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/memory.h>
55 #include <linux/compaction.h>
56 #include <trace/events/kmem.h>
57 #include <linux/ftrace_event.h>
58 #include <linux/memcontrol.h>
59 #include <linux/prefetch.h>
60 #include <linux/page-debug-flags.h>
62 #include <asm/tlbflush.h>
63 #include <asm/div64.h>
66 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
67 DEFINE_PER_CPU(int, numa_node
);
68 EXPORT_PER_CPU_SYMBOL(numa_node
);
71 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
73 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
74 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
75 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
76 * defined in <linux/topology.h>.
78 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
79 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
83 * Array of node states.
85 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
86 [N_POSSIBLE
] = NODE_MASK_ALL
,
87 [N_ONLINE
] = { { [0] = 1UL } },
89 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
91 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
93 [N_CPU
] = { { [0] = 1UL } },
96 EXPORT_SYMBOL(node_states
);
98 unsigned long totalram_pages __read_mostly
;
99 unsigned long totalreserve_pages __read_mostly
;
101 * When calculating the number of globally allowed dirty pages, there
102 * is a certain number of per-zone reserves that should not be
103 * considered dirtyable memory. This is the sum of those reserves
104 * over all existing zones that contribute dirtyable memory.
106 unsigned long dirty_balance_reserve __read_mostly
;
108 int percpu_pagelist_fraction
;
109 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
111 #ifdef CONFIG_PM_SLEEP
113 * The following functions are used by the suspend/hibernate code to temporarily
114 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
115 * while devices are suspended. To avoid races with the suspend/hibernate code,
116 * they should always be called with pm_mutex held (gfp_allowed_mask also should
117 * only be modified with pm_mutex held, unless the suspend/hibernate code is
118 * guaranteed not to run in parallel with that modification).
121 static gfp_t saved_gfp_mask
;
123 void pm_restore_gfp_mask(void)
125 WARN_ON(!mutex_is_locked(&pm_mutex
));
126 if (saved_gfp_mask
) {
127 gfp_allowed_mask
= saved_gfp_mask
;
132 void pm_restrict_gfp_mask(void)
134 WARN_ON(!mutex_is_locked(&pm_mutex
));
135 WARN_ON(saved_gfp_mask
);
136 saved_gfp_mask
= gfp_allowed_mask
;
137 gfp_allowed_mask
&= ~GFP_IOFS
;
140 bool pm_suspended_storage(void)
142 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
146 #endif /* CONFIG_PM_SLEEP */
148 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
149 int pageblock_order __read_mostly
;
152 static void __free_pages_ok(struct page
*page
, unsigned int order
);
155 * results with 256, 32 in the lowmem_reserve sysctl:
156 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
157 * 1G machine -> (16M dma, 784M normal, 224M high)
158 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
159 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
160 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
162 * TBD: should special case ZONE_DMA32 machines here - in those we normally
163 * don't need any ZONE_NORMAL reservation
165 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
166 #ifdef CONFIG_ZONE_DMA
169 #ifdef CONFIG_ZONE_DMA32
172 #ifdef CONFIG_HIGHMEM
178 EXPORT_SYMBOL(totalram_pages
);
180 static char * const zone_names
[MAX_NR_ZONES
] = {
181 #ifdef CONFIG_ZONE_DMA
184 #ifdef CONFIG_ZONE_DMA32
188 #ifdef CONFIG_HIGHMEM
194 int min_free_kbytes
= 1024;
196 static unsigned long __meminitdata nr_kernel_pages
;
197 static unsigned long __meminitdata nr_all_pages
;
198 static unsigned long __meminitdata dma_reserve
;
200 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
201 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
202 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
203 static unsigned long __initdata required_kernelcore
;
204 static unsigned long __initdata required_movablecore
;
205 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
207 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
209 EXPORT_SYMBOL(movable_zone
);
210 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
213 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
214 int nr_online_nodes __read_mostly
= 1;
215 EXPORT_SYMBOL(nr_node_ids
);
216 EXPORT_SYMBOL(nr_online_nodes
);
219 int page_group_by_mobility_disabled __read_mostly
;
221 static void set_pageblock_migratetype(struct page
*page
, int migratetype
)
224 if (unlikely(page_group_by_mobility_disabled
))
225 migratetype
= MIGRATE_UNMOVABLE
;
227 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
228 PB_migrate
, PB_migrate_end
);
231 bool oom_killer_disabled __read_mostly
;
233 #ifdef CONFIG_DEBUG_VM
234 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
238 unsigned long pfn
= page_to_pfn(page
);
241 seq
= zone_span_seqbegin(zone
);
242 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
244 else if (pfn
< zone
->zone_start_pfn
)
246 } while (zone_span_seqretry(zone
, seq
));
251 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
253 if (!pfn_valid_within(page_to_pfn(page
)))
255 if (zone
!= page_zone(page
))
261 * Temporary debugging check for pages not lying within a given zone.
263 static int bad_range(struct zone
*zone
, struct page
*page
)
265 if (page_outside_zone_boundaries(zone
, page
))
267 if (!page_is_consistent(zone
, page
))
273 static inline int bad_range(struct zone
*zone
, struct page
*page
)
279 static void bad_page(struct page
*page
)
281 static unsigned long resume
;
282 static unsigned long nr_shown
;
283 static unsigned long nr_unshown
;
285 /* Don't complain about poisoned pages */
286 if (PageHWPoison(page
)) {
287 reset_page_mapcount(page
); /* remove PageBuddy */
292 * Allow a burst of 60 reports, then keep quiet for that minute;
293 * or allow a steady drip of one report per second.
295 if (nr_shown
== 60) {
296 if (time_before(jiffies
, resume
)) {
302 "BUG: Bad page state: %lu messages suppressed\n",
309 resume
= jiffies
+ 60 * HZ
;
311 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
312 current
->comm
, page_to_pfn(page
));
318 /* Leave bad fields for debug, except PageBuddy could make trouble */
319 reset_page_mapcount(page
); /* remove PageBuddy */
320 add_taint(TAINT_BAD_PAGE
);
324 * Higher-order pages are called "compound pages". They are structured thusly:
326 * The first PAGE_SIZE page is called the "head page".
328 * The remaining PAGE_SIZE pages are called "tail pages".
330 * All pages have PG_compound set. All tail pages have their ->first_page
331 * pointing at the head page.
333 * The first tail page's ->lru.next holds the address of the compound page's
334 * put_page() function. Its ->lru.prev holds the order of allocation.
335 * This usage means that zero-order pages may not be compound.
338 static void free_compound_page(struct page
*page
)
340 __free_pages_ok(page
, compound_order(page
));
343 void prep_compound_page(struct page
*page
, unsigned long order
)
346 int nr_pages
= 1 << order
;
348 set_compound_page_dtor(page
, free_compound_page
);
349 set_compound_order(page
, order
);
351 for (i
= 1; i
< nr_pages
; i
++) {
352 struct page
*p
= page
+ i
;
354 set_page_count(p
, 0);
355 p
->first_page
= page
;
359 /* update __split_huge_page_refcount if you change this function */
360 static int destroy_compound_page(struct page
*page
, unsigned long order
)
363 int nr_pages
= 1 << order
;
366 if (unlikely(compound_order(page
) != order
) ||
367 unlikely(!PageHead(page
))) {
372 __ClearPageHead(page
);
374 for (i
= 1; i
< nr_pages
; i
++) {
375 struct page
*p
= page
+ i
;
377 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
387 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
392 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
393 * and __GFP_HIGHMEM from hard or soft interrupt context.
