if (page_to_nid(page
) != zone_to_nid(zone
))
4057 block_migratetype
= get_pageblock_migratetype(page
);
4059 /* Only test what is necessary when the reserves are not met */
4062 * Blocks with reserved pages will never free, skip
4065 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4066 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4069 /* If this block is reserved, account for it */
4070 if (block_migratetype
== MIGRATE_RESERVE
) {
4075 /* Suitable for reserving if this block is movable */
4076 if (block_migratetype
== MIGRATE_MOVABLE
) {
4077 set_pageblock_migratetype(page
,
4079 move_freepages_block(zone
, page
,
4084 } else if (!old_reserve
) {
4086 * At boot time we don't need to scan the whole zone
4087 * for turning off MIGRATE_RESERVE.
4093 * If the reserve is met and this is a previous reserved block,
4096 if (block_migratetype
== MIGRATE_RESERVE
) {
4097 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4098 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4104 * Initially all pages are reserved - free ones are freed
4105 * up by free_all_bootmem() once the early boot process is
4106 * done. Non-atomic initialization, single-pass.
4108 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4109 unsigned long start_pfn
, enum memmap_context context
)
4112 unsigned long end_pfn
= start_pfn
+ size
;
4116 if (highest_memmap_pfn
< end_pfn
- 1)
4117 highest_memmap_pfn
= end_pfn
- 1;
4119 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4120 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4122 * There can be holes in boot-time mem_map[]s
4123 * handed to this function. They do not
4124 * exist on hotplugged memory.
4126 if (context
== MEMMAP_EARLY
) {
4127 if (!early_pfn_valid(pfn
))
4129 if (!early_pfn_in_nid(pfn
, nid
))
4132 page
= pfn_to_page(pfn
);
4133 set_page_links(page
, zone
, nid
, pfn
);
4134 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4135 init_page_count(page
);
4136 page_mapcount_reset(page
);
4137 page_cpupid_reset_last(page
);
4138 SetPageReserved(page
);
4140 * Mark the block movable so that blocks are reserved for
4141 * movable at startup. This will force kernel allocations
4142 * to reserve their blocks rather than leaking throughout
4143 * the address space during boot when many long-lived
4144 * kernel allocations are made. Later some blocks near
4145 * the start are marked MIGRATE_RESERVE by
4146 * setup_zone_migrate_reserve()
4148 * bitmap is created for zone's valid pfn range. but memmap
4149 * can be created for invalid pages (for alignment)
4150 * check here not to call set_pageblock_migratetype() against
4153 if ((z
->zone_start_pfn
<= pfn
)
4154 && (pfn
< zone_end_pfn(z
))
4155 && !(pfn
& (pageblock_nr_pages
- 1)))
4156 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4158 INIT_LIST_HEAD(&page
->lru
);
4159 #ifdef WANT_PAGE_VIRTUAL
4160 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4161 if (!is_highmem_idx(zone
))
4162 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4167 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4169 unsigned int order
, t
;
4170 for_each_migratetype_order(order
, t
) {
4171 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4172 zone
->free_area
[order
].nr_free
= 0;
4176 #ifndef __HAVE_ARCH_MEMMAP_INIT
4177 #define memmap_init(size, nid, zone, start_pfn) \
4178 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4181 static int zone_batchsize(struct zone
*zone
)
4187 * The per-cpu-pages pools are set to around 1000th of the
4188 * size of the zone. But no more than 1/2 of a meg.
4190 * OK, so we don't know how big the cache is. So guess.
4192 batch
= zone
->managed_pages
/ 1024;
4193 if (batch
* PAGE_SIZE
> 512 * 1024)
4194 batch
= (512 * 1024) / PAGE_SIZE
;
4195 batch
/= 4; /* We effectively *= 4 below */
4200 * Clamp the batch to a 2^n - 1 value. Having a power
4201 * of 2 value was found to be more likely to have
4202 * suboptimal cache aliasing properties in some cases.
4204 * For example if 2 tasks are alternately allocating
4205 * batches of pages, one task can end up with a lot
4206 * of pages of one half of the possible page colors
4207 * and the other with pages of the other colors.
4209 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4214 /* The deferral and batching of frees should be suppressed under NOMMU
4217 * The problem is that NOMMU needs to be able to allocate large chunks
4218 * of contiguous memory as there's no hardware page translation to
4219 * assemble apparent contiguous memory from discontiguous pages.
4221 * Queueing large contiguous runs of pages for batching, however,
4222 * causes the pages to actually be freed in smaller chunks. As there
4223 * can be a significant delay between the individual batches being
4224 * recycled, this leads to the once large chunks of space being
4225 * fragmented and becoming unavailable for high-order allocations.
4232 * pcp->high and pcp->batch values are related and dependent on one another:
4233 * ->batch must never be higher then ->high.
4234 * The following function updates them in a safe manner without read side
4237 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4238 * those fields changing asynchronously (acording the the above rule).
4240 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4241 * outside of boot time (or some other assurance that no concurrent updaters
4244 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4245 unsigned long batch
)
4247 /* start with a fail safe value for batch */
4251 /* Update high, then batch, in order */
4258 /* a companion to pageset_set_high() */
4259 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4261 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4264 static void pageset_init(struct per_cpu_pageset
*p
)
4266 struct per_cpu_pages
*pcp
;
4269 memset(p
, 0, sizeof(*p
));
4273 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4274 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4277 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4280 pageset_set_batch(p
, batch
);
4284 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4285 * to the value high for the pageset p.
4287 static void pageset_set_high(struct per_cpu_pageset
*p
,
4290 unsigned long batch
= max(1UL, high
/ 4);
4291 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4292 batch
= PAGE_SHIFT
* 8;
4294 pageset_update(&p
->pcp
, high
, batch
);
4297 static void pageset_set_high_and_batch(struct zone
*zone
,
4298 struct per_cpu_pageset
*pcp
)
4300 if (percpu_pagelist_fraction
)
4301 pageset_set_high(pcp
,
4302 (zone
->managed_pages
/
4303 percpu_pagelist_fraction
));
4305 pageset_set_batch(pcp
, zone_batchsize(zone
));
4308 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4310 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4313 pageset_set_high_and_batch(zone
, pcp
);
4316 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4319 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4320 for_each_possible_cpu(cpu
)
4321 zone_pageset_init(zone
, cpu
);
4325 * Allocate per cpu pagesets and initialize them.
4326 * Before this call only boot pagesets were available.
4328 void __init
setup_per_cpu_pageset(void)
4332 for_each_populated_zone(zone
)
4333 setup_zone_pageset(zone
);
4336 static noinline __init_refok
4337 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4343 * The per-page waitqueue mechanism uses hashed waitqueues
4346 zone
->wait_table_hash_nr_entries
=
4347 wait_table_hash_nr_entries(zone_size_pages
);
4348 zone
->wait_table_bits
=
4349 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4350 alloc_size
= zone
->wait_table_hash_nr_entries
4351 * sizeof(wait_queue_head_t
);
4353 if (!slab_is_available()) {
4354 zone
->wait_table
= (wait_queue_head_t
*)
4355 memblock_virt_alloc_node_nopanic(
4356 alloc_size
, zone
->zone_pgdat
->node_id
);
4359 * This case means that a zone whose size was 0 gets new memory
4360 * via memory hot-add.
4361 * But it may be the case that a new node was hot-added. In
4362 * this case vmalloc() will not be able to use this new node's
4363 * memory - this wait_table must be initialized to use this new
4364 * node itself as well.
