2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/prefetch.h>
57 #include <linux/mm_inline.h>
58 #include <linux/migrate.h>
59 #include <linux/page-debug-flags.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
63 #include <asm/sections.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
69 static DEFINE_MUTEX(pcp_batch_high_lock
);
70 #define MIN_PERCPU_PAGELIST_FRACTION (8)
72 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
73 DEFINE_PER_CPU(int, numa_node
);
74 EXPORT_PER_CPU_SYMBOL(numa_node
);
77 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
79 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
80 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
81 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
82 * defined in <linux/topology.h>.
84 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
85 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
86 int _node_numa_mem_
[MAX_NUMNODES
];
90 * Array of node states.
92 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
93 [N_POSSIBLE
] = NODE_MASK_ALL
,
94 [N_ONLINE
] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY
] = { { [0] = 1UL } },
103 [N_CPU
] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states
);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock
);
111 unsigned long totalram_pages __read_mostly
;
112 unsigned long totalreserve_pages __read_mostly
;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly
;
121 int percpu_pagelist_fraction
;
122 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask
;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex
));
139 if (saved_gfp_mask
) {
140 gfp_allowed_mask
= saved_gfp_mask
;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex
));
148 WARN_ON(saved_gfp_mask
);
149 saved_gfp_mask
= gfp_allowed_mask
;
150 gfp_allowed_mask
&= ~GFP_IOFS
;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly
;
165 static void __free_pages_ok(struct page
*page
, unsigned int order
);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages
);
193 static char * const zone_names
[MAX_NR_ZONES
] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes
= 1024;
208 int user_min_free_kbytes
= -1;
210 static unsigned long __meminitdata nr_kernel_pages
;
211 static unsigned long __meminitdata nr_all_pages
;
212 static unsigned long __meminitdata dma_reserve
;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
217 static unsigned long __initdata required_kernelcore
;
218 static unsigned long __initdata required_movablecore
;
219 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone
);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
228 int nr_online_nodes __read_mostly
= 1;
229 EXPORT_SYMBOL(nr_node_ids
);
230 EXPORT_SYMBOL(nr_online_nodes
);
233 int page_group_by_mobility_disabled __read_mostly
;
235 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
237 if (unlikely(page_group_by_mobility_disabled
&&
238 migratetype
< MIGRATE_PCPTYPES
))
239 migratetype
= MIGRATE_UNMOVABLE
;
241 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
242 PB_migrate
, PB_migrate_end
);
245 bool oom_killer_disabled __read_mostly
;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
252 unsigned long pfn
= page_to_pfn(page
);
253 unsigned long sp
, start_pfn
;
256 seq
= zone_span_seqbegin(zone
);
257 start_pfn
= zone
->zone_start_pfn
;
258 sp
= zone
->spanned_pages
;
259 if (!zone_spans_pfn(zone
, pfn
))
261 } while (zone_span_seqretry(zone
, seq
));
264 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
265 pfn
, zone_to_nid(zone
), zone
->name
,
266 start_pfn
, start_pfn
+ sp
);
271 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
273 if (!pfn_valid_within(page_to_pfn(page
)))
275 if (zone
!= page_zone(page
))
281 * Temporary debugging check for pages not lying within a given zone.
283 static int bad_range(struct zone
*zone
, struct page
*page
)
285 if (page_outside_zone_boundaries(zone
, page
))
287 if (!page_is_consistent(zone
, page
))
293 static inline int bad_range(struct zone
*zone
, struct page
*page
)
299 static void bad_page(struct page
*page
, const char *reason
,
300 unsigned long bad_flags
)
302 static unsigned long resume
;
303 static unsigned long nr_shown
;
304 static unsigned long nr_unshown
;
306 /* Don't complain about poisoned pages */
307 if (PageHWPoison(page
)) {
308 page_mapcount_reset(page
); /* remove PageBuddy */
313 * Allow a burst of 60 reports, then keep quiet for that minute;
314 * or allow a steady drip of one report per second.
316 if (nr_shown
== 60) {
317 if (time_before(jiffies
, resume
)) {
323 "BUG: Bad page state: %lu messages suppressed\n",
330 resume
= jiffies
+ 60 * HZ
;
332 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
333 current
->comm
, page_to_pfn(page
));
334 dump_page_badflags(page
, reason
, bad_flags
);
339 /* Leave bad fields for debug, except PageBuddy could make trouble */
340 page_mapcount_reset(page
); /* remove PageBuddy */
341 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
345 * Higher-order pages are called "compound pages". They are structured thusly:
347 * The first PAGE_SIZE page is called the "head page".
349 * The remaining PAGE_SIZE pages are called "tail pages".
351 * All pages have PG_compound set. All tail pages have their ->first_page
352 * pointing at the head page.
354 * The first tail page's ->lru.next holds the address of the compound page's
355 * put_page() function. Its ->lru.prev holds the order of allocation.
356 * This usage means that zero-order pages may not be compound.
359 static void free_compound_page(struct page
*page
)
361 __free_pages_ok(page
, compound_order(page
));
364 void prep_compound_page(struct page
*page
, unsigned long order
)
367 int nr_pages
= 1 << order
;
369 set_compound_page_dtor(page
, free_compound_page
);
370 set_compound_order(page
, order
);
372 for (i
= 1; i
< nr_pages
; i
++) {
373 struct page
*p
= page
+ i
;
374 set_page_count(p
, 0);
375 p
->first_page
= page
;
376 /* Make sure p->first_page is always valid for PageTail() */
382 /* update __split_huge_page_refcount if you change this function */
383 static int destroy_compound_page(struct page
*page
, unsigned long order
)
386 int nr_pages
= 1 << order
;
389 if (unlikely(compound_order(page
) != order
)) {
390 bad_page(page
, "wrong compound order", 0);
394 __ClearPageHead(page
);
396 for (i
= 1; i
< nr_pages
; i
++) {
397 struct page
*p
= page
+ i
;
399 if (unlikely(!PageTail(p
))) {
400 bad_page(page
, "PageTail not set", 0);
402 } else if (unlikely(p
->first_page
!= page
)) {
403 bad_page(page
, "first_page not consistent", 0);
412 static inline void prep_zero_page(struct page
*page
, unsigned int order
,
418 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
419 * and __GFP_HIGHMEM from hard or soft interrupt context.
421 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
422 for (i
= 0; i
< (1 << order
); i
++)
423 clear_highpage(page
+ i
);
426 #ifdef CONFIG_DEBUG_PAGEALLOC
427 unsigned int _debug_guardpage_minorder
;
429 static int __init
debug_guardpage_minorder_setup(char *buf
)
433 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
434 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
437 _debug_guardpage_minorder
= res
;
438 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
441 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
443 static inline void set_page_guard_flag(struct page
*page
)
445 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
448 static inline void clear_page_guard_flag(struct page
*page
)
450 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
453 static inline void set_page_guard_flag(struct page
*page
) { }
454 static inline void clear_page_guard_flag(struct page
*page
) { }
457 static inline void set_page_order(struct page
*page
, unsigned int order
)
459 set_page_private(page
, order
);
460 __SetPageBuddy(page
);
463 static inline void rmv_page_order(struct page
*page
)
465 __ClearPageBuddy(page
);
466 set_page_private(page
, 0);
470 * This function checks whether a page is free && is the buddy
471 * we can do coalesce a page and its buddy if
472 * (a) the buddy is not in a hole &&
473 * (b) the buddy is in the buddy system &&
474 * (c) a page and its buddy have the same order &&
475 * (d) a page and its buddy are in the same zone.
477 * For recording whether a page is in the buddy system, we set ->_mapcount
478 * PAGE_BUDDY_MAPCOUNT_VALUE.
479 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
480 * serialized by zone->lock.
482 * For recording page's order, we use page_private(page).
484 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
487 if (!pfn_valid_within(page_to_pfn(buddy
)))
490 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
491 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
493 if (page_zone_id(page
) != page_zone_id(buddy
))
499 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
500 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
503 * zone check is done late to avoid uselessly
504 * calculating zone/node ids for pages that could
507 if (page_zone_id(page
) != page_zone_id(buddy
))
516 * Freeing function for a buddy system allocator.
518 * The concept of a buddy system is to maintain direct-mapped table
519 * (containing bit values) for memory blocks of various "orders".
520 * The bottom level table contains the map for the smallest allocatable
521 * units of memory (here, pages), and each level above it describes
522 * pairs of units from the levels below, hence, "buddies".
523 * At a high level, all that happens here is marking the table entry
524 * at the bottom level available, and propagating the changes upward
525 * as necessary, plus some accounting needed to play nicely with other
526 * parts of the VM system.
527 * At each level, we keep a list of pages, which are heads of continuous
528 * free pages of length of (1 << order) and marked with _mapcount
529 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
531 * So when we are allocating or freeing one, we can derive the state of the
532 * other. That is, if we allocate a small block, and both were
533 * free, the remainder of the region must be split into blocks.
534 * If a block is freed, and its buddy is also free, then this
535 * triggers coalescing into a block of larger size.
540 static inline void __free_one_page(struct page
*page
,
542 struct zone
*zone
, unsigned int order
,
545 unsigned long page_idx
;
546 unsigned long combined_idx
;
547 unsigned long uninitialized_var(buddy_idx
);
549 int max_order
= MAX_ORDER
;
551 VM_BUG_ON(!zone_is_initialized(zone
));
553 if (unlikely(PageCompound(page
)))
554 if (unlikely(destroy_compound_page(page
, order
)))
557 VM_BUG_ON(migratetype
== -1);
558 if (is_migrate_isolate(migratetype
)) {
560 * We restrict max order of merging to prevent merge
561 * between freepages on isolate pageblock and normal
562 * pageblock. Without this, pageblock isolation
563 * could cause incorrect freepage accounting.
565 max_order
= min(MAX_ORDER
, pageblock_order
+ 1);
567 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
570 page_idx
= pfn
& ((1 << max_order
) - 1);
572 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
573 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
575 while (order
< max_order
- 1) {
576 buddy_idx
= __find_buddy_index(page_idx
, order
);
577 buddy
= page
+ (buddy_idx
- page_idx
);
578 if (!page_is_buddy(page
, buddy
, order
))
581 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
582 * merge with it and move up one order.
584 if (page_is_guard(buddy
)) {
585 clear_page_guard_flag(buddy
);
586 set_page_private(page
, 0);
587 __mod_zone_freepage_state(zone
, 1 << order
,
590 list_del(&buddy
->lru
);
591 zone
->free_area
[order
].nr_free
--;
592 rmv_page_order(buddy
);
594 combined_idx
= buddy_idx
& page_idx
;
595 page
= page
+ (combined_idx
- page_idx
);
596 page_idx
= combined_idx
;
599 set_page_order(page
, order
);
602 * If this is not the largest possible page, check if the buddy
603 * of the next-highest order is free. If it is, it's possible
604 * that pages are being freed that will coalesce soon. In case,
605 * that is happening, add the free page to the tail of the list
606 * so it's less likely to be used soon and more likely to be merged
607 * as a higher order page
609 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
610 struct page
*higher_page
, *higher_buddy
;
611 combined_idx
= buddy_idx
& page_idx
;
612 higher_page
= page
+ (combined_idx
- page_idx
);
613 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
614 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
615 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
616 list_add_tail(&page
->lru
,
617 &zone
->free_area
[order
].free_list
[migratetype
]);
622 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
624 zone
->free_area
[order
].nr_free
++;
627 static inline int free_pages_check(struct page
*page
)
629 const char *bad_reason
= NULL
;
630 unsigned long bad_flags
= 0;
632 if (unlikely(page_mapcount(page
)))
633 bad_reason
= "nonzero mapcount";
634 if (unlikely(page
->mapping
!= NULL
))
635 bad_reason
= "non-NULL mapping";
636 if (unlikely(atomic_read(&page
->_count
) != 0))
637 bad_reason
= "nonzero _count";
638 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
639 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
640 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
642 if (unlikely(mem_cgroup_bad_page_check(page
)))
643 bad_reason
= "cgroup check failed";
644 if (unlikely(bad_reason
)) {
645 bad_page(page
, bad_reason
, bad_flags
);
648 page_cpupid_reset_last(page
);
649 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
650 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
655 * Frees a number of pages from the PCP lists
656 * Assumes all pages on list are in same zone, and of same order.
657 * count is the number of pages to free.
659 * If the zone was previously in an "all pages pinned" state then look to
660 * see if this freeing clears that state.
662 * And clear the zone's pages_scanned counter, to hold off the "all pages are
663 * pinned" detection logic.
665 static void free_pcppages_bulk(struct zone
*zone
, int count
,
666 struct per_cpu_pages
*pcp
)
671 unsigned long nr_scanned
;
673 spin_lock(&zone
->lock
);
674 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
676 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
680 struct list_head
*list
;
683 * Remove pages from lists in a round-robin fashion. A
684 * batch_free count is maintained that is incremented when an
685 * empty list is encountered. This is so more pages are freed
686 * off fuller lists instead of spinning excessively around empty
691 if (++migratetype
== MIGRATE_PCPTYPES
)
693 list
= &pcp
->lists
[migratetype
];
694 } while (list_empty(list
));
696 /* This is the only non-empty list. Free them all. */
697 if (batch_free
== MIGRATE_PCPTYPES
)
698 batch_free
= to_free
;
701 int mt
; /* migratetype of the to-be-freed page */
703 page
= list_entry(list
->prev
, struct page
, lru
);
704 /* must delete as __free_one_page list manipulates */
705 list_del(&page
->lru
);
706 mt
= get_freepage_migratetype(page
);
707 if (unlikely(has_isolate_pageblock(zone
)))
708 mt
= get_pageblock_migratetype(page
);
710 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
711 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
712 trace_mm_page_pcpu_drain(page
, 0, mt
);
713 } while (--to_free
&& --batch_free
&& !list_empty(list
));
715 spin_unlock(&zone
->lock
);
718 static void free_one_page(struct zone
*zone
,
719 struct page
*page
, unsigned long pfn
,
723 unsigned long nr_scanned
;
724 spin_lock(&zone
->lock
);
725 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
727 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
729 if (unlikely(has_isolate_pageblock(zone
) ||
730 is_migrate_isolate(migratetype
))) {
731 migratetype
= get_pfnblock_migratetype(page
, pfn
);
733 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
734 spin_unlock(&zone
->lock
);
737 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
742 trace_mm_page_free(page
, order
);
743 kmemcheck_free_shadow(page
, order
);
746 page
->mapping
= NULL
;
747 for (i
= 0; i
< (1 << order
); i
++)
748 bad
+= free_pages_check(page
+ i
);
752 if (!PageHighMem(page
)) {
753 debug_check_no_locks_freed(page_address(page
),
755 debug_check_no_obj_freed(page_address(page
),
758 arch_free_page(page
, order
);
759 kernel_map_pages(page
, 1 << order
, 0);
764 static void __free_pages_ok(struct page
*page
, unsigned int order
)
768 unsigned long pfn
= page_to_pfn(page
);
770 if (!free_pages_prepare(page
, order
))
773 migratetype
= get_pfnblock_migratetype(page
, pfn
);
774 local_irq_save(flags
);
775 __count_vm_events(PGFREE
, 1 << order
);
776 set_freepage_migratetype(page
, migratetype
);
777 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
778 local_irq_restore(flags
);
781 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
783 unsigned int nr_pages
= 1 << order
;
784 struct page
*p
= page
;
788 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
790 __ClearPageReserved(p
);
791 set_page_count(p
, 0);
793 __ClearPageReserved(p
);
794 set_page_count(p
, 0);
796 page_zone(page
)->managed_pages
+= nr_pages
;
797 set_page_refcounted(page
);
798 __free_pages(page
, order
);
802 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
803 void __init
init_cma_reserved_pageblock(struct page
*page
)
805 unsigned i
= pageblock_nr_pages
;
806 struct page
*p
= page
;
809 __ClearPageReserved(p
);
810 set_page_count(p
, 0);
813 set_pageblock_migratetype(page
, MIGRATE_CMA
);
815 if (pageblock_order
>= MAX_ORDER
) {
816 i
= pageblock_nr_pages
;
819 set_page_refcounted(p
);
820 __free_pages(p
, MAX_ORDER
- 1);
821 p
+= MAX_ORDER_NR_PAGES
;
822 } while (i
-= MAX_ORDER_NR_PAGES
);
824 set_page_refcounted(page
);
825 __free_pages(page
, pageblock_order
);
828 adjust_managed_page_count(page
, pageblock_nr_pages
);
833 * The order of subdivision here is critical for the IO subsystem.
834 * Please do not alter this order without good reasons and regression
835 * testing. Specifically, as large blocks of memory are subdivided,
836 * the order in which smaller blocks are delivered depends on the order
837 * they're subdivided in this function. This is the primary factor
838 * influencing the order in which pages are delivered to the IO
839 * subsystem according to empirical testing, and this is also justified
840 * by considering the behavior of a buddy system containing a single
841 * large block of memory acted on by a series of small allocations.
842 * This behavior is a critical factor in sglist merging's success.
846 static inline void expand(struct zone
*zone
, struct page
*page
,
847 int low
, int high
, struct free_area
*area
,
850 unsigned long size
= 1 << high
;
856 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
858 #ifdef CONFIG_DEBUG_PAGEALLOC
859 if (high
< debug_guardpage_minorder()) {
861 * Mark as guard pages (or page), that will allow to
862 * merge back to allocator when buddy will be freed.