395 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
396 for (i
= 0; i
< (1 << order
); i
++)
397 clear_highpage(page
+ i
);
400 #ifdef CONFIG_DEBUG_PAGEALLOC
401 unsigned int _debug_guardpage_minorder
;
403 static int __init
debug_guardpage_minorder_setup(char *buf
)
407 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
408 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
411 _debug_guardpage_minorder
= res
;
412 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
415 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
417 static inline void set_page_guard_flag(struct page
*page
)
419 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
422 static inline void clear_page_guard_flag(struct page
*page
)
424 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
427 static inline void set_page_guard_flag(struct page
*page
) { }
428 static inline void clear_page_guard_flag(struct page
*page
) { }
431 static inline void set_page_order(struct page
*page
, int order
)
433 set_page_private(page
, order
);
434 __SetPageBuddy(page
);
437 static inline void rmv_page_order(struct page
*page
)
439 __ClearPageBuddy(page
);
440 set_page_private(page
, 0);
444 * Locate the struct page for both the matching buddy in our
445 * pair (buddy1) and the combined O(n+1) page they form (page).
447 * 1) Any buddy B1 will have an order O twin B2 which satisfies
448 * the following equation:
450 * For example, if the starting buddy (buddy2) is #8 its order
452 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
454 * 2) Any buddy B will have an order O+1 parent P which
455 * satisfies the following equation:
458 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
460 static inline unsigned long
461 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
463 return page_idx
^ (1 << order
);
467 * This function checks whether a page is free && is the buddy
468 * we can do coalesce a page and its buddy if
469 * (a) the buddy is not in a hole &&
470 * (b) the buddy is in the buddy system &&
471 * (c) a page and its buddy have the same order &&
472 * (d) a page and its buddy are in the same zone.
474 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
475 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
477 * For recording page's order, we use page_private(page).
479 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
482 if (!pfn_valid_within(page_to_pfn(buddy
)))
485 if (page_zone_id(page
) != page_zone_id(buddy
))
488 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
489 VM_BUG_ON(page_count(buddy
) != 0);
493 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
494 VM_BUG_ON(page_count(buddy
) != 0);
501 * Freeing function for a buddy system allocator.
503 * The concept of a buddy system is to maintain direct-mapped table
504 * (containing bit values) for memory blocks of various "orders".
505 * The bottom level table contains the map for the smallest allocatable
506 * units of memory (here, pages), and each level above it describes
507 * pairs of units from the levels below, hence, "buddies".
508 * At a high level, all that happens here is marking the table entry
509 * at the bottom level available, and propagating the changes upward
510 * as necessary, plus some accounting needed to play nicely with other
511 * parts of the VM system.
512 * At each level, we keep a list of pages, which are heads of continuous
513 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
514 * order is recorded in page_private(page) field.
515 * So when we are allocating or freeing one, we can derive the state of the
516 * other. That is, if we allocate a small block, and both were
517 * free, the remainder of the region must be split into blocks.
518 * If a block is freed, and its buddy is also free, then this
519 * triggers coalescing into a block of larger size.
524 static inline void __free_one_page(struct page
*page
,
525 struct zone
*zone
, unsigned int order
,
528 unsigned long page_idx
;
529 unsigned long combined_idx
;
530 unsigned long uninitialized_var(buddy_idx
);
533 if (unlikely(PageCompound(page
)))
534 if (unlikely(destroy_compound_page(page
, order
)))
537 VM_BUG_ON(migratetype
== -1);
539 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
541 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
542 VM_BUG_ON(bad_range(zone
, page
));
544 while (order
< MAX_ORDER
-1) {
545 buddy_idx
= __find_buddy_index(page_idx
, order
);
546 buddy
= page
+ (buddy_idx
- page_idx
);
547 if (!page_is_buddy(page
, buddy
, order
))
550 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
551 * merge with it and move up one order.
553 if (page_is_guard(buddy
)) {
554 clear_page_guard_flag(buddy
);
555 set_page_private(page
, 0);
556 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
558 list_del(&buddy
->lru
);
559 zone
->free_area
[order
].nr_free
--;
560 rmv_page_order(buddy
);
562 combined_idx
= buddy_idx
& page_idx
;
563 page
= page
+ (combined_idx
- page_idx
);
564 page_idx
= combined_idx
;
567 set_page_order(page
, order
);
570 * If this is not the largest possible page, check if the buddy
571 * of the next-highest order is free. If it is, it's possible
572 * that pages are being freed that will coalesce soon. In case,
573 * that is happening, add the free page to the tail of the list
574 * so it's less likely to be used soon and more likely to be merged
575 * as a higher order page
577 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
578 struct page
*higher_page
, *higher_buddy
;
579 combined_idx
= buddy_idx
& page_idx
;
580 higher_page
= page
+ (combined_idx
- page_idx
);
581 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
582 higher_buddy
= page
+ (buddy_idx
- combined_idx
);
583 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
584 list_add_tail(&page
->lru
,
585 &zone
->free_area
[order
].free_list
[migratetype
]);
590 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
592 zone
->free_area
[order
].nr_free
++;
596 * free_page_mlock() -- clean up attempts to free and mlocked() page.
597 * Page should not be on lru, so no need to fix that up.
598 * free_pages_check() will verify...
600 static inline void free_page_mlock(struct page
*page
)
602 __dec_zone_page_state(page
, NR_MLOCK
);
603 __count_vm_event(UNEVICTABLE_MLOCKFREED
);
606 static inline int free_pages_check(struct page
*page
)
608 if (unlikely(page_mapcount(page
) |
609 (page
->mapping
!= NULL
) |
610 (atomic_read(&page
->_count
) != 0) |
611 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
) |
612 (mem_cgroup_bad_page_check(page
)))) {
616 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
617 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
622 * Frees a number of pages from the PCP lists
623 * Assumes all pages on list are in same zone, and of same order.
624 * count is the number of pages to free.
626 * If the zone was previously in an "all pages pinned" state then look to
627 * see if this freeing clears that state.
629 * And clear the zone's pages_scanned counter, to hold off the "all pages are
630 * pinned" detection logic.