4365 * To use this new node's memory, further consideration will be
4368 zone
->wait_table
= vmalloc(alloc_size
);
4370 if (!zone
->wait_table
)
4373 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4374 init_waitqueue_head(zone
->wait_table
+ i
);
4379 static __meminit
void zone_pcp_init(struct zone
*zone
)
4382 * per cpu subsystem is not up at this point. The following code
4383 * relies on the ability of the linker to provide the
4384 * offset of a (static) per cpu variable into the per cpu area.
4386 zone
->pageset
= &boot_pageset
;
4388 if (populated_zone(zone
))
4389 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4390 zone
->name
, zone
->present_pages
,
4391 zone_batchsize(zone
));
4394 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4395 unsigned long zone_start_pfn
,
4397 enum memmap_context context
)
4399 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4401 ret
= zone_wait_table_init(zone
, size
);
4404 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4406 zone
->zone_start_pfn
= zone_start_pfn
;
4408 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4409 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4411 (unsigned long)zone_idx(zone
),
4412 zone_start_pfn
, (zone_start_pfn
+ size
));
4414 zone_init_free_lists(zone
);
4419 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4420 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4422 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4424 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4426 unsigned long start_pfn
, end_pfn
;
4429 * NOTE: The following SMP-unsafe globals are only used early in boot
4430 * when the kernel is running single-threaded.
4432 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4433 static int __meminitdata last_nid
;
4435 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4438 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4440 last_start_pfn
= start_pfn
;
4441 last_end_pfn
= end_pfn
;
4447 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4449 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4453 nid
= __early_pfn_to_nid(pfn
);
4456 /* just returns 0 */
4460 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4461 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4465 nid
= __early_pfn_to_nid(pfn
);
4466 if (nid
>= 0 && nid
!= node
)
4473 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4474 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4475 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4477 * If an architecture guarantees that all ranges registered contain no holes
4478 * and may be freed, this this function may be used instead of calling
4479 * memblock_free_early_nid() manually.
4481 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4483 unsigned long start_pfn
, end_pfn
;
4486 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4487 start_pfn
= min(start_pfn
, max_low_pfn
);
4488 end_pfn
= min(end_pfn
, max_low_pfn
);
4490 if (start_pfn
< end_pfn
)
4491 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4492 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4498 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4499 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4501 * If an architecture guarantees that all ranges registered contain no holes and may
4502 * be freed, this function may be used instead of calling memory_present() manually.
4504 void __init
sparse_memory_present_with_active_regions(int nid
)
4506 unsigned long start_pfn
, end_pfn
;
4509 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4510 memory_present(this_nid
, start_pfn
, end_pfn
);
4514 * get_pfn_range_for_nid - Return the start and end page frames for a node
4515 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4516 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4517 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4519 * It returns the start and end page frame of a node based on information
4520 * provided by memblock_set_node(). If called for a node
4521 * with no available memory, a warning is printed and the start and end
4524 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4525 unsigned long *start_pfn
, unsigned long *end_pfn
)
4527 unsigned long this_start_pfn
, this_end_pfn
;
4533 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4534 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4535 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4538 if (*start_pfn
== -1UL)
4543 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4544 * assumption is made that zones within a node are ordered in monotonic
4545 * increasing memory addresses so that the "highest" populated zone is used
4547 static void __init
find_usable_zone_for_movable(void)
4550 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4551 if (zone_index
== ZONE_MOVABLE
)
4554 if (arch_zone_highest_possible_pfn
[zone_index
] >
4555 arch_zone_lowest_possible_pfn
[zone_index
])
4559 VM_BUG_ON(zone_index
== -1);
4560 movable_zone
= zone_index
;
4564 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4565 * because it is sized independent of architecture. Unlike the other zones,
4566 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4567 * in each node depending on the size of each node and how evenly kernelcore
4568 * is distributed. This helper function adjusts the zone ranges
4569 * provided by the architecture for a given node by using the end of the
4570 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4571 * zones within a node are in order of monotonic increases memory addresses
4573 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4574 unsigned long zone_type
,
4575 unsigned long node_start_pfn
,
4576 unsigned long node_end_pfn
,
4577 unsigned long *zone_start_pfn
,
4578 unsigned long *zone_end_pfn
)
4580 /* Only adjust if ZONE_MOVABLE is on this node */
4581 if (zone_movable_pfn
[nid
]) {
4582 /* Size ZONE_MOVABLE */
4583 if (zone_type
== ZONE_MOVABLE
) {
4584 *zone_start_pfn
= zone_movable_pfn
[nid
];
4585 *zone_end_pfn
= min(node_end_pfn
,
4586 arch_zone_highest_possible_pfn
[movable_zone
]);
4588 /* Adjust for ZONE_MOVABLE starting within this range */
4589 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4590 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4591 *zone_end_pfn
= zone_movable_pfn
[nid
];
4593 /* Check if this whole range is within ZONE_MOVABLE */
4594 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4595 *zone_start_pfn
= *zone_end_pfn
;
4600 * Return the number of pages a zone spans in a node, including holes
4601 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4603 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4604 unsigned long zone_type
,
4605 unsigned long node_start_pfn
,
4606 unsigned long node_end_pfn
,
4607 unsigned long *ignored
)
4609 unsigned long zone_start_pfn
, zone_end_pfn
;
4611 /* Get the start and end of the zone */
4612 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4613 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4614 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4615 node_start_pfn
, node_end_pfn
,
4616 &zone_start_pfn
, &zone_end_pfn
);
4618 /* Check that this node has pages within the zone's required range */
4619 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4622 /* Move the zone boundaries inside the node if necessary */
4623 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4624 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4626 /* Return the spanned pages */
4627 return zone_end_pfn
- zone_start_pfn
;
4631 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4632 * then all holes in the requested range will be accounted for.
4634 unsigned long __meminit
__absent_pages_in_range(int nid
,
4635 unsigned long range_start_pfn
,
4636 unsigned long range_end_pfn
)
4638 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4639 unsigned long start_pfn
, end_pfn
;
4642 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4643 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4644 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4645 nr_absent
-= end_pfn
- start_pfn
;
4651 * absent_pages_in_range - Return number of page frames in holes within a range
4652 * @start_pfn: The start PFN to start searching for holes
4653 * @end_pfn: The end PFN to stop searching for holes
4655 * It returns the number of pages frames in memory holes within a range.