863 * Corresponding page table entries will not be touched,
864 * pages will stay not present in virtual address space
866 INIT_LIST_HEAD(&page
[size
].lru
);
867 set_page_guard_flag(&page
[size
]);
868 set_page_private(&page
[size
], high
);
869 /* Guard pages are not available for any usage */
870 __mod_zone_freepage_state(zone
, -(1 << high
),
875 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
877 set_page_order(&page
[size
], high
);
882 * This page is about to be returned from the page allocator
884 static inline int check_new_page(struct page
*page
)
886 const char *bad_reason
= NULL
;
887 unsigned long bad_flags
= 0;
889 if (unlikely(page_mapcount(page
)))
890 bad_reason
= "nonzero mapcount";
891 if (unlikely(page
->mapping
!= NULL
))
892 bad_reason
= "non-NULL mapping";
893 if (unlikely(atomic_read(&page
->_count
) != 0))
894 bad_reason
= "nonzero _count";
895 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
896 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
897 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
899 if (unlikely(mem_cgroup_bad_page_check(page
)))
900 bad_reason
= "cgroup check failed";
901 if (unlikely(bad_reason
)) {
902 bad_page(page
, bad_reason
, bad_flags
);
908 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
)
912 for (i
= 0; i
< (1 << order
); i
++) {
913 struct page
*p
= page
+ i
;
914 if (unlikely(check_new_page(p
)))
918 set_page_private(page
, 0);
919 set_page_refcounted(page
);
921 arch_alloc_page(page
, order
);
922 kernel_map_pages(page
, 1 << order
, 1);
924 if (gfp_flags
& __GFP_ZERO
)
925 prep_zero_page(page
, order
, gfp_flags
);
927 if (order
&& (gfp_flags
& __GFP_COMP
))
928 prep_compound_page(page
, order
);
934 * Go through the free lists for the given migratetype and remove
935 * the smallest available page from the freelists
938 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
941 unsigned int current_order
;
942 struct free_area
*area
;
945 /* Find a page of the appropriate size in the preferred list */
946 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
947 area
= &(zone
->free_area
[current_order
]);
948 if (list_empty(&area
->free_list
[migratetype
]))
951 page
= list_entry(area
->free_list
[migratetype
].next
,
953 list_del(&page
->lru
);
954 rmv_page_order(page
);
956 expand(zone
, page
, order
, current_order
, area
, migratetype
);
957 set_freepage_migratetype(page
, migratetype
);
966 * This array describes the order lists are fallen back to when
967 * the free lists for the desirable migrate type are depleted
969 static int fallbacks
[MIGRATE_TYPES
][4] = {
970 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
971 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
973 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
974 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
976 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
978 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
979 #ifdef CONFIG_MEMORY_ISOLATION
980 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
985 * Move the free pages in a range to the free lists of the requested type.
986 * Note that start_page and end_pages are not aligned on a pageblock
987 * boundary. If alignment is required, use move_freepages_block()
989 int move_freepages(struct zone
*zone
,
990 struct page
*start_page
, struct page
*end_page
,
997 #ifndef CONFIG_HOLES_IN_ZONE
999 * page_zone is not safe to call in this context when
1000 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1001 * anyway as we check zone boundaries in move_freepages_block().
1002 * Remove at a later date when no bug reports exist related to
1003 * grouping pages by mobility
1005 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1008 for (page
= start_page
; page
<= end_page
;) {
1009 /* Make sure we are not inadvertently changing nodes */
1010 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1012 if (!pfn_valid_within(page_to_pfn(page
))) {
1017 if (!PageBuddy(page
)) {
1022 order
= page_order(page
);
1023 list_move(&page
->lru
,
1024 &zone
->free_area
[order
].free_list
[migratetype
]);
1025 set_freepage_migratetype(page
, migratetype
);
1027 pages_moved
+= 1 << order
;
1033 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1036 unsigned long start_pfn
, end_pfn
;
1037 struct page
*start_page
, *end_page
;
1039 start_pfn
= page_to_pfn(page
);
1040 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1041 start_page
= pfn_to_page(start_pfn
);
1042 end_page
= start_page
+ pageblock_nr_pages
- 1;
1043 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1045 /* Do not cross zone boundaries */
1046 if (!zone_spans_pfn(zone
, start_pfn
))
1048 if (!zone_spans_pfn(zone
, end_pfn
))
1051 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1054 static void change_pageblock_range(struct page
*pageblock_page
,
1055 int start_order
, int migratetype
)
1057 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1059 while (nr_pageblocks
--) {
1060 set_pageblock_migratetype(pageblock_page
, migratetype
);
1061 pageblock_page
+= pageblock_nr_pages
;
1066 * If breaking a large block of pages, move all free pages to the preferred
1067 * allocation list. If falling back for a reclaimable kernel allocation, be
1068 * more aggressive about taking ownership of free pages.
1070 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1071 * nor move CMA pages to different free lists. We don't want unmovable pages
1072 * to be allocated from MIGRATE_CMA areas.
1074 * Returns the new migratetype of the pageblock (or the same old migratetype
1075 * if it was unchanged).
1077 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1078 int start_type
, int fallback_type
)
1080 int current_order
= page_order(page
);
1083 * When borrowing from MIGRATE_CMA, we need to release the excess
1084 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1085 * is set to CMA so it is returned to the correct freelist in case
1086 * the page ends up being not actually allocated from the pcp lists.
1088 if (is_migrate_cma(fallback_type
))
1089 return fallback_type
;
1091 /* Take ownership for orders >= pageblock_order */
1092 if (current_order
>= pageblock_order
) {
1093 change_pageblock_range(page
, current_order
, start_type
);
1097 if (current_order
>= pageblock_order
/ 2 ||
1098 start_type
== MIGRATE_RECLAIMABLE
||
1099 page_group_by_mobility_disabled
) {
1102 pages
= move_freepages_block(zone
, page
, start_type
);
1104 /* Claim the whole block if over half of it is free */
1105 if (pages
>= (1 << (pageblock_order
-1)) ||
1106 page_group_by_mobility_disabled
) {
1108 set_pageblock_migratetype(page
, start_type
);
1114 return fallback_type
;
1117 /* Remove an element from the buddy allocator from the fallback list */
1118 static inline struct page
*
1119 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1121 struct free_area
*area
;
1122 unsigned int current_order
;
1124 int migratetype
, new_type
, i
;
1126 /* Find the largest possible block of pages in the other list */
1127 for (current_order
= MAX_ORDER
-1;
1128 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1131 migratetype
= fallbacks
[start_migratetype
][i
];
1133 /* MIGRATE_RESERVE handled later if necessary */
1134 if (migratetype
== MIGRATE_RESERVE
)
1137 area
= &(zone
->free_area
[current_order
]);
1138 if (list_empty(&area
->free_list
[migratetype
]))
1141 page
= list_entry(area
->free_list
[migratetype
].next
,
1145 new_type
= try_to_steal_freepages(zone
, page
,
1149 /* Remove the page from the freelists */
1150 list_del(&page
->lru
);
1151 rmv_page_order(page
);
1153 expand(zone
, page
, order
, current_order
, area
,
1155 /* The freepage_migratetype may differ from pageblock's
1156 * migratetype depending on the decisions in
1157 * try_to_steal_freepages. This is OK as long as it does
1158 * not differ for MIGRATE_CMA type.
1160 set_freepage_migratetype(page
, new_type
);
1162 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1163 start_migratetype
, migratetype
, new_type
);
1173 * Do the hard work of removing an element from the buddy allocator.
1174 * Call me with the zone->lock already held.
1176 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1182 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1184 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1185 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1188 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1189 * is used because __rmqueue_smallest is an inline function
1190 * and we want just one call site
1193 migratetype
= MIGRATE_RESERVE
;
1198 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1203 * Obtain a specified number of elements from the buddy allocator, all under
1204 * a single hold of the lock, for efficiency. Add them to the supplied list.
1205 * Returns the number of new pages which were placed at *list.
1207 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1208 unsigned long count
, struct list_head
*list
,
1209 int migratetype
, bool cold
)
1213 spin_lock(&zone
->lock
);
1214 for (i
= 0; i
< count
; ++i
) {
1215 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1216 if (unlikely(page
== NULL
))
1220 * Split buddy pages returned by expand() are received here
1221 * in physical page order. The page is added to the callers and
1222 * list and the list head then moves forward. From the callers
1223 * perspective, the linked list is ordered by page number in
1224 * some conditions. This is useful for IO devices that can
1225 * merge IO requests if the physical pages are ordered
1229 list_add(&page
->lru
, list
);
1231 list_add_tail(&page
->lru
, list
);
1233 if (is_migrate_cma(get_freepage_migratetype(page
)))
1234 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1237 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1238 spin_unlock(&zone
->lock
);
1244 * Called from the vmstat counter updater to drain pagesets of this
1245 * currently executing processor on remote nodes after they have
1248 * Note that this function must be called with the thread pinned to
1249 * a single processor.
1251 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1253 unsigned long flags
;
1254 int to_drain
, batch
;
1256 local_irq_save(flags
);
1257 batch
= ACCESS_ONCE(pcp
->batch
);
1258 to_drain
= min(pcp
->count
, batch
);
1260 free_pcppages_bulk(zone
, to_drain
, pcp
);
1261 pcp
->count
-= to_drain
;
1263 local_irq_restore(flags
);
1268 * Drain pages of the indicated processor.
1270 * The processor must either be the current processor and the
1271 * thread pinned to the current processor or a processor that
1274 static void drain_pages(unsigned int cpu
)
1276 unsigned long flags
;
1279 for_each_populated_zone(zone
) {
1280 struct per_cpu_pageset
*pset
;
1281 struct per_cpu_pages
*pcp
;
1283 local_irq_save(flags
);
1284 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1288 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1291 local_irq_restore(flags
);
1296 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1298 void drain_local_pages(void *arg
)
1300 drain_pages(smp_processor_id());
1304 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1306 * Note that this code is protected against sending an IPI to an offline
1307 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1308 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1309 * nothing keeps CPUs from showing up after we populated the cpumask and
1310 * before the call to on_each_cpu_mask().
1312 void drain_all_pages(void)
1315 struct per_cpu_pageset
*pcp
;
1319 * Allocate in the BSS so we wont require allocation in
1320 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1322 static cpumask_t cpus_with_pcps
;
1325 * We don't care about racing with CPU hotplug event
1326 * as offline notification will cause the notified
1327 * cpu to drain that CPU pcps and on_each_cpu_mask
1328 * disables preemption as part of its processing
1330 for_each_online_cpu(cpu
) {
1331 bool has_pcps
= false;
1332 for_each_populated_zone(zone
) {
1333 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1334 if (pcp
->pcp
.count
) {
1340 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1342 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1344 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1347 #ifdef CONFIG_HIBERNATION
1349 void mark_free_pages(struct zone
*zone
)
1351 unsigned long pfn
, max_zone_pfn
;
1352 unsigned long flags
;
1353 unsigned int order
, t
;
1354 struct list_head
*curr
;
1356 if (zone_is_empty(zone
))
1359 spin_lock_irqsave(&zone
->lock
, flags
);
1361 max_zone_pfn
= zone_end_pfn(zone
);
1362 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1363 if (pfn_valid(pfn
)) {
1364 struct page
*page
= pfn_to_page(pfn
);
1366 if (!swsusp_page_is_forbidden(page
))
1367 swsusp_unset_page_free(page
);
1370 for_each_migratetype_order(order
, t
) {
1371 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1374 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1375 for (i
= 0; i
< (1UL << order
); i
++)
1376 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1379 spin_unlock_irqrestore(&zone
->lock
, flags
);
1381 #endif /* CONFIG_PM */
1384 * Free a 0-order page
1385 * cold == true ? free a cold page : free a hot page
1387 void free_hot_cold_page(struct page
*page
, bool cold
)
1389 struct zone
*zone
= page_zone(page
);
1390 struct per_cpu_pages
*pcp
;
1391 unsigned long flags
;
1392 unsigned long pfn
= page_to_pfn(page
);
1395 if (!free_pages_prepare(page
, 0))
1398 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1399 set_freepage_migratetype(page
, migratetype
);
1400 local_irq_save(flags
);
1401 __count_vm_event(PGFREE
);
1404 * We only track unmovable, reclaimable and movable on pcp lists.
1405 * Free ISOLATE pages back to the allocator because they are being
1406 * offlined but treat RESERVE as movable pages so we can get those
1407 * areas back if necessary. Otherwise, we may have to free
1408 * excessively into the page allocator
1410 if (migratetype
>= MIGRATE_PCPTYPES
) {
1411 if (unlikely(is_migrate_isolate(migratetype
))) {
1412 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1415 migratetype
= MIGRATE_MOVABLE
;
1418 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1420 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1422 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1424 if (pcp
->count
>= pcp
->high
) {
1425 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1426 free_pcppages_bulk(zone
, batch
, pcp
);
1427 pcp
->count
-= batch
;
1431 local_irq_restore(flags
);
1435 * Free a list of 0-order pages
1437 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1439 struct page
*page
, *next
;
1441 list_for_each_entry_safe(page
, next
, list
, lru
) {
1442 trace_mm_page_free_batched(page
, cold
);
1443 free_hot_cold_page(page
, cold
);
1448 * split_page takes a non-compound higher-order page, and splits it into
1449 * n (1<<order) sub-pages: page[0..n]
1450 * Each sub-page must be freed individually.
1452 * Note: this is probably too low level an operation for use in drivers.
1453 * Please consult with lkml before using this in your driver.
1455 void split_page(struct page
*page
, unsigned int order
)
1459 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1460 VM_BUG_ON_PAGE(!page_count(page
), page
);
1462 #ifdef CONFIG_KMEMCHECK
1464 * Split shadow pages too, because free(page[0]) would
1465 * otherwise free the whole shadow.
1467 if (kmemcheck_page_is_tracked(page
))
1468 split_page(virt_to_page(page
[0].shadow
), order
);
1471 for (i
= 1; i
< (1 << order
); i
++)
1472 set_page_refcounted(page
+ i
);
1474 EXPORT_SYMBOL_GPL(split_page
);
1476 int __isolate_free_page(struct page
*page
, unsigned int order
)
1478 unsigned long watermark
;
1482 BUG_ON(!PageBuddy(page
));
1484 zone
= page_zone(page
);
1485 mt
= get_pageblock_migratetype(page
);
1487 if (!is_migrate_isolate(mt
)) {
1488 /* Obey watermarks as if the page was being allocated */
1489 watermark
= low_wmark_pages(zone
) + (1 << order
);
1490 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1493 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1496 /* Remove page from free list */
1497 list_del(&page
->lru
);
1498 zone
->free_area
[order
].nr_free
--;
1499 rmv_page_order(page
);
1501 /* Set the pageblock if the isolated page is at least a pageblock */
1502 if (order
>= pageblock_order
- 1) {
1503 struct page
*endpage
= page
+ (1 << order
) - 1;
1504 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1505 int mt
= get_pageblock_migratetype(page
);
1506 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1507 set_pageblock_migratetype(page
,
1512 return 1UL << order
;
1516 * Similar to split_page except the page is already free. As this is only
1517 * being used for migration, the migratetype of the block also changes.
1518 * As this is called with interrupts disabled, the caller is responsible
1519 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1522 * Note: this is probably too low level an operation for use in drivers.
1523 * Please consult with lkml before using this in your driver.
1525 int split_free_page(struct page
*page
)
1530 order
= page_order(page
);
1532 nr_pages
= __isolate_free_page(page
, order
);
1536 /* Split into individual pages */
1537 set_page_refcounted(page
);
1538 split_page(page
, order
);
1543 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1544 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1548 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1549 struct zone
*zone
, unsigned int order
,
1550 gfp_t gfp_flags
, int migratetype
)
1552 unsigned long flags
;
1554 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1557 if (likely(order
== 0)) {
1558 struct per_cpu_pages
*pcp
;
1559 struct list_head
*list
;
1561 local_irq_save(flags
);
1562 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1563 list
= &pcp
->lists
[migratetype
];
1564 if (list_empty(list
)) {
1565 pcp
->count
+= rmqueue_bulk(zone
, 0,
1568 if (unlikely(list_empty(list
)))
1573 page
= list_entry(list
->prev
, struct page
, lru
);
1575 page
= list_entry(list
->next
, struct page
, lru
);
1577 list_del(&page
->lru
);
1580 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1582 * __GFP_NOFAIL is not to be used in new code.
1584 * All __GFP_NOFAIL callers should be fixed so that they
1585 * properly detect and handle allocation failures.
1587 * We most definitely don't want callers attempting to
1588 * allocate greater than order-1 page units with
1591 WARN_ON_ONCE(order
> 1);
1593 spin_lock_irqsave(&zone
->lock
, flags
);
1594 page
= __rmqueue(zone
, order
, migratetype
);
1595 spin_unlock(&zone
->lock
);
1598 __mod_zone_freepage_state(zone
, -(1 << order
),
1599 get_freepage_migratetype(page
));
1602 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1603 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1604 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1605 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1607 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1608 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1609 local_irq_restore(flags
);
1611 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1612 if (prep_new_page(page
, order
, gfp_flags
))
1617 local_irq_restore(flags
);
1621 #ifdef CONFIG_FAIL_PAGE_ALLOC
1624 struct fault_attr attr
;
1626 u32 ignore_gfp_highmem
;
1627 u32 ignore_gfp_wait
;
1629 } fail_page_alloc
= {
1630 .attr
= FAULT_ATTR_INITIALIZER
,
1631 .ignore_gfp_wait
= 1,
1632 .ignore_gfp_highmem
= 1,
1636 static int __init
setup_fail_page_alloc(char *str
)
1638 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1640 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1642 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1644 if (order
< fail_page_alloc
.min_order
)
1646 if (gfp_mask
& __GFP_NOFAIL
)
1648 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1650 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1653 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1656 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1658 static int __init
fail_page_alloc_debugfs(void)
1660 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1663 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1664 &fail_page_alloc
.attr
);
1666 return PTR_ERR(dir
);
1668 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1669 &fail_page_alloc
.ignore_gfp_wait
))
1671 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1672 &fail_page_alloc
.ignore_gfp_highmem
))
1674 if (!debugfs_create_u32("min-order", mode
, dir
,
1675 &fail_page_alloc
.min_order
))
1680 debugfs_remove_recursive(dir
);
1685 late_initcall(fail_page_alloc_debugfs
);
1687 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1689 #else /* CONFIG_FAIL_PAGE_ALLOC */
1691 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1696 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1699 * Return true if free pages are above 'mark'. This takes into account the order
1700 * of the allocation.
1702 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1703 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1706 /* free_pages my go negative - that's OK */
1711 free_pages
-= (1 << order
) - 1;
1712 if (alloc_flags
& ALLOC_HIGH
)
1714 if (alloc_flags
& ALLOC_HARDER
)
1717 /* If allocation can't use CMA areas don't use free CMA pages */
1718 if (!(alloc_flags
& ALLOC_CMA
))
1719 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1722 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1724 for (o
= 0; o
< order
; o
++) {
1725 /* At the next order, this order's pages become unavailable */
1726 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1728 /* Require fewer higher order pages to be free */
1731 if (free_pages
<= min
)
1737 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1738 int classzone_idx
, int alloc_flags
)
1740 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1741 zone_page_state(z
, NR_FREE_PAGES
));
1744 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1745 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1747 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1749 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1750 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1752 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1758 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1759 * skip over zones that are not allowed by the cpuset, or that have
1760 * been recently (in last second) found to be nearly full. See further
1761 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1762 * that have to skip over a lot of full or unallowed zones.