632 static void free_pcppages_bulk(struct zone
*zone
, int count
,
633 struct per_cpu_pages
*pcp
)
639 spin_lock(&zone
->lock
);
640 zone
->all_unreclaimable
= 0;
641 zone
->pages_scanned
= 0;
645 struct list_head
*list
;
648 * Remove pages from lists in a round-robin fashion. A
649 * batch_free count is maintained that is incremented when an
650 * empty list is encountered. This is so more pages are freed
651 * off fuller lists instead of spinning excessively around empty
656 if (++migratetype
== MIGRATE_PCPTYPES
)
658 list
= &pcp
->lists
[migratetype
];
659 } while (list_empty(list
));
661 /* This is the only non-empty list. Free them all. */
662 if (batch_free
== MIGRATE_PCPTYPES
)
663 batch_free
= to_free
;
666 page
= list_entry(list
->prev
, struct page
, lru
);
667 /* must delete as __free_one_page list manipulates */
668 list_del(&page
->lru
);
669 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
670 __free_one_page(page
, zone
, 0, page_private(page
));
671 trace_mm_page_pcpu_drain(page
, 0, page_private(page
));
672 } while (--to_free
&& --batch_free
&& !list_empty(list
));
674 __mod_zone_page_state(zone
, NR_FREE_PAGES
, count
);
675 spin_unlock(&zone
->lock
);
678 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
681 spin_lock(&zone
->lock
);
682 zone
->all_unreclaimable
= 0;
683 zone
->pages_scanned
= 0;
685 __free_one_page(page
, zone
, order
, migratetype
);
686 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1 << order
);
687 spin_unlock(&zone
->lock
);
690 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
695 trace_mm_page_free(page
, order
);
696 kmemcheck_free_shadow(page
, order
);
699 page
->mapping
= NULL
;
700 for (i
= 0; i
< (1 << order
); i
++)
701 bad
+= free_pages_check(page
+ i
);
705 if (!PageHighMem(page
)) {
706 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
707 debug_check_no_obj_freed(page_address(page
),
710 arch_free_page(page
, order
);
711 kernel_map_pages(page
, 1 << order
, 0);
716 static void __free_pages_ok(struct page
*page
, unsigned int order
)
719 int wasMlocked
= __TestClearPageMlocked(page
);
721 if (!free_pages_prepare(page
, order
))
724 local_irq_save(flags
);
725 if (unlikely(wasMlocked
))
726 free_page_mlock(page
);
727 __count_vm_events(PGFREE
, 1 << order
);
728 free_one_page(page_zone(page
), page
, order
,
729 get_pageblock_migratetype(page
));
730 local_irq_restore(flags
);
733 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
735 unsigned int nr_pages
= 1 << order
;
739 for (loop
= 0; loop
< nr_pages
; loop
++) {
740 struct page
*p
= &page
[loop
];
742 if (loop
+ 1 < nr_pages
)
744 __ClearPageReserved(p
);
745 set_page_count(p
, 0);
748 set_page_refcounted(page
);
749 __free_pages(page
, order
);
754 * The order of subdivision here is critical for the IO subsystem.
755 * Please do not alter this order without good reasons and regression
756 * testing. Specifically, as large blocks of memory are subdivided,
757 * the order in which smaller blocks are delivered depends on the order
758 * they're subdivided in this function. This is the primary factor
759 * influencing the order in which pages are delivered to the IO
760 * subsystem according to empirical testing, and this is also justified
761 * by considering the behavior of a buddy system containing a single
762 * large block of memory acted on by a series of small allocations.
763 * This behavior is a critical factor in sglist merging's success.
767 static inline void expand(struct zone
*zone
, struct page
*page
,
768 int low
, int high
, struct free_area
*area
,
771 unsigned long size
= 1 << high
;
777 VM_BUG_ON(bad_range(zone
, &page
[size
]));
779 #ifdef CONFIG_DEBUG_PAGEALLOC
780 if (high
< debug_guardpage_minorder()) {
782 * Mark as guard pages (or page), that will allow to
783 * merge back to allocator when buddy will be freed.
784 * Corresponding page table entries will not be touched,
785 * pages will stay not present in virtual address space
787 INIT_LIST_HEAD(&page
[size
].lru
);
788 set_page_guard_flag(&page
[size
]);
789 set_page_private(&page
[size
], high
);
790 /* Guard pages are not available for any usage */
791 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << high
));
795 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
797 set_page_order(&page
[size
], high
);
802 * This page is about to be returned from the page allocator
804 static inline int check_new_page(struct page
*page
)
806 if (unlikely(page_mapcount(page
) |
807 (page
->mapping
!= NULL
) |
808 (atomic_read(&page
->_count
) != 0) |
809 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
) |
810 (mem_cgroup_bad_page_check(page
)))) {
817 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
821 for (i
= 0; i
< (1 << order
); i
++) {
822 struct page
*p
= page
+ i
;
823 if (unlikely(check_new_page(p
)))
827 set_page_private(page
, 0);
828 set_page_refcounted(page
);
830 arch_alloc_page(page
, order
);
831 kernel_map_pages(page
, 1 << order
, 1);
833 if (gfp_flags
& __GFP_ZERO
)
834 prep_zero_page(page
, order
, gfp_flags
);
836 if (order
&& (gfp_flags
& __GFP_COMP
))
837 prep_compound_page(page
, order
);
843 * Go through the free lists for the given migratetype and remove
844 * the smallest available page from the freelists
847 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
850 unsigned int current_order
;
851 struct free_area
* area
;
854 /* Find a page of the appropriate size in the preferred list */
855 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
856 area
= &(zone
->free_area
[current_order
]);
857 if (list_empty(&area
->free_list
[migratetype
]))
860 page
= list_entry(area
->free_list
[migratetype
].next
,
862 list_del(&page
->lru
);
863 rmv_page_order(page
);
865 expand(zone
, page
, order
, current_order
, area
, migratetype
);
874 * This array describes the order lists are fallen back to when
875 * the free lists for the desirable migrate type are depleted
877 static int fallbacks
[MIGRATE_TYPES
][MIGRATE_TYPES
-1] = {
878 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
879 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
880 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
881 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
, MIGRATE_RESERVE
, MIGRATE_RESERVE
}, /* Never used */
885 * Move the free pages in a range to the free lists of the requested type.
886 * Note that start_page and end_pages are not aligned on a pageblock
887 * boundary. If alignment is required, use move_freepages_block()
889 static int move_freepages(struct zone
*zone
,
890 struct page
*start_page
, struct page
*end_page
,
897 #ifndef CONFIG_HOLES_IN_ZONE
899 * page_zone is not safe to call in this context when
900 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
901 * anyway as we check zone boundaries in move_freepages_block().