4657 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4658 unsigned long end_pfn
)
4660 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4663 /* Return the number of page frames in holes in a zone on a node */
4664 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4665 unsigned long zone_type
,
4666 unsigned long node_start_pfn
,
4667 unsigned long node_end_pfn
,
4668 unsigned long *ignored
)
4670 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4671 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4672 unsigned long zone_start_pfn
, zone_end_pfn
;
4674 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4675 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4677 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4678 node_start_pfn
, node_end_pfn
,
4679 &zone_start_pfn
, &zone_end_pfn
);
4680 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4683 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4684 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4685 unsigned long zone_type
,
4686 unsigned long node_start_pfn
,
4687 unsigned long node_end_pfn
,
4688 unsigned long *zones_size
)
4690 return zones_size
[zone_type
];
4693 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4694 unsigned long zone_type
,
4695 unsigned long node_start_pfn
,
4696 unsigned long node_end_pfn
,
4697 unsigned long *zholes_size
)
4702 return zholes_size
[zone_type
];
4705 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4707 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4708 unsigned long node_start_pfn
,
4709 unsigned long node_end_pfn
,
4710 unsigned long *zones_size
,
4711 unsigned long *zholes_size
)
4713 unsigned long realtotalpages
, totalpages
= 0;
4716 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4717 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4721 pgdat
->node_spanned_pages
= totalpages
;
4723 realtotalpages
= totalpages
;
4724 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4726 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4727 node_start_pfn
, node_end_pfn
,
4729 pgdat
->node_present_pages
= realtotalpages
;
4730 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4734 #ifndef CONFIG_SPARSEMEM
4736 * Calculate the size of the zone->blockflags rounded to an unsigned long
4737 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4738 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4739 * round what is now in bits to nearest long in bits, then return it in
4742 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4744 unsigned long usemapsize
;
4746 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4747 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4748 usemapsize
= usemapsize
>> pageblock_order
;
4749 usemapsize
*= NR_PAGEBLOCK_BITS
;
4750 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4752 return usemapsize
/ 8;
4755 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4757 unsigned long zone_start_pfn
,
4758 unsigned long zonesize
)
4760 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4761 zone
->pageblock_flags
= NULL
;
4763 zone
->pageblock_flags
=
4764 memblock_virt_alloc_node_nopanic(usemapsize
,
4768 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4769 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4770 #endif /* CONFIG_SPARSEMEM */
4772 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4774 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4775 void __paginginit
set_pageblock_order(void)
4779 /* Check that pageblock_nr_pages has not already been setup */
4780 if (pageblock_order
)
4783 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4784 order
= HUGETLB_PAGE_ORDER
;
4786 order
= MAX_ORDER
- 1;
4789 * Assume the largest contiguous order of interest is a huge page.
4790 * This value may be variable depending on boot parameters on IA64 and
4793 pageblock_order
= order
;
4795 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4798 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4799 * is unused as pageblock_order is set at compile-time. See
4800 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4803 void __paginginit
set_pageblock_order(void)
4807 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4809 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4810 unsigned long present_pages
)
4812 unsigned long pages
= spanned_pages
;
4815 * Provide a more accurate estimation if there are holes within
4816 * the zone and SPARSEMEM is in use. If there are holes within the
4817 * zone, each populated memory region may cost us one or two extra
4818 * memmap pages due to alignment because memmap pages for each
4819 * populated regions may not naturally algined on page boundary.
4820 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4822 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4823 IS_ENABLED(CONFIG_SPARSEMEM
))
4824 pages
= present_pages
;
4826 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4830 * Set up the zone data structures:
4831 * - mark all pages reserved
4832 * - mark all memory queues empty
4833 * - clear the memory bitmaps
4835 * NOTE: pgdat should get zeroed by caller.
4837 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4838 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4839 unsigned long *zones_size
, unsigned long *zholes_size
)
4842 int nid
= pgdat
->node_id
;
4843 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4846 pgdat_resize_init(pgdat
);
4847 #ifdef CONFIG_NUMA_BALANCING
4848 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4849 pgdat
->numabalancing_migrate_nr_pages
= 0;
4850 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4852 init_waitqueue_head(&pgdat
->kswapd_wait
);
4853 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4854 pgdat_page_cgroup_init(pgdat
);
4856 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4857 struct zone
*zone
= pgdat
->node_zones
+ j
;
4858 unsigned long size
, realsize
, freesize
, memmap_pages
;
4860 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4861 node_end_pfn
, zones_size
);
4862 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4868 * Adjust freesize so that it accounts for how much memory
4869 * is used by this zone for memmap. This affects the watermark
4870 * and per-cpu initialisations
4872 memmap_pages
= calc_memmap_size(size
, realsize
);
4873 if (freesize
>= memmap_pages
) {
4874 freesize
-= memmap_pages
;
4877 " %s zone: %lu pages used for memmap\n",
4878 zone_names
[j
], memmap_pages
);
4881 " %s zone: %lu pages exceeds freesize %lu\n",
4882 zone_names
[j
], memmap_pages
, freesize
);
4884 /* Account for reserved pages */
4885 if (j
== 0 && freesize
> dma_reserve
) {
4886 freesize
-= dma_reserve
;
4887 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4888 zone_names
[0], dma_reserve
);
4891 if (!is_highmem_idx(j
))
4892 nr_kernel_pages
+= freesize
;
4893 /* Charge for highmem memmap if there are enough kernel pages */
4894 else if (nr_kernel_pages
> memmap_pages
* 2)
4895 nr_kernel_pages
-= memmap_pages
;
4896 nr_all_pages
+= freesize
;
4898 zone
->spanned_pages
= size
;
4899 zone
->present_pages
= realsize
;
4901 * Set an approximate value for lowmem here, it will be adjusted
4902 * when the bootmem allocator frees pages into the buddy system.
4903 * And all highmem pages will be managed by the buddy system.
4905 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4908 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4910 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4912 zone
->name
= zone_names
[j
];
4913 spin_lock_init(&zone
->lock
);
4914 spin_lock_init(&zone
->lru_lock
);
4915 zone_seqlock_init(zone
);
4916 zone
->zone_pgdat
= pgdat
;
4917 zone_pcp_init(zone
);
4919 /* For bootup, initialized properly in watermark setup */
4920 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4922 lruvec_init(&zone
->lruvec
);
4926 set_pageblock_order();
4927 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4928 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4929 size
, MEMMAP_EARLY
);
4931 memmap_init(size
, nid
, j
, zone_start_pfn
);
4932 zone_start_pfn
+= size
;
4936 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4938 /* Skip empty nodes */
4939 if (!pgdat
->node_spanned_pages
)
4942 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4943 /* ia64 gets its own node_mem_map, before this, without bootmem */
4944 if (!pgdat
->node_mem_map
) {
4945 unsigned long size
, start
, end
;
4949 * The zone's endpoints aren't required to be MAX_ORDER
4950 * aligned but the node_mem_map endpoints must be in order
4951 * for the buddy allocator to function correctly.
4953 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4954 end
= pgdat_end_pfn(pgdat
);
4955 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4956 size
= (end
- start
) * sizeof(struct page
);
4957 map
= alloc_remap(pgdat
->node_id
, size
);
4959 map
= memblock_virt_alloc_node_nopanic(size
,
4961 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4963 #ifndef CONFIG_NEED_MULTIPLE_NODES
4965 * With no DISCONTIG, the global mem_map is just set as node 0's
4967 if (pgdat
== NODE_DATA(0)) {
4968 mem_map
= NODE_DATA(0)->node_mem_map
;
4969 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4970 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4971 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4972 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4975 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4978 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4979 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4981 pg_data_t
*pgdat
= NODE_DATA(nid
);
4982 unsigned long start_pfn
= 0;
4983 unsigned long end_pfn
= 0;
4985 /* pg_data_t should be reset to zero when it's allocated */
4986 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4988 pgdat
->node_id
= nid
;
4989 pgdat
->node_start_pfn
= node_start_pfn
;
4990 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4991 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4992 printk(KERN_INFO
"Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid
,
4993 (u64
) start_pfn
<< PAGE_SHIFT
, (u64
) (end_pfn
<< PAGE_SHIFT
) - 1);
4995 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
4996 zones_size
, zholes_size
);
4998 alloc_node_mem_map(pgdat
);
4999 #ifdef CONFIG_FLAT_NODE_MEM_MAP
5000 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
5001 nid
, (unsigned long)pgdat
,
5002 (unsigned long)pgdat
->node_mem_map
);
5005 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
5006 zones_size
, zholes_size
);
5009 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5011 #if MAX_NUMNODES > 1
5013 * Figure out the number of possible node ids.