1764 * If the zonelist cache is present in the passed zonelist, then
1765 * returns a pointer to the allowed node mask (either the current
1766 * tasks mems_allowed, or node_states[N_MEMORY].)
1768 * If the zonelist cache is not available for this zonelist, does
1769 * nothing and returns NULL.
1771 * If the fullzones BITMAP in the zonelist cache is stale (more than
1772 * a second since last zap'd) then we zap it out (clear its bits.)
1774 * We hold off even calling zlc_setup, until after we've checked the
1775 * first zone in the zonelist, on the theory that most allocations will
1776 * be satisfied from that first zone, so best to examine that zone as
1777 * quickly as we can.
1779 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1781 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1782 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1784 zlc
= zonelist
->zlcache_ptr
;
1788 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1789 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1790 zlc
->last_full_zap
= jiffies
;
1793 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1794 &cpuset_current_mems_allowed
:
1795 &node_states
[N_MEMORY
];
1796 return allowednodes
;
1800 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1801 * if it is worth looking at further for free memory:
1802 * 1) Check that the zone isn't thought to be full (doesn't have its
1803 * bit set in the zonelist_cache fullzones BITMAP).
1804 * 2) Check that the zones node (obtained from the zonelist_cache
1805 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1806 * Return true (non-zero) if zone is worth looking at further, or
1807 * else return false (zero) if it is not.
1809 * This check -ignores- the distinction between various watermarks,
1810 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1811 * found to be full for any variation of these watermarks, it will
1812 * be considered full for up to one second by all requests, unless
1813 * we are so low on memory on all allowed nodes that we are forced
1814 * into the second scan of the zonelist.
1816 * In the second scan we ignore this zonelist cache and exactly
1817 * apply the watermarks to all zones, even it is slower to do so.
1818 * We are low on memory in the second scan, and should leave no stone
1819 * unturned looking for a free page.
1821 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1822 nodemask_t
*allowednodes
)
1824 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1825 int i
; /* index of *z in zonelist zones */
1826 int n
; /* node that zone *z is on */
1828 zlc
= zonelist
->zlcache_ptr
;
1832 i
= z
- zonelist
->_zonerefs
;
1835 /* This zone is worth trying if it is allowed but not full */
1836 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1840 * Given 'z' scanning a zonelist, set the corresponding bit in
1841 * zlc->fullzones, so that subsequent attempts to allocate a page
1842 * from that zone don't waste time re-examining it.
1844 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1846 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1847 int i
; /* index of *z in zonelist zones */
1849 zlc
= zonelist
->zlcache_ptr
;
1853 i
= z
- zonelist
->_zonerefs
;
1855 set_bit(i
, zlc
->fullzones
);
1859 * clear all zones full, called after direct reclaim makes progress so that
1860 * a zone that was recently full is not skipped over for up to a second
1862 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1864 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1866 zlc
= zonelist
->zlcache_ptr
;
1870 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1873 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1875 return local_zone
->node
== zone
->node
;
1878 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1880 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1884 #else /* CONFIG_NUMA */
1886 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1891 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1892 nodemask_t
*allowednodes
)
1897 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1901 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1905 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1910 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1915 #endif /* CONFIG_NUMA */
1917 static void reset_alloc_batches(struct zone
*preferred_zone
)
1919 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
1922 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
1923 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
1924 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
1925 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1926 } while (zone
++ != preferred_zone
);
1930 * get_page_from_freelist goes through the zonelist trying to allocate
1933 static struct page
*
1934 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1935 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1936 struct zone
*preferred_zone
, int classzone_idx
, int migratetype
)
1939 struct page
*page
= NULL
;
1941 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1942 int zlc_active
= 0; /* set if using zonelist_cache */
1943 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1944 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
1945 (gfp_mask
& __GFP_WRITE
);
1946 int nr_fair_skipped
= 0;
1947 bool zonelist_rescan
;
1950 zonelist_rescan
= false;
1953 * Scan zonelist, looking for a zone with enough free.
1954 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1956 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1957 high_zoneidx
, nodemask
) {
1960 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1961 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1963 if (cpusets_enabled() &&
1964 (alloc_flags
& ALLOC_CPUSET
) &&
1965 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1968 * Distribute pages in proportion to the individual
1969 * zone size to ensure fair page aging. The zone a
1970 * page was allocated in should have no effect on the
1971 * time the page has in memory before being reclaimed.
1973 if (alloc_flags
& ALLOC_FAIR
) {
1974 if (!zone_local(preferred_zone
, zone
))
1976 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
1982 * When allocating a page cache page for writing, we
1983 * want to get it from a zone that is within its dirty
1984 * limit, such that no single zone holds more than its
1985 * proportional share of globally allowed dirty pages.
1986 * The dirty limits take into account the zone's
1987 * lowmem reserves and high watermark so that kswapd
1988 * should be able to balance it without having to
1989 * write pages from its LRU list.
1991 * This may look like it could increase pressure on
1992 * lower zones by failing allocations in higher zones
1993 * before they are full. But the pages that do spill
1994 * over are limited as the lower zones are protected
1995 * by this very same mechanism. It should not become
1996 * a practical burden to them.
1998 * XXX: For now, allow allocations to potentially
1999 * exceed the per-zone dirty limit in the slowpath
2000 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2001 * which is important when on a NUMA setup the allowed
2002 * zones are together not big enough to reach the
2003 * global limit. The proper fix for these situations
2004 * will require awareness of zones in the
2005 * dirty-throttling and the flusher threads.
2007 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2010 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2011 if (!zone_watermark_ok(zone
, order
, mark
,
2012 classzone_idx
, alloc_flags
)) {
2015 /* Checked here to keep the fast path fast */
2016 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2017 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2020 if (IS_ENABLED(CONFIG_NUMA
) &&
2021 !did_zlc_setup
&& nr_online_nodes
> 1) {
2023 * we do zlc_setup if there are multiple nodes
2024 * and before considering the first zone allowed
2027 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2032 if (zone_reclaim_mode
== 0 ||
2033 !zone_allows_reclaim(preferred_zone
, zone
))
2034 goto this_zone_full
;
2037 * As we may have just activated ZLC, check if the first
2038 * eligible zone has failed zone_reclaim recently.
2040 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2041 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2044 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2046 case ZONE_RECLAIM_NOSCAN
:
2049 case ZONE_RECLAIM_FULL
:
2050 /* scanned but unreclaimable */
2053 /* did we reclaim enough */
2054 if (zone_watermark_ok(zone
, order
, mark
,
2055 classzone_idx
, alloc_flags
))
2059 * Failed to reclaim enough to meet watermark.
2060 * Only mark the zone full if checking the min
2061 * watermark or if we failed to reclaim just
2062 * 1<<order pages or else the page allocator
2063 * fastpath will prematurely mark zones full
2064 * when the watermark is between the low and
2067 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2068 ret
== ZONE_RECLAIM_SOME
)
2069 goto this_zone_full
;
2076 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2077 gfp_mask
, migratetype
);
2081 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2082 zlc_mark_zone_full(zonelist
, z
);
2087 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2088 * necessary to allocate the page. The expectation is
2089 * that the caller is taking steps that will free more
2090 * memory. The caller should avoid the page being used
2091 * for !PFMEMALLOC purposes.
2093 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2098 * The first pass makes sure allocations are spread fairly within the
2099 * local node. However, the local node might have free pages left
2100 * after the fairness batches are exhausted, and remote zones haven't
2101 * even been considered yet. Try once more without fairness, and
2102 * include remote zones now, before entering the slowpath and waking
2103 * kswapd: prefer spilling to a remote zone over swapping locally.
2105 if (alloc_flags
& ALLOC_FAIR
) {
2106 alloc_flags
&= ~ALLOC_FAIR
;
2107 if (nr_fair_skipped
) {
2108 zonelist_rescan
= true;
2109 reset_alloc_batches(preferred_zone
);
2111 if (nr_online_nodes
> 1)
2112 zonelist_rescan
= true;
2115 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2116 /* Disable zlc cache for second zonelist scan */
2118 zonelist_rescan
= true;
2121 if (zonelist_rescan
)
2128 * Large machines with many possible nodes should not always dump per-node
2129 * meminfo in irq context.
2131 static inline bool should_suppress_show_mem(void)
2136 ret
= in_interrupt();
2141 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2142 DEFAULT_RATELIMIT_INTERVAL
,
2143 DEFAULT_RATELIMIT_BURST
);
2145 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2147 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2149 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2150 debug_guardpage_minorder() > 0)
2154 * This documents exceptions given to allocations in certain
2155 * contexts that are allowed to allocate outside current's set
2158 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2159 if (test_thread_flag(TIF_MEMDIE
) ||
2160 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2161 filter
&= ~SHOW_MEM_FILTER_NODES
;
2162 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2163 filter
&= ~SHOW_MEM_FILTER_NODES
;
2166 struct va_format vaf
;
2169 va_start(args
, fmt
);
2174 pr_warn("%pV", &vaf
);
2179 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2180 current
->comm
, order
, gfp_mask
);
2183 if (!should_suppress_show_mem())
2188 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2189 unsigned long did_some_progress
,
2190 unsigned long pages_reclaimed
)
2192 /* Do not loop if specifically requested */
2193 if (gfp_mask
& __GFP_NORETRY
)
2196 /* Always retry if specifically requested */
2197 if (gfp_mask
& __GFP_NOFAIL
)
2201 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2202 * making forward progress without invoking OOM. Suspend also disables
2203 * storage devices so kswapd will not help. Bail if we are suspending.
2205 if (!did_some_progress
&& pm_suspended_storage())
2209 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2210 * means __GFP_NOFAIL, but that may not be true in other
2213 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2217 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2218 * specified, then we retry until we no longer reclaim any pages
2219 * (above), or we've reclaimed an order of pages at least as
2220 * large as the allocation's order. In both cases, if the
2221 * allocation still fails, we stop retrying.
2223 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2229 static inline struct page
*
2230 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2231 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2232 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2233 int classzone_idx
, int migratetype
)
2237 /* Acquire the per-zone oom lock for each zone */
2238 if (!oom_zonelist_trylock(zonelist
, gfp_mask
)) {
2239 schedule_timeout_uninterruptible(1);
2244 * PM-freezer should be notified that there might be an OOM killer on
2245 * its way to kill and wake somebody up. This is too early and we might
2246 * end up not killing anything but false positives are acceptable.
2247 * See freeze_processes.
2252 * Go through the zonelist yet one more time, keep very high watermark
2253 * here, this is only to catch a parallel oom killing, we must fail if
2254 * we're still under heavy pressure.
2256 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2257 order
, zonelist
, high_zoneidx
,
2258 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2259 preferred_zone
, classzone_idx
, migratetype
);
2263 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2264 /* The OOM killer will not help higher order allocs */
2265 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2267 /* The OOM killer does not needlessly kill tasks for lowmem */
2268 if (high_zoneidx
< ZONE_NORMAL
)
2271 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2272 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2273 * The caller should handle page allocation failure by itself if
2274 * it specifies __GFP_THISNODE.
2275 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2277 if (gfp_mask
& __GFP_THISNODE
)
2280 /* Exhausted what can be done so it's blamo time */
2281 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2284 oom_zonelist_unlock(zonelist
, gfp_mask
);
2288 #ifdef CONFIG_COMPACTION
2289 /* Try memory compaction for high-order allocations before reclaim */
2290 static struct page
*
2291 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2292 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2293 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2294 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2295 int *contended_compaction
, bool *deferred_compaction
)
2297 struct zone
*last_compact_zone
= NULL
;
2298 unsigned long compact_result
;
2304 current
->flags
|= PF_MEMALLOC
;
2305 compact_result
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2307 contended_compaction
,
2308 &last_compact_zone
);
2309 current
->flags
&= ~PF_MEMALLOC
;
2311 switch (compact_result
) {
2312 case COMPACT_DEFERRED
:
2313 *deferred_compaction
= true;
2315 case COMPACT_SKIPPED
:
2322 * At least in one zone compaction wasn't deferred or skipped, so let's
2323 * count a compaction stall
2325 count_vm_event(COMPACTSTALL
);
2327 /* Page migration frees to the PCP lists but we want merging */
2328 drain_pages(get_cpu());
2331 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2332 order
, zonelist
, high_zoneidx
,
2333 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2334 preferred_zone
, classzone_idx
, migratetype
);
2337 struct zone
*zone
= page_zone(page
);
2339 zone
->compact_blockskip_flush
= false;
2340 compaction_defer_reset(zone
, order
, true);
2341 count_vm_event(COMPACTSUCCESS
);
2346 * last_compact_zone is where try_to_compact_pages thought allocation
2347 * should succeed, so it did not defer compaction. But here we know
2348 * that it didn't succeed, so we do the defer.
2350 if (last_compact_zone
&& mode
!= MIGRATE_ASYNC
)
2351 defer_compaction(last_compact_zone
, order
);
2354 * It's bad if compaction run occurs and fails. The most likely reason
2355 * is that pages exist, but not enough to satisfy watermarks.
2357 count_vm_event(COMPACTFAIL
);
2364 static inline struct page
*
2365 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2366 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2367 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2368 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2369 int *contended_compaction
, bool *deferred_compaction
)
2373 #endif /* CONFIG_COMPACTION */
2375 /* Perform direct synchronous page reclaim */
2377 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2378 nodemask_t
*nodemask
)
2380 struct reclaim_state reclaim_state
;
2385 /* We now go into synchronous reclaim */
2386 cpuset_memory_pressure_bump();
2387 current
->flags
|= PF_MEMALLOC
;
2388 lockdep_set_current_reclaim_state(gfp_mask
);
2389 reclaim_state
.reclaimed_slab
= 0;
2390 current
->reclaim_state
= &reclaim_state
;
2392 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2394 current
->reclaim_state
= NULL
;
2395 lockdep_clear_current_reclaim_state();
2396 current
->flags
&= ~PF_MEMALLOC
;
2403 /* The really slow allocator path where we enter direct reclaim */
2404 static inline struct page
*
2405 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2406 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2407 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2408 int classzone_idx
, int migratetype
, unsigned long *did_some_progress
)
2410 struct page
*page
= NULL
;
2411 bool drained
= false;
2413 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2415 if (unlikely(!(*did_some_progress
)))
2418 /* After successful reclaim, reconsider all zones for allocation */
2419 if (IS_ENABLED(CONFIG_NUMA
))
2420 zlc_clear_zones_full(zonelist
);
2423 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2424 zonelist
, high_zoneidx
,
2425 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2426 preferred_zone
, classzone_idx
,
2430 * If an allocation failed after direct reclaim, it could be because
2431 * pages are pinned on the per-cpu lists. Drain them and try again
2433 if (!page
&& !drained
) {
2443 * This is called in the allocator slow-path if the allocation request is of
2444 * sufficient urgency to ignore watermarks and take other desperate measures
2446 static inline struct page
*
2447 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2448 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2449 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2450 int classzone_idx
, int migratetype
)
2455 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2456 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2457 preferred_zone
, classzone_idx
, migratetype
);
2459 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2460 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2461 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2466 static void wake_all_kswapds(unsigned int order
,
2467 struct zonelist
*zonelist
,
2468 enum zone_type high_zoneidx
,
2469 struct zone
*preferred_zone
,
2470 nodemask_t
*nodemask
)
2475 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2476 high_zoneidx
, nodemask
)
2477 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2481 gfp_to_alloc_flags(gfp_t gfp_mask
)
2483 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2484 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2486 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2487 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2490 * The caller may dip into page reserves a bit more if the caller
2491 * cannot run direct reclaim, or if the caller has realtime scheduling
2492 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2493 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2495 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2499 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2500 * if it can't schedule.
2502 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2503 alloc_flags
|= ALLOC_HARDER
;
2505 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2506 * comment for __cpuset_node_allowed_softwall().
2508 alloc_flags
&= ~ALLOC_CPUSET
;
2509 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2510 alloc_flags
|= ALLOC_HARDER
;
2512 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2513 if (gfp_mask
& __GFP_MEMALLOC
)
2514 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2515 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2516 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2517 else if (!in_interrupt() &&
2518 ((current
->flags
& PF_MEMALLOC
) ||
2519 unlikely(test_thread_flag(TIF_MEMDIE
))))
2520 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2523 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2524 alloc_flags
|= ALLOC_CMA
;
2529 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2531 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2534 static inline struct page
*
2535 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2536 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2537 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2538 int classzone_idx
, int migratetype
)
2540 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2541 struct page
*page
= NULL
;
2543 unsigned long pages_reclaimed
= 0;
2544 unsigned long did_some_progress
;
2545 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2546 bool deferred_compaction
= false;
2547 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2550 * In the slowpath, we sanity check order to avoid ever trying to
2551 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2552 * be using allocators in order of preference for an area that is
2555 if (order
>= MAX_ORDER
) {
2556 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2561 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2562 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2563 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2564 * using a larger set of nodes after it has established that the
2565 * allowed per node queues are empty and that nodes are
2568 if (IS_ENABLED(CONFIG_NUMA
) &&
2569 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2573 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2574 wake_all_kswapds(order
, zonelist
, high_zoneidx
,
2575 preferred_zone
, nodemask
);
2578 * OK, we're below the kswapd watermark and have kicked background
2579 * reclaim. Now things get more complex, so set up alloc_flags according
2580 * to how we want to proceed.
2582 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2585 * Find the true preferred zone if the allocation is unconstrained by
2588 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
) {
2589 struct zoneref
*preferred_zoneref
;
2590 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2591 NULL
, &preferred_zone
);
2592 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2596 /* This is the last chance, in general, before the goto nopage. */
2597 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2598 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2599 preferred_zone
, classzone_idx
, migratetype
);
2603 /* Allocate without watermarks if the context allows */
2604 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2606 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2607 * the allocation is high priority and these type of
2608 * allocations are system rather than user orientated
2610 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2612 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2613 zonelist
, high_zoneidx
, nodemask
,
2614 preferred_zone
, classzone_idx
, migratetype
);
2620 /* Atomic allocations - we can't balance anything */
2623 * All existing users of the deprecated __GFP_NOFAIL are
2624 * blockable, so warn of any new users that actually allow this
2625 * type of allocation to fail.
