2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/prefetch.h>
57 #include <linux/mm_inline.h>
58 #include <linux/migrate.h>
59 #include <linux/page-debug-flags.h>
60 #include <linux/hugetlb.h>
61 #include <linux/sched/rt.h>
63 #include <asm/sections.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
69 static DEFINE_MUTEX(pcp_batch_high_lock
);
70 #define MIN_PERCPU_PAGELIST_FRACTION (8)
72 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
73 DEFINE_PER_CPU(int, numa_node
);
74 EXPORT_PER_CPU_SYMBOL(numa_node
);
77 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
79 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
80 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
81 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
82 * defined in <linux/topology.h>.
84 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
85 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
86 int _node_numa_mem_
[MAX_NUMNODES
];
90 * Array of node states.
92 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
93 [N_POSSIBLE
] = NODE_MASK_ALL
,
94 [N_ONLINE
] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY
] = { { [0] = 1UL } },
103 [N_CPU
] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states
);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock
);
111 unsigned long totalram_pages __read_mostly
;
112 unsigned long totalreserve_pages __read_mostly
;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly
;
121 int percpu_pagelist_fraction
;
122 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask
;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex
));
139 if (saved_gfp_mask
) {
140 gfp_allowed_mask
= saved_gfp_mask
;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex
));
148 WARN_ON(saved_gfp_mask
);
149 saved_gfp_mask
= gfp_allowed_mask
;
150 gfp_allowed_mask
&= ~GFP_IOFS
;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly
;
165 static void __free_pages_ok(struct page
*page
, unsigned int order
);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages
);
193 static char * const zone_names
[MAX_NR_ZONES
] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes
= 1024;
208 int user_min_free_kbytes
= -1;
210 static unsigned long __meminitdata nr_kernel_pages
;
211 static unsigned long __meminitdata nr_all_pages
;
212 static unsigned long __meminitdata dma_reserve
;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
217 static unsigned long __initdata required_kernelcore
;
218 static unsigned long __initdata required_movablecore
;
219 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone
);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
228 int nr_online_nodes __read_mostly
= 1;
229 EXPORT_SYMBOL(nr_node_ids
);
230 EXPORT_SYMBOL(nr_online_nodes
);
233 int page_group_by_mobility_disabled __read_mostly
;
235 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
237 if (unlikely(page_group_by_mobility_disabled
&&
238 migratetype
< MIGRATE_PCPTYPES
))
239 migratetype
= MIGRATE_UNMOVABLE
;
241 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
242 PB_migrate
, PB_migrate_end
);
245 bool oom_killer_disabled __read_mostly
;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
252 unsigned long pfn
= page_to_pfn(page
);
253 unsigned long sp
, start_pfn
;
256 seq
= zone_span_seqbegin(zone
);
257 start_pfn
= zone
->zone_start_pfn
;
258 sp
= zone
->spanned_pages
;
259 if (!zone_spans_pfn(zone
, pfn
))
261 } while (zone_span_seqretry(zone
, seq
));
264 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
265 pfn
, zone_to_nid(zone
), zone
->name
,
266 start_pfn
, start_pfn
+ sp
);
271 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
273 if (!pfn_valid_within(page_to_pfn(page
)))
275 if (zone
!= page_zone(page
))
281 * Temporary debugging check for pages not lying within a given zone.
283 static int bad_range(struct zone
*zone
, struct page
*page
)
285 if (page_outside_zone_boundaries(zone
, page
))
287 if (!page_is_consistent(zone
, page
))
293 static inline int bad_range(struct zone
*zone
, struct page
*page
)
299 static void bad_page(struct page
*page
, const char *reason
,
300 unsigned long bad_flags
)
302 static unsigned long resume
;
303 static unsigned long nr_shown
;
304 static unsigned long nr_unshown
;
306 /* Don't complain about poisoned pages */
307 if (PageHWPoison(page
)) {
308 page_mapcount_reset(page
); /* remove PageBuddy */
313 * Allow a burst of 60 reports, then keep quiet for that minute;
314 * or allow a steady drip of one report per second.
316 if (nr_shown
== 60) {
317 if (time_before(jiffies
, resume
)) {
323 "BUG: Bad page state: %lu messages suppressed\n",
330 resume
= jiffies
+ 60 * HZ
;
332 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
333 current
->comm
, page_to_pfn(page
));
334 dump_page_badflags(page
, reason
, bad_flags
);
339 /* Leave bad fields for debug, except PageBuddy could make trouble */
340 page_mapcount_reset(page
); /* remove PageBuddy */
341 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
345 * Higher-order pages are called "compound pages". They are structured thusly:
347 * The first PAGE_SIZE page is called the "head page".
349 * The remaining PAGE_SIZE pages are called "tail pages".
351 * All pages have PG_compound set. All tail pages have their ->first_page
352 * pointing at the head page.
354 * The first tail page's ->lru.next holds the address of the compound page's
355 * put_page() function. Its ->lru.prev holds the order of allocation.
356 * This usage means that zero-order pages may not be compound.
359 static void free_compound_page(struct page
*page
)
361 __free_pages_ok(page
, compound_order(page
));
364 void prep_compound_page(struct page
*page
, unsigned long order
)
367 int nr_pages
= 1 << order
;
369 set_compound_page_dtor(page
, free_compound_page
);
370 set_compound_order(page
, order
);
372 for (i
= 1; i
< nr_pages
; i
++) {
373 struct page
*p
= page
+ i
;
374 set_page_count(p
, 0);
375 p
->first_page
= page
;
376 /* Make sure p->first_page is always valid for PageTail() */
382 /* update __split_huge_page_refcount if you change this function */
383 static int destroy_compound_page(struct page
*page
, unsigned long order
)
386 int nr_pages
= 1 << order
;
389 if (unlikely(compound_order(page
) != order
)) {
390 bad_page(page
, "wrong compound order", 0);
394 __ClearPageHead(page
);
396 for (i
= 1; i
< nr_pages
; i
++) {
397 struct page
*p
= page
+ i
;
399 if (unlikely(!PageTail(p
))) {
400 bad_page(page
, "PageTail not set", 0);
402 } else if (unlikely(p
->first_page
!= page
)) {
403 bad_page(page
, "first_page not consistent", 0);
412 static inline void prep_zero_page(struct page
*page
, unsigned int order
,
418 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
419 * and __GFP_HIGHMEM from hard or soft interrupt context.
421 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
422 for (i
= 0; i
< (1 << order
); i
++)
423 clear_highpage(page
+ i
);
426 #ifdef CONFIG_DEBUG_PAGEALLOC
427 unsigned int _debug_guardpage_minorder
;
429 static int __init
debug_guardpage_minorder_setup(char *buf
)
433 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
434 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
437 _debug_guardpage_minorder
= res
;
438 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
441 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
443 static inline void set_page_guard_flag(struct page
*page
)
445 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
448 static inline void clear_page_guard_flag(struct page
*page
)
450 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
453 static inline void set_page_guard_flag(struct page
*page
) { }
454 static inline void clear_page_guard_flag(struct page
*page
) { }
457 static inline void set_page_order(struct page
*page
, unsigned int order
)
459 set_page_private(page
, order
);
460 __SetPageBuddy(page
);
463 static inline void rmv_page_order(struct page
*page
)
465 __ClearPageBuddy(page
);
466 set_page_private(page
, 0);
470 * Locate the struct page for both the matching buddy in our
471 * pair (buddy1) and the combined O(n+1) page they form (page).
473 * 1) Any buddy B1 will have an order O twin B2 which satisfies
474 * the following equation:
476 * For example, if the starting buddy (buddy2) is #8 its order
478 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
480 * 2) Any buddy B will have an order O+1 parent P which
481 * satisfies the following equation:
484 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
486 static inline unsigned long
487 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
489 return page_idx
^ (1 << order
);
493 * This function checks whether a page is free && is the buddy
494 * we can do coalesce a page and its buddy if
495 * (a) the buddy is not in a hole &&
496 * (b) the buddy is in the buddy system &&
497 * (c) a page and its buddy have the same order &&
498 * (d) a page and its buddy are in the same zone.
500 * For recording whether a page is in the buddy system, we set ->_mapcount
501 * PAGE_BUDDY_MAPCOUNT_VALUE.
502 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
503 * serialized by zone->lock.
505 * For recording page's order, we use page_private(page).
507 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
510 if (!pfn_valid_within(page_to_pfn(buddy
)))
513 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
514 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
516 if (page_zone_id(page
) != page_zone_id(buddy
))
522 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
523 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
526 * zone check is done late to avoid uselessly
527 * calculating zone/node ids for pages that could
530 if (page_zone_id(page
) != page_zone_id(buddy
))
539 * Freeing function for a buddy system allocator.
541 * The concept of a buddy system is to maintain direct-mapped table
542 * (containing bit values) for memory blocks of various "orders".
543 * The bottom level table contains the map for the smallest allocatable
544 * units of memory (here, pages), and each level above it describes
545 * pairs of units from the levels below, hence, "buddies".
546 * At a high level, all that happens here is marking the table entry
547 * at the bottom level available, and propagating the changes upward
548 * as necessary, plus some accounting needed to play nicely with other
549 * parts of the VM system.
550 * At each level, we keep a list of pages, which are heads of continuous
551 * free pages of length of (1 << order) and marked with _mapcount
552 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
554 * So when we are allocating or freeing one, we can derive the state of the
555 * other. That is, if we allocate a small block, and both were
556 * free, the remainder of the region must be split into blocks.
557 * If a block is freed, and its buddy is also free, then this
558 * triggers coalescing into a block of larger size.
563 static inline void __free_one_page(struct page
*page
,
565 struct zone
*zone
, unsigned int order
,
568 unsigned long page_idx
;
569 unsigned long combined_idx
;
570 unsigned long uninitialized_var(buddy_idx
);
573 VM_BUG_ON(!zone_is_initialized(zone
));
575 if (unlikely(PageCompound(page
)))
576 if (unlikely(destroy_compound_page(page
, order
)))
579 VM_BUG_ON(migratetype
== -1);
580 if (!is_migrate_isolate(migratetype
))
581 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
583 page_idx
= pfn
& ((1 << MAX_ORDER
) - 1);
585 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
586 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
588 while (order
< MAX_ORDER
-1) {
589 buddy_idx
= __find_buddy_index(page_idx
, order
);
590 buddy
= page
+ (buddy_idx
- page_idx
);
591 if (!page_is_buddy(page
, buddy
, order
))
594 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
595 * merge with it and move up one order.
597 if (page_is_guard(buddy
)) {
598 clear_page_guard_flag(buddy
);
599 set_page_private(page
, 0);
600 __mod_zone_freepage_state(zone
, 1 << order
,
603 list_del(&buddy
->lru
);
604 zone
->free_area
[order
].nr_free
--;
605 rmv_page_order(buddy
);
607 combined_idx
= buddy_idx
& page_idx
;
608 page
= page
+ (combined_idx
- page_idx
);
609 page_idx
= combined_idx
;
612 set_page_order(page
, order
);
615 * If this is not the largest possible page, check if the buddy
616 * of the next-highest order is free. If it is, it's possible
617 * that pages are being freed that will coalesce soon. In case,
618 * that is happening, add the free page to the tail of the list
619 * so it's less likely to be used soon and more likely to be merged
620 * as a higher order page
622 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
623 struct page
*higher_page
, *higher_buddy
;
624 combined_idx
= buddy_idx
& page_idx
;
625 higher_page
= page
+ (combined_idx
- page_idx
);
626 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
627 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
628 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
629 list_add_tail(&page
->lru
,
630 &zone
->free_area
[order
].free_list
[migratetype
]);
635 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
637 zone
->free_area
[order
].nr_free
++;
640 static inline int free_pages_check(struct page
*page
)
642 const char *bad_reason
= NULL
;
643 unsigned long bad_flags
= 0;
645 if (unlikely(page_mapcount(page
)))
646 bad_reason
= "nonzero mapcount";
647 if (unlikely(page
->mapping
!= NULL
))
648 bad_reason
= "non-NULL mapping";
649 if (unlikely(atomic_read(&page
->_count
) != 0))
650 bad_reason
= "nonzero _count";
651 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
652 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
653 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
655 if (unlikely(mem_cgroup_bad_page_check(page
)))
656 bad_reason
= "cgroup check failed";
657 if (unlikely(bad_reason
)) {
658 bad_page(page
, bad_reason
, bad_flags
);
661 page_cpupid_reset_last(page
);
662 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
663 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
668 * Frees a number of pages from the PCP lists
669 * Assumes all pages on list are in same zone, and of same order.
670 * count is the number of pages to free.
672 * If the zone was previously in an "all pages pinned" state then look to
673 * see if this freeing clears that state.
675 * And clear the zone's pages_scanned counter, to hold off the "all pages are
676 * pinned" detection logic.
678 static void free_pcppages_bulk(struct zone
*zone
, int count
,
679 struct per_cpu_pages
*pcp
)
684 unsigned long nr_scanned
;
686 spin_lock(&zone
->lock
);
687 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
689 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
693 struct list_head
*list
;
696 * Remove pages from lists in a round-robin fashion. A
697 * batch_free count is maintained that is incremented when an
698 * empty list is encountered. This is so more pages are freed
699 * off fuller lists instead of spinning excessively around empty
704 if (++migratetype
== MIGRATE_PCPTYPES
)
706 list
= &pcp
->lists
[migratetype
];
707 } while (list_empty(list
));
709 /* This is the only non-empty list. Free them all. */
710 if (batch_free
== MIGRATE_PCPTYPES
)
711 batch_free
= to_free
;
714 int mt
; /* migratetype of the to-be-freed page */
716 page
= list_entry(list
->prev
, struct page
, lru
);
717 /* must delete as __free_one_page list manipulates */
718 list_del(&page
->lru
);
719 mt
= get_freepage_migratetype(page
);
720 if (unlikely(has_isolate_pageblock(zone
)))
721 mt
= get_pageblock_migratetype(page
);
723 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
724 __free_one_page(page
, page_to_pfn(page
), zone
, 0, mt
);
725 trace_mm_page_pcpu_drain(page
, 0, mt
);
726 } while (--to_free
&& --batch_free
&& !list_empty(list
));
728 spin_unlock(&zone
->lock
);
731 static void free_one_page(struct zone
*zone
,
732 struct page
*page
, unsigned long pfn
,
736 unsigned long nr_scanned
;
737 spin_lock(&zone
->lock
);
738 nr_scanned
= zone_page_state(zone
, NR_PAGES_SCANNED
);
740 __mod_zone_page_state(zone
, NR_PAGES_SCANNED
, -nr_scanned
);
742 if (unlikely(has_isolate_pageblock(zone
) ||
743 is_migrate_isolate(migratetype
))) {
744 migratetype
= get_pfnblock_migratetype(page
, pfn
);
746 __free_one_page(page
, pfn
, zone
, order
, migratetype
);
747 spin_unlock(&zone
->lock
);
750 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
755 trace_mm_page_free(page
, order
);
756 kmemcheck_free_shadow(page
, order
);
759 page
->mapping
= NULL
;
760 for (i
= 0; i
< (1 << order
); i
++)
761 bad
+= free_pages_check(page
+ i
);
765 if (!PageHighMem(page
)) {
766 debug_check_no_locks_freed(page_address(page
),
768 debug_check_no_obj_freed(page_address(page
),
771 arch_free_page(page
, order
);
772 kernel_map_pages(page
, 1 << order
, 0);
777 static void __free_pages_ok(struct page
*page
, unsigned int order
)
781 unsigned long pfn
= page_to_pfn(page
);
783 if (!free_pages_prepare(page
, order
))
786 migratetype
= get_pfnblock_migratetype(page
, pfn
);
787 local_irq_save(flags
);
788 __count_vm_events(PGFREE
, 1 << order
);
789 set_freepage_migratetype(page
, migratetype
);
790 free_one_page(page_zone(page
), page
, pfn
, order
, migratetype
);
791 local_irq_restore(flags
);
794 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
796 unsigned int nr_pages
= 1 << order
;
797 struct page
*p
= page
;
801 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
803 __ClearPageReserved(p
);
804 set_page_count(p
, 0);
806 __ClearPageReserved(p
);
807 set_page_count(p
, 0);
809 page_zone(page
)->managed_pages
+= nr_pages
;
810 set_page_refcounted(page
);
811 __free_pages(page
, order
);
815 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
816 void __init
init_cma_reserved_pageblock(struct page
*page
)
818 unsigned i
= pageblock_nr_pages
;
819 struct page
*p
= page
;
822 __ClearPageReserved(p
);
823 set_page_count(p
, 0);
826 set_pageblock_migratetype(page
, MIGRATE_CMA
);
828 if (pageblock_order
>= MAX_ORDER
) {
829 i
= pageblock_nr_pages
;
832 set_page_refcounted(p
);
833 __free_pages(p
, MAX_ORDER
- 1);
834 p
+= MAX_ORDER_NR_PAGES
;
835 } while (i
-= MAX_ORDER_NR_PAGES
);
837 set_page_refcounted(page
);
838 __free_pages(page
, pageblock_order
);
841 adjust_managed_page_count(page
, pageblock_nr_pages
);
846 * The order of subdivision here is critical for the IO subsystem.
847 * Please do not alter this order without good reasons and regression
848 * testing. Specifically, as large blocks of memory are subdivided,
849 * the order in which smaller blocks are delivered depends on the order
850 * they're subdivided in this function. This is the primary factor
851 * influencing the order in which pages are delivered to the IO
852 * subsystem according to empirical testing, and this is also justified
853 * by considering the behavior of a buddy system containing a single
854 * large block of memory acted on by a series of small allocations.
855 * This behavior is a critical factor in sglist merging's success.
859 static inline void expand(struct zone
*zone
, struct page
*page
,
860 int low
, int high
, struct free_area
*area
,
863 unsigned long size
= 1 << high
;
869 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
871 #ifdef CONFIG_DEBUG_PAGEALLOC
872 if (high
< debug_guardpage_minorder()) {
874 * Mark as guard pages (or page), that will allow to
875 * merge back to allocator when buddy will be freed.