5015 void __init
setup_nr_node_ids(void)
5018 unsigned int highest
= 0;
5020 for_each_node_mask(node
, node_possible_map
)
5022 nr_node_ids
= highest
+ 1;
5027 * node_map_pfn_alignment - determine the maximum internode alignment
5029 * This function should be called after node map is populated and sorted.
5030 * It calculates the maximum power of two alignment which can distinguish
5033 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5034 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5035 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5036 * shifted, 1GiB is enough and this function will indicate so.
5038 * This is used to test whether pfn -> nid mapping of the chosen memory
5039 * model has fine enough granularity to avoid incorrect mapping for the
5040 * populated node map.
5042 * Returns the determined alignment in pfn's. 0 if there is no alignment
5043 * requirement (single node).
5045 unsigned long __init
node_map_pfn_alignment(void)
5047 unsigned long accl_mask
= 0, last_end
= 0;
5048 unsigned long start
, end
, mask
;
5052 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5053 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5060 * Start with a mask granular enough to pin-point to the
5061 * start pfn and tick off bits one-by-one until it becomes
5062 * too coarse to separate the current node from the last.
5064 mask
= ~((1 << __ffs(start
)) - 1);
5065 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5068 /* accumulate all internode masks */
5072 /* convert mask to number of pages */
5073 return ~accl_mask
+ 1;
5076 /* Find the lowest pfn for a node */
5077 static unsigned long __init
find_min_pfn_for_node(int nid
)
5079 unsigned long min_pfn
= ULONG_MAX
;
5080 unsigned long start_pfn
;
5083 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5084 min_pfn
= min(min_pfn
, start_pfn
);
5086 if (min_pfn
== ULONG_MAX
) {
5088 "Could not find start_pfn for node %d\n", nid
);
5096 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5098 * It returns the minimum PFN based on information provided via
5099 * memblock_set_node().
5101 unsigned long __init
find_min_pfn_with_active_regions(void)
5103 return find_min_pfn_for_node(MAX_NUMNODES
);
5107 * early_calculate_totalpages()
5108 * Sum pages in active regions for movable zone.
5109 * Populate N_MEMORY for calculating usable_nodes.
5111 static unsigned long __init
early_calculate_totalpages(void)
5113 unsigned long totalpages
= 0;
5114 unsigned long start_pfn
, end_pfn
;
5117 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5118 unsigned long pages
= end_pfn
- start_pfn
;
5120 totalpages
+= pages
;
5122 node_set_state(nid
, N_MEMORY
);
5128 * Find the PFN the Movable zone begins in each node. Kernel memory
5129 * is spread evenly between nodes as long as the nodes have enough
5130 * memory. When they don't, some nodes will have more kernelcore than
5133 static void __init
find_zone_movable_pfns_for_nodes(void)
5136 unsigned long usable_startpfn
;
5137 unsigned long kernelcore_node
, kernelcore_remaining
;
5138 /* save the state before borrow the nodemask */
5139 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5140 unsigned long totalpages
= early_calculate_totalpages();
5141 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5142 struct memblock_region
*r
;
5144 /* Need to find movable_zone earlier when movable_node is specified. */
5145 find_usable_zone_for_movable();
5148 * If movable_node is specified, ignore kernelcore and movablecore
5151 if (movable_node_is_enabled()) {
5152 for_each_memblock(memory
, r
) {
5153 if (!memblock_is_hotpluggable(r
))
5158 usable_startpfn
= PFN_DOWN(r
->base
);
5159 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5160 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5168 * If movablecore=nn[KMG] was specified, calculate what size of
5169 * kernelcore that corresponds so that memory usable for
5170 * any allocation type is evenly spread. If both kernelcore
5171 * and movablecore are specified, then the value of kernelcore
5172 * will be used for required_kernelcore if it's greater than
5173 * what movablecore would have allowed.
5175 if (required_movablecore
) {
5176 unsigned long corepages
;
5179 * Round-up so that ZONE_MOVABLE is at least as large as what
5180 * was requested by the user
5182 required_movablecore
=
5183 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5184 corepages
= totalpages
- required_movablecore
;
5186 required_kernelcore
= max(required_kernelcore
, corepages
);
5189 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5190 if (!required_kernelcore
)
5193 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5194 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5197 /* Spread kernelcore memory as evenly as possible throughout nodes */
5198 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5199 for_each_node_state(nid
, N_MEMORY
) {
5200 unsigned long start_pfn
, end_pfn
;
5203 * Recalculate kernelcore_node if the division per node
5204 * now exceeds what is necessary to satisfy the requested
5205 * amount of memory for the kernel
5207 if (required_kernelcore
< kernelcore_node
)
5208 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5211 * As the map is walked, we track how much memory is usable
5212 * by the kernel using kernelcore_remaining. When it is
5213 * 0, the rest of the node is usable by ZONE_MOVABLE
5215 kernelcore_remaining
= kernelcore_node
;
5217 /* Go through each range of PFNs within this node */
5218 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5219 unsigned long size_pages
;
5221 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5222 if (start_pfn
>= end_pfn
)
5225 /* Account for what is only usable for kernelcore */
5226 if (start_pfn
< usable_startpfn
) {
5227 unsigned long kernel_pages
;
5228 kernel_pages
= min(end_pfn
, usable_startpfn
)
5231 kernelcore_remaining
-= min(kernel_pages
,
5232 kernelcore_remaining
);
5233 required_kernelcore
-= min(kernel_pages
,
5234 required_kernelcore
);
5236 /* Continue if range is now fully accounted */
5237 if (end_pfn
<= usable_startpfn
) {
5240 * Push zone_movable_pfn to the end so
5241 * that if we have to rebalance
5242 * kernelcore across nodes, we will
5243 * not double account here
5245 zone_movable_pfn
[nid
] = end_pfn
;
5248 start_pfn
= usable_startpfn
;
5252 * The usable PFN range for ZONE_MOVABLE is from
5253 * start_pfn->end_pfn. Calculate size_pages as the
5254 * number of pages used as kernelcore
5256 size_pages
= end_pfn
- start_pfn
;
5257 if (size_pages
> kernelcore_remaining
)
5258 size_pages
= kernelcore_remaining
;
5259 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5262 * Some kernelcore has been met, update counts and
5263 * break if the kernelcore for this node has been
5266 required_kernelcore
-= min(required_kernelcore
,
5268 kernelcore_remaining
-= size_pages
;
5269 if (!kernelcore_remaining
)
5275 * If there is still required_kernelcore, we do another pass with one
5276 * less node in the count. This will push zone_movable_pfn[nid] further
5277 * along on the nodes that still have memory until kernelcore is
5281 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5285 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5286 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5287 zone_movable_pfn
[nid
] =
5288 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5291 /* restore the node_state */
5292 node_states
[N_MEMORY
] = saved_node_state
;
5295 /* Any regular or high memory on that node ? */
5296 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5298 enum zone_type zone_type
;
5300 if (N_MEMORY
== N_NORMAL_MEMORY
)
5303 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5304 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5305 if (populated_zone(zone
)) {
5306 node_set_state(nid
, N_HIGH_MEMORY
);
5307 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5308 zone_type
<= ZONE_NORMAL
)
5309 node_set_state(nid
, N_NORMAL_MEMORY
);
5316 * free_area_init_nodes - Initialise all pg_data_t and zone data
5317 * @max_zone_pfn: an array of max PFNs for each zone
5319 * This will call free_area_init_node() for each active node in the system.