2627 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2631 /* Avoid recursion of direct reclaim */
2632 if (current
->flags
& PF_MEMALLOC
)
2635 /* Avoid allocations with no watermarks from looping endlessly */
2636 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2640 * Try direct compaction. The first pass is asynchronous. Subsequent
2641 * attempts after direct reclaim are synchronous
2643 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2644 high_zoneidx
, nodemask
, alloc_flags
,
2646 classzone_idx
, migratetype
,
2647 migration_mode
, &contended_compaction
,
2648 &deferred_compaction
);
2652 /* Checks for THP-specific high-order allocations */
2653 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2655 * If compaction is deferred for high-order allocations, it is
2656 * because sync compaction recently failed. If this is the case
2657 * and the caller requested a THP allocation, we do not want
2658 * to heavily disrupt the system, so we fail the allocation
2659 * instead of entering direct reclaim.
2661 if (deferred_compaction
)
2665 * In all zones where compaction was attempted (and not
2666 * deferred or skipped), lock contention has been detected.
2667 * For THP allocation we do not want to disrupt the others
2668 * so we fallback to base pages instead.
2670 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2674 * If compaction was aborted due to need_resched(), we do not
2675 * want to further increase allocation latency, unless it is
2676 * khugepaged trying to collapse.
2678 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2679 && !(current
->flags
& PF_KTHREAD
))
2684 * It can become very expensive to allocate transparent hugepages at
2685 * fault, so use asynchronous memory compaction for THP unless it is
2686 * khugepaged trying to collapse.
2688 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2689 (current
->flags
& PF_KTHREAD
))
2690 migration_mode
= MIGRATE_SYNC_LIGHT
;
2692 /* Try direct reclaim and then allocating */
2693 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2694 zonelist
, high_zoneidx
,
2696 alloc_flags
, preferred_zone
,
2697 classzone_idx
, migratetype
,
2698 &did_some_progress
);
2703 * If we failed to make any progress reclaiming, then we are
2704 * running out of options and have to consider going OOM
2706 if (!did_some_progress
) {
2707 if (oom_gfp_allowed(gfp_mask
)) {
2708 if (oom_killer_disabled
)
2710 /* Coredumps can quickly deplete all memory reserves */
2711 if ((current
->flags
& PF_DUMPCORE
) &&
2712 !(gfp_mask
& __GFP_NOFAIL
))
2714 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2715 zonelist
, high_zoneidx
,
2716 nodemask
, preferred_zone
,
2717 classzone_idx
, migratetype
);
2721 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2723 * The oom killer is not called for high-order
2724 * allocations that may fail, so if no progress
2725 * is being made, there are no other options and
2726 * retrying is unlikely to help.
2728 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2731 * The oom killer is not called for lowmem
2732 * allocations to prevent needlessly killing
2735 if (high_zoneidx
< ZONE_NORMAL
)
2743 /* Check if we should retry the allocation */
2744 pages_reclaimed
+= did_some_progress
;
2745 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2747 /* Wait for some write requests to complete then retry */
2748 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2752 * High-order allocations do not necessarily loop after
2753 * direct reclaim and reclaim/compaction depends on compaction
2754 * being called after reclaim so call directly if necessary
2756 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2757 high_zoneidx
, nodemask
, alloc_flags
,
2759 classzone_idx
, migratetype
,
2760 migration_mode
, &contended_compaction
,
2761 &deferred_compaction
);
2767 warn_alloc_failed(gfp_mask
, order
, NULL
);
2770 if (kmemcheck_enabled
)
2771 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2777 * This is the 'heart' of the zoned buddy allocator.
2780 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2781 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2783 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2784 struct zone
*preferred_zone
;
2785 struct zoneref
*preferred_zoneref
;
2786 struct page
*page
= NULL
;
2787 int migratetype
= gfpflags_to_migratetype(gfp_mask
);
2788 unsigned int cpuset_mems_cookie
;
2789 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2792 gfp_mask
&= gfp_allowed_mask
;
2794 lockdep_trace_alloc(gfp_mask
);
2796 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2798 if (should_fail_alloc_page(gfp_mask
, order
))
2802 * Check the zones suitable for the gfp_mask contain at least one
2803 * valid zone. It's possible to have an empty zonelist as a result
2804 * of GFP_THISNODE and a memoryless node
2806 if (unlikely(!zonelist
->_zonerefs
->zone
))
2809 if (IS_ENABLED(CONFIG_CMA
) && migratetype
== MIGRATE_MOVABLE
)
2810 alloc_flags
|= ALLOC_CMA
;
2813 cpuset_mems_cookie
= read_mems_allowed_begin();
2815 /* The preferred zone is used for statistics later */
2816 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2817 nodemask
? : &cpuset_current_mems_allowed
,
2819 if (!preferred_zone
)
2821 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2823 /* First allocation attempt */
2824 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2825 zonelist
, high_zoneidx
, alloc_flags
,
2826 preferred_zone
, classzone_idx
, migratetype
);
2827 if (unlikely(!page
)) {
2829 * Runtime PM, block IO and its error handling path
2830 * can deadlock because I/O on the device might not
2833 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2834 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2835 zonelist
, high_zoneidx
, nodemask
,
2836 preferred_zone
, classzone_idx
, migratetype
);
2839 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2843 * When updating a task's mems_allowed, it is possible to race with
2844 * parallel threads in such a way that an allocation can fail while
2845 * the mask is being updated. If a page allocation is about to fail,
2846 * check if the cpuset changed during allocation and if so, retry.
2848 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2853 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2856 * Common helper functions.
2858 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2863 * __get_free_pages() returns a 32-bit address, which cannot represent
2866 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2868 page
= alloc_pages(gfp_mask
, order
);
2871 return (unsigned long) page_address(page
);
2873 EXPORT_SYMBOL(__get_free_pages
);
2875 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2877 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2879 EXPORT_SYMBOL(get_zeroed_page
);
2881 void __free_pages(struct page
*page
, unsigned int order
)
2883 if (put_page_testzero(page
)) {
2885 free_hot_cold_page(page
, false);
2887 __free_pages_ok(page
, order
);
2891 EXPORT_SYMBOL(__free_pages
);
2893 void free_pages(unsigned long addr
, unsigned int order
)
2896 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2897 __free_pages(virt_to_page((void *)addr
), order
);
2901 EXPORT_SYMBOL(free_pages
);
2904 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2905 * of the current memory cgroup.
2907 * It should be used when the caller would like to use kmalloc, but since the
2908 * allocation is large, it has to fall back to the page allocator.
2910 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2913 struct mem_cgroup
*memcg
= NULL
;
2915 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2917 page
= alloc_pages(gfp_mask
, order
);
2918 memcg_kmem_commit_charge(page
, memcg
, order
);
2922 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2925 struct mem_cgroup
*memcg
= NULL
;
2927 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2929 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2930 memcg_kmem_commit_charge(page
, memcg
, order
);
2935 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2938 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2940 memcg_kmem_uncharge_pages(page
, order
);
2941 __free_pages(page
, order
);
2944 void free_kmem_pages(unsigned long addr
, unsigned int order
)
2947 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2948 __free_kmem_pages(virt_to_page((void *)addr
), order
);
2952 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2955 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2956 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2958 split_page(virt_to_page((void *)addr
), order
);
2959 while (used
< alloc_end
) {
2964 return (void *)addr
;
2968 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2969 * @size: the number of bytes to allocate
2970 * @gfp_mask: GFP flags for the allocation
2972 * This function is similar to alloc_pages(), except that it allocates the
2973 * minimum number of pages to satisfy the request. alloc_pages() can only
2974 * allocate memory in power-of-two pages.
2976 * This function is also limited by MAX_ORDER.
2978 * Memory allocated by this function must be released by free_pages_exact().
2980 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2982 unsigned int order
= get_order(size
);
2985 addr
= __get_free_pages(gfp_mask
, order
);
2986 return make_alloc_exact(addr
, order
, size
);
2988 EXPORT_SYMBOL(alloc_pages_exact
);
2991 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2993 * @nid: the preferred node ID where memory should be allocated
2994 * @size: the number of bytes to allocate
2995 * @gfp_mask: GFP flags for the allocation
2997 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2999 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3002 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3004 unsigned order
= get_order(size
);
3005 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3008 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3012 * free_pages_exact - release memory allocated via alloc_pages_exact()
3013 * @virt: the value returned by alloc_pages_exact.
3014 * @size: size of allocation, same value as passed to alloc_pages_exact().
3016 * Release the memory allocated by a previous call to alloc_pages_exact.
3018 void free_pages_exact(void *virt
, size_t size
)
3020 unsigned long addr
= (unsigned long)virt
;
3021 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3023 while (addr
< end
) {
3028 EXPORT_SYMBOL(free_pages_exact
);
3031 * nr_free_zone_pages - count number of pages beyond high watermark
3032 * @offset: The zone index of the highest zone
3034 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3035 * high watermark within all zones at or below a given zone index. For each
3036 * zone, the number of pages is calculated as:
3037 * managed_pages - high_pages
3039 static unsigned long nr_free_zone_pages(int offset
)
3044 /* Just pick one node, since fallback list is circular */
3045 unsigned long sum
= 0;
3047 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3049 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3050 unsigned long size
= zone
->managed_pages
;
3051 unsigned long high
= high_wmark_pages(zone
);
3060 * nr_free_buffer_pages - count number of pages beyond high watermark
3062 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3063 * watermark within ZONE_DMA and ZONE_NORMAL.
3065 unsigned long nr_free_buffer_pages(void)
3067 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3069 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3072 * nr_free_pagecache_pages - count number of pages beyond high watermark
3074 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3075 * high watermark within all zones.
3077 unsigned long nr_free_pagecache_pages(void)
3079 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3082 static inline void show_node(struct zone
*zone
)
3084 if (IS_ENABLED(CONFIG_NUMA
))
3085 printk("Node %d ", zone_to_nid(zone
));
3088 void si_meminfo(struct sysinfo
*val
)
3090 val
->totalram
= totalram_pages
;
3091 val
->sharedram
= global_page_state(NR_SHMEM
);
3092 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3093 val
->bufferram
= nr_blockdev_pages();
3094 val
->totalhigh
= totalhigh_pages
;
3095 val
->freehigh
= nr_free_highpages();
3096 val
->mem_unit
= PAGE_SIZE
;
3099 EXPORT_SYMBOL(si_meminfo
);
3102 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3104 int zone_type
; /* needs to be signed */
3105 unsigned long managed_pages
= 0;
3106 pg_data_t
*pgdat
= NODE_DATA(nid
);
3108 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3109 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3110 val
->totalram
= managed_pages
;
3111 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3112 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3113 #ifdef CONFIG_HIGHMEM
3114 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3115 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3121 val
->mem_unit
= PAGE_SIZE
;
3126 * Determine whether the node should be displayed or not, depending on whether
3127 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3129 bool skip_free_areas_node(unsigned int flags
, int nid
)
3132 unsigned int cpuset_mems_cookie
;
3134 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3138 cpuset_mems_cookie
= read_mems_allowed_begin();
3139 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3140 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3145 #define K(x) ((x) << (PAGE_SHIFT-10))
3147 static void show_migration_types(unsigned char type
)
3149 static const char types
[MIGRATE_TYPES
] = {
3150 [MIGRATE_UNMOVABLE
] = 'U',
3151 [MIGRATE_RECLAIMABLE
] = 'E',
3152 [MIGRATE_MOVABLE
] = 'M',
3153 [MIGRATE_RESERVE
] = 'R',
3155 [MIGRATE_CMA
] = 'C',
3157 #ifdef CONFIG_MEMORY_ISOLATION
3158 [MIGRATE_ISOLATE
] = 'I',
3161 char tmp
[MIGRATE_TYPES
+ 1];
3165 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3166 if (type
& (1 << i
))
3171 printk("(%s) ", tmp
);
3175 * Show free area list (used inside shift_scroll-lock stuff)
3176 * We also calculate the percentage fragmentation. We do this by counting the
3177 * memory on each free list with the exception of the first item on the list.
3178 * Suppresses nodes that are not allowed by current's cpuset if
3179 * SHOW_MEM_FILTER_NODES is passed.
3181 void show_free_areas(unsigned int filter
)
3186 for_each_populated_zone(zone
) {
3187 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3190 printk("%s per-cpu:\n", zone
->name
);
3192 for_each_online_cpu(cpu
) {
3193 struct per_cpu_pageset
*pageset
;
3195 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3197 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3198 cpu
, pageset
->pcp
.high
,
3199 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3203 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3204 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3206 " dirty:%lu writeback:%lu unstable:%lu\n"
3207 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3208 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3210 global_page_state(NR_ACTIVE_ANON
),
3211 global_page_state(NR_INACTIVE_ANON
),
3212 global_page_state(NR_ISOLATED_ANON
),
3213 global_page_state(NR_ACTIVE_FILE
),
3214 global_page_state(NR_INACTIVE_FILE
),
3215 global_page_state(NR_ISOLATED_FILE
),
3216 global_page_state(NR_UNEVICTABLE
),
3217 global_page_state(NR_FILE_DIRTY
),
3218 global_page_state(NR_WRITEBACK
),
3219 global_page_state(NR_UNSTABLE_NFS
),
3220 global_page_state(NR_FREE_PAGES
),
3221 global_page_state(NR_SLAB_RECLAIMABLE
),
3222 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3223 global_page_state(NR_FILE_MAPPED
),
3224 global_page_state(NR_SHMEM
),
3225 global_page_state(NR_PAGETABLE
),
3226 global_page_state(NR_BOUNCE
),
3227 global_page_state(NR_FREE_CMA_PAGES
));
3229 for_each_populated_zone(zone
) {
3232 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3240 " active_anon:%lukB"
3241 " inactive_anon:%lukB"
3242 " active_file:%lukB"
3243 " inactive_file:%lukB"
3244 " unevictable:%lukB"
3245 " isolated(anon):%lukB"
3246 " isolated(file):%lukB"
3254 " slab_reclaimable:%lukB"
3255 " slab_unreclaimable:%lukB"
3256 " kernel_stack:%lukB"
3261 " writeback_tmp:%lukB"
3262 " pages_scanned:%lu"
3263 " all_unreclaimable? %s"
3266 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3267 K(min_wmark_pages(zone
)),
3268 K(low_wmark_pages(zone
)),
3269 K(high_wmark_pages(zone
)),
3270 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3271 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3272 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3273 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3274 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3275 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3276 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3277 K(zone
->present_pages
),
3278 K(zone
->managed_pages
),
3279 K(zone_page_state(zone
, NR_MLOCK
)),
3280 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3281 K(zone_page_state(zone
, NR_WRITEBACK
)),
3282 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3283 K(zone_page_state(zone
, NR_SHMEM
)),
3284 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3285 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3286 zone_page_state(zone
, NR_KERNEL_STACK
) *
3288 K(zone_page_state(zone
, NR_PAGETABLE
)),
3289 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3290 K(zone_page_state(zone
, NR_BOUNCE
)),
3291 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3292 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3293 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3294 (!zone_reclaimable(zone
) ? "yes" : "no")
3296 printk("lowmem_reserve[]:");
3297 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3298 printk(" %ld", zone
->lowmem_reserve
[i
]);
3302 for_each_populated_zone(zone
) {
3303 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3304 unsigned char types
[MAX_ORDER
];
3306 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3309 printk("%s: ", zone
->name
);
3311 spin_lock_irqsave(&zone
->lock
, flags
);
3312 for (order
= 0; order
< MAX_ORDER
; order
++) {
3313 struct free_area
*area
= &zone
->free_area
[order
];
3316 nr
[order
] = area
->nr_free
;
3317 total
+= nr
[order
] << order
;
3320 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3321 if (!list_empty(&area
->free_list
[type
]))
3322 types
[order
] |= 1 << type
;
3325 spin_unlock_irqrestore(&zone
->lock
, flags
);
3326 for (order
= 0; order
< MAX_ORDER
; order
++) {
3327 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3329 show_migration_types(types
[order
]);
3331 printk("= %lukB\n", K(total
));
3334 hugetlb_show_meminfo();
3336 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3338 show_swap_cache_info();
3341 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3343 zoneref
->zone
= zone
;
3344 zoneref
->zone_idx
= zone_idx(zone
);
3348 * Builds allocation fallback zone lists.
3350 * Add all populated zones of a node to the zonelist.
3352 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3356 enum zone_type zone_type
= MAX_NR_ZONES
;
3360 zone
= pgdat
->node_zones
+ zone_type
;
3361 if (populated_zone(zone
)) {
3362 zoneref_set_zone(zone
,
3363 &zonelist
->_zonerefs
[nr_zones
++]);
3364 check_highest_zone(zone_type
);
3366 } while (zone_type
);
3374 * 0 = automatic detection of better ordering.
3375 * 1 = order by ([node] distance, -zonetype)
3376 * 2 = order by (-zonetype, [node] distance)
3378 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3379 * the same zonelist. So only NUMA can configure this param.
3381 #define ZONELIST_ORDER_DEFAULT 0
3382 #define ZONELIST_ORDER_NODE 1
3383 #define ZONELIST_ORDER_ZONE 2
3385 /* zonelist order in the kernel.
3386 * set_zonelist_order() will set this to NODE or ZONE.
3388 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3389 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3393 /* The value user specified ....changed by config */
3394 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3395 /* string for sysctl */
3396 #define NUMA_ZONELIST_ORDER_LEN 16
3397 char numa_zonelist_order
[16] = "default";
3400 * interface for configure zonelist ordering.
3401 * command line option "numa_zonelist_order"
3402 * = "[dD]efault - default, automatic configuration.