876 * Corresponding page table entries will not be touched,
877 * pages will stay not present in virtual address space
879 INIT_LIST_HEAD(&page
[size
].lru
);
880 set_page_guard_flag(&page
[size
]);
881 set_page_private(&page
[size
], high
);
882 /* Guard pages are not available for any usage */
883 __mod_zone_freepage_state(zone
, -(1 << high
),
888 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
890 set_page_order(&page
[size
], high
);
895 * This page is about to be returned from the page allocator
897 static inline int check_new_page(struct page
*page
)
899 const char *bad_reason
= NULL
;
900 unsigned long bad_flags
= 0;
902 if (unlikely(page_mapcount(page
)))
903 bad_reason
= "nonzero mapcount";
904 if (unlikely(page
->mapping
!= NULL
))
905 bad_reason
= "non-NULL mapping";
906 if (unlikely(atomic_read(&page
->_count
) != 0))
907 bad_reason
= "nonzero _count";
908 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
909 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
910 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
912 if (unlikely(mem_cgroup_bad_page_check(page
)))
913 bad_reason
= "cgroup check failed";
914 if (unlikely(bad_reason
)) {
915 bad_page(page
, bad_reason
, bad_flags
);
921 static int prep_new_page(struct page
*page
, unsigned int order
, gfp_t gfp_flags
)
925 for (i
= 0; i
< (1 << order
); i
++) {
926 struct page
*p
= page
+ i
;
927 if (unlikely(check_new_page(p
)))
931 set_page_private(page
, 0);
932 set_page_refcounted(page
);
934 arch_alloc_page(page
, order
);
935 kernel_map_pages(page
, 1 << order
, 1);
937 if (gfp_flags
& __GFP_ZERO
)
938 prep_zero_page(page
, order
, gfp_flags
);
940 if (order
&& (gfp_flags
& __GFP_COMP
))
941 prep_compound_page(page
, order
);
947 * Go through the free lists for the given migratetype and remove
948 * the smallest available page from the freelists
951 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
954 unsigned int current_order
;
955 struct free_area
*area
;
958 /* Find a page of the appropriate size in the preferred list */
959 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
960 area
= &(zone
->free_area
[current_order
]);
961 if (list_empty(&area
->free_list
[migratetype
]))
964 page
= list_entry(area
->free_list
[migratetype
].next
,
966 list_del(&page
->lru
);
967 rmv_page_order(page
);
969 expand(zone
, page
, order
, current_order
, area
, migratetype
);
970 set_freepage_migratetype(page
, migratetype
);
979 * This array describes the order lists are fallen back to when
980 * the free lists for the desirable migrate type are depleted
982 static int fallbacks
[MIGRATE_TYPES
][4] = {
983 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
984 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
986 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
987 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
989 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
991 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
992 #ifdef CONFIG_MEMORY_ISOLATION
993 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
998 * Move the free pages in a range to the free lists of the requested type.
999 * Note that start_page and end_pages are not aligned on a pageblock
1000 * boundary. If alignment is required, use move_freepages_block()
1002 int move_freepages(struct zone
*zone
,
1003 struct page
*start_page
, struct page
*end_page
,
1007 unsigned long order
;
1008 int pages_moved
= 0;
1010 #ifndef CONFIG_HOLES_IN_ZONE
1012 * page_zone is not safe to call in this context when
1013 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
1014 * anyway as we check zone boundaries in move_freepages_block().
1015 * Remove at a later date when no bug reports exist related to
1016 * grouping pages by mobility
1018 VM_BUG_ON(page_zone(start_page
) != page_zone(end_page
));
1021 for (page
= start_page
; page
<= end_page
;) {
1022 /* Make sure we are not inadvertently changing nodes */
1023 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
1025 if (!pfn_valid_within(page_to_pfn(page
))) {
1030 if (!PageBuddy(page
)) {
1035 order
= page_order(page
);
1036 list_move(&page
->lru
,
1037 &zone
->free_area
[order
].free_list
[migratetype
]);
1038 set_freepage_migratetype(page
, migratetype
);
1040 pages_moved
+= 1 << order
;
1046 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1049 unsigned long start_pfn
, end_pfn
;
1050 struct page
*start_page
, *end_page
;
1052 start_pfn
= page_to_pfn(page
);
1053 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1054 start_page
= pfn_to_page(start_pfn
);
1055 end_page
= start_page
+ pageblock_nr_pages
- 1;
1056 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1058 /* Do not cross zone boundaries */
1059 if (!zone_spans_pfn(zone
, start_pfn
))
1061 if (!zone_spans_pfn(zone
, end_pfn
))
1064 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1067 static void change_pageblock_range(struct page
*pageblock_page
,
1068 int start_order
, int migratetype
)
1070 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1072 while (nr_pageblocks
--) {
1073 set_pageblock_migratetype(pageblock_page
, migratetype
);
1074 pageblock_page
+= pageblock_nr_pages
;
1079 * If breaking a large block of pages, move all free pages to the preferred
1080 * allocation list. If falling back for a reclaimable kernel allocation, be
1081 * more aggressive about taking ownership of free pages.
1083 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1084 * nor move CMA pages to different free lists. We don't want unmovable pages
1085 * to be allocated from MIGRATE_CMA areas.
1087 * Returns the new migratetype of the pageblock (or the same old migratetype
1088 * if it was unchanged).
1090 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1091 int start_type
, int fallback_type
)
1093 int current_order
= page_order(page
);
1096 * When borrowing from MIGRATE_CMA, we need to release the excess
1097 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1098 * is set to CMA so it is returned to the correct freelist in case
1099 * the page ends up being not actually allocated from the pcp lists.
1101 if (is_migrate_cma(fallback_type
))
1102 return fallback_type
;
1104 /* Take ownership for orders >= pageblock_order */
1105 if (current_order
>= pageblock_order
) {
1106 change_pageblock_range(page
, current_order
, start_type
);
1110 if (current_order
>= pageblock_order
/ 2 ||
1111 start_type
== MIGRATE_RECLAIMABLE
||
1112 page_group_by_mobility_disabled
) {
1115 pages
= move_freepages_block(zone
, page
, start_type
);
1117 /* Claim the whole block if over half of it is free */
1118 if (pages
>= (1 << (pageblock_order
-1)) ||
1119 page_group_by_mobility_disabled
) {
1121 set_pageblock_migratetype(page
, start_type
);
1127 return fallback_type
;
1130 /* Remove an element from the buddy allocator from the fallback list */
1131 static inline struct page
*
1132 __rmqueue_fallback(struct zone
*zone
, unsigned int order
, int start_migratetype
)
1134 struct free_area
*area
;
1135 unsigned int current_order
;
1137 int migratetype
, new_type
, i
;
1139 /* Find the largest possible block of pages in the other list */
1140 for (current_order
= MAX_ORDER
-1;
1141 current_order
>= order
&& current_order
<= MAX_ORDER
-1;
1144 migratetype
= fallbacks
[start_migratetype
][i
];
1146 /* MIGRATE_RESERVE handled later if necessary */
1147 if (migratetype
== MIGRATE_RESERVE
)
1150 area
= &(zone
->free_area
[current_order
]);
1151 if (list_empty(&area
->free_list
[migratetype
]))
1154 page
= list_entry(area
->free_list
[migratetype
].next
,
1158 new_type
= try_to_steal_freepages(zone
, page
,
1162 /* Remove the page from the freelists */
1163 list_del(&page
->lru
);
1164 rmv_page_order(page
);
1166 expand(zone
, page
, order
, current_order
, area
,
1168 /* The freepage_migratetype may differ from pageblock's
1169 * migratetype depending on the decisions in
1170 * try_to_steal_freepages. This is OK as long as it does
1171 * not differ for MIGRATE_CMA type.
1173 set_freepage_migratetype(page
, new_type
);
1175 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1176 start_migratetype
, migratetype
, new_type
);
1186 * Do the hard work of removing an element from the buddy allocator.
1187 * Call me with the zone->lock already held.
1189 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1195 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1197 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1198 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1201 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1202 * is used because __rmqueue_smallest is an inline function
1203 * and we want just one call site
1206 migratetype
= MIGRATE_RESERVE
;
1211 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1216 * Obtain a specified number of elements from the buddy allocator, all under
1217 * a single hold of the lock, for efficiency. Add them to the supplied list.
1218 * Returns the number of new pages which were placed at *list.
1220 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1221 unsigned long count
, struct list_head
*list
,
1222 int migratetype
, bool cold
)
1226 spin_lock(&zone
->lock
);
1227 for (i
= 0; i
< count
; ++i
) {
1228 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1229 if (unlikely(page
== NULL
))
1233 * Split buddy pages returned by expand() are received here
1234 * in physical page order. The page is added to the callers and
1235 * list and the list head then moves forward. From the callers
1236 * perspective, the linked list is ordered by page number in
1237 * some conditions. This is useful for IO devices that can
1238 * merge IO requests if the physical pages are ordered
1242 list_add(&page
->lru
, list
);
1244 list_add_tail(&page
->lru
, list
);
1246 if (is_migrate_cma(get_freepage_migratetype(page
)))
1247 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1250 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1251 spin_unlock(&zone
->lock
);
1257 * Called from the vmstat counter updater to drain pagesets of this
1258 * currently executing processor on remote nodes after they have
1261 * Note that this function must be called with the thread pinned to
1262 * a single processor.
1264 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1266 unsigned long flags
;
1267 int to_drain
, batch
;
1269 local_irq_save(flags
);
1270 batch
= ACCESS_ONCE(pcp
->batch
);
1271 to_drain
= min(pcp
->count
, batch
);
1273 free_pcppages_bulk(zone
, to_drain
, pcp
);
1274 pcp
->count
-= to_drain
;
1276 local_irq_restore(flags
);
1281 * Drain pages of the indicated processor.
1283 * The processor must either be the current processor and the
1284 * thread pinned to the current processor or a processor that
1287 static void drain_pages(unsigned int cpu
)
1289 unsigned long flags
;
1292 for_each_populated_zone(zone
) {
1293 struct per_cpu_pageset
*pset
;
1294 struct per_cpu_pages
*pcp
;
1296 local_irq_save(flags
);
1297 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1301 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1304 local_irq_restore(flags
);
1309 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1311 void drain_local_pages(void *arg
)
1313 drain_pages(smp_processor_id());
1317 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1319 * Note that this code is protected against sending an IPI to an offline
1320 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1321 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1322 * nothing keeps CPUs from showing up after we populated the cpumask and
1323 * before the call to on_each_cpu_mask().
1325 void drain_all_pages(void)
1328 struct per_cpu_pageset
*pcp
;
1332 * Allocate in the BSS so we wont require allocation in
1333 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1335 static cpumask_t cpus_with_pcps
;
1338 * We don't care about racing with CPU hotplug event
1339 * as offline notification will cause the notified
1340 * cpu to drain that CPU pcps and on_each_cpu_mask
1341 * disables preemption as part of its processing
1343 for_each_online_cpu(cpu
) {
1344 bool has_pcps
= false;
1345 for_each_populated_zone(zone
) {
1346 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1347 if (pcp
->pcp
.count
) {
1353 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1355 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1357 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1360 #ifdef CONFIG_HIBERNATION
1362 void mark_free_pages(struct zone
*zone
)
1364 unsigned long pfn
, max_zone_pfn
;
1365 unsigned long flags
;
1366 unsigned int order
, t
;
1367 struct list_head
*curr
;
1369 if (zone_is_empty(zone
))
1372 spin_lock_irqsave(&zone
->lock
, flags
);
1374 max_zone_pfn
= zone_end_pfn(zone
);
1375 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1376 if (pfn_valid(pfn
)) {
1377 struct page
*page
= pfn_to_page(pfn
);
1379 if (!swsusp_page_is_forbidden(page
))
1380 swsusp_unset_page_free(page
);
1383 for_each_migratetype_order(order
, t
) {
1384 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1387 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1388 for (i
= 0; i
< (1UL << order
); i
++)
1389 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1392 spin_unlock_irqrestore(&zone
->lock
, flags
);
1394 #endif /* CONFIG_PM */
1397 * Free a 0-order page
1398 * cold == true ? free a cold page : free a hot page
1400 void free_hot_cold_page(struct page
*page
, bool cold
)
1402 struct zone
*zone
= page_zone(page
);
1403 struct per_cpu_pages
*pcp
;
1404 unsigned long flags
;
1405 unsigned long pfn
= page_to_pfn(page
);
1408 if (!free_pages_prepare(page
, 0))
1411 migratetype
= get_pfnblock_migratetype(page
, pfn
);
1412 set_freepage_migratetype(page
, migratetype
);
1413 local_irq_save(flags
);
1414 __count_vm_event(PGFREE
);
1417 * We only track unmovable, reclaimable and movable on pcp lists.
1418 * Free ISOLATE pages back to the allocator because they are being
1419 * offlined but treat RESERVE as movable pages so we can get those
1420 * areas back if necessary. Otherwise, we may have to free
1421 * excessively into the page allocator
1423 if (migratetype
>= MIGRATE_PCPTYPES
) {
1424 if (unlikely(is_migrate_isolate(migratetype
))) {
1425 free_one_page(zone
, page
, pfn
, 0, migratetype
);
1428 migratetype
= MIGRATE_MOVABLE
;
1431 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1433 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1435 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1437 if (pcp
->count
>= pcp
->high
) {
1438 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1439 free_pcppages_bulk(zone
, batch
, pcp
);
1440 pcp
->count
-= batch
;
1444 local_irq_restore(flags
);
1448 * Free a list of 0-order pages
1450 void free_hot_cold_page_list(struct list_head
*list
, bool cold
)
1452 struct page
*page
, *next
;
1454 list_for_each_entry_safe(page
, next
, list
, lru
) {
1455 trace_mm_page_free_batched(page
, cold
);
1456 free_hot_cold_page(page
, cold
);
1461 * split_page takes a non-compound higher-order page, and splits it into
1462 * n (1<<order) sub-pages: page[0..n]
1463 * Each sub-page must be freed individually.
1465 * Note: this is probably too low level an operation for use in drivers.
1466 * Please consult with lkml before using this in your driver.
1468 void split_page(struct page
*page
, unsigned int order
)
1472 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1473 VM_BUG_ON_PAGE(!page_count(page
), page
);
1475 #ifdef CONFIG_KMEMCHECK
1477 * Split shadow pages too, because free(page[0]) would
1478 * otherwise free the whole shadow.
1480 if (kmemcheck_page_is_tracked(page
))
1481 split_page(virt_to_page(page
[0].shadow
), order
);
1484 for (i
= 1; i
< (1 << order
); i
++)
1485 set_page_refcounted(page
+ i
);
1487 EXPORT_SYMBOL_GPL(split_page
);
1489 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1491 unsigned long watermark
;
1495 BUG_ON(!PageBuddy(page
));
1497 zone
= page_zone(page
);
1498 mt
= get_pageblock_migratetype(page
);
1500 if (!is_migrate_isolate(mt
)) {
1501 /* Obey watermarks as if the page was being allocated */
1502 watermark
= low_wmark_pages(zone
) + (1 << order
);
1503 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1506 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1509 /* Remove page from free list */
1510 list_del(&page
->lru
);
1511 zone
->free_area
[order
].nr_free
--;
1512 rmv_page_order(page
);
1514 /* Set the pageblock if the isolated page is at least a pageblock */
1515 if (order
>= pageblock_order
- 1) {
1516 struct page
*endpage
= page
+ (1 << order
) - 1;
1517 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1518 int mt
= get_pageblock_migratetype(page
);
1519 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1520 set_pageblock_migratetype(page
,
1525 return 1UL << order
;
1529 * Similar to split_page except the page is already free. As this is only
1530 * being used for migration, the migratetype of the block also changes.
1531 * As this is called with interrupts disabled, the caller is responsible
1532 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1535 * Note: this is probably too low level an operation for use in drivers.
1536 * Please consult with lkml before using this in your driver.
1538 int split_free_page(struct page
*page
)
1543 order
= page_order(page
);
1545 nr_pages
= __isolate_free_page(page
, order
);
1549 /* Split into individual pages */
1550 set_page_refcounted(page
);
1551 split_page(page
, order
);
1556 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1557 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1561 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1562 struct zone
*zone
, unsigned int order
,
1563 gfp_t gfp_flags
, int migratetype
)
1565 unsigned long flags
;
1567 bool cold
= ((gfp_flags
& __GFP_COLD
) != 0);
1570 if (likely(order
== 0)) {
1571 struct per_cpu_pages
*pcp
;
1572 struct list_head
*list
;
1574 local_irq_save(flags
);
1575 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1576 list
= &pcp
->lists
[migratetype
];
1577 if (list_empty(list
)) {
1578 pcp
->count
+= rmqueue_bulk(zone
, 0,
1581 if (unlikely(list_empty(list
)))
1586 page
= list_entry(list
->prev
, struct page
, lru
);
1588 page
= list_entry(list
->next
, struct page
, lru
);
1590 list_del(&page
->lru
);
1593 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1595 * __GFP_NOFAIL is not to be used in new code.
1597 * All __GFP_NOFAIL callers should be fixed so that they
1598 * properly detect and handle allocation failures.
1600 * We most definitely don't want callers attempting to
1601 * allocate greater than order-1 page units with
1604 WARN_ON_ONCE(order
> 1);
1606 spin_lock_irqsave(&zone
->lock
, flags
);
1607 page
= __rmqueue(zone
, order
, migratetype
);
1608 spin_unlock(&zone
->lock
);
1611 __mod_zone_freepage_state(zone
, -(1 << order
),
1612 get_freepage_migratetype(page
));
1615 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1616 if (atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]) <= 0 &&
1617 !test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
))
1618 set_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1620 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1621 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1622 local_irq_restore(flags
);
1624 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1625 if (prep_new_page(page
, order
, gfp_flags
))
1630 local_irq_restore(flags
);
1634 #ifdef CONFIG_FAIL_PAGE_ALLOC
1637 struct fault_attr attr
;
1639 u32 ignore_gfp_highmem
;
1640 u32 ignore_gfp_wait
;
1642 } fail_page_alloc
= {
1643 .attr
= FAULT_ATTR_INITIALIZER
,
1644 .ignore_gfp_wait
= 1,
1645 .ignore_gfp_highmem
= 1,
1649 static int __init
setup_fail_page_alloc(char *str
)
1651 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1653 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1655 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1657 if (order
< fail_page_alloc
.min_order
)
1659 if (gfp_mask
& __GFP_NOFAIL
)
1661 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1663 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1666 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1669 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1671 static int __init
fail_page_alloc_debugfs(void)
1673 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1676 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1677 &fail_page_alloc
.attr
);
1679 return PTR_ERR(dir
);
1681 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1682 &fail_page_alloc
.ignore_gfp_wait
))
1684 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1685 &fail_page_alloc
.ignore_gfp_highmem
))
1687 if (!debugfs_create_u32("min-order", mode
, dir
,
1688 &fail_page_alloc
.min_order
))
1693 debugfs_remove_recursive(dir
);
1698 late_initcall(fail_page_alloc_debugfs
);
1700 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1702 #else /* CONFIG_FAIL_PAGE_ALLOC */
1704 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1709 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1712 * Return true if free pages are above 'mark'. This takes into account the order
1713 * of the allocation.