5320 * Using the page ranges provided by memblock_set_node(), the size of each
5321 * zone in each node and their holes is calculated. If the maximum PFN
5322 * between two adjacent zones match, it is assumed that the zone is empty.
5323 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5324 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5325 * starts where the previous one ended. For example, ZONE_DMA32 starts
5326 * at arch_max_dma_pfn.
5328 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5330 unsigned long start_pfn
, end_pfn
;
5333 /* Record where the zone boundaries are */
5334 memset(arch_zone_lowest_possible_pfn
, 0,
5335 sizeof(arch_zone_lowest_possible_pfn
));
5336 memset(arch_zone_highest_possible_pfn
, 0,
5337 sizeof(arch_zone_highest_possible_pfn
));
5338 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5339 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5340 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5341 if (i
== ZONE_MOVABLE
)
5343 arch_zone_lowest_possible_pfn
[i
] =
5344 arch_zone_highest_possible_pfn
[i
-1];
5345 arch_zone_highest_possible_pfn
[i
] =
5346 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5348 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5349 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5351 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5352 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5353 find_zone_movable_pfns_for_nodes();
5355 /* Print out the zone ranges */
5356 printk("Zone ranges:\n");
5357 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5358 if (i
== ZONE_MOVABLE
)
5360 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5361 if (arch_zone_lowest_possible_pfn
[i
] ==
5362 arch_zone_highest_possible_pfn
[i
])
5363 printk(KERN_CONT
"empty\n");
5365 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5366 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5367 (arch_zone_highest_possible_pfn
[i
]
5368 << PAGE_SHIFT
) - 1);
5371 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5372 printk("Movable zone start for each node\n");
5373 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5374 if (zone_movable_pfn
[i
])
5375 printk(" Node %d: %#010lx\n", i
,
5376 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5379 /* Print out the early node map */
5380 printk("Early memory node ranges\n");
5381 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5382 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5383 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5385 /* Initialise every node */
5386 mminit_verify_pageflags_layout();
5387 setup_nr_node_ids();
5388 for_each_online_node(nid
) {
5389 pg_data_t
*pgdat
= NODE_DATA(nid
);
5390 free_area_init_node(nid
, NULL
,
5391 find_min_pfn_for_node(nid
), NULL
);
5393 /* Any memory on that node */
5394 if (pgdat
->node_present_pages
)
5395 node_set_state(nid
, N_MEMORY
);
5396 check_for_memory(pgdat
, nid
);
5400 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5402 unsigned long long coremem
;
5406 coremem
= memparse(p
, &p
);
5407 *core
= coremem
>> PAGE_SHIFT
;
5409 /* Paranoid check that UL is enough for the coremem value */
5410 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5416 * kernelcore=size sets the amount of memory for use for allocations that
5417 * cannot be reclaimed or migrated.
5419 static int __init
cmdline_parse_kernelcore(char *p
)
5421 return cmdline_parse_core(p
, &required_kernelcore
);
5425 * movablecore=size sets the amount of memory for use for allocations that
5426 * can be reclaimed or migrated.
5428 static int __init
cmdline_parse_movablecore(char *p
)
5430 return cmdline_parse_core(p
, &required_movablecore
);
5433 early_param("kernelcore", cmdline_parse_kernelcore
);
5434 early_param("movablecore", cmdline_parse_movablecore
);
5436 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5438 void adjust_managed_page_count(struct page
*page
, long count
)
5440 spin_lock(&managed_page_count_lock
);
5441 page_zone(page
)->managed_pages
+= count
;
5442 totalram_pages
+= count
;
5443 #ifdef CONFIG_HIGHMEM
5444 if (PageHighMem(page
))
5445 totalhigh_pages
+= count
;
5447 spin_unlock(&managed_page_count_lock
);
5449 EXPORT_SYMBOL(adjust_managed_page_count
);
5451 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5454 unsigned long pages
= 0;
5456 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5457 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5458 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5459 if ((unsigned int)poison
<= 0xFF)
5460 memset(pos
, poison
, PAGE_SIZE
);
5461 free_reserved_page(virt_to_page(pos
));
5465 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5466 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5470 EXPORT_SYMBOL(free_reserved_area
);
5472 #ifdef CONFIG_HIGHMEM
5473 void free_highmem_page(struct page
*page
)
5475 __free_reserved_page(page
);
5477 page_zone(page
)->managed_pages
++;
5483 void __init
mem_init_print_info(const char *str
)
5485 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5486 unsigned long init_code_size
, init_data_size
;
5488 physpages
= get_num_physpages();
5489 codesize
= _etext
- _stext
;
5490 datasize
= _edata
- _sdata
;
5491 rosize
= __end_rodata
- __start_rodata
;
5492 bss_size
= __bss_stop
- __bss_start
;
5493 init_data_size
= __init_end
- __init_begin
;
5494 init_code_size
= _einittext
- _sinittext
;
5497 * Detect special cases and adjust section sizes accordingly:
5498 * 1) .init.* may be embedded into .data sections
5499 * 2) .init.text.* may be out of [__init_begin, __init_end],
5500 * please refer to arch/tile/kernel/vmlinux.lds.S.
5501 * 3) .rodata.* may be embedded into .text or .data sections.
5503 #define adj_init_size(start, end, size, pos, adj) \
5505 if (start <= pos && pos < end && size > adj) \
5509 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5510 _sinittext
, init_code_size
);
5511 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5512 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5513 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5514 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5516 #undef adj_init_size
5518 printk("Memory: %luK/%luK available "
5519 "(%luK kernel code, %luK rwdata, %luK rodata, "
5520 "%luK init, %luK bss, %luK reserved"
5521 #ifdef CONFIG_HIGHMEM
5525 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5526 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5527 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5528 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5529 #ifdef CONFIG_HIGHMEM
5530 totalhigh_pages
<< (PAGE_SHIFT
-10),
5532 str
? ", " : "", str
? str
: "");
5536 * set_dma_reserve - set the specified number of pages reserved in the first zone
5537 * @new_dma_reserve: The number of pages to mark reserved
5539 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5540 * In the DMA zone, a significant percentage may be consumed by kernel image
5541 * and other unfreeable allocations which can skew the watermarks badly. This
5542 * function may optionally be used to account for unfreeable pages in the
5543 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5544 * smaller per-cpu batchsize.
5546 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5548 dma_reserve
= new_dma_reserve
;
5551 void __init
free_area_init(unsigned long *zones_size
)
5553 free_area_init_node(0, zones_size
,
5554 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5557 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5558 unsigned long action
, void *hcpu
)
5560 int cpu
= (unsigned long)hcpu
;
5562 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5563 lru_add_drain_cpu(cpu
);
5567 * Spill the event counters of the dead processor
5568 * into the current processors event counters.
5569 * This artificially elevates the count of the current
5572 vm_events_fold_cpu(cpu
);
5575 * Zero the differential counters of the dead processor
5576 * so that the vm statistics are consistent.
5578 * This is only okay since the processor is dead and cannot
5579 * race with what we are doing.
5581 cpu_vm_stats_fold(cpu
);
5586 void __init
page_alloc_init(void)
5588 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5592 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5593 * or min_free_kbytes changes.