902 * Remove at a later date when no bug reports exist related to
903 * grouping pages by mobility
905 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
908 for (page
= start_page
; page
<= end_page
;) {
909 /* Make sure we are not inadvertently changing nodes */
910 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
912 if (!pfn_valid_within(page_to_pfn(page
))) {
917 if (!PageBuddy(page
)) {
922 order
= page_order(page
);
923 list_move(&page
->lru
,
924 &zone
->free_area
[order
].free_list
[migratetype
]);
926 pages_moved
+= 1 << order
;
932 static int move_freepages_block(struct zone
*zone
, struct page
*page
,
935 unsigned long start_pfn
, end_pfn
;
936 struct page
*start_page
, *end_page
;
938 start_pfn
= page_to_pfn(page
);
939 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
940 start_page
= pfn_to_page(start_pfn
);
941 end_page
= start_page
+ pageblock_nr_pages
- 1;
942 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
944 /* Do not cross zone boundaries */
945 if (start_pfn
< zone
->zone_start_pfn
)
947 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
950 return move_freepages(zone
, start_page
, end_page
, migratetype
);
953 static void change_pageblock_range(struct page
*pageblock_page
,
954 int start_order
, int migratetype
)
956 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
958 while (nr_pageblocks
--) {
959 set_pageblock_migratetype(pageblock_page
, migratetype
);
960 pageblock_page
+= pageblock_nr_pages
;
964 /* Remove an element from the buddy allocator from the fallback list */
965 static inline struct page
*
966 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
968 struct free_area
* area
;
973 /* Find the largest possible block of pages in the other list */
974 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
976 for (i
= 0; i
< MIGRATE_TYPES
- 1; i
++) {
977 migratetype
= fallbacks
[start_migratetype
][i
];
979 /* MIGRATE_RESERVE handled later if necessary */
980 if (migratetype
== MIGRATE_RESERVE
)
983 area
= &(zone
->free_area
[current_order
]);
984 if (list_empty(&area
->free_list
[migratetype
]))
987 page
= list_entry(area
->free_list
[migratetype
].next
,
992 * If breaking a large block of pages, move all free
993 * pages to the preferred allocation list. If falling
994 * back for a reclaimable kernel allocation, be more
995 * aggressive about taking ownership of free pages
997 if (unlikely(current_order
>= (pageblock_order
>> 1)) ||
998 start_migratetype
== MIGRATE_RECLAIMABLE
||
999 page_group_by_mobility_disabled
) {
1000 unsigned long pages
;
1001 pages
= move_freepages_block(zone
, page
,
1004 /* Claim the whole block if over half of it is free */
1005 if (pages
>= (1 << (pageblock_order
-1)) ||
1006 page_group_by_mobility_disabled
)
1007 set_pageblock_migratetype(page
,
1010 migratetype
= start_migratetype
;
1013 /* Remove the page from the freelists */
1014 list_del(&page
->lru
);
1015 rmv_page_order(page
);
1017 /* Take ownership for orders >= pageblock_order */
1018 if (current_order
>= pageblock_order
)
1019 change_pageblock_range(page
, current_order
,
1022 expand(zone
, page
, order
, current_order
, area
, migratetype
);
1024 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1025 start_migratetype
, migratetype
);
1035 * Do the hard work of removing an element from the buddy allocator.
1036 * Call me with the zone->lock already held.
1038 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1044 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1046 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1047 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1050 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1051 * is used because __rmqueue_smallest is an inline function
1052 * and we want just one call site
1055 migratetype
= MIGRATE_RESERVE
;
1060 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1065 * Obtain a specified number of elements from the buddy allocator, all under
1066 * a single hold of the lock, for efficiency. Add them to the supplied list.
1067 * Returns the number of new pages which were placed at *list.
1069 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1070 unsigned long count
, struct list_head
*list
,
1071 int migratetype
, int cold
)
1075 spin_lock(&zone
->lock
);
1076 for (i
= 0; i
< count
; ++i
) {
1077 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1078 if (unlikely(page
== NULL
))
1082 * Split buddy pages returned by expand() are received here
1083 * in physical page order. The page is added to the callers and
1084 * list and the list head then moves forward. From the callers
1085 * perspective, the linked list is ordered by page number in
1086 * some conditions. This is useful for IO devices that can
1087 * merge IO requests if the physical pages are ordered
1090 if (likely(cold
== 0))
1091 list_add(&page
->lru
, list
);
1093 list_add_tail(&page
->lru
, list
);
1094 set_page_private(page
, migratetype
);
1097 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1098 spin_unlock(&zone
->lock
);
1104 * Called from the vmstat counter updater to drain pagesets of this
1105 * currently executing processor on remote nodes after they have
1108 * Note that this function must be called with the thread pinned to
1109 * a single processor.
1111 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1113 unsigned long flags
;
1116 local_irq_save(flags
);
1117 if (pcp
->count
>= pcp
->batch
)
1118 to_drain
= pcp
->batch
;
1120 to_drain
= pcp
->count
;
1121 free_pcppages_bulk(zone
, to_drain
, pcp
);
1122 pcp
->count
-= to_drain
;
1123 local_irq_restore(flags
);
1128 * Drain pages of the indicated processor.
1130 * The processor must either be the current processor and the
1131 * thread pinned to the current processor or a processor that
1134 static void drain_pages(unsigned int cpu
)
1136 unsigned long flags
;
1139 for_each_populated_zone(zone
) {
1140 struct per_cpu_pageset
*pset
;
1141 struct per_cpu_pages
*pcp
;
1143 local_irq_save(flags
);
1144 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1148 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1151 local_irq_restore(flags
);
1156 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1158 void drain_local_pages(void *arg
)
1160 drain_pages(smp_processor_id());
1164 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1166 void drain_all_pages(void)
1168 on_each_cpu(drain_local_pages
, NULL
, 1);
1171 #ifdef CONFIG_HIBERNATION
1173 void mark_free_pages(struct zone
*zone
)
1175 unsigned long pfn
, max_zone_pfn
;
1176 unsigned long flags
;
1178 struct list_head
*curr
;
1180 if (!zone
->spanned_pages
)
1183 spin_lock_irqsave(&zone
->lock
, flags
);
1185 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1186 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1187 if (pfn_valid(pfn
)) {
1188 struct page
*page
= pfn_to_page(pfn
);
1190 if (!swsusp_page_is_forbidden(page
))
1191 swsusp_unset_page_free(page
);
1194 for_each_migratetype_order(order
, t
) {
1195 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1198 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1199 for (i
= 0; i
< (1UL << order
); i
++)
1200 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1203 spin_unlock_irqrestore(&zone
->lock
, flags
);
1205 #endif /* CONFIG_PM */
1208 * Free a 0-order page
1209 * cold == 1 ? free a cold page : free a hot page
1211 void free_hot_cold_page(struct page
*page
, int cold
)
1213 struct zone
*zone
= page_zone(page
);
1214 struct per_cpu_pages
*pcp
;
1215 unsigned long flags
;
1217 int wasMlocked
= __TestClearPageMlocked(page
);
1219 if (!free_pages_prepare(page
, 0))
1222 migratetype
= get_pageblock_migratetype(page
);
1223 set_page_private(page
, migratetype
);
1224 local_irq_save(flags
);
1225 if (unlikely(wasMlocked
))
1226 free_page_mlock(page
);
1227 __count_vm_event(PGFREE
);
1230 * We only track unmovable, reclaimable and movable on pcp lists.
1231 * Free ISOLATE pages back to the allocator because they are being
1232 * offlined but treat RESERVE as movable pages so we can get those
1233 * areas back if necessary. Otherwise, we may have to free
1234 * excessively into the page allocator
1236 if (migratetype
>= MIGRATE_PCPTYPES
) {
1237 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1238 free_one_page(zone
, page
, 0, migratetype
);
1241 migratetype
= MIGRATE_MOVABLE
;
1244 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1246 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1248 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1250 if (pcp
->count
>= pcp
->high
) {
1251 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1252 pcp
->count
-= pcp
->batch
;
1256 local_irq_restore(flags
);
1260 * Free a list of 0-order pages
1262 void free_hot_cold_page_list(struct list_head
*list
, int cold
)
1264 struct page
*page
, *next
;
1266 list_for_each_entry_safe(page
, next
, list
, lru
) {
1267 trace_mm_page_free_batched(page
, cold
);
1268 free_hot_cold_page(page
, cold
);
1273 * split_page takes a non-compound higher-order page, and splits it into
1274 * n (1<<order) sub-pages: page[0..n]
1275 * Each sub-page must be freed individually.