1715 static bool __zone_watermark_ok(struct zone
*z
, unsigned int order
,
1716 unsigned long mark
, int classzone_idx
, int alloc_flags
,
1719 /* free_pages my go negative - that's OK */
1724 free_pages
-= (1 << order
) - 1;
1725 if (alloc_flags
& ALLOC_HIGH
)
1727 if (alloc_flags
& ALLOC_HARDER
)
1730 /* If allocation can't use CMA areas don't use free CMA pages */
1731 if (!(alloc_flags
& ALLOC_CMA
))
1732 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1735 if (free_pages
- free_cma
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1737 for (o
= 0; o
< order
; o
++) {
1738 /* At the next order, this order's pages become unavailable */
1739 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1741 /* Require fewer higher order pages to be free */
1744 if (free_pages
<= min
)
1750 bool zone_watermark_ok(struct zone
*z
, unsigned int order
, unsigned long mark
,
1751 int classzone_idx
, int alloc_flags
)
1753 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1754 zone_page_state(z
, NR_FREE_PAGES
));
1757 bool zone_watermark_ok_safe(struct zone
*z
, unsigned int order
,
1758 unsigned long mark
, int classzone_idx
, int alloc_flags
)
1760 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1762 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1763 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1765 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1771 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1772 * skip over zones that are not allowed by the cpuset, or that have
1773 * been recently (in last second) found to be nearly full. See further
1774 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1775 * that have to skip over a lot of full or unallowed zones.
1777 * If the zonelist cache is present in the passed zonelist, then
1778 * returns a pointer to the allowed node mask (either the current
1779 * tasks mems_allowed, or node_states[N_MEMORY].)
1781 * If the zonelist cache is not available for this zonelist, does
1782 * nothing and returns NULL.
1784 * If the fullzones BITMAP in the zonelist cache is stale (more than
1785 * a second since last zap'd) then we zap it out (clear its bits.)
1787 * We hold off even calling zlc_setup, until after we've checked the
1788 * first zone in the zonelist, on the theory that most allocations will
1789 * be satisfied from that first zone, so best to examine that zone as
1790 * quickly as we can.
1792 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1794 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1795 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1797 zlc
= zonelist
->zlcache_ptr
;
1801 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1802 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1803 zlc
->last_full_zap
= jiffies
;
1806 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1807 &cpuset_current_mems_allowed
:
1808 &node_states
[N_MEMORY
];
1809 return allowednodes
;
1813 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1814 * if it is worth looking at further for free memory:
1815 * 1) Check that the zone isn't thought to be full (doesn't have its
1816 * bit set in the zonelist_cache fullzones BITMAP).
1817 * 2) Check that the zones node (obtained from the zonelist_cache
1818 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1819 * Return true (non-zero) if zone is worth looking at further, or
1820 * else return false (zero) if it is not.
1822 * This check -ignores- the distinction between various watermarks,
1823 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1824 * found to be full for any variation of these watermarks, it will
1825 * be considered full for up to one second by all requests, unless
1826 * we are so low on memory on all allowed nodes that we are forced
1827 * into the second scan of the zonelist.
1829 * In the second scan we ignore this zonelist cache and exactly
1830 * apply the watermarks to all zones, even it is slower to do so.
1831 * We are low on memory in the second scan, and should leave no stone
1832 * unturned looking for a free page.
1834 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1835 nodemask_t
*allowednodes
)
1837 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1838 int i
; /* index of *z in zonelist zones */
1839 int n
; /* node that zone *z is on */
1841 zlc
= zonelist
->zlcache_ptr
;
1845 i
= z
- zonelist
->_zonerefs
;
1848 /* This zone is worth trying if it is allowed but not full */
1849 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1853 * Given 'z' scanning a zonelist, set the corresponding bit in
1854 * zlc->fullzones, so that subsequent attempts to allocate a page
1855 * from that zone don't waste time re-examining it.
1857 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1859 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1860 int i
; /* index of *z in zonelist zones */
1862 zlc
= zonelist
->zlcache_ptr
;
1866 i
= z
- zonelist
->_zonerefs
;
1868 set_bit(i
, zlc
->fullzones
);
1872 * clear all zones full, called after direct reclaim makes progress so that
1873 * a zone that was recently full is not skipped over for up to a second
1875 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1877 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1879 zlc
= zonelist
->zlcache_ptr
;
1883 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1886 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1888 return local_zone
->node
== zone
->node
;
1891 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1893 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1897 #else /* CONFIG_NUMA */
1899 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1904 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1905 nodemask_t
*allowednodes
)
1910 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1914 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1918 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1923 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1928 #endif /* CONFIG_NUMA */
1930 static void reset_alloc_batches(struct zone
*preferred_zone
)
1932 struct zone
*zone
= preferred_zone
->zone_pgdat
->node_zones
;
1935 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
1936 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
1937 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
1938 clear_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
);
1939 } while (zone
++ != preferred_zone
);
1943 * get_page_from_freelist goes through the zonelist trying to allocate
1946 static struct page
*
1947 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1948 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1949 struct zone
*preferred_zone
, int classzone_idx
, int migratetype
)
1952 struct page
*page
= NULL
;
1954 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1955 int zlc_active
= 0; /* set if using zonelist_cache */
1956 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1957 bool consider_zone_dirty
= (alloc_flags
& ALLOC_WMARK_LOW
) &&
1958 (gfp_mask
& __GFP_WRITE
);
1959 int nr_fair_skipped
= 0;
1960 bool zonelist_rescan
;
1963 zonelist_rescan
= false;
1966 * Scan zonelist, looking for a zone with enough free.
1967 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1969 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1970 high_zoneidx
, nodemask
) {
1973 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1974 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1976 if (cpusets_enabled() &&
1977 (alloc_flags
& ALLOC_CPUSET
) &&
1978 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1981 * Distribute pages in proportion to the individual
1982 * zone size to ensure fair page aging. The zone a
1983 * page was allocated in should have no effect on the
1984 * time the page has in memory before being reclaimed.
1986 if (alloc_flags
& ALLOC_FAIR
) {
1987 if (!zone_local(preferred_zone
, zone
))
1989 if (test_bit(ZONE_FAIR_DEPLETED
, &zone
->flags
)) {
1995 * When allocating a page cache page for writing, we
1996 * want to get it from a zone that is within its dirty
1997 * limit, such that no single zone holds more than its
1998 * proportional share of globally allowed dirty pages.
1999 * The dirty limits take into account the zone's
2000 * lowmem reserves and high watermark so that kswapd
2001 * should be able to balance it without having to
2002 * write pages from its LRU list.
2004 * This may look like it could increase pressure on
2005 * lower zones by failing allocations in higher zones
2006 * before they are full. But the pages that do spill
2007 * over are limited as the lower zones are protected
2008 * by this very same mechanism. It should not become
2009 * a practical burden to them.
2011 * XXX: For now, allow allocations to potentially
2012 * exceed the per-zone dirty limit in the slowpath
2013 * (ALLOC_WMARK_LOW unset) before going into reclaim,
2014 * which is important when on a NUMA setup the allowed
2015 * zones are together not big enough to reach the
2016 * global limit. The proper fix for these situations
2017 * will require awareness of zones in the
2018 * dirty-throttling and the flusher threads.
2020 if (consider_zone_dirty
&& !zone_dirty_ok(zone
))
2023 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
2024 if (!zone_watermark_ok(zone
, order
, mark
,
2025 classzone_idx
, alloc_flags
)) {
2028 /* Checked here to keep the fast path fast */
2029 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
2030 if (alloc_flags
& ALLOC_NO_WATERMARKS
)
2033 if (IS_ENABLED(CONFIG_NUMA
) &&
2034 !did_zlc_setup
&& nr_online_nodes
> 1) {
2036 * we do zlc_setup if there are multiple nodes
2037 * and before considering the first zone allowed
2040 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2045 if (zone_reclaim_mode
== 0 ||
2046 !zone_allows_reclaim(preferred_zone
, zone
))
2047 goto this_zone_full
;
2050 * As we may have just activated ZLC, check if the first
2051 * eligible zone has failed zone_reclaim recently.
2053 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2054 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2057 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2059 case ZONE_RECLAIM_NOSCAN
:
2062 case ZONE_RECLAIM_FULL
:
2063 /* scanned but unreclaimable */
2066 /* did we reclaim enough */
2067 if (zone_watermark_ok(zone
, order
, mark
,
2068 classzone_idx
, alloc_flags
))
2072 * Failed to reclaim enough to meet watermark.
2073 * Only mark the zone full if checking the min
2074 * watermark or if we failed to reclaim just
2075 * 1<<order pages or else the page allocator
2076 * fastpath will prematurely mark zones full
2077 * when the watermark is between the low and
2080 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2081 ret
== ZONE_RECLAIM_SOME
)
2082 goto this_zone_full
;
2089 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2090 gfp_mask
, migratetype
);
2094 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
)
2095 zlc_mark_zone_full(zonelist
, z
);
2100 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2101 * necessary to allocate the page. The expectation is
2102 * that the caller is taking steps that will free more
2103 * memory. The caller should avoid the page being used
2104 * for !PFMEMALLOC purposes.
2106 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2111 * The first pass makes sure allocations are spread fairly within the
2112 * local node. However, the local node might have free pages left
2113 * after the fairness batches are exhausted, and remote zones haven't
2114 * even been considered yet. Try once more without fairness, and
2115 * include remote zones now, before entering the slowpath and waking
2116 * kswapd: prefer spilling to a remote zone over swapping locally.
2118 if (alloc_flags
& ALLOC_FAIR
) {
2119 alloc_flags
&= ~ALLOC_FAIR
;
2120 if (nr_fair_skipped
) {
2121 zonelist_rescan
= true;
2122 reset_alloc_batches(preferred_zone
);
2124 if (nr_online_nodes
> 1)
2125 zonelist_rescan
= true;
2128 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && zlc_active
)) {
2129 /* Disable zlc cache for second zonelist scan */
2131 zonelist_rescan
= true;
2134 if (zonelist_rescan
)
2141 * Large machines with many possible nodes should not always dump per-node
2142 * meminfo in irq context.
2144 static inline bool should_suppress_show_mem(void)
2149 ret
= in_interrupt();
2154 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2155 DEFAULT_RATELIMIT_INTERVAL
,
2156 DEFAULT_RATELIMIT_BURST
);
2158 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2160 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2162 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2163 debug_guardpage_minorder() > 0)
2167 * This documents exceptions given to allocations in certain
2168 * contexts that are allowed to allocate outside current's set
2171 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2172 if (test_thread_flag(TIF_MEMDIE
) ||
2173 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2174 filter
&= ~SHOW_MEM_FILTER_NODES
;
2175 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2176 filter
&= ~SHOW_MEM_FILTER_NODES
;
2179 struct va_format vaf
;
2182 va_start(args
, fmt
);
2187 pr_warn("%pV", &vaf
);
2192 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2193 current
->comm
, order
, gfp_mask
);
2196 if (!should_suppress_show_mem())
2201 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2202 unsigned long did_some_progress
,
2203 unsigned long pages_reclaimed
)
2205 /* Do not loop if specifically requested */
2206 if (gfp_mask
& __GFP_NORETRY
)
2209 /* Always retry if specifically requested */
2210 if (gfp_mask
& __GFP_NOFAIL
)
2214 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2215 * making forward progress without invoking OOM. Suspend also disables
2216 * storage devices so kswapd will not help. Bail if we are suspending.
2218 if (!did_some_progress
&& pm_suspended_storage())
2222 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2223 * means __GFP_NOFAIL, but that may not be true in other
2226 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2230 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2231 * specified, then we retry until we no longer reclaim any pages
2232 * (above), or we've reclaimed an order of pages at least as
2233 * large as the allocation's order. In both cases, if the
2234 * allocation still fails, we stop retrying.
2236 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2242 static inline struct page
*
2243 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2244 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2245 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2246 int classzone_idx
, int migratetype
)
2250 /* Acquire the per-zone oom lock for each zone */
2251 if (!oom_zonelist_trylock(zonelist
, gfp_mask
)) {
2252 schedule_timeout_uninterruptible(1);
2257 * PM-freezer should be notified that there might be an OOM killer on
2258 * its way to kill and wake somebody up. This is too early and we might
2259 * end up not killing anything but false positives are acceptable.
2260 * See freeze_processes.
2265 * Go through the zonelist yet one more time, keep very high watermark
2266 * here, this is only to catch a parallel oom killing, we must fail if
2267 * we're still under heavy pressure.
2269 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2270 order
, zonelist
, high_zoneidx
,
2271 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2272 preferred_zone
, classzone_idx
, migratetype
);
2276 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2277 /* The OOM killer will not help higher order allocs */
2278 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2280 /* The OOM killer does not needlessly kill tasks for lowmem */
2281 if (high_zoneidx
< ZONE_NORMAL
)
2284 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2285 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2286 * The caller should handle page allocation failure by itself if
2287 * it specifies __GFP_THISNODE.
2288 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2290 if (gfp_mask
& __GFP_THISNODE
)
2293 /* Exhausted what can be done so it's blamo time */
2294 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2297 oom_zonelist_unlock(zonelist
, gfp_mask
);
2301 #ifdef CONFIG_COMPACTION
2302 /* Try memory compaction for high-order allocations before reclaim */
2303 static struct page
*
2304 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2305 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2306 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2307 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2308 int *contended_compaction
, bool *deferred_compaction
)
2310 struct zone
*last_compact_zone
= NULL
;
2311 unsigned long compact_result
;
2317 current
->flags
|= PF_MEMALLOC
;
2318 compact_result
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2320 contended_compaction
,
2321 &last_compact_zone
);
2322 current
->flags
&= ~PF_MEMALLOC
;
2324 switch (compact_result
) {
2325 case COMPACT_DEFERRED
:
2326 *deferred_compaction
= true;
2328 case COMPACT_SKIPPED
:
2335 * At least in one zone compaction wasn't deferred or skipped, so let's
2336 * count a compaction stall
2338 count_vm_event(COMPACTSTALL
);
2340 /* Page migration frees to the PCP lists but we want merging */
2341 drain_pages(get_cpu());
2344 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2345 order
, zonelist
, high_zoneidx
,
2346 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2347 preferred_zone
, classzone_idx
, migratetype
);
2350 struct zone
*zone
= page_zone(page
);
2352 zone
->compact_blockskip_flush
= false;
2353 compaction_defer_reset(zone
, order
, true);
2354 count_vm_event(COMPACTSUCCESS
);
2359 * last_compact_zone is where try_to_compact_pages thought allocation
2360 * should succeed, so it did not defer compaction. But here we know
2361 * that it didn't succeed, so we do the defer.
2363 if (last_compact_zone
&& mode
!= MIGRATE_ASYNC
)
2364 defer_compaction(last_compact_zone
, order
);
2367 * It's bad if compaction run occurs and fails. The most likely reason
2368 * is that pages exist, but not enough to satisfy watermarks.
2370 count_vm_event(COMPACTFAIL
);
2377 static inline struct page
*
2378 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2379 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2380 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2381 int classzone_idx
, int migratetype
, enum migrate_mode mode
,
2382 int *contended_compaction
, bool *deferred_compaction
)
2386 #endif /* CONFIG_COMPACTION */
2388 /* Perform direct synchronous page reclaim */
2390 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2391 nodemask_t
*nodemask
)
2393 struct reclaim_state reclaim_state
;
2398 /* We now go into synchronous reclaim */
2399 cpuset_memory_pressure_bump();
2400 current
->flags
|= PF_MEMALLOC
;
2401 lockdep_set_current_reclaim_state(gfp_mask
);
2402 reclaim_state
.reclaimed_slab
= 0;
2403 current
->reclaim_state
= &reclaim_state
;
2405 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2407 current
->reclaim_state
= NULL
;
2408 lockdep_clear_current_reclaim_state();
2409 current
->flags
&= ~PF_MEMALLOC
;
2416 /* The really slow allocator path where we enter direct reclaim */
2417 static inline struct page
*
2418 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2419 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2420 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2421 int classzone_idx
, int migratetype
, unsigned long *did_some_progress
)
2423 struct page
*page
= NULL
;
2424 bool drained
= false;
2426 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2428 if (unlikely(!(*did_some_progress
)))
2431 /* After successful reclaim, reconsider all zones for allocation */
2432 if (IS_ENABLED(CONFIG_NUMA
))
2433 zlc_clear_zones_full(zonelist
);
2436 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2437 zonelist
, high_zoneidx
,
2438 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2439 preferred_zone
, classzone_idx
,
2443 * If an allocation failed after direct reclaim, it could be because
2444 * pages are pinned on the per-cpu lists. Drain them and try again
2446 if (!page
&& !drained
) {
2456 * This is called in the allocator slow-path if the allocation request is of
2457 * sufficient urgency to ignore watermarks and take other desperate measures
2459 static inline struct page
*
2460 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2461 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2462 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2463 int classzone_idx
, int migratetype
)
2468 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2469 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2470 preferred_zone
, classzone_idx
, migratetype
);
2472 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2473 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2474 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2479 static void wake_all_kswapds(unsigned int order
,
2480 struct zonelist
*zonelist
,
2481 enum zone_type high_zoneidx
,
2482 struct zone
*preferred_zone
,
2483 nodemask_t
*nodemask
)
2488 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
2489 high_zoneidx
, nodemask
)
2490 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2494 gfp_to_alloc_flags(gfp_t gfp_mask
)
2496 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2497 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2499 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2500 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2503 * The caller may dip into page reserves a bit more if the caller
2504 * cannot run direct reclaim, or if the caller has realtime scheduling
2505 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2506 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2508 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2512 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2513 * if it can't schedule.
2515 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2516 alloc_flags
|= ALLOC_HARDER
;
2518 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2519 * comment for __cpuset_node_allowed_softwall().
2521 alloc_flags
&= ~ALLOC_CPUSET
;
2522 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2523 alloc_flags
|= ALLOC_HARDER
;
2525 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2526 if (gfp_mask
& __GFP_MEMALLOC
)
2527 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2528 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2529 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2530 else if (!in_interrupt() &&
2531 ((current
->flags
& PF_MEMALLOC
) ||
2532 unlikely(test_thread_flag(TIF_MEMDIE
))))
2533 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2536 if (gfpflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2537 alloc_flags
|= ALLOC_CMA
;
2542 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2544 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2547 static inline struct page
*
2548 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2549 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2550 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2551 int classzone_idx
, int migratetype
)
2553 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2554 struct page
*page
= NULL
;
2556 unsigned long pages_reclaimed
= 0;
2557 unsigned long did_some_progress
;
2558 enum migrate_mode migration_mode
= MIGRATE_ASYNC
;
2559 bool deferred_compaction
= false;
2560 int contended_compaction
= COMPACT_CONTENDED_NONE
;
2563 * In the slowpath, we sanity check order to avoid ever trying to
2564 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2565 * be using allocators in order of preference for an area that is
2568 if (order
>= MAX_ORDER
) {
2569 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2574 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2575 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2576 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2577 * using a larger set of nodes after it has established that the
2578 * allowed per node queues are empty and that nodes are
2581 if (IS_ENABLED(CONFIG_NUMA
) &&
2582 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2586 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2587 wake_all_kswapds(order
, zonelist
, high_zoneidx
,
2588 preferred_zone
, nodemask
);
2591 * OK, we're below the kswapd watermark and have kicked background
2592 * reclaim. Now things get more complex, so set up alloc_flags according
2593 * to how we want to proceed.