5595 static void calculate_totalreserve_pages(void)
5597 struct pglist_data
*pgdat
;
5598 unsigned long reserve_pages
= 0;
5599 enum zone_type i
, j
;
5601 for_each_online_pgdat(pgdat
) {
5602 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5603 struct zone
*zone
= pgdat
->node_zones
+ i
;
5606 /* Find valid and maximum lowmem_reserve in the zone */
5607 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5608 if (zone
->lowmem_reserve
[j
] > max
)
5609 max
= zone
->lowmem_reserve
[j
];
5612 /* we treat the high watermark as reserved pages. */
5613 max
+= high_wmark_pages(zone
);
5615 if (max
> zone
->managed_pages
)
5616 max
= zone
->managed_pages
;
5617 reserve_pages
+= max
;
5619 * Lowmem reserves are not available to
5620 * GFP_HIGHUSER page cache allocations and
5621 * kswapd tries to balance zones to their high
5622 * watermark. As a result, neither should be
5623 * regarded as dirtyable memory, to prevent a
5624 * situation where reclaim has to clean pages
5625 * in order to balance the zones.
5627 zone
->dirty_balance_reserve
= max
;
5630 dirty_balance_reserve
= reserve_pages
;
5631 totalreserve_pages
= reserve_pages
;
5635 * setup_per_zone_lowmem_reserve - called whenever
5636 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5637 * has a correct pages reserved value, so an adequate number of
5638 * pages are left in the zone after a successful __alloc_pages().
5640 static void setup_per_zone_lowmem_reserve(void)
5642 struct pglist_data
*pgdat
;
5643 enum zone_type j
, idx
;
5645 for_each_online_pgdat(pgdat
) {
5646 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5647 struct zone
*zone
= pgdat
->node_zones
+ j
;
5648 unsigned long managed_pages
= zone
->managed_pages
;
5650 zone
->lowmem_reserve
[j
] = 0;
5654 struct zone
*lower_zone
;
5658 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5659 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5661 lower_zone
= pgdat
->node_zones
+ idx
;
5662 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5663 sysctl_lowmem_reserve_ratio
[idx
];
5664 managed_pages
+= lower_zone
->managed_pages
;
5669 /* update totalreserve_pages */
5670 calculate_totalreserve_pages();
5673 static void __setup_per_zone_wmarks(void)
5675 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5676 unsigned long lowmem_pages
= 0;
5678 unsigned long flags
;
5680 /* Calculate total number of !ZONE_HIGHMEM pages */
5681 for_each_zone(zone
) {
5682 if (!is_highmem(zone
))
5683 lowmem_pages
+= zone
->managed_pages
;
5686 for_each_zone(zone
) {
5689 spin_lock_irqsave(&zone
->lock
, flags
);
5690 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5691 do_div(tmp
, lowmem_pages
);
5692 if (is_highmem(zone
)) {
5694 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5695 * need highmem pages, so cap pages_min to a small
5698 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5699 * deltas controls asynch page reclaim, and so should
5700 * not be capped for highmem.
5702 unsigned long min_pages
;
5704 min_pages
= zone
->managed_pages
/ 1024;
5705 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5706 zone
->watermark
[WMARK_MIN
] = min_pages
;
5709 * If it's a lowmem zone, reserve a number of pages
5710 * proportionate to the zone's size.
5712 zone
->watermark
[WMARK_MIN
] = tmp
;
5715 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5716 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5718 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5719 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5720 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5722 setup_zone_migrate_reserve(zone
);
5723 spin_unlock_irqrestore(&zone
->lock
, flags
);
5726 /* update totalreserve_pages */
5727 calculate_totalreserve_pages();
5731 * setup_per_zone_wmarks - called when min_free_kbytes changes
5732 * or when memory is hot-{added|removed}
5734 * Ensures that the watermark[min,low,high] values for each zone are set
5735 * correctly with respect to min_free_kbytes.
5737 void setup_per_zone_wmarks(void)
5739 mutex_lock(&zonelists_mutex
);
5740 __setup_per_zone_wmarks();
5741 mutex_unlock(&zonelists_mutex
);
5745 * The inactive anon list should be small enough that the VM never has to
5746 * do too much work, but large enough that each inactive page has a chance
5747 * to be referenced again before it is swapped out.
5749 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5750 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5751 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5752 * the anonymous pages are kept on the inactive list.
5755 * memory ratio inactive anon
5756 * -------------------------------------
5765 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5767 unsigned int gb
, ratio
;
5769 /* Zone size in gigabytes */
5770 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5772 ratio
= int_sqrt(10 * gb
);
5776 zone
->inactive_ratio
= ratio
;
5779 static void __meminit
setup_per_zone_inactive_ratio(void)
5784 calculate_zone_inactive_ratio(zone
);
5788 * Initialise min_free_kbytes.
5790 * For small machines we want it small (128k min). For large machines
5791 * we want it large (64MB max). But it is not linear, because network
5792 * bandwidth does not increase linearly with machine size. We use
5794 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5795 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5811 int __meminit
init_per_zone_wmark_min(void)
5813 unsigned long lowmem_kbytes
;
5814 int new_min_free_kbytes
;
5816 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5817 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5819 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5820 min_free_kbytes
= new_min_free_kbytes
;
5821 if (min_free_kbytes
< 128)
5822 min_free_kbytes
= 128;
5823 if (min_free_kbytes
> 65536)
5824 min_free_kbytes
= 65536;
5826 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5827 new_min_free_kbytes
, user_min_free_kbytes
);
5829 setup_per_zone_wmarks();
5830 refresh_zone_stat_thresholds();
5831 setup_per_zone_lowmem_reserve();
5832 setup_per_zone_inactive_ratio();
5835 module_init(init_per_zone_wmark_min
)
5838 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5839 * that we can call two helper functions whenever min_free_kbytes
5842 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5843 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5847 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5852 user_min_free_kbytes
= min_free_kbytes
;
5853 setup_per_zone_wmarks();
5859 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5860 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5865 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5870 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5871 sysctl_min_unmapped_ratio
) / 100;
5875 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5876 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5881 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5886 zone
->min_slab_pages
= (zone
->managed_pages
*
5887 sysctl_min_slab_ratio
) / 100;
5893 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5894 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5895 * whenever sysctl_lowmem_reserve_ratio changes.
5897 * The reserve ratio obviously has absolutely no relation with the
5898 * minimum watermarks. The lowmem reserve ratio can only make sense
5899 * if in function of the boot time zone sizes.
5901 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5902 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5904 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5905 setup_per_zone_lowmem_reserve();
5910 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5911 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5912 * pagelist can have before it gets flushed back to buddy allocator.