1277 * Note: this is probably too low level an operation for use in drivers.
1278 * Please consult with lkml before using this in your driver.
1280 void split_page(struct page
*page
, unsigned int order
)
1284 VM_BUG_ON(PageCompound(page
));
1285 VM_BUG_ON(!page_count(page
));
1287 #ifdef CONFIG_KMEMCHECK
1289 * Split shadow pages too, because free(page[0]) would
1290 * otherwise free the whole shadow.
1292 if (kmemcheck_page_is_tracked(page
))
1293 split_page(virt_to_page(page
[0].shadow
), order
);
1296 for (i
= 1; i
< (1 << order
); i
++)
1297 set_page_refcounted(page
+ i
);
1301 * Similar to split_page except the page is already free. As this is only
1302 * being used for migration, the migratetype of the block also changes.
1303 * As this is called with interrupts disabled, the caller is responsible
1304 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1307 * Note: this is probably too low level an operation for use in drivers.
1308 * Please consult with lkml before using this in your driver.
1310 int split_free_page(struct page
*page
)
1313 unsigned long watermark
;
1316 BUG_ON(!PageBuddy(page
));
1318 zone
= page_zone(page
);
1319 order
= page_order(page
);
1321 /* Obey watermarks as if the page was being allocated */
1322 watermark
= low_wmark_pages(zone
) + (1 << order
);
1323 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1326 /* Remove page from free list */
1327 list_del(&page
->lru
);
1328 zone
->free_area
[order
].nr_free
--;
1329 rmv_page_order(page
);
1330 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1UL << order
));
1332 /* Split into individual pages */
1333 set_page_refcounted(page
);
1334 split_page(page
, order
);
1336 if (order
>= pageblock_order
- 1) {
1337 struct page
*endpage
= page
+ (1 << order
) - 1;
1338 for (; page
< endpage
; page
+= pageblock_nr_pages
)
1339 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
1346 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1347 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1351 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1352 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1355 unsigned long flags
;
1357 int cold
= !!(gfp_flags
& __GFP_COLD
);
1360 if (likely(order
== 0)) {
1361 struct per_cpu_pages
*pcp
;
1362 struct list_head
*list
;
1364 local_irq_save(flags
);
1365 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1366 list
= &pcp
->lists
[migratetype
];
1367 if (list_empty(list
)) {
1368 pcp
->count
+= rmqueue_bulk(zone
, 0,
1371 if (unlikely(list_empty(list
)))
1376 page
= list_entry(list
->prev
, struct page
, lru
);
1378 page
= list_entry(list
->next
, struct page
, lru
);
1380 list_del(&page
->lru
);
1383 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1385 * __GFP_NOFAIL is not to be used in new code.
1387 * All __GFP_NOFAIL callers should be fixed so that they
1388 * properly detect and handle allocation failures.
1390 * We most definitely don't want callers attempting to
1391 * allocate greater than order-1 page units with
1394 WARN_ON_ONCE(order
> 1);
1396 spin_lock_irqsave(&zone
->lock
, flags
);
1397 page
= __rmqueue(zone
, order
, migratetype
);
1398 spin_unlock(&zone
->lock
);
1401 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(1 << order
));
1404 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1405 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1406 local_irq_restore(flags
);
1408 VM_BUG_ON(bad_range(zone
, page
));
1409 if (prep_new_page(page
, order
, gfp_flags
))
1414 local_irq_restore(flags
);
1418 /* The ALLOC_WMARK bits are used as an index to zone->watermark */
1419 #define ALLOC_WMARK_MIN WMARK_MIN
1420 #define ALLOC_WMARK_LOW WMARK_LOW
1421 #define ALLOC_WMARK_HIGH WMARK_HIGH
1422 #define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1424 /* Mask to get the watermark bits */
1425 #define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1427 #define ALLOC_HARDER 0x10 /* try to alloc harder */
1428 #define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1429 #define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1431 #ifdef CONFIG_FAIL_PAGE_ALLOC
1434 struct fault_attr attr
;
1436 u32 ignore_gfp_highmem
;
1437 u32 ignore_gfp_wait
;
1439 } fail_page_alloc
= {
1440 .attr
= FAULT_ATTR_INITIALIZER
,
1441 .ignore_gfp_wait
= 1,
1442 .ignore_gfp_highmem
= 1,
1446 static int __init
setup_fail_page_alloc(char *str
)
1448 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1450 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1452 static int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1454 if (order
< fail_page_alloc
.min_order
)
1456 if (gfp_mask
& __GFP_NOFAIL
)
1458 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1460 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1463 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1466 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1468 static int __init
fail_page_alloc_debugfs(void)
1470 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1473 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1474 &fail_page_alloc
.attr
);
1476 return PTR_ERR(dir
);
1478 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1479 &fail_page_alloc
.ignore_gfp_wait
))
1481 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1482 &fail_page_alloc
.ignore_gfp_highmem
))
1484 if (!debugfs_create_u32("min-order", mode
, dir
,
1485 &fail_page_alloc
.min_order
))
1490 debugfs_remove_recursive(dir
);
1495 late_initcall(fail_page_alloc_debugfs
);
1497 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1499 #else /* CONFIG_FAIL_PAGE_ALLOC */
1501 static inline int should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1506 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1509 * Return true if free pages are above 'mark'. This takes into account the order
1510 * of the allocation.
1512 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1513 int classzone_idx
, int alloc_flags
, long free_pages
)
1515 /* free_pages my go negative - that's OK */
1519 free_pages
-= (1 << order
) - 1;
1520 if (alloc_flags
& ALLOC_HIGH
)
1522 if (alloc_flags
& ALLOC_HARDER
)
1525 if (free_pages
<= min
+ z
->lowmem_reserve
[classzone_idx
])
1527 for (o
= 0; o
< order
; o
++) {
1528 /* At the next order, this order's pages become unavailable */
1529 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1531 /* Require fewer higher order pages to be free */
1534 if (free_pages
<= min
)
1540 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1541 int classzone_idx
, int alloc_flags
)
1543 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1544 zone_page_state(z
, NR_FREE_PAGES
));
1547 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1548 int classzone_idx
, int alloc_flags
)
1550 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1552 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1553 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1555 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1561 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1562 * skip over zones that are not allowed by the cpuset, or that have
1563 * been recently (in last second) found to be nearly full. See further
1564 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1565 * that have to skip over a lot of full or unallowed zones.
1567 * If the zonelist cache is present in the passed in zonelist, then
1568 * returns a pointer to the allowed node mask (either the current
1569 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1571 * If the zonelist cache is not available for this zonelist, does
1572 * nothing and returns NULL.