3403 * = "[nN]ode - order by node locality, then by zone within node
3404 * = "[zZ]one - order by zone, then by locality within zone
3407 static int __parse_numa_zonelist_order(char *s
)
3409 if (*s
== 'd' || *s
== 'D') {
3410 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3411 } else if (*s
== 'n' || *s
== 'N') {
3412 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3413 } else if (*s
== 'z' || *s
== 'Z') {
3414 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3417 "Ignoring invalid numa_zonelist_order value: "
3424 static __init
int setup_numa_zonelist_order(char *s
)
3431 ret
= __parse_numa_zonelist_order(s
);
3433 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3437 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3440 * sysctl handler for numa_zonelist_order
3442 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3443 void __user
*buffer
, size_t *length
,
3446 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3448 static DEFINE_MUTEX(zl_order_mutex
);
3450 mutex_lock(&zl_order_mutex
);
3452 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3456 strcpy(saved_string
, (char *)table
->data
);
3458 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3462 int oldval
= user_zonelist_order
;
3464 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3467 * bogus value. restore saved string
3469 strncpy((char *)table
->data
, saved_string
,
3470 NUMA_ZONELIST_ORDER_LEN
);
3471 user_zonelist_order
= oldval
;
3472 } else if (oldval
!= user_zonelist_order
) {
3473 mutex_lock(&zonelists_mutex
);
3474 build_all_zonelists(NULL
, NULL
);
3475 mutex_unlock(&zonelists_mutex
);
3479 mutex_unlock(&zl_order_mutex
);
3484 #define MAX_NODE_LOAD (nr_online_nodes)
3485 static int node_load
[MAX_NUMNODES
];
3488 * find_next_best_node - find the next node that should appear in a given node's fallback list
3489 * @node: node whose fallback list we're appending
3490 * @used_node_mask: nodemask_t of already used nodes
3492 * We use a number of factors to determine which is the next node that should
3493 * appear on a given node's fallback list. The node should not have appeared
3494 * already in @node's fallback list, and it should be the next closest node
3495 * according to the distance array (which contains arbitrary distance values
3496 * from each node to each node in the system), and should also prefer nodes
3497 * with no CPUs, since presumably they'll have very little allocation pressure
3498 * on them otherwise.
3499 * It returns -1 if no node is found.
3501 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3504 int min_val
= INT_MAX
;
3505 int best_node
= NUMA_NO_NODE
;
3506 const struct cpumask
*tmp
= cpumask_of_node(0);
3508 /* Use the local node if we haven't already */
3509 if (!node_isset(node
, *used_node_mask
)) {
3510 node_set(node
, *used_node_mask
);
3514 for_each_node_state(n
, N_MEMORY
) {
3516 /* Don't want a node to appear more than once */
3517 if (node_isset(n
, *used_node_mask
))
3520 /* Use the distance array to find the distance */
3521 val
= node_distance(node
, n
);
3523 /* Penalize nodes under us ("prefer the next node") */
3526 /* Give preference to headless and unused nodes */
3527 tmp
= cpumask_of_node(n
);
3528 if (!cpumask_empty(tmp
))
3529 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3531 /* Slight preference for less loaded node */
3532 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3533 val
+= node_load
[n
];
3535 if (val
< min_val
) {
3542 node_set(best_node
, *used_node_mask
);
3549 * Build zonelists ordered by node and zones within node.
3550 * This results in maximum locality--normal zone overflows into local
3551 * DMA zone, if any--but risks exhausting DMA zone.
3553 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3556 struct zonelist
*zonelist
;
3558 zonelist
= &pgdat
->node_zonelists
[0];
3559 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3561 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3562 zonelist
->_zonerefs
[j
].zone
= NULL
;
3563 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3567 * Build gfp_thisnode zonelists
3569 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3572 struct zonelist
*zonelist
;
3574 zonelist
= &pgdat
->node_zonelists
[1];
3575 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3576 zonelist
->_zonerefs
[j
].zone
= NULL
;
3577 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3581 * Build zonelists ordered by zone and nodes within zones.
3582 * This results in conserving DMA zone[s] until all Normal memory is
3583 * exhausted, but results in overflowing to remote node while memory
3584 * may still exist in local DMA zone.
3586 static int node_order
[MAX_NUMNODES
];
3588 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3591 int zone_type
; /* needs to be signed */
3593 struct zonelist
*zonelist
;
3595 zonelist
= &pgdat
->node_zonelists
[0];
3597 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3598 for (j
= 0; j
< nr_nodes
; j
++) {
3599 node
= node_order
[j
];
3600 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3601 if (populated_zone(z
)) {
3603 &zonelist
->_zonerefs
[pos
++]);
3604 check_highest_zone(zone_type
);
3608 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3609 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3612 #if defined(CONFIG_64BIT)
3614 * Devices that require DMA32/DMA are relatively rare and do not justify a
3615 * penalty to every machine in case the specialised case applies. Default
3616 * to Node-ordering on 64-bit NUMA machines
3618 static int default_zonelist_order(void)
3620 return ZONELIST_ORDER_NODE
;
3624 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3625 * by the kernel. If processes running on node 0 deplete the low memory zone
3626 * then reclaim will occur more frequency increasing stalls and potentially
3627 * be easier to OOM if a large percentage of the zone is under writeback or
3628 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3629 * Hence, default to zone ordering on 32-bit.
3631 static int default_zonelist_order(void)
3633 return ZONELIST_ORDER_ZONE
;
3635 #endif /* CONFIG_64BIT */
3637 static void set_zonelist_order(void)
3639 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3640 current_zonelist_order
= default_zonelist_order();
3642 current_zonelist_order
= user_zonelist_order
;
3645 static void build_zonelists(pg_data_t
*pgdat
)
3649 nodemask_t used_mask
;
3650 int local_node
, prev_node
;
3651 struct zonelist
*zonelist
;
3652 int order
= current_zonelist_order
;
3654 /* initialize zonelists */
3655 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3656 zonelist
= pgdat
->node_zonelists
+ i
;
3657 zonelist
->_zonerefs
[0].zone
= NULL
;
3658 zonelist
->_zonerefs
[0].zone_idx
= 0;
3661 /* NUMA-aware ordering of nodes */
3662 local_node
= pgdat
->node_id
;
3663 load
= nr_online_nodes
;
3664 prev_node
= local_node
;
3665 nodes_clear(used_mask
);
3667 memset(node_order
, 0, sizeof(node_order
));
3670 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3672 * We don't want to pressure a particular node.
3673 * So adding penalty to the first node in same
3674 * distance group to make it round-robin.
3676 if (node_distance(local_node
, node
) !=
3677 node_distance(local_node
, prev_node
))
3678 node_load
[node
] = load
;
3682 if (order
== ZONELIST_ORDER_NODE
)
3683 build_zonelists_in_node_order(pgdat
, node
);
3685 node_order
[j
++] = node
; /* remember order */
3688 if (order
== ZONELIST_ORDER_ZONE
) {
3689 /* calculate node order -- i.e., DMA last! */
3690 build_zonelists_in_zone_order(pgdat
, j
);
3693 build_thisnode_zonelists(pgdat
);
3696 /* Construct the zonelist performance cache - see further mmzone.h */
3697 static void build_zonelist_cache(pg_data_t
*pgdat
)
3699 struct zonelist
*zonelist
;
3700 struct zonelist_cache
*zlc
;
3703 zonelist
= &pgdat
->node_zonelists
[0];
3704 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3705 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3706 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3707 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3710 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3712 * Return node id of node used for "local" allocations.
3713 * I.e., first node id of first zone in arg node's generic zonelist.
3714 * Used for initializing percpu 'numa_mem', which is used primarily
3715 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3717 int local_memory_node(int node
)
3721 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3722 gfp_zone(GFP_KERNEL
),
3729 #else /* CONFIG_NUMA */
3731 static void set_zonelist_order(void)
3733 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3736 static void build_zonelists(pg_data_t
*pgdat
)
3738 int node
, local_node
;
3740 struct zonelist
*zonelist
;
3742 local_node
= pgdat
->node_id
;
3744 zonelist
= &pgdat
->node_zonelists
[0];
3745 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3748 * Now we build the zonelist so that it contains the zones
3749 * of all the other nodes.
3750 * We don't want to pressure a particular node, so when
3751 * building the zones for node N, we make sure that the
3752 * zones coming right after the local ones are those from
3753 * node N+1 (modulo N)
3755 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3756 if (!node_online(node
))
3758 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3760 for (node
= 0; node
< local_node
; node
++) {
3761 if (!node_online(node
))
3763 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3766 zonelist
->_zonerefs
[j
].zone
= NULL
;
3767 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3770 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3771 static void build_zonelist_cache(pg_data_t
*pgdat
)
3773 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3776 #endif /* CONFIG_NUMA */
3779 * Boot pageset table. One per cpu which is going to be used for all
3780 * zones and all nodes. The parameters will be set in such a way
3781 * that an item put on a list will immediately be handed over to
3782 * the buddy list. This is safe since pageset manipulation is done
3783 * with interrupts disabled.
3785 * The boot_pagesets must be kept even after bootup is complete for
3786 * unused processors and/or zones. They do play a role for bootstrapping
3787 * hotplugged processors.
3789 * zoneinfo_show() and maybe other functions do
3790 * not check if the processor is online before following the pageset pointer.
3791 * Other parts of the kernel may not check if the zone is available.
3793 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3794 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3795 static void setup_zone_pageset(struct zone
*zone
);
3798 * Global mutex to protect against size modification of zonelists
3799 * as well as to serialize pageset setup for the new populated zone.
3801 DEFINE_MUTEX(zonelists_mutex
);
3803 /* return values int ....just for stop_machine() */
3804 static int __build_all_zonelists(void *data
)
3808 pg_data_t
*self
= data
;
3811 memset(node_load
, 0, sizeof(node_load
));
3814 if (self
&& !node_online(self
->node_id
)) {
3815 build_zonelists(self
);
3816 build_zonelist_cache(self
);
3819 for_each_online_node(nid
) {
3820 pg_data_t
*pgdat
= NODE_DATA(nid
);
3822 build_zonelists(pgdat
);
3823 build_zonelist_cache(pgdat
);
3827 * Initialize the boot_pagesets that are going to be used
3828 * for bootstrapping processors. The real pagesets for
3829 * each zone will be allocated later when the per cpu
3830 * allocator is available.
3832 * boot_pagesets are used also for bootstrapping offline
3833 * cpus if the system is already booted because the pagesets
3834 * are needed to initialize allocators on a specific cpu too.
3835 * F.e. the percpu allocator needs the page allocator which
3836 * needs the percpu allocator in order to allocate its pagesets
3837 * (a chicken-egg dilemma).
3839 for_each_possible_cpu(cpu
) {
3840 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3842 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3844 * We now know the "local memory node" for each node--
3845 * i.e., the node of the first zone in the generic zonelist.
3846 * Set up numa_mem percpu variable for on-line cpus. During
3847 * boot, only the boot cpu should be on-line; we'll init the
3848 * secondary cpus' numa_mem as they come on-line. During
3849 * node/memory hotplug, we'll fixup all on-line cpus.
3851 if (cpu_online(cpu
))
3852 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3860 * Called with zonelists_mutex held always
3861 * unless system_state == SYSTEM_BOOTING.
3863 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3865 set_zonelist_order();
3867 if (system_state
== SYSTEM_BOOTING
) {
3868 __build_all_zonelists(NULL
);
3869 mminit_verify_zonelist();
3870 cpuset_init_current_mems_allowed();
3872 #ifdef CONFIG_MEMORY_HOTPLUG
3874 setup_zone_pageset(zone
);
3876 /* we have to stop all cpus to guarantee there is no user
3878 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3879 /* cpuset refresh routine should be here */
3881 vm_total_pages
= nr_free_pagecache_pages();
3883 * Disable grouping by mobility if the number of pages in the
3884 * system is too low to allow the mechanism to work. It would be
3885 * more accurate, but expensive to check per-zone. This check is
3886 * made on memory-hotadd so a system can start with mobility
3887 * disabled and enable it later
3889 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3890 page_group_by_mobility_disabled
= 1;
3892 page_group_by_mobility_disabled
= 0;
3894 printk("Built %i zonelists in %s order, mobility grouping %s. "
3895 "Total pages: %ld\n",
3897 zonelist_order_name
[current_zonelist_order
],
3898 page_group_by_mobility_disabled
? "off" : "on",
3901 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3906 * Helper functions to size the waitqueue hash table.
3907 * Essentially these want to choose hash table sizes sufficiently
3908 * large so that collisions trying to wait on pages are rare.
3909 * But in fact, the number of active page waitqueues on typical
3910 * systems is ridiculously low, less than 200. So this is even
3911 * conservative, even though it seems large.
3913 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3914 * waitqueues, i.e. the size of the waitq table given the number of pages.
3916 #define PAGES_PER_WAITQUEUE 256
3918 #ifndef CONFIG_MEMORY_HOTPLUG
3919 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3921 unsigned long size
= 1;
3923 pages
/= PAGES_PER_WAITQUEUE
;
3925 while (size
< pages
)
3929 * Once we have dozens or even hundreds of threads sleeping
3930 * on IO we've got bigger problems than wait queue collision.
3931 * Limit the size of the wait table to a reasonable size.
3933 size
= min(size
, 4096UL);
3935 return max(size
, 4UL);
3939 * A zone's size might be changed by hot-add, so it is not possible to determine
3940 * a suitable size for its wait_table. So we use the maximum size now.
3942 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3944 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3945 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3946 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3948 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3949 * or more by the traditional way. (See above). It equals:
3951 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3952 * ia64(16K page size) : = ( 8G + 4M)byte.
3953 * powerpc (64K page size) : = (32G +16M)byte.
3955 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3962 * This is an integer logarithm so that shifts can be used later
3963 * to extract the more random high bits from the multiplicative
3964 * hash function before the remainder is taken.
3966 static inline unsigned long wait_table_bits(unsigned long size
)
3972 * Check if a pageblock contains reserved pages
3974 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3978 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3979 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3986 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3987 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3988 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3989 * higher will lead to a bigger reserve which will get freed as contiguous
3990 * blocks as reclaim kicks in
3992 static void setup_zone_migrate_reserve(struct zone
*zone
)
3994 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3996 unsigned long block_migratetype
;
4001 * Get the start pfn, end pfn and the number of blocks to reserve
4002 * We have to be careful to be aligned to pageblock_nr_pages to
4003 * make sure that we always check pfn_valid for the first page in
4006 start_pfn
= zone
->zone_start_pfn
;
4007 end_pfn
= zone_end_pfn(zone
);
4008 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4009 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4013 * Reserve blocks are generally in place to help high-order atomic
4014 * allocations that are short-lived. A min_free_kbytes value that
4015 * would result in more than 2 reserve blocks for atomic allocations
4016 * is assumed to be in place to help anti-fragmentation for the
4017 * future allocation of hugepages at runtime.
4019 reserve
= min(2, reserve
);
4020 old_reserve
= zone
->nr_migrate_reserve_block
;
4022 /* When memory hot-add, we almost always need to do nothing */
4023 if (reserve
== old_reserve
)
4025 zone
->nr_migrate_reserve_block
= reserve
;
4027 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4028 if (!pfn_valid(pfn
))
4030 page
= pfn_to_page(pfn
);
4032 /* Watch out for overlapping nodes */
4033 if (page_to_nid(page
) != zone_to_nid(zone
))
4036 block_migratetype
= get_pageblock_migratetype(page
);
4038 /* Only test what is necessary when the reserves are not met */
4041 * Blocks with reserved pages will never free, skip
4044 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4045 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4048 /* If this block is reserved, account for it */
4049 if (block_migratetype
== MIGRATE_RESERVE
) {
4054 /* Suitable for reserving if this block is movable */
4055 if (block_migratetype
== MIGRATE_MOVABLE
) {
4056 set_pageblock_migratetype(page
,
4058 move_freepages_block(zone
, page
,
4063 } else if (!old_reserve
) {
4065 * At boot time we don't need to scan the whole zone
4066 * for turning off MIGRATE_RESERVE.
4072 * If the reserve is met and this is a previous reserved block,
4075 if (block_migratetype
== MIGRATE_RESERVE
) {
4076 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4077 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4083 * Initially all pages are reserved - free ones are freed
4084 * up by free_all_bootmem() once the early boot process is
4085 * done. Non-atomic initialization, single-pass.
4087 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4088 unsigned long start_pfn
, enum memmap_context context
)
4091 unsigned long end_pfn
= start_pfn
+ size
;
4095 if (highest_memmap_pfn
< end_pfn
- 1)
4096 highest_memmap_pfn
= end_pfn
- 1;
4098 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4099 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4101 * There can be holes in boot-time mem_map[]s
4102 * handed to this function. They do not
4103 * exist on hotplugged memory.
4105 if (context
== MEMMAP_EARLY
) {
4106 if (!early_pfn_valid(pfn
))
4108 if (!early_pfn_in_nid(pfn
, nid
))
4111 page
= pfn_to_page(pfn
);
4112 set_page_links(page
, zone
, nid
, pfn
);
4113 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4114 init_page_count(page
);
4115 page_mapcount_reset(page
);
4116 page_cpupid_reset_last(page
);
4117 SetPageReserved(page
);
4119 * Mark the block movable so that blocks are reserved for
4120 * movable at startup. This will force kernel allocations
4121 * to reserve their blocks rather than leaking throughout
4122 * the address space during boot when many long-lived
4123 * kernel allocations are made. Later some blocks near
4124 * the start are marked MIGRATE_RESERVE by
4125 * setup_zone_migrate_reserve()
4127 * bitmap is created for zone's valid pfn range. but memmap
4128 * can be created for invalid pages (for alignment)
4129 * check here not to call set_pageblock_migratetype() against
4132 if ((z
->zone_start_pfn
<= pfn
)
4133 && (pfn
< zone_end_pfn(z
))
4134 && !(pfn
& (pageblock_nr_pages
- 1)))
4135 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4137 INIT_LIST_HEAD(&page
->lru
);
4138 #ifdef WANT_PAGE_VIRTUAL
4139 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4140 if (!is_highmem_idx(zone
))
4141 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4146 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4148 unsigned int order
, t
;
4149 for_each_migratetype_order(order
, t
) {
4150 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4151 zone
->free_area
[order
].nr_free
= 0;
4155 #ifndef __HAVE_ARCH_MEMMAP_INIT
4156 #define memmap_init(size, nid, zone, start_pfn) \
4157 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4160 static int zone_batchsize(struct zone
*zone
)
4166 * The per-cpu-pages pools are set to around 1000th of the
4167 * size of the zone. But no more than 1/2 of a meg.
4169 * OK, so we don't know how big the cache is. So guess.