2595 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2598 * Find the true preferred zone if the allocation is unconstrained by
2601 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
) {
2602 struct zoneref
*preferred_zoneref
;
2603 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2604 NULL
, &preferred_zone
);
2605 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2609 /* This is the last chance, in general, before the goto nopage. */
2610 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2611 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2612 preferred_zone
, classzone_idx
, migratetype
);
2616 /* Allocate without watermarks if the context allows */
2617 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2619 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2620 * the allocation is high priority and these type of
2621 * allocations are system rather than user orientated
2623 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2625 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2626 zonelist
, high_zoneidx
, nodemask
,
2627 preferred_zone
, classzone_idx
, migratetype
);
2633 /* Atomic allocations - we can't balance anything */
2636 * All existing users of the deprecated __GFP_NOFAIL are
2637 * blockable, so warn of any new users that actually allow this
2638 * type of allocation to fail.
2640 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2644 /* Avoid recursion of direct reclaim */
2645 if (current
->flags
& PF_MEMALLOC
)
2648 /* Avoid allocations with no watermarks from looping endlessly */
2649 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2653 * Try direct compaction. The first pass is asynchronous. Subsequent
2654 * attempts after direct reclaim are synchronous
2656 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2657 high_zoneidx
, nodemask
, alloc_flags
,
2659 classzone_idx
, migratetype
,
2660 migration_mode
, &contended_compaction
,
2661 &deferred_compaction
);
2665 /* Checks for THP-specific high-order allocations */
2666 if ((gfp_mask
& GFP_TRANSHUGE
) == GFP_TRANSHUGE
) {
2668 * If compaction is deferred for high-order allocations, it is
2669 * because sync compaction recently failed. If this is the case
2670 * and the caller requested a THP allocation, we do not want
2671 * to heavily disrupt the system, so we fail the allocation
2672 * instead of entering direct reclaim.
2674 if (deferred_compaction
)
2678 * In all zones where compaction was attempted (and not
2679 * deferred or skipped), lock contention has been detected.
2680 * For THP allocation we do not want to disrupt the others
2681 * so we fallback to base pages instead.
2683 if (contended_compaction
== COMPACT_CONTENDED_LOCK
)
2687 * If compaction was aborted due to need_resched(), we do not
2688 * want to further increase allocation latency, unless it is
2689 * khugepaged trying to collapse.
2691 if (contended_compaction
== COMPACT_CONTENDED_SCHED
2692 && !(current
->flags
& PF_KTHREAD
))
2697 * It can become very expensive to allocate transparent hugepages at
2698 * fault, so use asynchronous memory compaction for THP unless it is
2699 * khugepaged trying to collapse.
2701 if ((gfp_mask
& GFP_TRANSHUGE
) != GFP_TRANSHUGE
||
2702 (current
->flags
& PF_KTHREAD
))
2703 migration_mode
= MIGRATE_SYNC_LIGHT
;
2705 /* Try direct reclaim and then allocating */
2706 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2707 zonelist
, high_zoneidx
,
2709 alloc_flags
, preferred_zone
,
2710 classzone_idx
, migratetype
,
2711 &did_some_progress
);
2716 * If we failed to make any progress reclaiming, then we are
2717 * running out of options and have to consider going OOM
2719 if (!did_some_progress
) {
2720 if (oom_gfp_allowed(gfp_mask
)) {
2721 if (oom_killer_disabled
)
2723 /* Coredumps can quickly deplete all memory reserves */
2724 if ((current
->flags
& PF_DUMPCORE
) &&
2725 !(gfp_mask
& __GFP_NOFAIL
))
2727 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2728 zonelist
, high_zoneidx
,
2729 nodemask
, preferred_zone
,
2730 classzone_idx
, migratetype
);
2734 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2736 * The oom killer is not called for high-order
2737 * allocations that may fail, so if no progress
2738 * is being made, there are no other options and
2739 * retrying is unlikely to help.
2741 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2744 * The oom killer is not called for lowmem
2745 * allocations to prevent needlessly killing
2748 if (high_zoneidx
< ZONE_NORMAL
)
2756 /* Check if we should retry the allocation */
2757 pages_reclaimed
+= did_some_progress
;
2758 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2760 /* Wait for some write requests to complete then retry */
2761 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2765 * High-order allocations do not necessarily loop after
2766 * direct reclaim and reclaim/compaction depends on compaction
2767 * being called after reclaim so call directly if necessary
2769 page
= __alloc_pages_direct_compact(gfp_mask
, order
, zonelist
,
2770 high_zoneidx
, nodemask
, alloc_flags
,
2772 classzone_idx
, migratetype
,
2773 migration_mode
, &contended_compaction
,
2774 &deferred_compaction
);
2780 warn_alloc_failed(gfp_mask
, order
, NULL
);
2783 if (kmemcheck_enabled
)
2784 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2790 * This is the 'heart' of the zoned buddy allocator.
2793 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2794 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2796 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2797 struct zone
*preferred_zone
;
2798 struct zoneref
*preferred_zoneref
;
2799 struct page
*page
= NULL
;
2800 int migratetype
= gfpflags_to_migratetype(gfp_mask
);
2801 unsigned int cpuset_mems_cookie
;
2802 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2805 gfp_mask
&= gfp_allowed_mask
;
2807 lockdep_trace_alloc(gfp_mask
);
2809 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2811 if (should_fail_alloc_page(gfp_mask
, order
))
2815 * Check the zones suitable for the gfp_mask contain at least one
2816 * valid zone. It's possible to have an empty zonelist as a result
2817 * of GFP_THISNODE and a memoryless node
2819 if (unlikely(!zonelist
->_zonerefs
->zone
))
2822 if (IS_ENABLED(CONFIG_CMA
) && migratetype
== MIGRATE_MOVABLE
)
2823 alloc_flags
|= ALLOC_CMA
;
2826 cpuset_mems_cookie
= read_mems_allowed_begin();
2828 /* The preferred zone is used for statistics later */
2829 preferred_zoneref
= first_zones_zonelist(zonelist
, high_zoneidx
,
2830 nodemask
? : &cpuset_current_mems_allowed
,
2832 if (!preferred_zone
)
2834 classzone_idx
= zonelist_zone_idx(preferred_zoneref
);
2836 /* First allocation attempt */
2837 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2838 zonelist
, high_zoneidx
, alloc_flags
,
2839 preferred_zone
, classzone_idx
, migratetype
);
2840 if (unlikely(!page
)) {
2842 * Runtime PM, block IO and its error handling path
2843 * can deadlock because I/O on the device might not
2846 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2847 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2848 zonelist
, high_zoneidx
, nodemask
,
2849 preferred_zone
, classzone_idx
, migratetype
);
2852 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2856 * When updating a task's mems_allowed, it is possible to race with
2857 * parallel threads in such a way that an allocation can fail while
2858 * the mask is being updated. If a page allocation is about to fail,
2859 * check if the cpuset changed during allocation and if so, retry.
2861 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2866 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2869 * Common helper functions.
2871 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2876 * __get_free_pages() returns a 32-bit address, which cannot represent
2879 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2881 page
= alloc_pages(gfp_mask
, order
);
2884 return (unsigned long) page_address(page
);
2886 EXPORT_SYMBOL(__get_free_pages
);
2888 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2890 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2892 EXPORT_SYMBOL(get_zeroed_page
);
2894 void __free_pages(struct page
*page
, unsigned int order
)
2896 if (put_page_testzero(page
)) {
2898 free_hot_cold_page(page
, false);
2900 __free_pages_ok(page
, order
);
2904 EXPORT_SYMBOL(__free_pages
);
2906 void free_pages(unsigned long addr
, unsigned int order
)
2909 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2910 __free_pages(virt_to_page((void *)addr
), order
);
2914 EXPORT_SYMBOL(free_pages
);
2917 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2918 * of the current memory cgroup.
2920 * It should be used when the caller would like to use kmalloc, but since the
2921 * allocation is large, it has to fall back to the page allocator.
2923 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2926 struct mem_cgroup
*memcg
= NULL
;
2928 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2930 page
= alloc_pages(gfp_mask
, order
);
2931 memcg_kmem_commit_charge(page
, memcg
, order
);
2935 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2938 struct mem_cgroup
*memcg
= NULL
;
2940 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2942 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2943 memcg_kmem_commit_charge(page
, memcg
, order
);
2948 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2951 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2953 memcg_kmem_uncharge_pages(page
, order
);
2954 __free_pages(page
, order
);
2957 void free_kmem_pages(unsigned long addr
, unsigned int order
)
2960 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2961 __free_kmem_pages(virt_to_page((void *)addr
), order
);
2965 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2968 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2969 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2971 split_page(virt_to_page((void *)addr
), order
);
2972 while (used
< alloc_end
) {
2977 return (void *)addr
;
2981 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2982 * @size: the number of bytes to allocate
2983 * @gfp_mask: GFP flags for the allocation
2985 * This function is similar to alloc_pages(), except that it allocates the
2986 * minimum number of pages to satisfy the request. alloc_pages() can only
2987 * allocate memory in power-of-two pages.
2989 * This function is also limited by MAX_ORDER.
2991 * Memory allocated by this function must be released by free_pages_exact().
2993 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2995 unsigned int order
= get_order(size
);
2998 addr
= __get_free_pages(gfp_mask
, order
);
2999 return make_alloc_exact(addr
, order
, size
);
3001 EXPORT_SYMBOL(alloc_pages_exact
);
3004 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
3006 * @nid: the preferred node ID where memory should be allocated
3007 * @size: the number of bytes to allocate
3008 * @gfp_mask: GFP flags for the allocation
3010 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
3012 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
3015 void * __meminit
alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
3017 unsigned order
= get_order(size
);
3018 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
3021 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
3025 * free_pages_exact - release memory allocated via alloc_pages_exact()
3026 * @virt: the value returned by alloc_pages_exact.
3027 * @size: size of allocation, same value as passed to alloc_pages_exact().
3029 * Release the memory allocated by a previous call to alloc_pages_exact.
3031 void free_pages_exact(void *virt
, size_t size
)
3033 unsigned long addr
= (unsigned long)virt
;
3034 unsigned long end
= addr
+ PAGE_ALIGN(size
);
3036 while (addr
< end
) {
3041 EXPORT_SYMBOL(free_pages_exact
);
3044 * nr_free_zone_pages - count number of pages beyond high watermark
3045 * @offset: The zone index of the highest zone
3047 * nr_free_zone_pages() counts the number of counts pages which are beyond the
3048 * high watermark within all zones at or below a given zone index. For each
3049 * zone, the number of pages is calculated as:
3050 * managed_pages - high_pages
3052 static unsigned long nr_free_zone_pages(int offset
)
3057 /* Just pick one node, since fallback list is circular */
3058 unsigned long sum
= 0;
3060 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
3062 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
3063 unsigned long size
= zone
->managed_pages
;
3064 unsigned long high
= high_wmark_pages(zone
);
3073 * nr_free_buffer_pages - count number of pages beyond high watermark
3075 * nr_free_buffer_pages() counts the number of pages which are beyond the high
3076 * watermark within ZONE_DMA and ZONE_NORMAL.
3078 unsigned long nr_free_buffer_pages(void)
3080 return nr_free_zone_pages(gfp_zone(GFP_USER
));
3082 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
3085 * nr_free_pagecache_pages - count number of pages beyond high watermark
3087 * nr_free_pagecache_pages() counts the number of pages which are beyond the
3088 * high watermark within all zones.
3090 unsigned long nr_free_pagecache_pages(void)
3092 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
3095 static inline void show_node(struct zone
*zone
)
3097 if (IS_ENABLED(CONFIG_NUMA
))
3098 printk("Node %d ", zone_to_nid(zone
));
3101 void si_meminfo(struct sysinfo
*val
)
3103 val
->totalram
= totalram_pages
;
3104 val
->sharedram
= global_page_state(NR_SHMEM
);
3105 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3106 val
->bufferram
= nr_blockdev_pages();
3107 val
->totalhigh
= totalhigh_pages
;
3108 val
->freehigh
= nr_free_highpages();
3109 val
->mem_unit
= PAGE_SIZE
;
3112 EXPORT_SYMBOL(si_meminfo
);
3115 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3117 int zone_type
; /* needs to be signed */
3118 unsigned long managed_pages
= 0;
3119 pg_data_t
*pgdat
= NODE_DATA(nid
);
3121 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3122 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3123 val
->totalram
= managed_pages
;
3124 val
->sharedram
= node_page_state(nid
, NR_SHMEM
);
3125 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3126 #ifdef CONFIG_HIGHMEM
3127 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3128 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3134 val
->mem_unit
= PAGE_SIZE
;
3139 * Determine whether the node should be displayed or not, depending on whether
3140 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3142 bool skip_free_areas_node(unsigned int flags
, int nid
)
3145 unsigned int cpuset_mems_cookie
;
3147 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3151 cpuset_mems_cookie
= read_mems_allowed_begin();
3152 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3153 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3158 #define K(x) ((x) << (PAGE_SHIFT-10))
3160 static void show_migration_types(unsigned char type
)
3162 static const char types
[MIGRATE_TYPES
] = {
3163 [MIGRATE_UNMOVABLE
] = 'U',
3164 [MIGRATE_RECLAIMABLE
] = 'E',
3165 [MIGRATE_MOVABLE
] = 'M',
3166 [MIGRATE_RESERVE
] = 'R',
3168 [MIGRATE_CMA
] = 'C',
3170 #ifdef CONFIG_MEMORY_ISOLATION
3171 [MIGRATE_ISOLATE
] = 'I',
3174 char tmp
[MIGRATE_TYPES
+ 1];
3178 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3179 if (type
& (1 << i
))
3184 printk("(%s) ", tmp
);
3188 * Show free area list (used inside shift_scroll-lock stuff)
3189 * We also calculate the percentage fragmentation. We do this by counting the
3190 * memory on each free list with the exception of the first item on the list.
3191 * Suppresses nodes that are not allowed by current's cpuset if
3192 * SHOW_MEM_FILTER_NODES is passed.
3194 void show_free_areas(unsigned int filter
)
3199 for_each_populated_zone(zone
) {
3200 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3203 printk("%s per-cpu:\n", zone
->name
);
3205 for_each_online_cpu(cpu
) {
3206 struct per_cpu_pageset
*pageset
;
3208 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3210 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3211 cpu
, pageset
->pcp
.high
,
3212 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3216 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3217 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3219 " dirty:%lu writeback:%lu unstable:%lu\n"
3220 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3221 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3223 global_page_state(NR_ACTIVE_ANON
),
3224 global_page_state(NR_INACTIVE_ANON
),
3225 global_page_state(NR_ISOLATED_ANON
),
3226 global_page_state(NR_ACTIVE_FILE
),
3227 global_page_state(NR_INACTIVE_FILE
),
3228 global_page_state(NR_ISOLATED_FILE
),
3229 global_page_state(NR_UNEVICTABLE
),
3230 global_page_state(NR_FILE_DIRTY
),
3231 global_page_state(NR_WRITEBACK
),
3232 global_page_state(NR_UNSTABLE_NFS
),
3233 global_page_state(NR_FREE_PAGES
),
3234 global_page_state(NR_SLAB_RECLAIMABLE
),
3235 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3236 global_page_state(NR_FILE_MAPPED
),
3237 global_page_state(NR_SHMEM
),
3238 global_page_state(NR_PAGETABLE
),
3239 global_page_state(NR_BOUNCE
),
3240 global_page_state(NR_FREE_CMA_PAGES
));
3242 for_each_populated_zone(zone
) {
3245 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3253 " active_anon:%lukB"
3254 " inactive_anon:%lukB"
3255 " active_file:%lukB"
3256 " inactive_file:%lukB"
3257 " unevictable:%lukB"
3258 " isolated(anon):%lukB"
3259 " isolated(file):%lukB"
3267 " slab_reclaimable:%lukB"
3268 " slab_unreclaimable:%lukB"
3269 " kernel_stack:%lukB"
3274 " writeback_tmp:%lukB"
3275 " pages_scanned:%lu"
3276 " all_unreclaimable? %s"
3279 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3280 K(min_wmark_pages(zone
)),
3281 K(low_wmark_pages(zone
)),
3282 K(high_wmark_pages(zone
)),
3283 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3284 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3285 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3286 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3287 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3288 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3289 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3290 K(zone
->present_pages
),
3291 K(zone
->managed_pages
),
3292 K(zone_page_state(zone
, NR_MLOCK
)),
3293 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3294 K(zone_page_state(zone
, NR_WRITEBACK
)),
3295 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3296 K(zone_page_state(zone
, NR_SHMEM
)),
3297 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3298 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3299 zone_page_state(zone
, NR_KERNEL_STACK
) *
3301 K(zone_page_state(zone
, NR_PAGETABLE
)),
3302 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3303 K(zone_page_state(zone
, NR_BOUNCE
)),
3304 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3305 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3306 K(zone_page_state(zone
, NR_PAGES_SCANNED
)),
3307 (!zone_reclaimable(zone
) ? "yes" : "no")
3309 printk("lowmem_reserve[]:");
3310 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3311 printk(" %ld", zone
->lowmem_reserve
[i
]);
3315 for_each_populated_zone(zone
) {
3316 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3317 unsigned char types
[MAX_ORDER
];
3319 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3322 printk("%s: ", zone
->name
);
3324 spin_lock_irqsave(&zone
->lock
, flags
);
3325 for (order
= 0; order
< MAX_ORDER
; order
++) {
3326 struct free_area
*area
= &zone
->free_area
[order
];
3329 nr
[order
] = area
->nr_free
;
3330 total
+= nr
[order
] << order
;
3333 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3334 if (!list_empty(&area
->free_list
[type
]))
3335 types
[order
] |= 1 << type
;
3338 spin_unlock_irqrestore(&zone
->lock
, flags
);
3339 for (order
= 0; order
< MAX_ORDER
; order
++) {
3340 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3342 show_migration_types(types
[order
]);
3344 printk("= %lukB\n", K(total
));
3347 hugetlb_show_meminfo();
3349 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3351 show_swap_cache_info();
3354 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3356 zoneref
->zone
= zone
;
3357 zoneref
->zone_idx
= zone_idx(zone
);
3361 * Builds allocation fallback zone lists.
3363 * Add all populated zones of a node to the zonelist.
3365 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3369 enum zone_type zone_type
= MAX_NR_ZONES
;
3373 zone
= pgdat
->node_zones
+ zone_type
;
3374 if (populated_zone(zone
)) {
3375 zoneref_set_zone(zone
,
3376 &zonelist
->_zonerefs
[nr_zones
++]);
3377 check_highest_zone(zone_type
);
3379 } while (zone_type
);
3387 * 0 = automatic detection of better ordering.
3388 * 1 = order by ([node] distance, -zonetype)
3389 * 2 = order by (-zonetype, [node] distance)
3391 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3392 * the same zonelist. So only NUMA can configure this param.
3394 #define ZONELIST_ORDER_DEFAULT 0
3395 #define ZONELIST_ORDER_NODE 1
3396 #define ZONELIST_ORDER_ZONE 2
3398 /* zonelist order in the kernel.
3399 * set_zonelist_order() will set this to NODE or ZONE.