1574 * If the fullzones BITMAP in the zonelist cache is stale (more than
1575 * a second since last zap'd) then we zap it out (clear its bits.)
1577 * We hold off even calling zlc_setup, until after we've checked the
1578 * first zone in the zonelist, on the theory that most allocations will
1579 * be satisfied from that first zone, so best to examine that zone as
1580 * quickly as we can.
1582 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1584 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1585 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1587 zlc
= zonelist
->zlcache_ptr
;
1591 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1592 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1593 zlc
->last_full_zap
= jiffies
;
1596 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1597 &cpuset_current_mems_allowed
:
1598 &node_states
[N_HIGH_MEMORY
];
1599 return allowednodes
;
1603 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1604 * if it is worth looking at further for free memory:
1605 * 1) Check that the zone isn't thought to be full (doesn't have its
1606 * bit set in the zonelist_cache fullzones BITMAP).
1607 * 2) Check that the zones node (obtained from the zonelist_cache
1608 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1609 * Return true (non-zero) if zone is worth looking at further, or
1610 * else return false (zero) if it is not.
1612 * This check -ignores- the distinction between various watermarks,
1613 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1614 * found to be full for any variation of these watermarks, it will
1615 * be considered full for up to one second by all requests, unless
1616 * we are so low on memory on all allowed nodes that we are forced
1617 * into the second scan of the zonelist.
1619 * In the second scan we ignore this zonelist cache and exactly
1620 * apply the watermarks to all zones, even it is slower to do so.
1621 * We are low on memory in the second scan, and should leave no stone
1622 * unturned looking for a free page.
1624 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1625 nodemask_t
*allowednodes
)
1627 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1628 int i
; /* index of *z in zonelist zones */
1629 int n
; /* node that zone *z is on */
1631 zlc
= zonelist
->zlcache_ptr
;
1635 i
= z
- zonelist
->_zonerefs
;
1638 /* This zone is worth trying if it is allowed but not full */
1639 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1643 * Given 'z' scanning a zonelist, set the corresponding bit in
1644 * zlc->fullzones, so that subsequent attempts to allocate a page
1645 * from that zone don't waste time re-examining it.
1647 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1649 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1650 int i
; /* index of *z in zonelist zones */
1652 zlc
= zonelist
->zlcache_ptr
;
1656 i
= z
- zonelist
->_zonerefs
;
1658 set_bit(i
, zlc
->fullzones
);
1662 * clear all zones full, called after direct reclaim makes progress so that
1663 * a zone that was recently full is not skipped over for up to a second
1665 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1667 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1669 zlc
= zonelist
->zlcache_ptr
;
1673 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1676 #else /* CONFIG_NUMA */
1678 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1683 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1684 nodemask_t
*allowednodes
)
1689 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1693 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1696 #endif /* CONFIG_NUMA */
1699 * get_page_from_freelist goes through the zonelist trying to allocate
1702 static struct page
*
1703 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1704 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1705 struct zone
*preferred_zone
, int migratetype
)
1708 struct page
*page
= NULL
;
1711 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1712 int zlc_active
= 0; /* set if using zonelist_cache */
1713 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1715 classzone_idx
= zone_idx(preferred_zone
);
1718 * Scan zonelist, looking for a zone with enough free.
1719 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1721 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1722 high_zoneidx
, nodemask
) {
1723 if (NUMA_BUILD
&& zlc_active
&&
1724 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1726 if ((alloc_flags
& ALLOC_CPUSET
) &&
1727 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1730 * When allocating a page cache page for writing, we
1731 * want to get it from a zone that is within its dirty
1732 * limit, such that no single zone holds more than its
1733 * proportional share of globally allowed dirty pages.
1734 * The dirty limits take into account the zone's
1735 * lowmem reserves and high watermark so that kswapd
1736 * should be able to balance it without having to
1737 * write pages from its LRU list.
1739 * This may look like it could increase pressure on
1740 * lower zones by failing allocations in higher zones
1741 * before they are full. But the pages that do spill
1742 * over are limited as the lower zones are protected
1743 * by this very same mechanism. It should not become
1744 * a practical burden to them.
1746 * XXX: For now, allow allocations to potentially
1747 * exceed the per-zone dirty limit in the slowpath
1748 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1749 * which is important when on a NUMA setup the allowed
1750 * zones are together not big enough to reach the
1751 * global limit. The proper fix for these situations
1752 * will require awareness of zones in the
1753 * dirty-throttling and the flusher threads.
1755 if ((alloc_flags
& ALLOC_WMARK_LOW
) &&
1756 (gfp_mask
& __GFP_WRITE
) && !zone_dirty_ok(zone
))
1757 goto this_zone_full
;
1759 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1760 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1764 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1765 if (zone_watermark_ok(zone
, order
, mark
,
1766 classzone_idx
, alloc_flags
))
1769 if (NUMA_BUILD
&& !did_zlc_setup
&& nr_online_nodes
> 1) {
1771 * we do zlc_setup if there are multiple nodes
1772 * and before considering the first zone allowed
1775 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1780 if (zone_reclaim_mode
== 0)
1781 goto this_zone_full
;
1784 * As we may have just activated ZLC, check if the first
1785 * eligible zone has failed zone_reclaim recently.
1787 if (NUMA_BUILD
&& zlc_active
&&
1788 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1791 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1793 case ZONE_RECLAIM_NOSCAN
:
1796 case ZONE_RECLAIM_FULL
:
1797 /* scanned but unreclaimable */
1800 /* did we reclaim enough */
1801 if (!zone_watermark_ok(zone
, order
, mark
,
1802 classzone_idx
, alloc_flags
))
1803 goto this_zone_full
;
1808 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1809 gfp_mask
, migratetype
);
1814 zlc_mark_zone_full(zonelist
, z
);
1817 if (unlikely(NUMA_BUILD
&& page
== NULL
&& zlc_active
)) {
1818 /* Disable zlc cache for second zonelist scan */
1826 * Large machines with many possible nodes should not always dump per-node
1827 * meminfo in irq context.
1829 static inline bool should_suppress_show_mem(void)
1834 ret
= in_interrupt();
1839 static DEFINE_RATELIMIT_STATE(nopage_rs
,
1840 DEFAULT_RATELIMIT_INTERVAL
,
1841 DEFAULT_RATELIMIT_BURST
);
1843 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
1845 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
1847 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
1848 debug_guardpage_minorder() > 0)
1852 * This documents exceptions given to allocations in certain
1853 * contexts that are allowed to allocate outside current's set
1856 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
1857 if (test_thread_flag(TIF_MEMDIE
) ||
1858 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
1859 filter
&= ~SHOW_MEM_FILTER_NODES
;
1860 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
1861 filter
&= ~SHOW_MEM_FILTER_NODES
;
1864 struct va_format vaf
;
1867 va_start(args
, fmt
);
1872 pr_warn("%pV", &vaf
);
1877 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
1878 current
->comm
, order
, gfp_mask
);
1881 if (!should_suppress_show_mem())
1886 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
1887 unsigned long did_some_progress
,
1888 unsigned long pages_reclaimed
)
1890 /* Do not loop if specifically requested */
1891 if (gfp_mask
& __GFP_NORETRY
)
1894 /* Always retry if specifically requested */
1895 if (gfp_mask
& __GFP_NOFAIL
)
1899 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
1900 * making forward progress without invoking OOM. Suspend also disables
1901 * storage devices so kswapd will not help. Bail if we are suspending.