4171 batch
= zone
->managed_pages
/ 1024;
4172 if (batch
* PAGE_SIZE
> 512 * 1024)
4173 batch
= (512 * 1024) / PAGE_SIZE
;
4174 batch
/= 4; /* We effectively *= 4 below */
4179 * Clamp the batch to a 2^n - 1 value. Having a power
4180 * of 2 value was found to be more likely to have
4181 * suboptimal cache aliasing properties in some cases.
4183 * For example if 2 tasks are alternately allocating
4184 * batches of pages, one task can end up with a lot
4185 * of pages of one half of the possible page colors
4186 * and the other with pages of the other colors.
4188 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4193 /* The deferral and batching of frees should be suppressed under NOMMU
4196 * The problem is that NOMMU needs to be able to allocate large chunks
4197 * of contiguous memory as there's no hardware page translation to
4198 * assemble apparent contiguous memory from discontiguous pages.
4200 * Queueing large contiguous runs of pages for batching, however,
4201 * causes the pages to actually be freed in smaller chunks. As there
4202 * can be a significant delay between the individual batches being
4203 * recycled, this leads to the once large chunks of space being
4204 * fragmented and becoming unavailable for high-order allocations.
4211 * pcp->high and pcp->batch values are related and dependent on one another:
4212 * ->batch must never be higher then ->high.
4213 * The following function updates them in a safe manner without read side
4216 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4217 * those fields changing asynchronously (acording the the above rule).
4219 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4220 * outside of boot time (or some other assurance that no concurrent updaters
4223 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4224 unsigned long batch
)
4226 /* start with a fail safe value for batch */
4230 /* Update high, then batch, in order */
4237 /* a companion to pageset_set_high() */
4238 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4240 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4243 static void pageset_init(struct per_cpu_pageset
*p
)
4245 struct per_cpu_pages
*pcp
;
4248 memset(p
, 0, sizeof(*p
));
4252 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4253 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4256 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4259 pageset_set_batch(p
, batch
);
4263 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4264 * to the value high for the pageset p.
4266 static void pageset_set_high(struct per_cpu_pageset
*p
,
4269 unsigned long batch
= max(1UL, high
/ 4);
4270 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4271 batch
= PAGE_SHIFT
* 8;
4273 pageset_update(&p
->pcp
, high
, batch
);
4276 static void pageset_set_high_and_batch(struct zone
*zone
,
4277 struct per_cpu_pageset
*pcp
)
4279 if (percpu_pagelist_fraction
)
4280 pageset_set_high(pcp
,
4281 (zone
->managed_pages
/
4282 percpu_pagelist_fraction
));
4284 pageset_set_batch(pcp
, zone_batchsize(zone
));
4287 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4289 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4292 pageset_set_high_and_batch(zone
, pcp
);
4295 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4298 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4299 for_each_possible_cpu(cpu
)
4300 zone_pageset_init(zone
, cpu
);
4304 * Allocate per cpu pagesets and initialize them.
4305 * Before this call only boot pagesets were available.
4307 void __init
setup_per_cpu_pageset(void)
4311 for_each_populated_zone(zone
)
4312 setup_zone_pageset(zone
);
4315 static noinline __init_refok
4316 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4322 * The per-page waitqueue mechanism uses hashed waitqueues
4325 zone
->wait_table_hash_nr_entries
=
4326 wait_table_hash_nr_entries(zone_size_pages
);
4327 zone
->wait_table_bits
=
4328 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4329 alloc_size
= zone
->wait_table_hash_nr_entries
4330 * sizeof(wait_queue_head_t
);
4332 if (!slab_is_available()) {
4333 zone
->wait_table
= (wait_queue_head_t
*)
4334 memblock_virt_alloc_node_nopanic(
4335 alloc_size
, zone
->zone_pgdat
->node_id
);
4338 * This case means that a zone whose size was 0 gets new memory
4339 * via memory hot-add.
4340 * But it may be the case that a new node was hot-added. In
4341 * this case vmalloc() will not be able to use this new node's
4342 * memory - this wait_table must be initialized to use this new
4343 * node itself as well.
4344 * To use this new node's memory, further consideration will be
4347 zone
->wait_table
= vmalloc(alloc_size
);
4349 if (!zone
->wait_table
)
4352 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4353 init_waitqueue_head(zone
->wait_table
+ i
);
4358 static __meminit
void zone_pcp_init(struct zone
*zone
)
4361 * per cpu subsystem is not up at this point. The following code
4362 * relies on the ability of the linker to provide the
4363 * offset of a (static) per cpu variable into the per cpu area.
4365 zone
->pageset
= &boot_pageset
;
4367 if (populated_zone(zone
))
4368 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4369 zone
->name
, zone
->present_pages
,
4370 zone_batchsize(zone
));
4373 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4374 unsigned long zone_start_pfn
,
4376 enum memmap_context context
)
4378 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4380 ret
= zone_wait_table_init(zone
, size
);
4383 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4385 zone
->zone_start_pfn
= zone_start_pfn
;
4387 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4388 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4390 (unsigned long)zone_idx(zone
),
4391 zone_start_pfn
, (zone_start_pfn
+ size
));
4393 zone_init_free_lists(zone
);
4398 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4399 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4401 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4403 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4405 unsigned long start_pfn
, end_pfn
;
4408 * NOTE: The following SMP-unsafe globals are only used early in boot
4409 * when the kernel is running single-threaded.
4411 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4412 static int __meminitdata last_nid
;
4414 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4417 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4419 last_start_pfn
= start_pfn
;
4420 last_end_pfn
= end_pfn
;
4426 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4428 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4432 nid
= __early_pfn_to_nid(pfn
);
4435 /* just returns 0 */
4439 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4440 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4444 nid
= __early_pfn_to_nid(pfn
);
4445 if (nid
>= 0 && nid
!= node
)
4452 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4453 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4454 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4456 * If an architecture guarantees that all ranges registered contain no holes
4457 * and may be freed, this this function may be used instead of calling
4458 * memblock_free_early_nid() manually.
4460 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4462 unsigned long start_pfn
, end_pfn
;
4465 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4466 start_pfn
= min(start_pfn
, max_low_pfn
);
4467 end_pfn
= min(end_pfn
, max_low_pfn
);
4469 if (start_pfn
< end_pfn
)
4470 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4471 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4477 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4478 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4480 * If an architecture guarantees that all ranges registered contain no holes and may
4481 * be freed, this function may be used instead of calling memory_present() manually.
4483 void __init
sparse_memory_present_with_active_regions(int nid
)
4485 unsigned long start_pfn
, end_pfn
;
4488 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4489 memory_present(this_nid
, start_pfn
, end_pfn
);
4493 * get_pfn_range_for_nid - Return the start and end page frames for a node
4494 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4495 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4496 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4498 * It returns the start and end page frame of a node based on information
4499 * provided by memblock_set_node(). If called for a node
4500 * with no available memory, a warning is printed and the start and end
4503 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4504 unsigned long *start_pfn
, unsigned long *end_pfn
)
4506 unsigned long this_start_pfn
, this_end_pfn
;
4512 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4513 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4514 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4517 if (*start_pfn
== -1UL)
4522 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4523 * assumption is made that zones within a node are ordered in monotonic
4524 * increasing memory addresses so that the "highest" populated zone is used
4526 static void __init
find_usable_zone_for_movable(void)
4529 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4530 if (zone_index
== ZONE_MOVABLE
)
4533 if (arch_zone_highest_possible_pfn
[zone_index
] >
4534 arch_zone_lowest_possible_pfn
[zone_index
])
4538 VM_BUG_ON(zone_index
== -1);
4539 movable_zone
= zone_index
;
4543 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4544 * because it is sized independent of architecture. Unlike the other zones,
4545 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4546 * in each node depending on the size of each node and how evenly kernelcore
4547 * is distributed. This helper function adjusts the zone ranges
4548 * provided by the architecture for a given node by using the end of the
4549 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4550 * zones within a node are in order of monotonic increases memory addresses
4552 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4553 unsigned long zone_type
,
4554 unsigned long node_start_pfn
,
4555 unsigned long node_end_pfn
,
4556 unsigned long *zone_start_pfn
,
4557 unsigned long *zone_end_pfn
)
4559 /* Only adjust if ZONE_MOVABLE is on this node */
4560 if (zone_movable_pfn
[nid
]) {
4561 /* Size ZONE_MOVABLE */
4562 if (zone_type
== ZONE_MOVABLE
) {
4563 *zone_start_pfn
= zone_movable_pfn
[nid
];
4564 *zone_end_pfn
= min(node_end_pfn
,
4565 arch_zone_highest_possible_pfn
[movable_zone
]);
4567 /* Adjust for ZONE_MOVABLE starting within this range */
4568 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4569 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4570 *zone_end_pfn
= zone_movable_pfn
[nid
];
4572 /* Check if this whole range is within ZONE_MOVABLE */
4573 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4574 *zone_start_pfn
= *zone_end_pfn
;
4579 * Return the number of pages a zone spans in a node, including holes
4580 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4582 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4583 unsigned long zone_type
,
4584 unsigned long node_start_pfn
,
4585 unsigned long node_end_pfn
,
4586 unsigned long *ignored
)
4588 unsigned long zone_start_pfn
, zone_end_pfn
;
4590 /* Get the start and end of the zone */
4591 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4592 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4593 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4594 node_start_pfn
, node_end_pfn
,
4595 &zone_start_pfn
, &zone_end_pfn
);
4597 /* Check that this node has pages within the zone's required range */
4598 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4601 /* Move the zone boundaries inside the node if necessary */
4602 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4603 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4605 /* Return the spanned pages */
4606 return zone_end_pfn
- zone_start_pfn
;
4610 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4611 * then all holes in the requested range will be accounted for.
4613 unsigned long __meminit
__absent_pages_in_range(int nid
,
4614 unsigned long range_start_pfn
,
4615 unsigned long range_end_pfn
)
4617 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4618 unsigned long start_pfn
, end_pfn
;
4621 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4622 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4623 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4624 nr_absent
-= end_pfn
- start_pfn
;
4630 * absent_pages_in_range - Return number of page frames in holes within a range
4631 * @start_pfn: The start PFN to start searching for holes
4632 * @end_pfn: The end PFN to stop searching for holes
4634 * It returns the number of pages frames in memory holes within a range.
4636 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4637 unsigned long end_pfn
)
4639 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4642 /* Return the number of page frames in holes in a zone on a node */
4643 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4644 unsigned long zone_type
,
4645 unsigned long node_start_pfn
,
4646 unsigned long node_end_pfn
,
4647 unsigned long *ignored
)
4649 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4650 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4651 unsigned long zone_start_pfn
, zone_end_pfn
;
4653 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4654 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4656 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4657 node_start_pfn
, node_end_pfn
,
4658 &zone_start_pfn
, &zone_end_pfn
);
4659 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4662 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4663 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4664 unsigned long zone_type
,
4665 unsigned long node_start_pfn
,
4666 unsigned long node_end_pfn
,
4667 unsigned long *zones_size
)
4669 return zones_size
[zone_type
];
4672 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4673 unsigned long zone_type
,
4674 unsigned long node_start_pfn
,
4675 unsigned long node_end_pfn
,
4676 unsigned long *zholes_size
)
4681 return zholes_size
[zone_type
];
4684 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4686 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4687 unsigned long node_start_pfn
,
4688 unsigned long node_end_pfn
,
4689 unsigned long *zones_size
,
4690 unsigned long *zholes_size
)
4692 unsigned long realtotalpages
, totalpages
= 0;
4695 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4696 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4700 pgdat
->node_spanned_pages
= totalpages
;
4702 realtotalpages
= totalpages
;
4703 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4705 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4706 node_start_pfn
, node_end_pfn
,
4708 pgdat
->node_present_pages
= realtotalpages
;
4709 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4713 #ifndef CONFIG_SPARSEMEM
4715 * Calculate the size of the zone->blockflags rounded to an unsigned long
4716 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4717 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4718 * round what is now in bits to nearest long in bits, then return it in
4721 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4723 unsigned long usemapsize
;
4725 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4726 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4727 usemapsize
= usemapsize
>> pageblock_order
;
4728 usemapsize
*= NR_PAGEBLOCK_BITS
;
4729 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4731 return usemapsize
/ 8;
4734 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4736 unsigned long zone_start_pfn
,
4737 unsigned long zonesize
)
4739 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4740 zone
->pageblock_flags
= NULL
;
4742 zone
->pageblock_flags
=
4743 memblock_virt_alloc_node_nopanic(usemapsize
,
4747 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4748 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4749 #endif /* CONFIG_SPARSEMEM */
4751 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4753 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4754 void __paginginit
set_pageblock_order(void)
4758 /* Check that pageblock_nr_pages has not already been setup */
4759 if (pageblock_order
)
4762 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4763 order
= HUGETLB_PAGE_ORDER
;
4765 order
= MAX_ORDER
- 1;
4768 * Assume the largest contiguous order of interest is a huge page.
4769 * This value may be variable depending on boot parameters on IA64 and
4772 pageblock_order
= order
;
4774 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4777 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4778 * is unused as pageblock_order is set at compile-time. See
4779 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4782 void __paginginit
set_pageblock_order(void)
4786 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4788 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4789 unsigned long present_pages
)
4791 unsigned long pages
= spanned_pages
;
4794 * Provide a more accurate estimation if there are holes within
4795 * the zone and SPARSEMEM is in use. If there are holes within the
4796 * zone, each populated memory region may cost us one or two extra
4797 * memmap pages due to alignment because memmap pages for each
4798 * populated regions may not naturally algined on page boundary.
4799 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4801 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4802 IS_ENABLED(CONFIG_SPARSEMEM
))
4803 pages
= present_pages
;
4805 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4809 * Set up the zone data structures:
4810 * - mark all pages reserved
4811 * - mark all memory queues empty
4812 * - clear the memory bitmaps
4814 * NOTE: pgdat should get zeroed by caller.
4816 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4817 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4818 unsigned long *zones_size
, unsigned long *zholes_size
)
4821 int nid
= pgdat
->node_id
;
4822 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4825 pgdat_resize_init(pgdat
);
4826 #ifdef CONFIG_NUMA_BALANCING
4827 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4828 pgdat
->numabalancing_migrate_nr_pages
= 0;
4829 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4831 init_waitqueue_head(&pgdat
->kswapd_wait
);
4832 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4833 pgdat_page_cgroup_init(pgdat
);
4835 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4836 struct zone
*zone
= pgdat
->node_zones
+ j
;
4837 unsigned long size
, realsize
, freesize
, memmap_pages
;
4839 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4840 node_end_pfn
, zones_size
);
4841 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4847 * Adjust freesize so that it accounts for how much memory
4848 * is used by this zone for memmap. This affects the watermark
4849 * and per-cpu initialisations
4851 memmap_pages
= calc_memmap_size(size
, realsize
);
4852 if (freesize
>= memmap_pages
) {
4853 freesize
-= memmap_pages
;
4856 " %s zone: %lu pages used for memmap\n",
4857 zone_names
[j
], memmap_pages
);
4860 " %s zone: %lu pages exceeds freesize %lu\n",
4861 zone_names
[j
], memmap_pages
, freesize
);
4863 /* Account for reserved pages */
4864 if (j
== 0 && freesize
> dma_reserve
) {
4865 freesize
-= dma_reserve
;
4866 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4867 zone_names
[0], dma_reserve
);
4870 if (!is_highmem_idx(j
))
4871 nr_kernel_pages
+= freesize
;
4872 /* Charge for highmem memmap if there are enough kernel pages */
4873 else if (nr_kernel_pages
> memmap_pages
* 2)
4874 nr_kernel_pages
-= memmap_pages
;
4875 nr_all_pages
+= freesize
;
4877 zone
->spanned_pages
= size
;
4878 zone
->present_pages
= realsize
;
4880 * Set an approximate value for lowmem here, it will be adjusted
4881 * when the bootmem allocator frees pages into the buddy system.
4882 * And all highmem pages will be managed by the buddy system.
4884 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4887 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4889 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4891 zone
->name
= zone_names
[j
];
4892 spin_lock_init(&zone
->lock
);
4893 spin_lock_init(&zone
->lru_lock
);
4894 zone_seqlock_init(zone
);
4895 zone
->zone_pgdat
= pgdat
;
4896 zone_pcp_init(zone
);
4898 /* For bootup, initialized properly in watermark setup */
4899 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4901 lruvec_init(&zone
->lruvec
);
4905 set_pageblock_order();
4906 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4907 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4908 size
, MEMMAP_EARLY
);
4910 memmap_init(size
, nid
, j
, zone_start_pfn
);
4911 zone_start_pfn
+= size
;
4915 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4917 /* Skip empty nodes */
4918 if (!pgdat
->node_spanned_pages
)
4921 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4922 /* ia64 gets its own node_mem_map, before this, without bootmem */
4923 if (!pgdat
->node_mem_map
) {
4924 unsigned long size
, start
, end
;
4928 * The zone's endpoints aren't required to be MAX_ORDER
4929 * aligned but the node_mem_map endpoints must be in order
4930 * for the buddy allocator to function correctly.
4932 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4933 end
= pgdat_end_pfn(pgdat
);
4934 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4935 size
= (end
- start
) * sizeof(struct page
);
4936 map
= alloc_remap(pgdat
->node_id
, size
);
4938 map
= memblock_virt_alloc_node_nopanic(size
,
4940 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4942 #ifndef CONFIG_NEED_MULTIPLE_NODES
4944 * With no DISCONTIG, the global mem_map is just set as node 0's
4946 if (pgdat
== NODE_DATA(0)) {
4947 mem_map
= NODE_DATA(0)->node_mem_map
;
4948 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4949 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4950 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4951 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4954 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4957 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4958 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4960 pg_data_t
*pgdat
= NODE_DATA(nid
);
4961 unsigned long start_pfn
= 0;
4962 unsigned long end_pfn
= 0;
4964 /* pg_data_t should be reset to zero when it's allocated */
4965 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4967 pgdat
->node_id
= nid
;
4968 pgdat
->node_start_pfn
= node_start_pfn
;
4969 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4970 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4971 printk(KERN_INFO
"Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid
,
4972 (u64
) start_pfn
<< PAGE_SHIFT
, (u64
) (end_pfn
<< PAGE_SHIFT
) - 1);
4974 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
4975 zones_size
, zholes_size
);
4977 alloc_node_mem_map(pgdat
);
4978 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4979 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4980 nid
, (unsigned long)pgdat
,
4981 (unsigned long)pgdat
->node_mem_map
);
4984 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
4985 zones_size
, zholes_size
);
4988 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4990 #if MAX_NUMNODES > 1
4992 * Figure out the number of possible node ids.