3401 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3402 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3406 /* The value user specified ....changed by config */
3407 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3408 /* string for sysctl */
3409 #define NUMA_ZONELIST_ORDER_LEN 16
3410 char numa_zonelist_order
[16] = "default";
3413 * interface for configure zonelist ordering.
3414 * command line option "numa_zonelist_order"
3415 * = "[dD]efault - default, automatic configuration.
3416 * = "[nN]ode - order by node locality, then by zone within node
3417 * = "[zZ]one - order by zone, then by locality within zone
3420 static int __parse_numa_zonelist_order(char *s
)
3422 if (*s
== 'd' || *s
== 'D') {
3423 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3424 } else if (*s
== 'n' || *s
== 'N') {
3425 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3426 } else if (*s
== 'z' || *s
== 'Z') {
3427 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3430 "Ignoring invalid numa_zonelist_order value: "
3437 static __init
int setup_numa_zonelist_order(char *s
)
3444 ret
= __parse_numa_zonelist_order(s
);
3446 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3450 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3453 * sysctl handler for numa_zonelist_order
3455 int numa_zonelist_order_handler(struct ctl_table
*table
, int write
,
3456 void __user
*buffer
, size_t *length
,
3459 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3461 static DEFINE_MUTEX(zl_order_mutex
);
3463 mutex_lock(&zl_order_mutex
);
3465 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3469 strcpy(saved_string
, (char *)table
->data
);
3471 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3475 int oldval
= user_zonelist_order
;
3477 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3480 * bogus value. restore saved string
3482 strncpy((char *)table
->data
, saved_string
,
3483 NUMA_ZONELIST_ORDER_LEN
);
3484 user_zonelist_order
= oldval
;
3485 } else if (oldval
!= user_zonelist_order
) {
3486 mutex_lock(&zonelists_mutex
);
3487 build_all_zonelists(NULL
, NULL
);
3488 mutex_unlock(&zonelists_mutex
);
3492 mutex_unlock(&zl_order_mutex
);
3497 #define MAX_NODE_LOAD (nr_online_nodes)
3498 static int node_load
[MAX_NUMNODES
];
3501 * find_next_best_node - find the next node that should appear in a given node's fallback list
3502 * @node: node whose fallback list we're appending
3503 * @used_node_mask: nodemask_t of already used nodes
3505 * We use a number of factors to determine which is the next node that should
3506 * appear on a given node's fallback list. The node should not have appeared
3507 * already in @node's fallback list, and it should be the next closest node
3508 * according to the distance array (which contains arbitrary distance values
3509 * from each node to each node in the system), and should also prefer nodes
3510 * with no CPUs, since presumably they'll have very little allocation pressure
3511 * on them otherwise.
3512 * It returns -1 if no node is found.
3514 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3517 int min_val
= INT_MAX
;
3518 int best_node
= NUMA_NO_NODE
;
3519 const struct cpumask
*tmp
= cpumask_of_node(0);
3521 /* Use the local node if we haven't already */
3522 if (!node_isset(node
, *used_node_mask
)) {
3523 node_set(node
, *used_node_mask
);
3527 for_each_node_state(n
, N_MEMORY
) {
3529 /* Don't want a node to appear more than once */
3530 if (node_isset(n
, *used_node_mask
))
3533 /* Use the distance array to find the distance */
3534 val
= node_distance(node
, n
);
3536 /* Penalize nodes under us ("prefer the next node") */
3539 /* Give preference to headless and unused nodes */
3540 tmp
= cpumask_of_node(n
);
3541 if (!cpumask_empty(tmp
))
3542 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3544 /* Slight preference for less loaded node */
3545 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3546 val
+= node_load
[n
];
3548 if (val
< min_val
) {
3555 node_set(best_node
, *used_node_mask
);
3562 * Build zonelists ordered by node and zones within node.
3563 * This results in maximum locality--normal zone overflows into local
3564 * DMA zone, if any--but risks exhausting DMA zone.
3566 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3569 struct zonelist
*zonelist
;
3571 zonelist
= &pgdat
->node_zonelists
[0];
3572 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3574 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3575 zonelist
->_zonerefs
[j
].zone
= NULL
;
3576 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3580 * Build gfp_thisnode zonelists
3582 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3585 struct zonelist
*zonelist
;
3587 zonelist
= &pgdat
->node_zonelists
[1];
3588 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3589 zonelist
->_zonerefs
[j
].zone
= NULL
;
3590 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3594 * Build zonelists ordered by zone and nodes within zones.
3595 * This results in conserving DMA zone[s] until all Normal memory is
3596 * exhausted, but results in overflowing to remote node while memory
3597 * may still exist in local DMA zone.
3599 static int node_order
[MAX_NUMNODES
];
3601 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3604 int zone_type
; /* needs to be signed */
3606 struct zonelist
*zonelist
;
3608 zonelist
= &pgdat
->node_zonelists
[0];
3610 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3611 for (j
= 0; j
< nr_nodes
; j
++) {
3612 node
= node_order
[j
];
3613 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3614 if (populated_zone(z
)) {
3616 &zonelist
->_zonerefs
[pos
++]);
3617 check_highest_zone(zone_type
);
3621 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3622 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3625 #if defined(CONFIG_64BIT)
3627 * Devices that require DMA32/DMA are relatively rare and do not justify a
3628 * penalty to every machine in case the specialised case applies. Default
3629 * to Node-ordering on 64-bit NUMA machines
3631 static int default_zonelist_order(void)
3633 return ZONELIST_ORDER_NODE
;
3637 * On 32-bit, the Normal zone needs to be preserved for allocations accessible
3638 * by the kernel. If processes running on node 0 deplete the low memory zone
3639 * then reclaim will occur more frequency increasing stalls and potentially
3640 * be easier to OOM if a large percentage of the zone is under writeback or
3641 * dirty. The problem is significantly worse if CONFIG_HIGHPTE is not set.
3642 * Hence, default to zone ordering on 32-bit.
3644 static int default_zonelist_order(void)
3646 return ZONELIST_ORDER_ZONE
;
3648 #endif /* CONFIG_64BIT */
3650 static void set_zonelist_order(void)
3652 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3653 current_zonelist_order
= default_zonelist_order();
3655 current_zonelist_order
= user_zonelist_order
;
3658 static void build_zonelists(pg_data_t
*pgdat
)
3662 nodemask_t used_mask
;
3663 int local_node
, prev_node
;
3664 struct zonelist
*zonelist
;
3665 int order
= current_zonelist_order
;
3667 /* initialize zonelists */
3668 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3669 zonelist
= pgdat
->node_zonelists
+ i
;
3670 zonelist
->_zonerefs
[0].zone
= NULL
;
3671 zonelist
->_zonerefs
[0].zone_idx
= 0;
3674 /* NUMA-aware ordering of nodes */
3675 local_node
= pgdat
->node_id
;
3676 load
= nr_online_nodes
;
3677 prev_node
= local_node
;
3678 nodes_clear(used_mask
);
3680 memset(node_order
, 0, sizeof(node_order
));
3683 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3685 * We don't want to pressure a particular node.
3686 * So adding penalty to the first node in same
3687 * distance group to make it round-robin.
3689 if (node_distance(local_node
, node
) !=
3690 node_distance(local_node
, prev_node
))
3691 node_load
[node
] = load
;
3695 if (order
== ZONELIST_ORDER_NODE
)
3696 build_zonelists_in_node_order(pgdat
, node
);
3698 node_order
[j
++] = node
; /* remember order */
3701 if (order
== ZONELIST_ORDER_ZONE
) {
3702 /* calculate node order -- i.e., DMA last! */
3703 build_zonelists_in_zone_order(pgdat
, j
);
3706 build_thisnode_zonelists(pgdat
);
3709 /* Construct the zonelist performance cache - see further mmzone.h */
3710 static void build_zonelist_cache(pg_data_t
*pgdat
)
3712 struct zonelist
*zonelist
;
3713 struct zonelist_cache
*zlc
;
3716 zonelist
= &pgdat
->node_zonelists
[0];
3717 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3718 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3719 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3720 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3723 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3725 * Return node id of node used for "local" allocations.
3726 * I.e., first node id of first zone in arg node's generic zonelist.
3727 * Used for initializing percpu 'numa_mem', which is used primarily
3728 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3730 int local_memory_node(int node
)
3734 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3735 gfp_zone(GFP_KERNEL
),
3742 #else /* CONFIG_NUMA */
3744 static void set_zonelist_order(void)
3746 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3749 static void build_zonelists(pg_data_t
*pgdat
)
3751 int node
, local_node
;
3753 struct zonelist
*zonelist
;
3755 local_node
= pgdat
->node_id
;
3757 zonelist
= &pgdat
->node_zonelists
[0];
3758 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3761 * Now we build the zonelist so that it contains the zones
3762 * of all the other nodes.
3763 * We don't want to pressure a particular node, so when
3764 * building the zones for node N, we make sure that the
3765 * zones coming right after the local ones are those from
3766 * node N+1 (modulo N)
3768 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3769 if (!node_online(node
))
3771 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3773 for (node
= 0; node
< local_node
; node
++) {
3774 if (!node_online(node
))
3776 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3779 zonelist
->_zonerefs
[j
].zone
= NULL
;
3780 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3783 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3784 static void build_zonelist_cache(pg_data_t
*pgdat
)
3786 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3789 #endif /* CONFIG_NUMA */
3792 * Boot pageset table. One per cpu which is going to be used for all
3793 * zones and all nodes. The parameters will be set in such a way
3794 * that an item put on a list will immediately be handed over to
3795 * the buddy list. This is safe since pageset manipulation is done
3796 * with interrupts disabled.
3798 * The boot_pagesets must be kept even after bootup is complete for
3799 * unused processors and/or zones. They do play a role for bootstrapping
3800 * hotplugged processors.
3802 * zoneinfo_show() and maybe other functions do
3803 * not check if the processor is online before following the pageset pointer.
3804 * Other parts of the kernel may not check if the zone is available.
3806 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3807 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3808 static void setup_zone_pageset(struct zone
*zone
);
3811 * Global mutex to protect against size modification of zonelists
3812 * as well as to serialize pageset setup for the new populated zone.
3814 DEFINE_MUTEX(zonelists_mutex
);
3816 /* return values int ....just for stop_machine() */
3817 static int __build_all_zonelists(void *data
)
3821 pg_data_t
*self
= data
;
3824 memset(node_load
, 0, sizeof(node_load
));
3827 if (self
&& !node_online(self
->node_id
)) {
3828 build_zonelists(self
);
3829 build_zonelist_cache(self
);
3832 for_each_online_node(nid
) {
3833 pg_data_t
*pgdat
= NODE_DATA(nid
);
3835 build_zonelists(pgdat
);
3836 build_zonelist_cache(pgdat
);
3840 * Initialize the boot_pagesets that are going to be used
3841 * for bootstrapping processors. The real pagesets for
3842 * each zone will be allocated later when the per cpu
3843 * allocator is available.
3845 * boot_pagesets are used also for bootstrapping offline
3846 * cpus if the system is already booted because the pagesets
3847 * are needed to initialize allocators on a specific cpu too.
3848 * F.e. the percpu allocator needs the page allocator which
3849 * needs the percpu allocator in order to allocate its pagesets
3850 * (a chicken-egg dilemma).
3852 for_each_possible_cpu(cpu
) {
3853 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3855 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3857 * We now know the "local memory node" for each node--
3858 * i.e., the node of the first zone in the generic zonelist.
3859 * Set up numa_mem percpu variable for on-line cpus. During
3860 * boot, only the boot cpu should be on-line; we'll init the
3861 * secondary cpus' numa_mem as they come on-line. During
3862 * node/memory hotplug, we'll fixup all on-line cpus.
3864 if (cpu_online(cpu
))
3865 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3873 * Called with zonelists_mutex held always
3874 * unless system_state == SYSTEM_BOOTING.
3876 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3878 set_zonelist_order();
3880 if (system_state
== SYSTEM_BOOTING
) {
3881 __build_all_zonelists(NULL
);
3882 mminit_verify_zonelist();
3883 cpuset_init_current_mems_allowed();
3885 #ifdef CONFIG_MEMORY_HOTPLUG
3887 setup_zone_pageset(zone
);
3889 /* we have to stop all cpus to guarantee there is no user
3891 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3892 /* cpuset refresh routine should be here */
3894 vm_total_pages
= nr_free_pagecache_pages();
3896 * Disable grouping by mobility if the number of pages in the
3897 * system is too low to allow the mechanism to work. It would be
3898 * more accurate, but expensive to check per-zone. This check is
3899 * made on memory-hotadd so a system can start with mobility
3900 * disabled and enable it later
3902 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3903 page_group_by_mobility_disabled
= 1;
3905 page_group_by_mobility_disabled
= 0;
3907 printk("Built %i zonelists in %s order, mobility grouping %s. "
3908 "Total pages: %ld\n",
3910 zonelist_order_name
[current_zonelist_order
],
3911 page_group_by_mobility_disabled
? "off" : "on",
3914 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3919 * Helper functions to size the waitqueue hash table.
3920 * Essentially these want to choose hash table sizes sufficiently
3921 * large so that collisions trying to wait on pages are rare.
3922 * But in fact, the number of active page waitqueues on typical
3923 * systems is ridiculously low, less than 200. So this is even
3924 * conservative, even though it seems large.
3926 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3927 * waitqueues, i.e. the size of the waitq table given the number of pages.
3929 #define PAGES_PER_WAITQUEUE 256
3931 #ifndef CONFIG_MEMORY_HOTPLUG
3932 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3934 unsigned long size
= 1;
3936 pages
/= PAGES_PER_WAITQUEUE
;
3938 while (size
< pages
)
3942 * Once we have dozens or even hundreds of threads sleeping
3943 * on IO we've got bigger problems than wait queue collision.
3944 * Limit the size of the wait table to a reasonable size.
3946 size
= min(size
, 4096UL);
3948 return max(size
, 4UL);
3952 * A zone's size might be changed by hot-add, so it is not possible to determine
3953 * a suitable size for its wait_table. So we use the maximum size now.
3955 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3957 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3958 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3959 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3961 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3962 * or more by the traditional way. (See above). It equals:
3964 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3965 * ia64(16K page size) : = ( 8G + 4M)byte.
3966 * powerpc (64K page size) : = (32G +16M)byte.
3968 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3975 * This is an integer logarithm so that shifts can be used later
3976 * to extract the more random high bits from the multiplicative
3977 * hash function before the remainder is taken.
3979 static inline unsigned long wait_table_bits(unsigned long size
)
3985 * Check if a pageblock contains reserved pages
3987 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3991 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3992 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3999 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
4000 * of blocks reserved is based on min_wmark_pages(zone). The memory within
4001 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
4002 * higher will lead to a bigger reserve which will get freed as contiguous
4003 * blocks as reclaim kicks in
4005 static void setup_zone_migrate_reserve(struct zone
*zone
)
4007 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
4009 unsigned long block_migratetype
;
4014 * Get the start pfn, end pfn and the number of blocks to reserve
4015 * We have to be careful to be aligned to pageblock_nr_pages to
4016 * make sure that we always check pfn_valid for the first page in
4019 start_pfn
= zone
->zone_start_pfn
;
4020 end_pfn
= zone_end_pfn(zone
);
4021 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
4022 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
4026 * Reserve blocks are generally in place to help high-order atomic
4027 * allocations that are short-lived. A min_free_kbytes value that
4028 * would result in more than 2 reserve blocks for atomic allocations
4029 * is assumed to be in place to help anti-fragmentation for the
4030 * future allocation of hugepages at runtime.
4032 reserve
= min(2, reserve
);
4033 old_reserve
= zone
->nr_migrate_reserve_block
;
4035 /* When memory hot-add, we almost always need to do nothing */
4036 if (reserve
== old_reserve
)
4038 zone
->nr_migrate_reserve_block
= reserve
;
4040 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
4041 if (!pfn_valid(pfn
))
4043 page
= pfn_to_page(pfn
);
4045 /* Watch out for overlapping nodes */
4046 if (page_to_nid(page
) != zone_to_nid(zone
))
4049 block_migratetype
= get_pageblock_migratetype(page
);
4051 /* Only test what is necessary when the reserves are not met */
4054 * Blocks with reserved pages will never free, skip
4057 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
4058 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4061 /* If this block is reserved, account for it */
4062 if (block_migratetype
== MIGRATE_RESERVE
) {
4067 /* Suitable for reserving if this block is movable */
4068 if (block_migratetype
== MIGRATE_MOVABLE
) {
4069 set_pageblock_migratetype(page
,
4071 move_freepages_block(zone
, page
,
4076 } else if (!old_reserve
) {
4078 * At boot time we don't need to scan the whole zone
4079 * for turning off MIGRATE_RESERVE.
4085 * If the reserve is met and this is a previous reserved block,
4088 if (block_migratetype
== MIGRATE_RESERVE
) {
4089 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4090 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4096 * Initially all pages are reserved - free ones are freed
4097 * up by free_all_bootmem() once the early boot process is
4098 * done. Non-atomic initialization, single-pass.
4100 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4101 unsigned long start_pfn
, enum memmap_context context
)
4104 unsigned long end_pfn
= start_pfn
+ size
;
4108 if (highest_memmap_pfn
< end_pfn
- 1)
4109 highest_memmap_pfn
= end_pfn
- 1;
4111 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4112 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4114 * There can be holes in boot-time mem_map[]s
4115 * handed to this function. They do not
4116 * exist on hotplugged memory.
4118 if (context
== MEMMAP_EARLY
) {
4119 if (!early_pfn_valid(pfn
))
4121 if (!early_pfn_in_nid(pfn
, nid
))
4124 page
= pfn_to_page(pfn
);
4125 set_page_links(page
, zone
, nid
, pfn
);
4126 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4127 init_page_count(page
);
4128 page_mapcount_reset(page
);
4129 page_cpupid_reset_last(page
);
4130 SetPageReserved(page
);
4132 * Mark the block movable so that blocks are reserved for
4133 * movable at startup. This will force kernel allocations
4134 * to reserve their blocks rather than leaking throughout
4135 * the address space during boot when many long-lived
4136 * kernel allocations are made. Later some blocks near
4137 * the start are marked MIGRATE_RESERVE by
4138 * setup_zone_migrate_reserve()
4140 * bitmap is created for zone's valid pfn range. but memmap
4141 * can be created for invalid pages (for alignment)
4142 * check here not to call set_pageblock_migratetype() against
4145 if ((z
->zone_start_pfn
<= pfn
)
4146 && (pfn
< zone_end_pfn(z
))
4147 && !(pfn
& (pageblock_nr_pages
- 1)))
4148 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4150 INIT_LIST_HEAD(&page
->lru
);
4151 #ifdef WANT_PAGE_VIRTUAL
4152 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4153 if (!is_highmem_idx(zone
))
4154 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4159 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4161 unsigned int order
, t
;
4162 for_each_migratetype_order(order
, t
) {
4163 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4164 zone
->free_area
[order
].nr_free
= 0;
4168 #ifndef __HAVE_ARCH_MEMMAP_INIT
4169 #define memmap_init(size, nid, zone, start_pfn) \
4170 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4173 static int zone_batchsize(struct zone
*zone
)
4179 * The per-cpu-pages pools are set to around 1000th of the
4180 * size of the zone. But no more than 1/2 of a meg.