1903 if (!did_some_progress
&& pm_suspended_storage())
1907 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1908 * means __GFP_NOFAIL, but that may not be true in other
1911 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
1915 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1916 * specified, then we retry until we no longer reclaim any pages
1917 * (above), or we've reclaimed an order of pages at least as
1918 * large as the allocation's order. In both cases, if the
1919 * allocation still fails, we stop retrying.
1921 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
1927 static inline struct page
*
1928 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
1929 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1930 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
1935 /* Acquire the OOM killer lock for the zones in zonelist */
1936 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
1937 schedule_timeout_uninterruptible(1);
1942 * Go through the zonelist yet one more time, keep very high watermark
1943 * here, this is only to catch a parallel oom killing, we must fail if
1944 * we're still under heavy pressure.
1946 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
1947 order
, zonelist
, high_zoneidx
,
1948 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
1949 preferred_zone
, migratetype
);
1953 if (!(gfp_mask
& __GFP_NOFAIL
)) {
1954 /* The OOM killer will not help higher order allocs */
1955 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
1957 /* The OOM killer does not needlessly kill tasks for lowmem */
1958 if (high_zoneidx
< ZONE_NORMAL
)
1961 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1962 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1963 * The caller should handle page allocation failure by itself if
1964 * it specifies __GFP_THISNODE.
1965 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1967 if (gfp_mask
& __GFP_THISNODE
)
1970 /* Exhausted what can be done so it's blamo time */
1971 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
1974 clear_zonelist_oom(zonelist
, gfp_mask
);
1978 #ifdef CONFIG_COMPACTION
1979 /* Try memory compaction for high-order allocations before reclaim */
1980 static struct page
*
1981 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
1982 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
1983 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
1984 int migratetype
, bool sync_migration
,
1985 bool *deferred_compaction
,
1986 unsigned long *did_some_progress
)
1993 if (compaction_deferred(preferred_zone
, order
)) {
1994 *deferred_compaction
= true;
1998 current
->flags
|= PF_MEMALLOC
;
1999 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2000 nodemask
, sync_migration
);
2001 current
->flags
&= ~PF_MEMALLOC
;
2002 if (*did_some_progress
!= COMPACT_SKIPPED
) {
2004 /* Page migration frees to the PCP lists but we want merging */
2005 drain_pages(get_cpu());
2008 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2009 order
, zonelist
, high_zoneidx
,
2010 alloc_flags
, preferred_zone
,
2013 preferred_zone
->compact_considered
= 0;
2014 preferred_zone
->compact_defer_shift
= 0;
2015 if (order
>= preferred_zone
->compact_order_failed
)
2016 preferred_zone
->compact_order_failed
= order
+ 1;
2017 count_vm_event(COMPACTSUCCESS
);
2022 * It's bad if compaction run occurs and fails.
2023 * The most likely reason is that pages exist,
2024 * but not enough to satisfy watermarks.
2026 count_vm_event(COMPACTFAIL
);
2029 * As async compaction considers a subset of pageblocks, only
2030 * defer if the failure was a sync compaction failure.
2033 defer_compaction(preferred_zone
, order
);
2041 static inline struct page
*
2042 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2043 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2044 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2045 int migratetype
, bool sync_migration
,
2046 bool *deferred_compaction
,
2047 unsigned long *did_some_progress
)
2051 #endif /* CONFIG_COMPACTION */
2053 /* The really slow allocator path where we enter direct reclaim */
2054 static inline struct page
*
2055 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2056 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2057 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2058 int migratetype
, unsigned long *did_some_progress
)
2060 struct page
*page
= NULL
;
2061 struct reclaim_state reclaim_state
;
2062 bool drained
= false;
2066 /* We now go into synchronous reclaim */
2067 cpuset_memory_pressure_bump();
2068 current
->flags
|= PF_MEMALLOC
;
2069 lockdep_set_current_reclaim_state(gfp_mask
);
2070 reclaim_state
.reclaimed_slab
= 0;
2071 current
->reclaim_state
= &reclaim_state
;
2073 *did_some_progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2075 current
->reclaim_state
= NULL
;
2076 lockdep_clear_current_reclaim_state();
2077 current
->flags
&= ~PF_MEMALLOC
;
2081 if (unlikely(!(*did_some_progress
)))
2084 /* After successful reclaim, reconsider all zones for allocation */
2086 zlc_clear_zones_full(zonelist
);
2089 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2090 zonelist
, high_zoneidx
,
2091 alloc_flags
, preferred_zone
,
2095 * If an allocation failed after direct reclaim, it could be because
2096 * pages are pinned on the per-cpu lists. Drain them and try again
2098 if (!page
&& !drained
) {
2108 * This is called in the allocator slow-path if the allocation request is of
2109 * sufficient urgency to ignore watermarks and take other desperate measures
2111 static inline struct page
*
2112 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2113 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2114 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2120 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2121 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2122 preferred_zone
, migratetype
);
2124 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2125 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2126 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2132 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
2133 enum zone_type high_zoneidx
,
2134 enum zone_type classzone_idx
)
2139 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
2140 wakeup_kswapd(zone
, order
, classzone_idx
);
2144 gfp_to_alloc_flags(gfp_t gfp_mask
)
2146 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2147 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2149 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2150 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2153 * The caller may dip into page reserves a bit more if the caller
2154 * cannot run direct reclaim, or if the caller has realtime scheduling
2155 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2156 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2158 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2162 * Not worth trying to allocate harder for
2163 * __GFP_NOMEMALLOC even if it can't schedule.
2165 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2166 alloc_flags
|= ALLOC_HARDER
;
2168 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2169 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2171 alloc_flags
&= ~ALLOC_CPUSET
;
2172 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2173 alloc_flags
|= ALLOC_HARDER
;
2175 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2176 if (!in_interrupt() &&
2177 ((current
->flags
& PF_MEMALLOC
) ||
2178 unlikely(test_thread_flag(TIF_MEMDIE
))))
2179 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2185 static inline struct page
*
2186 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2187 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2188 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2191 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2192 struct page
*page
= NULL
;
2194 unsigned long pages_reclaimed
= 0;
2195 unsigned long did_some_progress
;
2196 bool sync_migration
= false;
2197 bool deferred_compaction
= false;
2200 * In the slowpath, we sanity check order to avoid ever trying to
2201 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2202 * be using allocators in order of preference for an area that is
2205 if (order
>= MAX_ORDER
) {
2206 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2211 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2212 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2213 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2214 * using a larger set of nodes after it has established that the
2215 * allowed per node queues are empty and that nodes are
2218 if (NUMA_BUILD
&& (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2222 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2223 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2224 zone_idx(preferred_zone
));
2227 * OK, we're below the kswapd watermark and have kicked background
2228 * reclaim. Now things get more complex, so set up alloc_flags according
2229 * to how we want to proceed.