4994 void __init
setup_nr_node_ids(void)
4997 unsigned int highest
= 0;
4999 for_each_node_mask(node
, node_possible_map
)
5001 nr_node_ids
= highest
+ 1;
5006 * node_map_pfn_alignment - determine the maximum internode alignment
5008 * This function should be called after node map is populated and sorted.
5009 * It calculates the maximum power of two alignment which can distinguish
5012 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5013 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5014 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5015 * shifted, 1GiB is enough and this function will indicate so.
5017 * This is used to test whether pfn -> nid mapping of the chosen memory
5018 * model has fine enough granularity to avoid incorrect mapping for the
5019 * populated node map.
5021 * Returns the determined alignment in pfn's. 0 if there is no alignment
5022 * requirement (single node).
5024 unsigned long __init
node_map_pfn_alignment(void)
5026 unsigned long accl_mask
= 0, last_end
= 0;
5027 unsigned long start
, end
, mask
;
5031 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5032 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5039 * Start with a mask granular enough to pin-point to the
5040 * start pfn and tick off bits one-by-one until it becomes
5041 * too coarse to separate the current node from the last.
5043 mask
= ~((1 << __ffs(start
)) - 1);
5044 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5047 /* accumulate all internode masks */
5051 /* convert mask to number of pages */
5052 return ~accl_mask
+ 1;
5055 /* Find the lowest pfn for a node */
5056 static unsigned long __init
find_min_pfn_for_node(int nid
)
5058 unsigned long min_pfn
= ULONG_MAX
;
5059 unsigned long start_pfn
;
5062 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5063 min_pfn
= min(min_pfn
, start_pfn
);
5065 if (min_pfn
== ULONG_MAX
) {
5067 "Could not find start_pfn for node %d\n", nid
);
5075 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5077 * It returns the minimum PFN based on information provided via
5078 * memblock_set_node().
5080 unsigned long __init
find_min_pfn_with_active_regions(void)
5082 return find_min_pfn_for_node(MAX_NUMNODES
);
5086 * early_calculate_totalpages()
5087 * Sum pages in active regions for movable zone.
5088 * Populate N_MEMORY for calculating usable_nodes.
5090 static unsigned long __init
early_calculate_totalpages(void)
5092 unsigned long totalpages
= 0;
5093 unsigned long start_pfn
, end_pfn
;
5096 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5097 unsigned long pages
= end_pfn
- start_pfn
;
5099 totalpages
+= pages
;
5101 node_set_state(nid
, N_MEMORY
);
5107 * Find the PFN the Movable zone begins in each node. Kernel memory
5108 * is spread evenly between nodes as long as the nodes have enough
5109 * memory. When they don't, some nodes will have more kernelcore than
5112 static void __init
find_zone_movable_pfns_for_nodes(void)
5115 unsigned long usable_startpfn
;
5116 unsigned long kernelcore_node
, kernelcore_remaining
;
5117 /* save the state before borrow the nodemask */
5118 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5119 unsigned long totalpages
= early_calculate_totalpages();
5120 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5121 struct memblock_region
*r
;
5123 /* Need to find movable_zone earlier when movable_node is specified. */
5124 find_usable_zone_for_movable();
5127 * If movable_node is specified, ignore kernelcore and movablecore
5130 if (movable_node_is_enabled()) {
5131 for_each_memblock(memory
, r
) {
5132 if (!memblock_is_hotpluggable(r
))
5137 usable_startpfn
= PFN_DOWN(r
->base
);
5138 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5139 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5147 * If movablecore=nn[KMG] was specified, calculate what size of
5148 * kernelcore that corresponds so that memory usable for
5149 * any allocation type is evenly spread. If both kernelcore
5150 * and movablecore are specified, then the value of kernelcore
5151 * will be used for required_kernelcore if it's greater than
5152 * what movablecore would have allowed.
5154 if (required_movablecore
) {
5155 unsigned long corepages
;
5158 * Round-up so that ZONE_MOVABLE is at least as large as what
5159 * was requested by the user
5161 required_movablecore
=
5162 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5163 corepages
= totalpages
- required_movablecore
;
5165 required_kernelcore
= max(required_kernelcore
, corepages
);
5168 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5169 if (!required_kernelcore
)
5172 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5173 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5176 /* Spread kernelcore memory as evenly as possible throughout nodes */
5177 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5178 for_each_node_state(nid
, N_MEMORY
) {
5179 unsigned long start_pfn
, end_pfn
;
5182 * Recalculate kernelcore_node if the division per node
5183 * now exceeds what is necessary to satisfy the requested
5184 * amount of memory for the kernel
5186 if (required_kernelcore
< kernelcore_node
)
5187 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5190 * As the map is walked, we track how much memory is usable
5191 * by the kernel using kernelcore_remaining. When it is
5192 * 0, the rest of the node is usable by ZONE_MOVABLE
5194 kernelcore_remaining
= kernelcore_node
;
5196 /* Go through each range of PFNs within this node */
5197 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5198 unsigned long size_pages
;
5200 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5201 if (start_pfn
>= end_pfn
)
5204 /* Account for what is only usable for kernelcore */
5205 if (start_pfn
< usable_startpfn
) {
5206 unsigned long kernel_pages
;
5207 kernel_pages
= min(end_pfn
, usable_startpfn
)
5210 kernelcore_remaining
-= min(kernel_pages
,
5211 kernelcore_remaining
);
5212 required_kernelcore
-= min(kernel_pages
,
5213 required_kernelcore
);
5215 /* Continue if range is now fully accounted */
5216 if (end_pfn
<= usable_startpfn
) {
5219 * Push zone_movable_pfn to the end so
5220 * that if we have to rebalance
5221 * kernelcore across nodes, we will
5222 * not double account here
5224 zone_movable_pfn
[nid
] = end_pfn
;
5227 start_pfn
= usable_startpfn
;
5231 * The usable PFN range for ZONE_MOVABLE is from
5232 * start_pfn->end_pfn. Calculate size_pages as the
5233 * number of pages used as kernelcore
5235 size_pages
= end_pfn
- start_pfn
;
5236 if (size_pages
> kernelcore_remaining
)
5237 size_pages
= kernelcore_remaining
;
5238 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5241 * Some kernelcore has been met, update counts and
5242 * break if the kernelcore for this node has been
5245 required_kernelcore
-= min(required_kernelcore
,
5247 kernelcore_remaining
-= size_pages
;
5248 if (!kernelcore_remaining
)
5254 * If there is still required_kernelcore, we do another pass with one
5255 * less node in the count. This will push zone_movable_pfn[nid] further
5256 * along on the nodes that still have memory until kernelcore is
5260 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5264 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5265 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5266 zone_movable_pfn
[nid
] =
5267 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5270 /* restore the node_state */
5271 node_states
[N_MEMORY
] = saved_node_state
;
5274 /* Any regular or high memory on that node ? */
5275 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5277 enum zone_type zone_type
;
5279 if (N_MEMORY
== N_NORMAL_MEMORY
)
5282 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5283 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5284 if (populated_zone(zone
)) {
5285 node_set_state(nid
, N_HIGH_MEMORY
);
5286 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5287 zone_type
<= ZONE_NORMAL
)
5288 node_set_state(nid
, N_NORMAL_MEMORY
);
5295 * free_area_init_nodes - Initialise all pg_data_t and zone data
5296 * @max_zone_pfn: an array of max PFNs for each zone
5298 * This will call free_area_init_node() for each active node in the system.
5299 * Using the page ranges provided by memblock_set_node(), the size of each
5300 * zone in each node and their holes is calculated. If the maximum PFN
5301 * between two adjacent zones match, it is assumed that the zone is empty.
5302 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5303 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5304 * starts where the previous one ended. For example, ZONE_DMA32 starts
5305 * at arch_max_dma_pfn.
5307 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5309 unsigned long start_pfn
, end_pfn
;
5312 /* Record where the zone boundaries are */
5313 memset(arch_zone_lowest_possible_pfn
, 0,
5314 sizeof(arch_zone_lowest_possible_pfn
));
5315 memset(arch_zone_highest_possible_pfn
, 0,
5316 sizeof(arch_zone_highest_possible_pfn
));
5317 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5318 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5319 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5320 if (i
== ZONE_MOVABLE
)
5322 arch_zone_lowest_possible_pfn
[i
] =
5323 arch_zone_highest_possible_pfn
[i
-1];
5324 arch_zone_highest_possible_pfn
[i
] =
5325 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5327 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5328 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5330 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5331 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5332 find_zone_movable_pfns_for_nodes();
5334 /* Print out the zone ranges */
5335 printk("Zone ranges:\n");
5336 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5337 if (i
== ZONE_MOVABLE
)
5339 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5340 if (arch_zone_lowest_possible_pfn
[i
] ==
5341 arch_zone_highest_possible_pfn
[i
])
5342 printk(KERN_CONT
"empty\n");
5344 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5345 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5346 (arch_zone_highest_possible_pfn
[i
]
5347 << PAGE_SHIFT
) - 1);
5350 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5351 printk("Movable zone start for each node\n");
5352 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5353 if (zone_movable_pfn
[i
])
5354 printk(" Node %d: %#010lx\n", i
,
5355 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5358 /* Print out the early node map */
5359 printk("Early memory node ranges\n");
5360 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5361 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5362 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5364 /* Initialise every node */
5365 mminit_verify_pageflags_layout();
5366 setup_nr_node_ids();
5367 for_each_online_node(nid
) {
5368 pg_data_t
*pgdat
= NODE_DATA(nid
);
5369 free_area_init_node(nid
, NULL
,
5370 find_min_pfn_for_node(nid
), NULL
);
5372 /* Any memory on that node */
5373 if (pgdat
->node_present_pages
)
5374 node_set_state(nid
, N_MEMORY
);
5375 check_for_memory(pgdat
, nid
);
5379 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5381 unsigned long long coremem
;
5385 coremem
= memparse(p
, &p
);
5386 *core
= coremem
>> PAGE_SHIFT
;
5388 /* Paranoid check that UL is enough for the coremem value */
5389 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5395 * kernelcore=size sets the amount of memory for use for allocations that
5396 * cannot be reclaimed or migrated.
5398 static int __init
cmdline_parse_kernelcore(char *p
)
5400 return cmdline_parse_core(p
, &required_kernelcore
);
5404 * movablecore=size sets the amount of memory for use for allocations that
5405 * can be reclaimed or migrated.
5407 static int __init
cmdline_parse_movablecore(char *p
)
5409 return cmdline_parse_core(p
, &required_movablecore
);
5412 early_param("kernelcore", cmdline_parse_kernelcore
);
5413 early_param("movablecore", cmdline_parse_movablecore
);
5415 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5417 void adjust_managed_page_count(struct page
*page
, long count
)
5419 spin_lock(&managed_page_count_lock
);
5420 page_zone(page
)->managed_pages
+= count
;
5421 totalram_pages
+= count
;
5422 #ifdef CONFIG_HIGHMEM
5423 if (PageHighMem(page
))
5424 totalhigh_pages
+= count
;
5426 spin_unlock(&managed_page_count_lock
);
5428 EXPORT_SYMBOL(adjust_managed_page_count
);
5430 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5433 unsigned long pages
= 0;
5435 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5436 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5437 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5438 if ((unsigned int)poison
<= 0xFF)
5439 memset(pos
, poison
, PAGE_SIZE
);
5440 free_reserved_page(virt_to_page(pos
));
5444 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5445 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5449 EXPORT_SYMBOL(free_reserved_area
);
5451 #ifdef CONFIG_HIGHMEM
5452 void free_highmem_page(struct page
*page
)
5454 __free_reserved_page(page
);
5456 page_zone(page
)->managed_pages
++;
5462 void __init
mem_init_print_info(const char *str
)
5464 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5465 unsigned long init_code_size
, init_data_size
;
5467 physpages
= get_num_physpages();
5468 codesize
= _etext
- _stext
;
5469 datasize
= _edata
- _sdata
;
5470 rosize
= __end_rodata
- __start_rodata
;
5471 bss_size
= __bss_stop
- __bss_start
;
5472 init_data_size
= __init_end
- __init_begin
;
5473 init_code_size
= _einittext
- _sinittext
;
5476 * Detect special cases and adjust section sizes accordingly:
5477 * 1) .init.* may be embedded into .data sections
5478 * 2) .init.text.* may be out of [__init_begin, __init_end],
5479 * please refer to arch/tile/kernel/vmlinux.lds.S.
5480 * 3) .rodata.* may be embedded into .text or .data sections.
5482 #define adj_init_size(start, end, size, pos, adj) \
5484 if (start <= pos && pos < end && size > adj) \
5488 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5489 _sinittext
, init_code_size
);
5490 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5491 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5492 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5493 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5495 #undef adj_init_size
5497 printk("Memory: %luK/%luK available "
5498 "(%luK kernel code, %luK rwdata, %luK rodata, "
5499 "%luK init, %luK bss, %luK reserved"
5500 #ifdef CONFIG_HIGHMEM
5504 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5505 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5506 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5507 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5508 #ifdef CONFIG_HIGHMEM
5509 totalhigh_pages
<< (PAGE_SHIFT
-10),
5511 str
? ", " : "", str
? str
: "");
5515 * set_dma_reserve - set the specified number of pages reserved in the first zone
5516 * @new_dma_reserve: The number of pages to mark reserved
5518 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5519 * In the DMA zone, a significant percentage may be consumed by kernel image
5520 * and other unfreeable allocations which can skew the watermarks badly. This
5521 * function may optionally be used to account for unfreeable pages in the
5522 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5523 * smaller per-cpu batchsize.
5525 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5527 dma_reserve
= new_dma_reserve
;
5530 void __init
free_area_init(unsigned long *zones_size
)
5532 free_area_init_node(0, zones_size
,
5533 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5536 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5537 unsigned long action
, void *hcpu
)
5539 int cpu
= (unsigned long)hcpu
;
5541 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5542 lru_add_drain_cpu(cpu
);
5546 * Spill the event counters of the dead processor
5547 * into the current processors event counters.
5548 * This artificially elevates the count of the current
5551 vm_events_fold_cpu(cpu
);
5554 * Zero the differential counters of the dead processor
5555 * so that the vm statistics are consistent.
5557 * This is only okay since the processor is dead and cannot
5558 * race with what we are doing.
5560 cpu_vm_stats_fold(cpu
);
5565 void __init
page_alloc_init(void)
5567 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5571 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5572 * or min_free_kbytes changes.
5574 static void calculate_totalreserve_pages(void)
5576 struct pglist_data
*pgdat
;
5577 unsigned long reserve_pages
= 0;
5578 enum zone_type i
, j
;
5580 for_each_online_pgdat(pgdat
) {
5581 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5582 struct zone
*zone
= pgdat
->node_zones
+ i
;
5585 /* Find valid and maximum lowmem_reserve in the zone */
5586 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5587 if (zone
->lowmem_reserve
[j
] > max
)
5588 max
= zone
->lowmem_reserve
[j
];
5591 /* we treat the high watermark as reserved pages. */
5592 max
+= high_wmark_pages(zone
);
5594 if (max
> zone
->managed_pages
)
5595 max
= zone
->managed_pages
;
5596 reserve_pages
+= max
;
5598 * Lowmem reserves are not available to
5599 * GFP_HIGHUSER page cache allocations and
5600 * kswapd tries to balance zones to their high
5601 * watermark. As a result, neither should be
5602 * regarded as dirtyable memory, to prevent a
5603 * situation where reclaim has to clean pages
5604 * in order to balance the zones.
5606 zone
->dirty_balance_reserve
= max
;
5609 dirty_balance_reserve
= reserve_pages
;
5610 totalreserve_pages
= reserve_pages
;
5614 * setup_per_zone_lowmem_reserve - called whenever
5615 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5616 * has a correct pages reserved value, so an adequate number of
5617 * pages are left in the zone after a successful __alloc_pages().
5619 static void setup_per_zone_lowmem_reserve(void)
5621 struct pglist_data
*pgdat
;
5622 enum zone_type j
, idx
;
5624 for_each_online_pgdat(pgdat
) {
5625 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5626 struct zone
*zone
= pgdat
->node_zones
+ j
;
5627 unsigned long managed_pages
= zone
->managed_pages
;
5629 zone
->lowmem_reserve
[j
] = 0;
5633 struct zone
*lower_zone
;
5637 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5638 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5640 lower_zone
= pgdat
->node_zones
+ idx
;
5641 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5642 sysctl_lowmem_reserve_ratio
[idx
];
5643 managed_pages
+= lower_zone
->managed_pages
;
5648 /* update totalreserve_pages */
5649 calculate_totalreserve_pages();
5652 static void __setup_per_zone_wmarks(void)
5654 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5655 unsigned long lowmem_pages
= 0;
5657 unsigned long flags
;
5659 /* Calculate total number of !ZONE_HIGHMEM pages */
5660 for_each_zone(zone
) {
5661 if (!is_highmem(zone
))
5662 lowmem_pages
+= zone
->managed_pages
;
5665 for_each_zone(zone
) {
5668 spin_lock_irqsave(&zone
->lock
, flags
);
5669 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5670 do_div(tmp
, lowmem_pages
);
5671 if (is_highmem(zone
)) {
5673 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5674 * need highmem pages, so cap pages_min to a small
5677 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5678 * deltas controls asynch page reclaim, and so should
5679 * not be capped for highmem.
5681 unsigned long min_pages
;
5683 min_pages
= zone
->managed_pages
/ 1024;
5684 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5685 zone
->watermark
[WMARK_MIN
] = min_pages
;
5688 * If it's a lowmem zone, reserve a number of pages
5689 * proportionate to the zone's size.
5691 zone
->watermark
[WMARK_MIN
] = tmp
;
5694 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5695 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5697 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5698 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5699 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5701 setup_zone_migrate_reserve(zone
);
5702 spin_unlock_irqrestore(&zone
->lock
, flags
);
5705 /* update totalreserve_pages */
5706 calculate_totalreserve_pages();
5710 * setup_per_zone_wmarks - called when min_free_kbytes changes
5711 * or when memory is hot-{added|removed}
5713 * Ensures that the watermark[min,low,high] values for each zone are set
5714 * correctly with respect to min_free_kbytes.