4182 * OK, so we don't know how big the cache is. So guess.
4184 batch
= zone
->managed_pages
/ 1024;
4185 if (batch
* PAGE_SIZE
> 512 * 1024)
4186 batch
= (512 * 1024) / PAGE_SIZE
;
4187 batch
/= 4; /* We effectively *= 4 below */
4192 * Clamp the batch to a 2^n - 1 value. Having a power
4193 * of 2 value was found to be more likely to have
4194 * suboptimal cache aliasing properties in some cases.
4196 * For example if 2 tasks are alternately allocating
4197 * batches of pages, one task can end up with a lot
4198 * of pages of one half of the possible page colors
4199 * and the other with pages of the other colors.
4201 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4206 /* The deferral and batching of frees should be suppressed under NOMMU
4209 * The problem is that NOMMU needs to be able to allocate large chunks
4210 * of contiguous memory as there's no hardware page translation to
4211 * assemble apparent contiguous memory from discontiguous pages.
4213 * Queueing large contiguous runs of pages for batching, however,
4214 * causes the pages to actually be freed in smaller chunks. As there
4215 * can be a significant delay between the individual batches being
4216 * recycled, this leads to the once large chunks of space being
4217 * fragmented and becoming unavailable for high-order allocations.
4224 * pcp->high and pcp->batch values are related and dependent on one another:
4225 * ->batch must never be higher then ->high.
4226 * The following function updates them in a safe manner without read side
4229 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4230 * those fields changing asynchronously (acording the the above rule).
4232 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4233 * outside of boot time (or some other assurance that no concurrent updaters
4236 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4237 unsigned long batch
)
4239 /* start with a fail safe value for batch */
4243 /* Update high, then batch, in order */
4250 /* a companion to pageset_set_high() */
4251 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4253 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4256 static void pageset_init(struct per_cpu_pageset
*p
)
4258 struct per_cpu_pages
*pcp
;
4261 memset(p
, 0, sizeof(*p
));
4265 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4266 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4269 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4272 pageset_set_batch(p
, batch
);
4276 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4277 * to the value high for the pageset p.
4279 static void pageset_set_high(struct per_cpu_pageset
*p
,
4282 unsigned long batch
= max(1UL, high
/ 4);
4283 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4284 batch
= PAGE_SHIFT
* 8;
4286 pageset_update(&p
->pcp
, high
, batch
);
4289 static void pageset_set_high_and_batch(struct zone
*zone
,
4290 struct per_cpu_pageset
*pcp
)
4292 if (percpu_pagelist_fraction
)
4293 pageset_set_high(pcp
,
4294 (zone
->managed_pages
/
4295 percpu_pagelist_fraction
));
4297 pageset_set_batch(pcp
, zone_batchsize(zone
));
4300 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4302 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4305 pageset_set_high_and_batch(zone
, pcp
);
4308 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4311 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4312 for_each_possible_cpu(cpu
)
4313 zone_pageset_init(zone
, cpu
);
4317 * Allocate per cpu pagesets and initialize them.
4318 * Before this call only boot pagesets were available.
4320 void __init
setup_per_cpu_pageset(void)
4324 for_each_populated_zone(zone
)
4325 setup_zone_pageset(zone
);
4328 static noinline __init_refok
4329 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4335 * The per-page waitqueue mechanism uses hashed waitqueues
4338 zone
->wait_table_hash_nr_entries
=
4339 wait_table_hash_nr_entries(zone_size_pages
);
4340 zone
->wait_table_bits
=
4341 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4342 alloc_size
= zone
->wait_table_hash_nr_entries
4343 * sizeof(wait_queue_head_t
);
4345 if (!slab_is_available()) {
4346 zone
->wait_table
= (wait_queue_head_t
*)
4347 memblock_virt_alloc_node_nopanic(
4348 alloc_size
, zone
->zone_pgdat
->node_id
);
4351 * This case means that a zone whose size was 0 gets new memory
4352 * via memory hot-add.
4353 * But it may be the case that a new node was hot-added. In
4354 * this case vmalloc() will not be able to use this new node's
4355 * memory - this wait_table must be initialized to use this new
4356 * node itself as well.
4357 * To use this new node's memory, further consideration will be
4360 zone
->wait_table
= vmalloc(alloc_size
);
4362 if (!zone
->wait_table
)
4365 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4366 init_waitqueue_head(zone
->wait_table
+ i
);
4371 static __meminit
void zone_pcp_init(struct zone
*zone
)
4374 * per cpu subsystem is not up at this point. The following code
4375 * relies on the ability of the linker to provide the
4376 * offset of a (static) per cpu variable into the per cpu area.
4378 zone
->pageset
= &boot_pageset
;
4380 if (populated_zone(zone
))
4381 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4382 zone
->name
, zone
->present_pages
,
4383 zone_batchsize(zone
));
4386 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4387 unsigned long zone_start_pfn
,
4389 enum memmap_context context
)
4391 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4393 ret
= zone_wait_table_init(zone
, size
);
4396 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4398 zone
->zone_start_pfn
= zone_start_pfn
;
4400 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4401 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4403 (unsigned long)zone_idx(zone
),
4404 zone_start_pfn
, (zone_start_pfn
+ size
));
4406 zone_init_free_lists(zone
);
4411 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4412 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4414 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4416 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4418 unsigned long start_pfn
, end_pfn
;
4421 * NOTE: The following SMP-unsafe globals are only used early in boot
4422 * when the kernel is running single-threaded.
4424 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4425 static int __meminitdata last_nid
;
4427 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4430 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4432 last_start_pfn
= start_pfn
;
4433 last_end_pfn
= end_pfn
;
4439 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4441 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4445 nid
= __early_pfn_to_nid(pfn
);
4448 /* just returns 0 */
4452 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4453 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4457 nid
= __early_pfn_to_nid(pfn
);
4458 if (nid
>= 0 && nid
!= node
)
4465 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4466 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4467 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4469 * If an architecture guarantees that all ranges registered contain no holes
4470 * and may be freed, this this function may be used instead of calling
4471 * memblock_free_early_nid() manually.
4473 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4475 unsigned long start_pfn
, end_pfn
;
4478 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4479 start_pfn
= min(start_pfn
, max_low_pfn
);
4480 end_pfn
= min(end_pfn
, max_low_pfn
);
4482 if (start_pfn
< end_pfn
)
4483 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4484 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4490 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4491 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4493 * If an architecture guarantees that all ranges registered contain no holes and may
4494 * be freed, this function may be used instead of calling memory_present() manually.
4496 void __init
sparse_memory_present_with_active_regions(int nid
)
4498 unsigned long start_pfn
, end_pfn
;
4501 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4502 memory_present(this_nid
, start_pfn
, end_pfn
);
4506 * get_pfn_range_for_nid - Return the start and end page frames for a node
4507 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4508 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4509 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4511 * It returns the start and end page frame of a node based on information
4512 * provided by memblock_set_node(). If called for a node
4513 * with no available memory, a warning is printed and the start and end
4516 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4517 unsigned long *start_pfn
, unsigned long *end_pfn
)
4519 unsigned long this_start_pfn
, this_end_pfn
;
4525 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4526 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4527 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4530 if (*start_pfn
== -1UL)
4535 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4536 * assumption is made that zones within a node are ordered in monotonic
4537 * increasing memory addresses so that the "highest" populated zone is used
4539 static void __init
find_usable_zone_for_movable(void)
4542 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4543 if (zone_index
== ZONE_MOVABLE
)
4546 if (arch_zone_highest_possible_pfn
[zone_index
] >
4547 arch_zone_lowest_possible_pfn
[zone_index
])
4551 VM_BUG_ON(zone_index
== -1);
4552 movable_zone
= zone_index
;
4556 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4557 * because it is sized independent of architecture. Unlike the other zones,
4558 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4559 * in each node depending on the size of each node and how evenly kernelcore
4560 * is distributed. This helper function adjusts the zone ranges
4561 * provided by the architecture for a given node by using the end of the
4562 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4563 * zones within a node are in order of monotonic increases memory addresses
4565 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4566 unsigned long zone_type
,
4567 unsigned long node_start_pfn
,
4568 unsigned long node_end_pfn
,
4569 unsigned long *zone_start_pfn
,
4570 unsigned long *zone_end_pfn
)
4572 /* Only adjust if ZONE_MOVABLE is on this node */
4573 if (zone_movable_pfn
[nid
]) {
4574 /* Size ZONE_MOVABLE */
4575 if (zone_type
== ZONE_MOVABLE
) {
4576 *zone_start_pfn
= zone_movable_pfn
[nid
];
4577 *zone_end_pfn
= min(node_end_pfn
,
4578 arch_zone_highest_possible_pfn
[movable_zone
]);
4580 /* Adjust for ZONE_MOVABLE starting within this range */
4581 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4582 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4583 *zone_end_pfn
= zone_movable_pfn
[nid
];
4585 /* Check if this whole range is within ZONE_MOVABLE */
4586 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4587 *zone_start_pfn
= *zone_end_pfn
;
4592 * Return the number of pages a zone spans in a node, including holes
4593 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4595 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4596 unsigned long zone_type
,
4597 unsigned long node_start_pfn
,
4598 unsigned long node_end_pfn
,
4599 unsigned long *ignored
)
4601 unsigned long zone_start_pfn
, zone_end_pfn
;
4603 /* Get the start and end of the zone */
4604 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4605 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4606 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4607 node_start_pfn
, node_end_pfn
,
4608 &zone_start_pfn
, &zone_end_pfn
);
4610 /* Check that this node has pages within the zone's required range */
4611 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4614 /* Move the zone boundaries inside the node if necessary */
4615 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4616 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4618 /* Return the spanned pages */
4619 return zone_end_pfn
- zone_start_pfn
;
4623 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4624 * then all holes in the requested range will be accounted for.
4626 unsigned long __meminit
__absent_pages_in_range(int nid
,
4627 unsigned long range_start_pfn
,
4628 unsigned long range_end_pfn
)
4630 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4631 unsigned long start_pfn
, end_pfn
;
4634 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4635 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4636 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4637 nr_absent
-= end_pfn
- start_pfn
;
4643 * absent_pages_in_range - Return number of page frames in holes within a range
4644 * @start_pfn: The start PFN to start searching for holes
4645 * @end_pfn: The end PFN to stop searching for holes
4647 * It returns the number of pages frames in memory holes within a range.
4649 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4650 unsigned long end_pfn
)
4652 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4655 /* Return the number of page frames in holes in a zone on a node */
4656 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4657 unsigned long zone_type
,
4658 unsigned long node_start_pfn
,
4659 unsigned long node_end_pfn
,
4660 unsigned long *ignored
)
4662 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4663 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4664 unsigned long zone_start_pfn
, zone_end_pfn
;
4666 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4667 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4669 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4670 node_start_pfn
, node_end_pfn
,
4671 &zone_start_pfn
, &zone_end_pfn
);
4672 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4675 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4676 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4677 unsigned long zone_type
,
4678 unsigned long node_start_pfn
,
4679 unsigned long node_end_pfn
,
4680 unsigned long *zones_size
)
4682 return zones_size
[zone_type
];
4685 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4686 unsigned long zone_type
,
4687 unsigned long node_start_pfn
,
4688 unsigned long node_end_pfn
,
4689 unsigned long *zholes_size
)
4694 return zholes_size
[zone_type
];
4697 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4699 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4700 unsigned long node_start_pfn
,
4701 unsigned long node_end_pfn
,
4702 unsigned long *zones_size
,
4703 unsigned long *zholes_size
)
4705 unsigned long realtotalpages
, totalpages
= 0;
4708 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4709 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4713 pgdat
->node_spanned_pages
= totalpages
;
4715 realtotalpages
= totalpages
;
4716 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4718 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4719 node_start_pfn
, node_end_pfn
,
4721 pgdat
->node_present_pages
= realtotalpages
;
4722 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4726 #ifndef CONFIG_SPARSEMEM
4728 * Calculate the size of the zone->blockflags rounded to an unsigned long
4729 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4730 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4731 * round what is now in bits to nearest long in bits, then return it in
4734 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4736 unsigned long usemapsize
;
4738 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4739 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4740 usemapsize
= usemapsize
>> pageblock_order
;
4741 usemapsize
*= NR_PAGEBLOCK_BITS
;
4742 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4744 return usemapsize
/ 8;
4747 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4749 unsigned long zone_start_pfn
,
4750 unsigned long zonesize
)
4752 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4753 zone
->pageblock_flags
= NULL
;
4755 zone
->pageblock_flags
=
4756 memblock_virt_alloc_node_nopanic(usemapsize
,
4760 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4761 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4762 #endif /* CONFIG_SPARSEMEM */
4764 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4766 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4767 void __paginginit
set_pageblock_order(void)
4771 /* Check that pageblock_nr_pages has not already been setup */
4772 if (pageblock_order
)
4775 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4776 order
= HUGETLB_PAGE_ORDER
;
4778 order
= MAX_ORDER
- 1;
4781 * Assume the largest contiguous order of interest is a huge page.
4782 * This value may be variable depending on boot parameters on IA64 and
4785 pageblock_order
= order
;
4787 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4790 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4791 * is unused as pageblock_order is set at compile-time. See
4792 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4795 void __paginginit
set_pageblock_order(void)
4799 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4801 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4802 unsigned long present_pages
)
4804 unsigned long pages
= spanned_pages
;
4807 * Provide a more accurate estimation if there are holes within
4808 * the zone and SPARSEMEM is in use. If there are holes within the
4809 * zone, each populated memory region may cost us one or two extra
4810 * memmap pages due to alignment because memmap pages for each
4811 * populated regions may not naturally algined on page boundary.
4812 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4814 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4815 IS_ENABLED(CONFIG_SPARSEMEM
))
4816 pages
= present_pages
;
4818 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4822 * Set up the zone data structures:
4823 * - mark all pages reserved
4824 * - mark all memory queues empty
4825 * - clear the memory bitmaps
4827 * NOTE: pgdat should get zeroed by caller.
4829 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4830 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4831 unsigned long *zones_size
, unsigned long *zholes_size
)
4834 int nid
= pgdat
->node_id
;
4835 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4838 pgdat_resize_init(pgdat
);
4839 #ifdef CONFIG_NUMA_BALANCING
4840 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4841 pgdat
->numabalancing_migrate_nr_pages
= 0;
4842 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4844 init_waitqueue_head(&pgdat
->kswapd_wait
);
4845 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4846 pgdat_page_cgroup_init(pgdat
);
4848 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4849 struct zone
*zone
= pgdat
->node_zones
+ j
;
4850 unsigned long size
, realsize
, freesize
, memmap_pages
;
4852 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4853 node_end_pfn
, zones_size
);
4854 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4860 * Adjust freesize so that it accounts for how much memory
4861 * is used by this zone for memmap. This affects the watermark
4862 * and per-cpu initialisations
4864 memmap_pages
= calc_memmap_size(size
, realsize
);
4865 if (freesize
>= memmap_pages
) {
4866 freesize
-= memmap_pages
;
4869 " %s zone: %lu pages used for memmap\n",
4870 zone_names
[j
], memmap_pages
);
4873 " %s zone: %lu pages exceeds freesize %lu\n",
4874 zone_names
[j
], memmap_pages
, freesize
);
4876 /* Account for reserved pages */
4877 if (j
== 0 && freesize
> dma_reserve
) {
4878 freesize
-= dma_reserve
;
4879 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4880 zone_names
[0], dma_reserve
);
4883 if (!is_highmem_idx(j
))
4884 nr_kernel_pages
+= freesize
;
4885 /* Charge for highmem memmap if there are enough kernel pages */
4886 else if (nr_kernel_pages
> memmap_pages
* 2)
4887 nr_kernel_pages
-= memmap_pages
;
4888 nr_all_pages
+= freesize
;
4890 zone
->spanned_pages
= size
;
4891 zone
->present_pages
= realsize
;
4893 * Set an approximate value for lowmem here, it will be adjusted
4894 * when the bootmem allocator frees pages into the buddy system.
4895 * And all highmem pages will be managed by the buddy system.
4897 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4900 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4902 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4904 zone
->name
= zone_names
[j
];
4905 spin_lock_init(&zone
->lock
);
4906 spin_lock_init(&zone
->lru_lock
);
4907 zone_seqlock_init(zone
);
4908 zone
->zone_pgdat
= pgdat
;
4909 zone_pcp_init(zone
);
4911 /* For bootup, initialized properly in watermark setup */
4912 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4914 lruvec_init(&zone
->lruvec
);
4918 set_pageblock_order();
4919 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4920 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4921 size
, MEMMAP_EARLY
);
4923 memmap_init(size
, nid
, j
, zone_start_pfn
);
4924 zone_start_pfn
+= size
;
4928 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4930 /* Skip empty nodes */
4931 if (!pgdat
->node_spanned_pages
)
4934 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4935 /* ia64 gets its own node_mem_map, before this, without bootmem */
4936 if (!pgdat
->node_mem_map
) {
4937 unsigned long size
, start
, end
;
4941 * The zone's endpoints aren't required to be MAX_ORDER
4942 * aligned but the node_mem_map endpoints must be in order
4943 * for the buddy allocator to function correctly.
4945 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4946 end
= pgdat_end_pfn(pgdat
);
4947 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4948 size
= (end
- start
) * sizeof(struct page
);
4949 map
= alloc_remap(pgdat
->node_id
, size
);
4951 map
= memblock_virt_alloc_node_nopanic(size
,
4953 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4955 #ifndef CONFIG_NEED_MULTIPLE_NODES
4957 * With no DISCONTIG, the global mem_map is just set as node 0's
4959 if (pgdat
== NODE_DATA(0)) {
4960 mem_map
= NODE_DATA(0)->node_mem_map
;
4961 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4962 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4963 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4964 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4967 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4970 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4971 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4973 pg_data_t
*pgdat
= NODE_DATA(nid
);
4974 unsigned long start_pfn
= 0;
4975 unsigned long end_pfn
= 0;
4977 /* pg_data_t should be reset to zero when it's allocated */
4978 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4980 pgdat
->node_id
= nid
;
4981 pgdat
->node_start_pfn
= node_start_pfn
;
4982 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4983 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4984 printk(KERN_INFO
"Initmem setup node %d [mem %#010Lx-%#010Lx]\n", nid
,
4985 (u64
) start_pfn
<< PAGE_SHIFT
, (u64
) (end_pfn
<< PAGE_SHIFT
) - 1);
4987 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
4988 zones_size
, zholes_size
);
4990 alloc_node_mem_map(pgdat
);
4991 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4992 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4993 nid
, (unsigned long)pgdat
,
4994 (unsigned long)pgdat
->node_mem_map
);
4997 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
4998 zones_size
, zholes_size
);
5001 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
5003 #if MAX_NUMNODES > 1
5005 * Figure out the number of possible node ids.