5914 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5915 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5918 int old_percpu_pagelist_fraction
;
5921 mutex_lock(&pcp_batch_high_lock
);
5922 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5924 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5925 if (!write
|| ret
< 0)
5928 /* Sanity checking to avoid pcp imbalance */
5929 if (percpu_pagelist_fraction
&&
5930 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5931 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5937 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
5940 for_each_populated_zone(zone
) {
5943 for_each_possible_cpu(cpu
)
5944 pageset_set_high_and_batch(zone
,
5945 per_cpu_ptr(zone
->pageset
, cpu
));
5948 mutex_unlock(&pcp_batch_high_lock
);
5952 int hashdist
= HASHDIST_DEFAULT
;
5955 static int __init
set_hashdist(char *str
)
5959 hashdist
= simple_strtoul(str
, &str
, 0);
5962 __setup("hashdist=", set_hashdist
);
5966 * allocate a large system hash table from bootmem
5967 * - it is assumed that the hash table must contain an exact power-of-2
5968 * quantity of entries
5969 * - limit is the number of hash buckets, not the total allocation size
5971 void *__init
alloc_large_system_hash(const char *tablename
,
5972 unsigned long bucketsize
,
5973 unsigned long numentries
,
5976 unsigned int *_hash_shift
,
5977 unsigned int *_hash_mask
,
5978 unsigned long low_limit
,
5979 unsigned long high_limit
)
5981 unsigned long long max
= high_limit
;
5982 unsigned long log2qty
, size
;
5985 /* allow the kernel cmdline to have a say */
5987 /* round applicable memory size up to nearest megabyte */
5988 numentries
= nr_kernel_pages
;
5990 /* It isn't necessary when PAGE_SIZE >= 1MB */
5991 if (PAGE_SHIFT
< 20)
5992 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
5994 /* limit to 1 bucket per 2^scale bytes of low memory */
5995 if (scale
> PAGE_SHIFT
)
5996 numentries
>>= (scale
- PAGE_SHIFT
);
5998 numentries
<<= (PAGE_SHIFT
- scale
);
6000 /* Make sure we've got at least a 0-order allocation.. */
6001 if (unlikely(flags
& HASH_SMALL
)) {
6002 /* Makes no sense without HASH_EARLY */
6003 WARN_ON(!(flags
& HASH_EARLY
));
6004 if (!(numentries
>> *_hash_shift
)) {
6005 numentries
= 1UL << *_hash_shift
;
6006 BUG_ON(!numentries
);
6008 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6009 numentries
= PAGE_SIZE
/ bucketsize
;
6011 numentries
= roundup_pow_of_two(numentries
);
6013 /* limit allocation size to 1/16 total memory by default */
6015 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6016 do_div(max
, bucketsize
);
6018 max
= min(max
, 0x80000000ULL
);
6020 if (numentries
< low_limit
)
6021 numentries
= low_limit
;
6022 if (numentries
> max
)
6025 log2qty
= ilog2(numentries
);
6028 size
= bucketsize
<< log2qty
;
6029 if (flags
& HASH_EARLY
)
6030 table
= memblock_virt_alloc_nopanic(size
, 0);
6032 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6035 * If bucketsize is not a power-of-two, we may free
6036 * some pages at the end of hash table which
6037 * alloc_pages_exact() automatically does
6039 if (get_order(size
) < MAX_ORDER
) {
6040 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6041 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6044 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6047 panic("Failed to allocate %s hash table\n", tablename
);
6049 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6052 ilog2(size
) - PAGE_SHIFT
,
6056 *_hash_shift
= log2qty
;
6058 *_hash_mask
= (1 << log2qty
) - 1;
6063 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6064 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6067 #ifdef CONFIG_SPARSEMEM
6068 return __pfn_to_section(pfn
)->pageblock_flags
;
6070 return zone
->pageblock_flags
;
6071 #endif /* CONFIG_SPARSEMEM */
6074 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6076 #ifdef CONFIG_SPARSEMEM
6077 pfn
&= (PAGES_PER_SECTION
-1);
6078 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6080 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6081 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6082 #endif /* CONFIG_SPARSEMEM */
6086 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6087 * @page: The page within the block of interest
6088 * @pfn: The target page frame number
6089 * @end_bitidx: The last bit of interest to retrieve
6090 * @mask: mask of bits that the caller is interested in
6092 * Return: pageblock_bits flags
6094 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6095 unsigned long end_bitidx
,
6099 unsigned long *bitmap
;
6100 unsigned long bitidx
, word_bitidx
;
6103 zone
= page_zone(page
);
6104 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6105 bitidx
= pfn_to_bitidx(zone
, pfn
);
6106 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6107 bitidx
&= (BITS_PER_LONG
-1);
6109 word
= bitmap
[word_bitidx
];
6110 bitidx
+= end_bitidx
;
6111 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6115 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6116 * @page: The page within the block of interest
6117 * @flags: The flags to set
6118 * @pfn: The target page frame number
6119 * @end_bitidx: The last bit of interest
6120 * @mask: mask of bits that the caller is interested in
6122 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6124 unsigned long end_bitidx
,
6128 unsigned long *bitmap
;
6129 unsigned long bitidx
, word_bitidx
;
6130 unsigned long old_word
, word
;
6132 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6134 zone
= page_zone(page
);
6135 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6136 bitidx
= pfn_to_bitidx(zone
, pfn
);
6137 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6138 bitidx
&= (BITS_PER_LONG
-1);
6140 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6142 bitidx
+= end_bitidx
;
6143 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6144 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6146 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6148 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6149 if (word
== old_word
)
6156 * This function checks whether pageblock includes unmovable pages or not.
6157 * If @count is not zero, it is okay to include less @count unmovable pages
6159 * PageLRU check without isolation or lru_lock could race so that
6160 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6161 * expect this function should be exact.
6163 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6164 bool skip_hwpoisoned_pages
)
6166 unsigned long pfn
, iter
, found
;
6170 * For avoiding noise data, lru_add_drain_all() should be called
6171 * If ZONE_MOVABLE, the zone never contains unmovable pages
6173 if (zone_idx(zone
) == ZONE_MOVABLE
)
6175 mt
= get_pageblock_migratetype(page
);
6176 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6179 pfn
= page_to_pfn(page
);
6180 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6181 unsigned long check
= pfn
+ iter
;
6183 if (!pfn_valid_within(check
))
6186 page
= pfn_to_page(check
);
6189 * Hugepages are not in LRU lists, but they're movable.
6190 * We need not scan over tail pages bacause we don't
6191 * handle each tail page individually in migration.
6193 if (PageHuge(page
)) {
6194 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6199 * We can't use page_count without pin a page
6200 * because another CPU can free compound page.
6201 * This check already skips compound tails of THP
6202 * because their page->_count is zero at all time.
6204 if (!atomic_read(&page
->_count
)) {
6205 if (PageBuddy(page
))
6206 iter
+= (1 << page_order(page
)) - 1;
6211 * The HWPoisoned page may be not in buddy system, and
6212 * page_count() is not 0.
6214 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6220 * If there are RECLAIMABLE pages, we need to check it.
6221 * But now, memory offline itself doesn't call shrink_slab()
6222 * and it still to be fixed.
6225 * If the page is not RAM, page_count()should be 0.
6226 * we don't need more check. This is an _used_ not-movable page.
6228 * The problematic thing here is PG_reserved pages. PG_reserved
6229 * is set to both of a memory hole page and a _used_ kernel
6238 bool is_pageblock_removable_nolock(struct page
*page
)
6244 * We have to be careful here because we are iterating over memory
6245 * sections which are not zone aware so we might end up outside of
6246 * the zone but still within the section.
6247 * We have to take care about the node as well. If the node is offline
6248 * its NODE_DATA will be NULL - see page_zone.