5716 void setup_per_zone_wmarks(void)
5718 mutex_lock(&zonelists_mutex
);
5719 __setup_per_zone_wmarks();
5720 mutex_unlock(&zonelists_mutex
);
5724 * The inactive anon list should be small enough that the VM never has to
5725 * do too much work, but large enough that each inactive page has a chance
5726 * to be referenced again before it is swapped out.
5728 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5729 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5730 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5731 * the anonymous pages are kept on the inactive list.
5734 * memory ratio inactive anon
5735 * -------------------------------------
5744 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5746 unsigned int gb
, ratio
;
5748 /* Zone size in gigabytes */
5749 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5751 ratio
= int_sqrt(10 * gb
);
5755 zone
->inactive_ratio
= ratio
;
5758 static void __meminit
setup_per_zone_inactive_ratio(void)
5763 calculate_zone_inactive_ratio(zone
);
5767 * Initialise min_free_kbytes.
5769 * For small machines we want it small (128k min). For large machines
5770 * we want it large (64MB max). But it is not linear, because network
5771 * bandwidth does not increase linearly with machine size. We use
5773 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5774 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5790 int __meminit
init_per_zone_wmark_min(void)
5792 unsigned long lowmem_kbytes
;
5793 int new_min_free_kbytes
;
5795 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5796 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5798 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5799 min_free_kbytes
= new_min_free_kbytes
;
5800 if (min_free_kbytes
< 128)
5801 min_free_kbytes
= 128;
5802 if (min_free_kbytes
> 65536)
5803 min_free_kbytes
= 65536;
5805 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5806 new_min_free_kbytes
, user_min_free_kbytes
);
5808 setup_per_zone_wmarks();
5809 refresh_zone_stat_thresholds();
5810 setup_per_zone_lowmem_reserve();
5811 setup_per_zone_inactive_ratio();
5814 module_init(init_per_zone_wmark_min
)
5817 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5818 * that we can call two helper functions whenever min_free_kbytes
5821 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5822 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5826 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5831 user_min_free_kbytes
= min_free_kbytes
;
5832 setup_per_zone_wmarks();
5838 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5839 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5844 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5849 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5850 sysctl_min_unmapped_ratio
) / 100;
5854 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5855 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5860 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5865 zone
->min_slab_pages
= (zone
->managed_pages
*
5866 sysctl_min_slab_ratio
) / 100;
5872 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5873 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5874 * whenever sysctl_lowmem_reserve_ratio changes.
5876 * The reserve ratio obviously has absolutely no relation with the
5877 * minimum watermarks. The lowmem reserve ratio can only make sense
5878 * if in function of the boot time zone sizes.
5880 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5881 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5883 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5884 setup_per_zone_lowmem_reserve();
5889 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5890 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5891 * pagelist can have before it gets flushed back to buddy allocator.
5893 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5894 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5897 int old_percpu_pagelist_fraction
;
5900 mutex_lock(&pcp_batch_high_lock
);
5901 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5903 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5904 if (!write
|| ret
< 0)
5907 /* Sanity checking to avoid pcp imbalance */
5908 if (percpu_pagelist_fraction
&&
5909 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5910 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5916 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
5919 for_each_populated_zone(zone
) {
5922 for_each_possible_cpu(cpu
)
5923 pageset_set_high_and_batch(zone
,
5924 per_cpu_ptr(zone
->pageset
, cpu
));
5927 mutex_unlock(&pcp_batch_high_lock
);
5931 int hashdist
= HASHDIST_DEFAULT
;
5934 static int __init
set_hashdist(char *str
)
5938 hashdist
= simple_strtoul(str
, &str
, 0);
5941 __setup("hashdist=", set_hashdist
);
5945 * allocate a large system hash table from bootmem
5946 * - it is assumed that the hash table must contain an exact power-of-2
5947 * quantity of entries
5948 * - limit is the number of hash buckets, not the total allocation size
5950 void *__init
alloc_large_system_hash(const char *tablename
,
5951 unsigned long bucketsize
,
5952 unsigned long numentries
,
5955 unsigned int *_hash_shift
,
5956 unsigned int *_hash_mask
,
5957 unsigned long low_limit
,
5958 unsigned long high_limit
)
5960 unsigned long long max
= high_limit
;
5961 unsigned long log2qty
, size
;
5964 /* allow the kernel cmdline to have a say */
5966 /* round applicable memory size up to nearest megabyte */
5967 numentries
= nr_kernel_pages
;
5969 /* It isn't necessary when PAGE_SIZE >= 1MB */
5970 if (PAGE_SHIFT
< 20)
5971 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
5973 /* limit to 1 bucket per 2^scale bytes of low memory */
5974 if (scale
> PAGE_SHIFT
)
5975 numentries
>>= (scale
- PAGE_SHIFT
);
5977 numentries
<<= (PAGE_SHIFT
- scale
);
5979 /* Make sure we've got at least a 0-order allocation.. */
5980 if (unlikely(flags
& HASH_SMALL
)) {
5981 /* Makes no sense without HASH_EARLY */
5982 WARN_ON(!(flags
& HASH_EARLY
));
5983 if (!(numentries
>> *_hash_shift
)) {
5984 numentries
= 1UL << *_hash_shift
;
5985 BUG_ON(!numentries
);
5987 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5988 numentries
= PAGE_SIZE
/ bucketsize
;
5990 numentries
= roundup_pow_of_two(numentries
);
5992 /* limit allocation size to 1/16 total memory by default */
5994 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5995 do_div(max
, bucketsize
);
5997 max
= min(max
, 0x80000000ULL
);
5999 if (numentries
< low_limit
)
6000 numentries
= low_limit
;
6001 if (numentries
> max
)
6004 log2qty
= ilog2(numentries
);
6007 size
= bucketsize
<< log2qty
;
6008 if (flags
& HASH_EARLY
)
6009 table
= memblock_virt_alloc_nopanic(size
, 0);
6011 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6014 * If bucketsize is not a power-of-two, we may free
6015 * some pages at the end of hash table which
6016 * alloc_pages_exact() automatically does
6018 if (get_order(size
) < MAX_ORDER
) {
6019 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6020 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6023 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6026 panic("Failed to allocate %s hash table\n", tablename
);
6028 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6031 ilog2(size
) - PAGE_SHIFT
,
6035 *_hash_shift
= log2qty
;
6037 *_hash_mask
= (1 << log2qty
) - 1;
6042 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6043 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6046 #ifdef CONFIG_SPARSEMEM
6047 return __pfn_to_section(pfn
)->pageblock_flags
;
6049 return zone
->pageblock_flags
;
6050 #endif /* CONFIG_SPARSEMEM */
6053 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6055 #ifdef CONFIG_SPARSEMEM
6056 pfn
&= (PAGES_PER_SECTION
-1);
6057 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6059 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6060 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6061 #endif /* CONFIG_SPARSEMEM */
6065 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6066 * @page: The page within the block of interest
6067 * @pfn: The target page frame number
6068 * @end_bitidx: The last bit of interest to retrieve
6069 * @mask: mask of bits that the caller is interested in
6071 * Return: pageblock_bits flags
6073 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6074 unsigned long end_bitidx
,
6078 unsigned long *bitmap
;
6079 unsigned long bitidx
, word_bitidx
;
6082 zone
= page_zone(page
);
6083 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6084 bitidx
= pfn_to_bitidx(zone
, pfn
);
6085 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6086 bitidx
&= (BITS_PER_LONG
-1);
6088 word
= bitmap
[word_bitidx
];
6089 bitidx
+= end_bitidx
;
6090 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6094 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6095 * @page: The page within the block of interest
6096 * @flags: The flags to set
6097 * @pfn: The target page frame number
6098 * @end_bitidx: The last bit of interest
6099 * @mask: mask of bits that the caller is interested in
6101 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6103 unsigned long end_bitidx
,
6107 unsigned long *bitmap
;
6108 unsigned long bitidx
, word_bitidx
;
6109 unsigned long old_word
, word
;
6111 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6113 zone
= page_zone(page
);
6114 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6115 bitidx
= pfn_to_bitidx(zone
, pfn
);
6116 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6117 bitidx
&= (BITS_PER_LONG
-1);
6119 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6121 bitidx
+= end_bitidx
;
6122 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6123 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6125 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6127 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6128 if (word
== old_word
)
6135 * This function checks whether pageblock includes unmovable pages or not.
6136 * If @count is not zero, it is okay to include less @count unmovable pages
6138 * PageLRU check without isolation or lru_lock could race so that
6139 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6140 * expect this function should be exact.
6142 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6143 bool skip_hwpoisoned_pages
)
6145 unsigned long pfn
, iter
, found
;
6149 * For avoiding noise data, lru_add_drain_all() should be called
6150 * If ZONE_MOVABLE, the zone never contains unmovable pages
6152 if (zone_idx(zone
) == ZONE_MOVABLE
)
6154 mt
= get_pageblock_migratetype(page
);
6155 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6158 pfn
= page_to_pfn(page
);
6159 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6160 unsigned long check
= pfn
+ iter
;
6162 if (!pfn_valid_within(check
))
6165 page
= pfn_to_page(check
);
6168 * Hugepages are not in LRU lists, but they're movable.
6169 * We need not scan over tail pages bacause we don't
6170 * handle each tail page individually in migration.
6172 if (PageHuge(page
)) {
6173 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6178 * We can't use page_count without pin a page
6179 * because another CPU can free compound page.
6180 * This check already skips compound tails of THP
6181 * because their page->_count is zero at all time.
6183 if (!atomic_read(&page
->_count
)) {
6184 if (PageBuddy(page
))
6185 iter
+= (1 << page_order(page
)) - 1;
6190 * The HWPoisoned page may be not in buddy system, and
6191 * page_count() is not 0.
6193 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6199 * If there are RECLAIMABLE pages, we need to check it.
6200 * But now, memory offline itself doesn't call shrink_slab()
6201 * and it still to be fixed.
6204 * If the page is not RAM, page_count()should be 0.
6205 * we don't need more check. This is an _used_ not-movable page.
6207 * The problematic thing here is PG_reserved pages. PG_reserved
6208 * is set to both of a memory hole page and a _used_ kernel
6217 bool is_pageblock_removable_nolock(struct page
*page
)
6223 * We have to be careful here because we are iterating over memory
6224 * sections which are not zone aware so we might end up outside of
6225 * the zone but still within the section.
6226 * We have to take care about the node as well. If the node is offline
6227 * its NODE_DATA will be NULL - see page_zone.
6229 if (!node_online(page_to_nid(page
)))
6232 zone
= page_zone(page
);
6233 pfn
= page_to_pfn(page
);
6234 if (!zone_spans_pfn(zone
, pfn
))
6237 return !has_unmovable_pages(zone
, page
, 0, true);
6242 static unsigned long pfn_max_align_down(unsigned long pfn
)
6244 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6245 pageblock_nr_pages
) - 1);
6248 static unsigned long pfn_max_align_up(unsigned long pfn
)
6250 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6251 pageblock_nr_pages
));
6254 /* [start, end) must belong to a single zone. */
6255 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6256 unsigned long start
, unsigned long end
)
6258 /* This function is based on compact_zone() from compaction.c. */
6259 unsigned long nr_reclaimed
;
6260 unsigned long pfn
= start
;
6261 unsigned int tries
= 0;
6266 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6267 if (fatal_signal_pending(current
)) {
6272 if (list_empty(&cc
->migratepages
)) {
6273 cc
->nr_migratepages
= 0;
6274 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6280 } else if (++tries
== 5) {
6281 ret
= ret
< 0 ? ret
: -EBUSY
;
6285 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6287 cc
->nr_migratepages
-= nr_reclaimed
;
6289 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6290 NULL
, 0, cc
->mode
, MR_CMA
);
6293 putback_movable_pages(&cc
->migratepages
);
6300 * alloc_contig_range() -- tries to allocate given range of pages
6301 * @start: start PFN to allocate
6302 * @end: one-past-the-last PFN to allocate
6303 * @migratetype: migratetype of the underlaying pageblocks (either
6304 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6305 * in range must have the same migratetype and it must
6306 * be either of the two.
6308 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6309 * aligned, however it's the caller's responsibility to guarantee that
6310 * we are the only thread that changes migrate type of pageblocks the
6313 * The PFN range must belong to a single zone.
6315 * Returns zero on success or negative error code. On success all
6316 * pages which PFN is in [start, end) are allocated for the caller and
6317 * need to be freed with free_contig_range().
6319 int alloc_contig_range(unsigned long start
, unsigned long end
,
6320 unsigned migratetype
)
6322 unsigned long outer_start
, outer_end
;
6325 struct compact_control cc
= {
6326 .nr_migratepages
= 0,
6328 .zone
= page_zone(pfn_to_page(start
)),
6329 .mode
= MIGRATE_SYNC
,
6330 .ignore_skip_hint
= true,
6332 INIT_LIST_HEAD(&cc
.migratepages
);
6335 * What we do here is we mark all pageblocks in range as
6336 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6337 * have different sizes, and due to the way page allocator
6338 * work, we align the range to biggest of the two pages so
6339 * that page allocator won't try to merge buddies from
6340 * different pageblocks and change MIGRATE_ISOLATE to some
6341 * other migration type.
6343 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6344 * migrate the pages from an unaligned range (ie. pages that
6345 * we are interested in). This will put all the pages in
6346 * range back to page allocator as MIGRATE_ISOLATE.
6348 * When this is done, we take the pages in range from page
6349 * allocator removing them from the buddy system. This way
6350 * page allocator will never consider using them.
6352 * This lets us mark the pageblocks back as
6353 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6354 * aligned range but not in the unaligned, original range are
6355 * put back to page allocator so that buddy can use them.
6358 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6359 pfn_max_align_up(end
), migratetype
,
6364 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6369 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6370 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6371 * more, all pages in [start, end) are free in page allocator.
6372 * What we are going to do is to allocate all pages from
6373 * [start, end) (that is remove them from page allocator).
6375 * The only problem is that pages at the beginning and at the
6376 * end of interesting range may be not aligned with pages that
6377 * page allocator holds, ie. they can be part of higher order
6378 * pages. Because of this, we reserve the bigger range and
6379 * once this is done free the pages we are not interested in.
6381 * We don't have to hold zone->lock here because the pages are
6382 * isolated thus they won't get removed from buddy.
6385 lru_add_drain_all();
6389 outer_start
= start
;
6390 while (!PageBuddy(pfn_to_page(outer_start
))) {
6391 if (++order
>= MAX_ORDER
) {
6395 outer_start
&= ~0UL << order
;
6398 /* Make sure the range is really isolated. */
6399 if (test_pages_isolated(outer_start
, end
, false)) {
6400 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6407 /* Grab isolated pages from freelists. */
6408 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6414 /* Free head and tail (if any) */
6415 if (start
!= outer_start
)
6416 free_contig_range(outer_start
, start
- outer_start
);
6417 if (end
!= outer_end
)
6418 free_contig_range(end
, outer_end
- end
);
6421 undo_isolate_page_range(pfn_max_align_down(start
),
6422 pfn_max_align_up(end
), migratetype
);
6426 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6428 unsigned int count
= 0;
6430 for (; nr_pages
--; pfn
++) {
6431 struct page
*page
= pfn_to_page(pfn
);
6433 count
+= page_count(page
) != 1;
6436 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6440 #ifdef CONFIG_MEMORY_HOTPLUG
6442 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6443 * page high values need to be recalulated.
6445 void __meminit
zone_pcp_update(struct zone
*zone
)
6448 mutex_lock(&pcp_batch_high_lock
);
6449 for_each_possible_cpu(cpu
)
6450 pageset_set_high_and_batch(zone
,
6451 per_cpu_ptr(zone
->pageset
, cpu
));
6452 mutex_unlock(&pcp_batch_high_lock
);
6456 void zone_pcp_reset(struct zone
*zone
)
6458 unsigned long flags
;
6460 struct per_cpu_pageset
*pset
;
6462 /* avoid races with drain_pages() */
6463 local_irq_save(flags
);
6464 if (zone
->pageset
!= &boot_pageset
) {
6465 for_each_online_cpu(cpu
) {
6466 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6467 drain_zonestat(zone
, pset
);
6469 free_percpu(zone
->pageset
);
6470 zone
->pageset
= &boot_pageset
;
6472 local_irq_restore(flags
);
6475 #ifdef CONFIG_MEMORY_HOTREMOVE
6477 * All pages in the range must be isolated before calling this.
6480 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6484 unsigned int order
, i
;
6486 unsigned long flags
;
6487 /* find the first valid pfn */
6488 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6493 zone
= page_zone(pfn_to_page(pfn
));
6494 spin_lock_irqsave(&zone
->lock
, flags
);
6496 while (pfn
< end_pfn
) {
6497 if (!pfn_valid(pfn
)) {
6501 page
= pfn_to_page(pfn
);
6503 * The HWPoisoned page may be not in buddy system, and
6504 * page_count() is not 0.
6506 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6508 SetPageReserved(page
);
6512 BUG_ON(page_count(page
));
6513 BUG_ON(!PageBuddy(page
));
6514 order
= page_order(page
);
6515 #ifdef CONFIG_DEBUG_VM
6516 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6517 pfn
, 1 << order
, end_pfn
);
6519 list_del(&page
->lru
);
6520 rmv_page_order(page
);
6521 zone
->free_area
[order
].nr_free
--;
6522 for (i
= 0; i
< (1 << order
); i
++)
6523 SetPageReserved((page
+i
));
6524 pfn
+= (1 << order
);
6526 spin_unlock_irqrestore(&zone
->lock
, flags
);
6530 #ifdef CONFIG_MEMORY_FAILURE
6531 bool is_free_buddy_page(struct page
*page
)
6533 struct zone
*zone
= page_zone(page
);
6534 unsigned long pfn
= page_to_pfn(page
);
6535 unsigned long flags
;
6538 spin_lock_irqsave(&zone
->lock
, flags
);
6539 for (order
= 0; order
< MAX_ORDER
; order
++) {
6540 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6542 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6545 spin_unlock_irqrestore(&zone
->lock
, flags
);
6547 return order
< MAX_ORDER
;