5007 void __init
setup_nr_node_ids(void)
5010 unsigned int highest
= 0;
5012 for_each_node_mask(node
, node_possible_map
)
5014 nr_node_ids
= highest
+ 1;
5019 * node_map_pfn_alignment - determine the maximum internode alignment
5021 * This function should be called after node map is populated and sorted.
5022 * It calculates the maximum power of two alignment which can distinguish
5025 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
5026 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
5027 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
5028 * shifted, 1GiB is enough and this function will indicate so.
5030 * This is used to test whether pfn -> nid mapping of the chosen memory
5031 * model has fine enough granularity to avoid incorrect mapping for the
5032 * populated node map.
5034 * Returns the determined alignment in pfn's. 0 if there is no alignment
5035 * requirement (single node).
5037 unsigned long __init
node_map_pfn_alignment(void)
5039 unsigned long accl_mask
= 0, last_end
= 0;
5040 unsigned long start
, end
, mask
;
5044 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
5045 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
5052 * Start with a mask granular enough to pin-point to the
5053 * start pfn and tick off bits one-by-one until it becomes
5054 * too coarse to separate the current node from the last.
5056 mask
= ~((1 << __ffs(start
)) - 1);
5057 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5060 /* accumulate all internode masks */
5064 /* convert mask to number of pages */
5065 return ~accl_mask
+ 1;
5068 /* Find the lowest pfn for a node */
5069 static unsigned long __init
find_min_pfn_for_node(int nid
)
5071 unsigned long min_pfn
= ULONG_MAX
;
5072 unsigned long start_pfn
;
5075 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5076 min_pfn
= min(min_pfn
, start_pfn
);
5078 if (min_pfn
== ULONG_MAX
) {
5080 "Could not find start_pfn for node %d\n", nid
);
5088 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5090 * It returns the minimum PFN based on information provided via
5091 * memblock_set_node().
5093 unsigned long __init
find_min_pfn_with_active_regions(void)
5095 return find_min_pfn_for_node(MAX_NUMNODES
);
5099 * early_calculate_totalpages()
5100 * Sum pages in active regions for movable zone.
5101 * Populate N_MEMORY for calculating usable_nodes.
5103 static unsigned long __init
early_calculate_totalpages(void)
5105 unsigned long totalpages
= 0;
5106 unsigned long start_pfn
, end_pfn
;
5109 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5110 unsigned long pages
= end_pfn
- start_pfn
;
5112 totalpages
+= pages
;
5114 node_set_state(nid
, N_MEMORY
);
5120 * Find the PFN the Movable zone begins in each node. Kernel memory
5121 * is spread evenly between nodes as long as the nodes have enough
5122 * memory. When they don't, some nodes will have more kernelcore than
5125 static void __init
find_zone_movable_pfns_for_nodes(void)
5128 unsigned long usable_startpfn
;
5129 unsigned long kernelcore_node
, kernelcore_remaining
;
5130 /* save the state before borrow the nodemask */
5131 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5132 unsigned long totalpages
= early_calculate_totalpages();
5133 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5134 struct memblock_region
*r
;
5136 /* Need to find movable_zone earlier when movable_node is specified. */
5137 find_usable_zone_for_movable();
5140 * If movable_node is specified, ignore kernelcore and movablecore
5143 if (movable_node_is_enabled()) {
5144 for_each_memblock(memory
, r
) {
5145 if (!memblock_is_hotpluggable(r
))
5150 usable_startpfn
= PFN_DOWN(r
->base
);
5151 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5152 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5160 * If movablecore=nn[KMG] was specified, calculate what size of
5161 * kernelcore that corresponds so that memory usable for
5162 * any allocation type is evenly spread. If both kernelcore
5163 * and movablecore are specified, then the value of kernelcore
5164 * will be used for required_kernelcore if it's greater than
5165 * what movablecore would have allowed.
5167 if (required_movablecore
) {
5168 unsigned long corepages
;
5171 * Round-up so that ZONE_MOVABLE is at least as large as what
5172 * was requested by the user
5174 required_movablecore
=
5175 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5176 corepages
= totalpages
- required_movablecore
;
5178 required_kernelcore
= max(required_kernelcore
, corepages
);
5181 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5182 if (!required_kernelcore
)
5185 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5186 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5189 /* Spread kernelcore memory as evenly as possible throughout nodes */
5190 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5191 for_each_node_state(nid
, N_MEMORY
) {
5192 unsigned long start_pfn
, end_pfn
;
5195 * Recalculate kernelcore_node if the division per node
5196 * now exceeds what is necessary to satisfy the requested
5197 * amount of memory for the kernel
5199 if (required_kernelcore
< kernelcore_node
)
5200 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5203 * As the map is walked, we track how much memory is usable
5204 * by the kernel using kernelcore_remaining. When it is
5205 * 0, the rest of the node is usable by ZONE_MOVABLE
5207 kernelcore_remaining
= kernelcore_node
;
5209 /* Go through each range of PFNs within this node */
5210 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5211 unsigned long size_pages
;
5213 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5214 if (start_pfn
>= end_pfn
)
5217 /* Account for what is only usable for kernelcore */
5218 if (start_pfn
< usable_startpfn
) {
5219 unsigned long kernel_pages
;
5220 kernel_pages
= min(end_pfn
, usable_startpfn
)
5223 kernelcore_remaining
-= min(kernel_pages
,
5224 kernelcore_remaining
);
5225 required_kernelcore
-= min(kernel_pages
,
5226 required_kernelcore
);
5228 /* Continue if range is now fully accounted */
5229 if (end_pfn
<= usable_startpfn
) {
5232 * Push zone_movable_pfn to the end so
5233 * that if we have to rebalance
5234 * kernelcore across nodes, we will
5235 * not double account here
5237 zone_movable_pfn
[nid
] = end_pfn
;
5240 start_pfn
= usable_startpfn
;
5244 * The usable PFN range for ZONE_MOVABLE is from
5245 * start_pfn->end_pfn. Calculate size_pages as the
5246 * number of pages used as kernelcore
5248 size_pages
= end_pfn
- start_pfn
;
5249 if (size_pages
> kernelcore_remaining
)
5250 size_pages
= kernelcore_remaining
;
5251 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5254 * Some kernelcore has been met, update counts and
5255 * break if the kernelcore for this node has been
5258 required_kernelcore
-= min(required_kernelcore
,
5260 kernelcore_remaining
-= size_pages
;
5261 if (!kernelcore_remaining
)
5267 * If there is still required_kernelcore, we do another pass with one
5268 * less node in the count. This will push zone_movable_pfn[nid] further
5269 * along on the nodes that still have memory until kernelcore is
5273 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5277 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5278 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5279 zone_movable_pfn
[nid
] =
5280 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5283 /* restore the node_state */
5284 node_states
[N_MEMORY
] = saved_node_state
;
5287 /* Any regular or high memory on that node ? */
5288 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5290 enum zone_type zone_type
;
5292 if (N_MEMORY
== N_NORMAL_MEMORY
)
5295 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5296 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5297 if (populated_zone(zone
)) {
5298 node_set_state(nid
, N_HIGH_MEMORY
);
5299 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5300 zone_type
<= ZONE_NORMAL
)
5301 node_set_state(nid
, N_NORMAL_MEMORY
);
5308 * free_area_init_nodes - Initialise all pg_data_t and zone data
5309 * @max_zone_pfn: an array of max PFNs for each zone
5311 * This will call free_area_init_node() for each active node in the system.
5312 * Using the page ranges provided by memblock_set_node(), the size of each
5313 * zone in each node and their holes is calculated. If the maximum PFN
5314 * between two adjacent zones match, it is assumed that the zone is empty.
5315 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5316 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5317 * starts where the previous one ended. For example, ZONE_DMA32 starts
5318 * at arch_max_dma_pfn.
5320 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5322 unsigned long start_pfn
, end_pfn
;
5325 /* Record where the zone boundaries are */
5326 memset(arch_zone_lowest_possible_pfn
, 0,
5327 sizeof(arch_zone_lowest_possible_pfn
));
5328 memset(arch_zone_highest_possible_pfn
, 0,
5329 sizeof(arch_zone_highest_possible_pfn
));
5330 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5331 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5332 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5333 if (i
== ZONE_MOVABLE
)
5335 arch_zone_lowest_possible_pfn
[i
] =
5336 arch_zone_highest_possible_pfn
[i
-1];
5337 arch_zone_highest_possible_pfn
[i
] =
5338 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5340 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5341 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5343 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5344 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5345 find_zone_movable_pfns_for_nodes();
5347 /* Print out the zone ranges */
5348 printk("Zone ranges:\n");
5349 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5350 if (i
== ZONE_MOVABLE
)
5352 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5353 if (arch_zone_lowest_possible_pfn
[i
] ==
5354 arch_zone_highest_possible_pfn
[i
])
5355 printk(KERN_CONT
"empty\n");
5357 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5358 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5359 (arch_zone_highest_possible_pfn
[i
]
5360 << PAGE_SHIFT
) - 1);
5363 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5364 printk("Movable zone start for each node\n");
5365 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5366 if (zone_movable_pfn
[i
])
5367 printk(" Node %d: %#010lx\n", i
,
5368 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5371 /* Print out the early node map */
5372 printk("Early memory node ranges\n");
5373 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5374 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5375 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5377 /* Initialise every node */
5378 mminit_verify_pageflags_layout();
5379 setup_nr_node_ids();
5380 for_each_online_node(nid
) {
5381 pg_data_t
*pgdat
= NODE_DATA(nid
);
5382 free_area_init_node(nid
, NULL
,
5383 find_min_pfn_for_node(nid
), NULL
);
5385 /* Any memory on that node */
5386 if (pgdat
->node_present_pages
)
5387 node_set_state(nid
, N_MEMORY
);
5388 check_for_memory(pgdat
, nid
);
5392 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5394 unsigned long long coremem
;
5398 coremem
= memparse(p
, &p
);
5399 *core
= coremem
>> PAGE_SHIFT
;
5401 /* Paranoid check that UL is enough for the coremem value */
5402 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5408 * kernelcore=size sets the amount of memory for use for allocations that
5409 * cannot be reclaimed or migrated.
5411 static int __init
cmdline_parse_kernelcore(char *p
)
5413 return cmdline_parse_core(p
, &required_kernelcore
);
5417 * movablecore=size sets the amount of memory for use for allocations that
5418 * can be reclaimed or migrated.
5420 static int __init
cmdline_parse_movablecore(char *p
)
5422 return cmdline_parse_core(p
, &required_movablecore
);
5425 early_param("kernelcore", cmdline_parse_kernelcore
);
5426 early_param("movablecore", cmdline_parse_movablecore
);
5428 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5430 void adjust_managed_page_count(struct page
*page
, long count
)
5432 spin_lock(&managed_page_count_lock
);
5433 page_zone(page
)->managed_pages
+= count
;
5434 totalram_pages
+= count
;
5435 #ifdef CONFIG_HIGHMEM
5436 if (PageHighMem(page
))
5437 totalhigh_pages
+= count
;
5439 spin_unlock(&managed_page_count_lock
);
5441 EXPORT_SYMBOL(adjust_managed_page_count
);
5443 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5446 unsigned long pages
= 0;
5448 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5449 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5450 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5451 if ((unsigned int)poison
<= 0xFF)
5452 memset(pos
, poison
, PAGE_SIZE
);
5453 free_reserved_page(virt_to_page(pos
));
5457 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5458 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5462 EXPORT_SYMBOL(free_reserved_area
);
5464 #ifdef CONFIG_HIGHMEM
5465 void free_highmem_page(struct page
*page
)
5467 __free_reserved_page(page
);
5469 page_zone(page
)->managed_pages
++;
5475 void __init
mem_init_print_info(const char *str
)
5477 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5478 unsigned long init_code_size
, init_data_size
;
5480 physpages
= get_num_physpages();
5481 codesize
= _etext
- _stext
;
5482 datasize
= _edata
- _sdata
;
5483 rosize
= __end_rodata
- __start_rodata
;
5484 bss_size
= __bss_stop
- __bss_start
;
5485 init_data_size
= __init_end
- __init_begin
;
5486 init_code_size
= _einittext
- _sinittext
;
5489 * Detect special cases and adjust section sizes accordingly:
5490 * 1) .init.* may be embedded into .data sections
5491 * 2) .init.text.* may be out of [__init_begin, __init_end],
5492 * please refer to arch/tile/kernel/vmlinux.lds.S.
5493 * 3) .rodata.* may be embedded into .text or .data sections.
5495 #define adj_init_size(start, end, size, pos, adj) \
5497 if (start <= pos && pos < end && size > adj) \
5501 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5502 _sinittext
, init_code_size
);
5503 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5504 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5505 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5506 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5508 #undef adj_init_size
5510 printk("Memory: %luK/%luK available "
5511 "(%luK kernel code, %luK rwdata, %luK rodata, "
5512 "%luK init, %luK bss, %luK reserved"
5513 #ifdef CONFIG_HIGHMEM
5517 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5518 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5519 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5520 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5521 #ifdef CONFIG_HIGHMEM
5522 totalhigh_pages
<< (PAGE_SHIFT
-10),
5524 str
? ", " : "", str
? str
: "");
5528 * set_dma_reserve - set the specified number of pages reserved in the first zone
5529 * @new_dma_reserve: The number of pages to mark reserved
5531 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5532 * In the DMA zone, a significant percentage may be consumed by kernel image
5533 * and other unfreeable allocations which can skew the watermarks badly. This
5534 * function may optionally be used to account for unfreeable pages in the
5535 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5536 * smaller per-cpu batchsize.
5538 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5540 dma_reserve
= new_dma_reserve
;
5543 void __init
free_area_init(unsigned long *zones_size
)
5545 free_area_init_node(0, zones_size
,
5546 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5549 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5550 unsigned long action
, void *hcpu
)
5552 int cpu
= (unsigned long)hcpu
;
5554 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5555 lru_add_drain_cpu(cpu
);
5559 * Spill the event counters of the dead processor
5560 * into the current processors event counters.
5561 * This artificially elevates the count of the current
5564 vm_events_fold_cpu(cpu
);
5567 * Zero the differential counters of the dead processor
5568 * so that the vm statistics are consistent.
5570 * This is only okay since the processor is dead and cannot
5571 * race with what we are doing.
5573 cpu_vm_stats_fold(cpu
);
5578 void __init
page_alloc_init(void)
5580 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5584 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5585 * or min_free_kbytes changes.
5587 static void calculate_totalreserve_pages(void)
5589 struct pglist_data
*pgdat
;
5590 unsigned long reserve_pages
= 0;
5591 enum zone_type i
, j
;
5593 for_each_online_pgdat(pgdat
) {
5594 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5595 struct zone
*zone
= pgdat
->node_zones
+ i
;
5598 /* Find valid and maximum lowmem_reserve in the zone */
5599 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5600 if (zone
->lowmem_reserve
[j
] > max
)
5601 max
= zone
->lowmem_reserve
[j
];
5604 /* we treat the high watermark as reserved pages. */
5605 max
+= high_wmark_pages(zone
);
5607 if (max
> zone
->managed_pages
)
5608 max
= zone
->managed_pages
;
5609 reserve_pages
+= max
;
5611 * Lowmem reserves are not available to
5612 * GFP_HIGHUSER page cache allocations and
5613 * kswapd tries to balance zones to their high
5614 * watermark. As a result, neither should be
5615 * regarded as dirtyable memory, to prevent a
5616 * situation where reclaim has to clean pages
5617 * in order to balance the zones.
5619 zone
->dirty_balance_reserve
= max
;
5622 dirty_balance_reserve
= reserve_pages
;
5623 totalreserve_pages
= reserve_pages
;
5627 * setup_per_zone_lowmem_reserve - called whenever
5628 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5629 * has a correct pages reserved value, so an adequate number of
5630 * pages are left in the zone after a successful __alloc_pages().
5632 static void setup_per_zone_lowmem_reserve(void)
5634 struct pglist_data
*pgdat
;
5635 enum zone_type j
, idx
;
5637 for_each_online_pgdat(pgdat
) {
5638 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5639 struct zone
*zone
= pgdat
->node_zones
+ j
;
5640 unsigned long managed_pages
= zone
->managed_pages
;
5642 zone
->lowmem_reserve
[j
] = 0;
5646 struct zone
*lower_zone
;
5650 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5651 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5653 lower_zone
= pgdat
->node_zones
+ idx
;
5654 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5655 sysctl_lowmem_reserve_ratio
[idx
];
5656 managed_pages
+= lower_zone
->managed_pages
;
5661 /* update totalreserve_pages */
5662 calculate_totalreserve_pages();
5665 static void __setup_per_zone_wmarks(void)
5667 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5668 unsigned long lowmem_pages
= 0;
5670 unsigned long flags
;
5672 /* Calculate total number of !ZONE_HIGHMEM pages */
5673 for_each_zone(zone
) {
5674 if (!is_highmem(zone
))
5675 lowmem_pages
+= zone
->managed_pages
;
5678 for_each_zone(zone
) {
5681 spin_lock_irqsave(&zone
->lock
, flags
);
5682 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5683 do_div(tmp
, lowmem_pages
);
5684 if (is_highmem(zone
)) {
5686 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5687 * need highmem pages, so cap pages_min to a small
5690 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5691 * deltas controls asynch page reclaim, and so should
5692 * not be capped for highmem.
5694 unsigned long min_pages
;
5696 min_pages
= zone
->managed_pages
/ 1024;
5697 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5698 zone
->watermark
[WMARK_MIN
] = min_pages
;
5701 * If it's a lowmem zone, reserve a number of pages
5702 * proportionate to the zone's size.
5704 zone
->watermark
[WMARK_MIN
] = tmp
;
5707 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5708 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5710 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5711 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
5712 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
5714 setup_zone_migrate_reserve(zone
);
5715 spin_unlock_irqrestore(&zone
->lock
, flags
);
5718 /* update totalreserve_pages */
5719 calculate_totalreserve_pages();
5723 * setup_per_zone_wmarks - called when min_free_kbytes changes
5724 * or when memory is hot-{added|removed}
5726 * Ensures that the watermark[min,low,high] values for each zone are set
5727 * correctly with respect to min_free_kbytes.
5729 void setup_per_zone_wmarks(void)
5731 mutex_lock(&zonelists_mutex
);
5732 __setup_per_zone_wmarks();
5733 mutex_unlock(&zonelists_mutex
);
5737 * The inactive anon list should be small enough that the VM never has to
5738 * do too much work, but large enough that each inactive page has a chance
5739 * to be referenced again before it is swapped out.
5741 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5742 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5743 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5744 * the anonymous pages are kept on the inactive list.