6250 if (!node_online(page_to_nid(page
)))
6253 zone
= page_zone(page
);
6254 pfn
= page_to_pfn(page
);
6255 if (!zone_spans_pfn(zone
, pfn
))
6258 return !has_unmovable_pages(zone
, page
, 0, true);
6263 static unsigned long pfn_max_align_down(unsigned long pfn
)
6265 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6266 pageblock_nr_pages
) - 1);
6269 static unsigned long pfn_max_align_up(unsigned long pfn
)
6271 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6272 pageblock_nr_pages
));
6275 /* [start, end) must belong to a single zone. */
6276 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6277 unsigned long start
, unsigned long end
)
6279 /* This function is based on compact_zone() from compaction.c. */
6280 unsigned long nr_reclaimed
;
6281 unsigned long pfn
= start
;
6282 unsigned int tries
= 0;
6287 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6288 if (fatal_signal_pending(current
)) {
6293 if (list_empty(&cc
->migratepages
)) {
6294 cc
->nr_migratepages
= 0;
6295 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6301 } else if (++tries
== 5) {
6302 ret
= ret
< 0 ? ret
: -EBUSY
;
6306 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6308 cc
->nr_migratepages
-= nr_reclaimed
;
6310 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6311 NULL
, 0, cc
->mode
, MR_CMA
);
6314 putback_movable_pages(&cc
->migratepages
);
6321 * alloc_contig_range() -- tries to allocate given range of pages
6322 * @start: start PFN to allocate
6323 * @end: one-past-the-last PFN to allocate
6324 * @migratetype: migratetype of the underlaying pageblocks (either
6325 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6326 * in range must have the same migratetype and it must
6327 * be either of the two.
6329 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6330 * aligned, however it's the caller's responsibility to guarantee that
6331 * we are the only thread that changes migrate type of pageblocks the
6334 * The PFN range must belong to a single zone.
6336 * Returns zero on success or negative error code. On success all
6337 * pages which PFN is in [start, end) are allocated for the caller and
6338 * need to be freed with free_contig_range().
6340 int alloc_contig_range(unsigned long start
, unsigned long end
,
6341 unsigned migratetype
)
6343 unsigned long outer_start
, outer_end
;
6346 struct compact_control cc
= {
6347 .nr_migratepages
= 0,
6349 .zone
= page_zone(pfn_to_page(start
)),
6350 .mode
= MIGRATE_SYNC
,
6351 .ignore_skip_hint
= true,
6353 INIT_LIST_HEAD(&cc
.migratepages
);
6356 * What we do here is we mark all pageblocks in range as
6357 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6358 * have different sizes, and due to the way page allocator
6359 * work, we align the range to biggest of the two pages so
6360 * that page allocator won't try to merge buddies from
6361 * different pageblocks and change MIGRATE_ISOLATE to some
6362 * other migration type.
6364 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6365 * migrate the pages from an unaligned range (ie. pages that
6366 * we are interested in). This will put all the pages in
6367 * range back to page allocator as MIGRATE_ISOLATE.
6369 * When this is done, we take the pages in range from page
6370 * allocator removing them from the buddy system. This way
6371 * page allocator will never consider using them.
6373 * This lets us mark the pageblocks back as
6374 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6375 * aligned range but not in the unaligned, original range are
6376 * put back to page allocator so that buddy can use them.
6379 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6380 pfn_max_align_up(end
), migratetype
,
6385 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6390 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6391 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6392 * more, all pages in [start, end) are free in page allocator.
6393 * What we are going to do is to allocate all pages from
6394 * [start, end) (that is remove them from page allocator).
6396 * The only problem is that pages at the beginning and at the
6397 * end of interesting range may be not aligned with pages that
6398 * page allocator holds, ie. they can be part of higher order
6399 * pages. Because of this, we reserve the bigger range and
6400 * once this is done free the pages we are not interested in.
6402 * We don't have to hold zone->lock here because the pages are
6403 * isolated thus they won't get removed from buddy.
6406 lru_add_drain_all();
6410 outer_start
= start
;
6411 while (!PageBuddy(pfn_to_page(outer_start
))) {
6412 if (++order
>= MAX_ORDER
) {
6416 outer_start
&= ~0UL << order
;
6419 /* Make sure the range is really isolated. */
6420 if (test_pages_isolated(outer_start
, end
, false)) {
6421 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6428 /* Grab isolated pages from freelists. */
6429 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6435 /* Free head and tail (if any) */
6436 if (start
!= outer_start
)
6437 free_contig_range(outer_start
, start
- outer_start
);
6438 if (end
!= outer_end
)
6439 free_contig_range(end
, outer_end
- end
);
6442 undo_isolate_page_range(pfn_max_align_down(start
),
6443 pfn_max_align_up(end
), migratetype
);
6447 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6449 unsigned int count
= 0;
6451 for (; nr_pages
--; pfn
++) {
6452 struct page
*page
= pfn_to_page(pfn
);
6454 count
+= page_count(page
) != 1;
6457 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6461 #ifdef CONFIG_MEMORY_HOTPLUG
6463 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6464 * page high values need to be recalulated.
6466 void __meminit
zone_pcp_update(struct zone
*zone
)
6469 mutex_lock(&pcp_batch_high_lock
);
6470 for_each_possible_cpu(cpu
)
6471 pageset_set_high_and_batch(zone
,
6472 per_cpu_ptr(zone
->pageset
, cpu
));
6473 mutex_unlock(&pcp_batch_high_lock
);
6477 void zone_pcp_reset(struct zone
*zone
)
6479 unsigned long flags
;
6481 struct per_cpu_pageset
*pset
;
6483 /* avoid races with drain_pages() */
6484 local_irq_save(flags
);
6485 if (zone
->pageset
!= &boot_pageset
) {
6486 for_each_online_cpu(cpu
) {
6487 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6488 drain_zonestat(zone
, pset
);
6490 free_percpu(zone
->pageset
);
6491 zone
->pageset
= &boot_pageset
;
6493 local_irq_restore(flags
);
6496 #ifdef CONFIG_MEMORY_HOTREMOVE
6498 * All pages in the range must be isolated before calling this.
6501 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6505 unsigned int order
, i
;
6507 unsigned long flags
;
6508 /* find the first valid pfn */
6509 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6514 zone
= page_zone(pfn_to_page(pfn
));
6515 spin_lock_irqsave(&zone
->lock
, flags
);
6517 while (pfn
< end_pfn
) {
6518 if (!pfn_valid(pfn
)) {
6522 page
= pfn_to_page(pfn
);
6524 * The HWPoisoned page may be not in buddy system, and
6525 * page_count() is not 0.
6527 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6529 SetPageReserved(page
);
6533 BUG_ON(page_count(page
));
6534 BUG_ON(!PageBuddy(page
));
6535 order
= page_order(page
);
6536 #ifdef CONFIG_DEBUG_VM
6537 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6538 pfn
, 1 << order
, end_pfn
);
6540 list_del(&page
->lru
);
6541 rmv_page_order(page
);
6542 zone
->free_area
[order
].nr_free
--;
6543 for (i
= 0; i
< (1 << order
); i
++)
6544 SetPageReserved((page
+i
));
6545 pfn
+= (1 << order
);
6547 spin_unlock_irqrestore(&zone
->lock
, flags
);
6551 #ifdef CONFIG_MEMORY_FAILURE
6552 bool is_free_buddy_page(struct page
*page
)
6554 struct zone
*zone
= page_zone(page
);
6555 unsigned long pfn
= page_to_pfn(page
);
6556 unsigned long flags
;
6559 spin_lock_irqsave(&zone
->lock
, flags
);
6560 for (order
= 0; order
< MAX_ORDER
; order
++) {
6561 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6563 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6566 spin_unlock_irqrestore(&zone
->lock
, flags
);
6568 return order
< MAX_ORDER
;