5747 * memory ratio inactive anon
5748 * -------------------------------------
5757 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5759 unsigned int gb
, ratio
;
5761 /* Zone size in gigabytes */
5762 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5764 ratio
= int_sqrt(10 * gb
);
5768 zone
->inactive_ratio
= ratio
;
5771 static void __meminit
setup_per_zone_inactive_ratio(void)
5776 calculate_zone_inactive_ratio(zone
);
5780 * Initialise min_free_kbytes.
5782 * For small machines we want it small (128k min). For large machines
5783 * we want it large (64MB max). But it is not linear, because network
5784 * bandwidth does not increase linearly with machine size. We use
5786 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5787 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5803 int __meminit
init_per_zone_wmark_min(void)
5805 unsigned long lowmem_kbytes
;
5806 int new_min_free_kbytes
;
5808 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5809 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5811 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5812 min_free_kbytes
= new_min_free_kbytes
;
5813 if (min_free_kbytes
< 128)
5814 min_free_kbytes
= 128;
5815 if (min_free_kbytes
> 65536)
5816 min_free_kbytes
= 65536;
5818 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5819 new_min_free_kbytes
, user_min_free_kbytes
);
5821 setup_per_zone_wmarks();
5822 refresh_zone_stat_thresholds();
5823 setup_per_zone_lowmem_reserve();
5824 setup_per_zone_inactive_ratio();
5827 module_init(init_per_zone_wmark_min
)
5830 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5831 * that we can call two helper functions whenever min_free_kbytes
5834 int min_free_kbytes_sysctl_handler(struct ctl_table
*table
, int write
,
5835 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5839 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5844 user_min_free_kbytes
= min_free_kbytes
;
5845 setup_per_zone_wmarks();
5851 int sysctl_min_unmapped_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5852 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5857 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5862 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5863 sysctl_min_unmapped_ratio
) / 100;
5867 int sysctl_min_slab_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5868 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5873 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5878 zone
->min_slab_pages
= (zone
->managed_pages
*
5879 sysctl_min_slab_ratio
) / 100;
5885 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5886 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5887 * whenever sysctl_lowmem_reserve_ratio changes.
5889 * The reserve ratio obviously has absolutely no relation with the
5890 * minimum watermarks. The lowmem reserve ratio can only make sense
5891 * if in function of the boot time zone sizes.
5893 int lowmem_reserve_ratio_sysctl_handler(struct ctl_table
*table
, int write
,
5894 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5896 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5897 setup_per_zone_lowmem_reserve();
5902 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5903 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5904 * pagelist can have before it gets flushed back to buddy allocator.
5906 int percpu_pagelist_fraction_sysctl_handler(struct ctl_table
*table
, int write
,
5907 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5910 int old_percpu_pagelist_fraction
;
5913 mutex_lock(&pcp_batch_high_lock
);
5914 old_percpu_pagelist_fraction
= percpu_pagelist_fraction
;
5916 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5917 if (!write
|| ret
< 0)
5920 /* Sanity checking to avoid pcp imbalance */
5921 if (percpu_pagelist_fraction
&&
5922 percpu_pagelist_fraction
< MIN_PERCPU_PAGELIST_FRACTION
) {
5923 percpu_pagelist_fraction
= old_percpu_pagelist_fraction
;
5929 if (percpu_pagelist_fraction
== old_percpu_pagelist_fraction
)
5932 for_each_populated_zone(zone
) {
5935 for_each_possible_cpu(cpu
)
5936 pageset_set_high_and_batch(zone
,
5937 per_cpu_ptr(zone
->pageset
, cpu
));
5940 mutex_unlock(&pcp_batch_high_lock
);
5944 int hashdist
= HASHDIST_DEFAULT
;
5947 static int __init
set_hashdist(char *str
)
5951 hashdist
= simple_strtoul(str
, &str
, 0);
5954 __setup("hashdist=", set_hashdist
);
5958 * allocate a large system hash table from bootmem
5959 * - it is assumed that the hash table must contain an exact power-of-2
5960 * quantity of entries
5961 * - limit is the number of hash buckets, not the total allocation size
5963 void *__init
alloc_large_system_hash(const char *tablename
,
5964 unsigned long bucketsize
,
5965 unsigned long numentries
,
5968 unsigned int *_hash_shift
,
5969 unsigned int *_hash_mask
,
5970 unsigned long low_limit
,
5971 unsigned long high_limit
)
5973 unsigned long long max
= high_limit
;
5974 unsigned long log2qty
, size
;
5977 /* allow the kernel cmdline to have a say */
5979 /* round applicable memory size up to nearest megabyte */
5980 numentries
= nr_kernel_pages
;
5982 /* It isn't necessary when PAGE_SIZE >= 1MB */
5983 if (PAGE_SHIFT
< 20)
5984 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
5986 /* limit to 1 bucket per 2^scale bytes of low memory */
5987 if (scale
> PAGE_SHIFT
)
5988 numentries
>>= (scale
- PAGE_SHIFT
);
5990 numentries
<<= (PAGE_SHIFT
- scale
);
5992 /* Make sure we've got at least a 0-order allocation.. */
5993 if (unlikely(flags
& HASH_SMALL
)) {
5994 /* Makes no sense without HASH_EARLY */
5995 WARN_ON(!(flags
& HASH_EARLY
));
5996 if (!(numentries
>> *_hash_shift
)) {
5997 numentries
= 1UL << *_hash_shift
;
5998 BUG_ON(!numentries
);
6000 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
6001 numentries
= PAGE_SIZE
/ bucketsize
;
6003 numentries
= roundup_pow_of_two(numentries
);
6005 /* limit allocation size to 1/16 total memory by default */
6007 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
6008 do_div(max
, bucketsize
);
6010 max
= min(max
, 0x80000000ULL
);
6012 if (numentries
< low_limit
)
6013 numentries
= low_limit
;
6014 if (numentries
> max
)
6017 log2qty
= ilog2(numentries
);
6020 size
= bucketsize
<< log2qty
;
6021 if (flags
& HASH_EARLY
)
6022 table
= memblock_virt_alloc_nopanic(size
, 0);
6024 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
6027 * If bucketsize is not a power-of-two, we may free
6028 * some pages at the end of hash table which
6029 * alloc_pages_exact() automatically does
6031 if (get_order(size
) < MAX_ORDER
) {
6032 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
6033 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
6036 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
6039 panic("Failed to allocate %s hash table\n", tablename
);
6041 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
6044 ilog2(size
) - PAGE_SHIFT
,
6048 *_hash_shift
= log2qty
;
6050 *_hash_mask
= (1 << log2qty
) - 1;
6055 /* Return a pointer to the bitmap storing bits affecting a block of pages */
6056 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
6059 #ifdef CONFIG_SPARSEMEM
6060 return __pfn_to_section(pfn
)->pageblock_flags
;
6062 return zone
->pageblock_flags
;
6063 #endif /* CONFIG_SPARSEMEM */
6066 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
6068 #ifdef CONFIG_SPARSEMEM
6069 pfn
&= (PAGES_PER_SECTION
-1);
6070 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6072 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6073 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6074 #endif /* CONFIG_SPARSEMEM */
6078 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
6079 * @page: The page within the block of interest
6080 * @pfn: The target page frame number
6081 * @end_bitidx: The last bit of interest to retrieve
6082 * @mask: mask of bits that the caller is interested in
6084 * Return: pageblock_bits flags
6086 unsigned long get_pfnblock_flags_mask(struct page
*page
, unsigned long pfn
,
6087 unsigned long end_bitidx
,
6091 unsigned long *bitmap
;
6092 unsigned long bitidx
, word_bitidx
;
6095 zone
= page_zone(page
);
6096 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6097 bitidx
= pfn_to_bitidx(zone
, pfn
);
6098 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6099 bitidx
&= (BITS_PER_LONG
-1);
6101 word
= bitmap
[word_bitidx
];
6102 bitidx
+= end_bitidx
;
6103 return (word
>> (BITS_PER_LONG
- bitidx
- 1)) & mask
;
6107 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
6108 * @page: The page within the block of interest
6109 * @flags: The flags to set
6110 * @pfn: The target page frame number
6111 * @end_bitidx: The last bit of interest
6112 * @mask: mask of bits that the caller is interested in
6114 void set_pfnblock_flags_mask(struct page
*page
, unsigned long flags
,
6116 unsigned long end_bitidx
,
6120 unsigned long *bitmap
;
6121 unsigned long bitidx
, word_bitidx
;
6122 unsigned long old_word
, word
;
6124 BUILD_BUG_ON(NR_PAGEBLOCK_BITS
!= 4);
6126 zone
= page_zone(page
);
6127 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6128 bitidx
= pfn_to_bitidx(zone
, pfn
);
6129 word_bitidx
= bitidx
/ BITS_PER_LONG
;
6130 bitidx
&= (BITS_PER_LONG
-1);
6132 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6134 bitidx
+= end_bitidx
;
6135 mask
<<= (BITS_PER_LONG
- bitidx
- 1);
6136 flags
<<= (BITS_PER_LONG
- bitidx
- 1);
6138 word
= ACCESS_ONCE(bitmap
[word_bitidx
]);
6140 old_word
= cmpxchg(&bitmap
[word_bitidx
], word
, (word
& ~mask
) | flags
);
6141 if (word
== old_word
)
6148 * This function checks whether pageblock includes unmovable pages or not.
6149 * If @count is not zero, it is okay to include less @count unmovable pages
6151 * PageLRU check without isolation or lru_lock could race so that
6152 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6153 * expect this function should be exact.
6155 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6156 bool skip_hwpoisoned_pages
)
6158 unsigned long pfn
, iter
, found
;
6162 * For avoiding noise data, lru_add_drain_all() should be called
6163 * If ZONE_MOVABLE, the zone never contains unmovable pages
6165 if (zone_idx(zone
) == ZONE_MOVABLE
)
6167 mt
= get_pageblock_migratetype(page
);
6168 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6171 pfn
= page_to_pfn(page
);
6172 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6173 unsigned long check
= pfn
+ iter
;
6175 if (!pfn_valid_within(check
))
6178 page
= pfn_to_page(check
);
6181 * Hugepages are not in LRU lists, but they're movable.
6182 * We need not scan over tail pages bacause we don't
6183 * handle each tail page individually in migration.
6185 if (PageHuge(page
)) {
6186 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6191 * We can't use page_count without pin a page
6192 * because another CPU can free compound page.
6193 * This check already skips compound tails of THP
6194 * because their page->_count is zero at all time.
6196 if (!atomic_read(&page
->_count
)) {
6197 if (PageBuddy(page
))
6198 iter
+= (1 << page_order(page
)) - 1;
6203 * The HWPoisoned page may be not in buddy system, and
6204 * page_count() is not 0.
6206 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6212 * If there are RECLAIMABLE pages, we need to check it.
6213 * But now, memory offline itself doesn't call shrink_slab()
6214 * and it still to be fixed.
6217 * If the page is not RAM, page_count()should be 0.
6218 * we don't need more check. This is an _used_ not-movable page.
6220 * The problematic thing here is PG_reserved pages. PG_reserved
6221 * is set to both of a memory hole page and a _used_ kernel
6230 bool is_pageblock_removable_nolock(struct page
*page
)
6236 * We have to be careful here because we are iterating over memory
6237 * sections which are not zone aware so we might end up outside of
6238 * the zone but still within the section.
6239 * We have to take care about the node as well. If the node is offline
6240 * its NODE_DATA will be NULL - see page_zone.
6242 if (!node_online(page_to_nid(page
)))
6245 zone
= page_zone(page
);
6246 pfn
= page_to_pfn(page
);
6247 if (!zone_spans_pfn(zone
, pfn
))
6250 return !has_unmovable_pages(zone
, page
, 0, true);
6255 static unsigned long pfn_max_align_down(unsigned long pfn
)
6257 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6258 pageblock_nr_pages
) - 1);
6261 static unsigned long pfn_max_align_up(unsigned long pfn
)
6263 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6264 pageblock_nr_pages
));
6267 /* [start, end) must belong to a single zone. */
6268 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6269 unsigned long start
, unsigned long end
)
6271 /* This function is based on compact_zone() from compaction.c. */
6272 unsigned long nr_reclaimed
;
6273 unsigned long pfn
= start
;
6274 unsigned int tries
= 0;
6279 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6280 if (fatal_signal_pending(current
)) {
6285 if (list_empty(&cc
->migratepages
)) {
6286 cc
->nr_migratepages
= 0;
6287 pfn
= isolate_migratepages_range(cc
, pfn
, end
);
6293 } else if (++tries
== 5) {
6294 ret
= ret
< 0 ? ret
: -EBUSY
;
6298 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6300 cc
->nr_migratepages
-= nr_reclaimed
;
6302 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6303 NULL
, 0, cc
->mode
, MR_CMA
);
6306 putback_movable_pages(&cc
->migratepages
);
6313 * alloc_contig_range() -- tries to allocate given range of pages
6314 * @start: start PFN to allocate
6315 * @end: one-past-the-last PFN to allocate
6316 * @migratetype: migratetype of the underlaying pageblocks (either
6317 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6318 * in range must have the same migratetype and it must
6319 * be either of the two.
6321 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6322 * aligned, however it's the caller's responsibility to guarantee that
6323 * we are the only thread that changes migrate type of pageblocks the
6326 * The PFN range must belong to a single zone.
6328 * Returns zero on success or negative error code. On success all
6329 * pages which PFN is in [start, end) are allocated for the caller and
6330 * need to be freed with free_contig_range().
6332 int alloc_contig_range(unsigned long start
, unsigned long end
,
6333 unsigned migratetype
)
6335 unsigned long outer_start
, outer_end
;
6338 struct compact_control cc
= {
6339 .nr_migratepages
= 0,
6341 .zone
= page_zone(pfn_to_page(start
)),
6342 .mode
= MIGRATE_SYNC
,
6343 .ignore_skip_hint
= true,
6345 INIT_LIST_HEAD(&cc
.migratepages
);
6348 * What we do here is we mark all pageblocks in range as
6349 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6350 * have different sizes, and due to the way page allocator
6351 * work, we align the range to biggest of the two pages so
6352 * that page allocator won't try to merge buddies from
6353 * different pageblocks and change MIGRATE_ISOLATE to some
6354 * other migration type.
6356 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6357 * migrate the pages from an unaligned range (ie. pages that
6358 * we are interested in). This will put all the pages in
6359 * range back to page allocator as MIGRATE_ISOLATE.
6361 * When this is done, we take the pages in range from page
6362 * allocator removing them from the buddy system. This way
6363 * page allocator will never consider using them.
6365 * This lets us mark the pageblocks back as
6366 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6367 * aligned range but not in the unaligned, original range are
6368 * put back to page allocator so that buddy can use them.
6371 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6372 pfn_max_align_up(end
), migratetype
,
6377 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6382 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6383 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6384 * more, all pages in [start, end) are free in page allocator.
6385 * What we are going to do is to allocate all pages from
6386 * [start, end) (that is remove them from page allocator).
6388 * The only problem is that pages at the beginning and at the
6389 * end of interesting range may be not aligned with pages that
6390 * page allocator holds, ie. they can be part of higher order
6391 * pages. Because of this, we reserve the bigger range and
6392 * once this is done free the pages we are not interested in.
6394 * We don't have to hold zone->lock here because the pages are
6395 * isolated thus they won't get removed from buddy.
6398 lru_add_drain_all();
6402 outer_start
= start
;
6403 while (!PageBuddy(pfn_to_page(outer_start
))) {
6404 if (++order
>= MAX_ORDER
) {
6408 outer_start
&= ~0UL << order
;
6411 /* Make sure the range is really isolated. */
6412 if (test_pages_isolated(outer_start
, end
, false)) {
6413 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6420 /* Grab isolated pages from freelists. */
6421 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6427 /* Free head and tail (if any) */
6428 if (start
!= outer_start
)
6429 free_contig_range(outer_start
, start
- outer_start
);
6430 if (end
!= outer_end
)
6431 free_contig_range(end
, outer_end
- end
);
6434 undo_isolate_page_range(pfn_max_align_down(start
),
6435 pfn_max_align_up(end
), migratetype
);
6439 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6441 unsigned int count
= 0;
6443 for (; nr_pages
--; pfn
++) {
6444 struct page
*page
= pfn_to_page(pfn
);
6446 count
+= page_count(page
) != 1;
6449 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6453 #ifdef CONFIG_MEMORY_HOTPLUG
6455 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6456 * page high values need to be recalulated.
6458 void __meminit
zone_pcp_update(struct zone
*zone
)
6461 mutex_lock(&pcp_batch_high_lock
);
6462 for_each_possible_cpu(cpu
)
6463 pageset_set_high_and_batch(zone
,
6464 per_cpu_ptr(zone
->pageset
, cpu
));
6465 mutex_unlock(&pcp_batch_high_lock
);
6469 void zone_pcp_reset(struct zone
*zone
)
6471 unsigned long flags
;
6473 struct per_cpu_pageset
*pset
;
6475 /* avoid races with drain_pages() */
6476 local_irq_save(flags
);
6477 if (zone
->pageset
!= &boot_pageset
) {
6478 for_each_online_cpu(cpu
) {
6479 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6480 drain_zonestat(zone
, pset
);
6482 free_percpu(zone
->pageset
);
6483 zone
->pageset
= &boot_pageset
;
6485 local_irq_restore(flags
);
6488 #ifdef CONFIG_MEMORY_HOTREMOVE
6490 * All pages in the range must be isolated before calling this.
6493 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6497 unsigned int order
, i
;
6499 unsigned long flags
;
6500 /* find the first valid pfn */
6501 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6506 zone
= page_zone(pfn_to_page(pfn
));
6507 spin_lock_irqsave(&zone
->lock
, flags
);
6509 while (pfn
< end_pfn
) {
6510 if (!pfn_valid(pfn
)) {
6514 page
= pfn_to_page(pfn
);
6516 * The HWPoisoned page may be not in buddy system, and
6517 * page_count() is not 0.
6519 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6521 SetPageReserved(page
);
6525 BUG_ON(page_count(page
));
6526 BUG_ON(!PageBuddy(page
));
6527 order
= page_order(page
);
6528 #ifdef CONFIG_DEBUG_VM
6529 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6530 pfn
, 1 << order
, end_pfn
);
6532 list_del(&page
->lru
);
6533 rmv_page_order(page
);
6534 zone
->free_area
[order
].nr_free
--;
6535 for (i
= 0; i
< (1 << order
); i
++)
6536 SetPageReserved((page
+i
));
6537 pfn
+= (1 << order
);
6539 spin_unlock_irqrestore(&zone
->lock
, flags
);
6543 #ifdef CONFIG_MEMORY_FAILURE
6544 bool is_free_buddy_page(struct page
*page
)
6546 struct zone
*zone
= page_zone(page
);
6547 unsigned long pfn
= page_to_pfn(page
);
6548 unsigned long flags
;
6551 spin_lock_irqsave(&zone
->lock
, flags
);
6552 for (order
= 0; order
< MAX_ORDER
; order
++) {
6553 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6555 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6558 spin_unlock_irqrestore(&zone
->lock
, flags
);
6560 return order
< MAX_ORDER
;