2231 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2234 * Find the true preferred zone if the allocation is unconstrained by
2237 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2238 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2242 /* This is the last chance, in general, before the goto nopage. */
2243 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2244 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2245 preferred_zone
, migratetype
);
2249 /* Allocate without watermarks if the context allows */
2250 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2251 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2252 zonelist
, high_zoneidx
, nodemask
,
2253 preferred_zone
, migratetype
);
2258 /* Atomic allocations - we can't balance anything */
2262 /* Avoid recursion of direct reclaim */
2263 if (current
->flags
& PF_MEMALLOC
)
2266 /* Avoid allocations with no watermarks from looping endlessly */
2267 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2271 * Try direct compaction. The first pass is asynchronous. Subsequent
2272 * attempts after direct reclaim are synchronous
2274 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2275 zonelist
, high_zoneidx
,
2277 alloc_flags
, preferred_zone
,
2278 migratetype
, sync_migration
,
2279 &deferred_compaction
,
2280 &did_some_progress
);
2283 sync_migration
= true;
2286 * If compaction is deferred for high-order allocations, it is because
2287 * sync compaction recently failed. In this is the case and the caller
2288 * has requested the system not be heavily disrupted, fail the
2289 * allocation now instead of entering direct reclaim
2291 if (deferred_compaction
&& (gfp_mask
& __GFP_NO_KSWAPD
))
2294 /* Try direct reclaim and then allocating */
2295 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2296 zonelist
, high_zoneidx
,
2298 alloc_flags
, preferred_zone
,
2299 migratetype
, &did_some_progress
);
2304 * If we failed to make any progress reclaiming, then we are
2305 * running out of options and have to consider going OOM
2307 if (!did_some_progress
) {
2308 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2309 if (oom_killer_disabled
)
2311 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2312 zonelist
, high_zoneidx
,
2313 nodemask
, preferred_zone
,
2318 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2320 * The oom killer is not called for high-order
2321 * allocations that may fail, so if no progress
2322 * is being made, there are no other options and
2323 * retrying is unlikely to help.
2325 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2328 * The oom killer is not called for lowmem
2329 * allocations to prevent needlessly killing
2332 if (high_zoneidx
< ZONE_NORMAL
)
2340 /* Check if we should retry the allocation */
2341 pages_reclaimed
+= did_some_progress
;
2342 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2344 /* Wait for some write requests to complete then retry */
2345 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2349 * High-order allocations do not necessarily loop after
2350 * direct reclaim and reclaim/compaction depends on compaction
2351 * being called after reclaim so call directly if necessary
2353 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2354 zonelist
, high_zoneidx
,
2356 alloc_flags
, preferred_zone
,
2357 migratetype
, sync_migration
,
2358 &deferred_compaction
,
2359 &did_some_progress
);
2365 warn_alloc_failed(gfp_mask
, order
, NULL
);
2368 if (kmemcheck_enabled
)
2369 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2375 * This is the 'heart' of the zoned buddy allocator.
2378 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2379 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2381 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2382 struct zone
*preferred_zone
;
2383 struct page
*page
= NULL
;
2384 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2385 unsigned int cpuset_mems_cookie
;
2387 gfp_mask
&= gfp_allowed_mask
;
2389 lockdep_trace_alloc(gfp_mask
);
2391 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2393 if (should_fail_alloc_page(gfp_mask
, order
))
2397 * Check the zones suitable for the gfp_mask contain at least one
2398 * valid zone. It's possible to have an empty zonelist as a result
2399 * of GFP_THISNODE and a memoryless node
2401 if (unlikely(!zonelist
->_zonerefs
->zone
))
2405 cpuset_mems_cookie
= get_mems_allowed();
2407 /* The preferred zone is used for statistics later */
2408 first_zones_zonelist(zonelist
, high_zoneidx
,
2409 nodemask
? : &cpuset_current_mems_allowed
,
2411 if (!preferred_zone
)
2414 /* First allocation attempt */
2415 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2416 zonelist
, high_zoneidx
, ALLOC_WMARK_LOW
|ALLOC_CPUSET
,
2417 preferred_zone
, migratetype
);
2418 if (unlikely(!page
))
2419 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2420 zonelist
, high_zoneidx
, nodemask
,
2421 preferred_zone
, migratetype
);
2423 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2427 * When updating a task's mems_allowed, it is possible to race with
2428 * parallel threads in such a way that an allocation can fail while
2429 * the mask is being updated. If a page allocation is about to fail,
2430 * check if the cpuset changed during allocation and if so, retry.
2432 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2437 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2440 * Common helper functions.
2442 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2447 * __get_free_pages() returns a 32-bit address, which cannot represent
2450 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2452 page
= alloc_pages(gfp_mask
, order
);
2455 return (unsigned long) page_address(page
);
2457 EXPORT_SYMBOL(__get_free_pages
);
2459 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2461 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2463 EXPORT_SYMBOL(get_zeroed_page
);
2465 void __free_pages(struct page
*page
, unsigned int order
)
2467 if (put_page_testzero(page
)) {
2469 free_hot_cold_page(page
, 0);
2471 __free_pages_ok(page
, order
);
2475 EXPORT_SYMBOL(__free_pages
);
2477 void free_pages(unsigned long addr
, unsigned int order
)
2480 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2481 __free_pages(virt_to_page((void *)addr
), order
);
2485 EXPORT_SYMBOL(free_pages
);
2487 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2490 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2491 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2493 split_page(virt_to_page((void *)addr
), order
);
2494 while (used
< alloc_end
) {
2499 return (void *)addr
;
2503 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2504 * @size: the number of bytes to allocate
2505 * @gfp_mask: GFP flags for the allocation
2507 * This function is similar to alloc_pages(), except that it allocates the
2508 * minimum number of pages to satisfy the request. alloc_pages() can only
2509 * allocate memory in power-of-two pages.
2511 * This function is also limited by MAX_ORDER.
2513 * Memory allocated by this function must be released by free_pages_exact().
2515 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2517 unsigned int order
= get_order(size
);
2520 addr
= __get_free_pages(gfp_mask
, order
);
2521 return make_alloc_exact(addr
, order
, size
);
2523 EXPORT_SYMBOL(alloc_pages_exact
);
2526 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2528 * @nid: the preferred node ID where memory should be allocated
2529 * @size: the number of bytes to allocate
2530 * @gfp_mask: GFP flags for the allocation
2532 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2534 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2537 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2539 unsigned order
= get_order(size
);
2540 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2543 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2545 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2548 * free_pages_exact - release memory allocated via alloc_pages_exact()
2549 * @virt: the value returned by alloc_pages_exact.
2550 * @size: size of allocation, same value as passed to alloc_pages_exact().
2552 * Release the memory allocated by a previous call to alloc_pages_exact.
2554 void free_pages_exact(void *virt
, size_t size
)
2556 unsigned long addr
= (unsigned long)virt
;
2557 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2559 while (addr
< end
) {
2564 EXPORT_SYMBOL(free_pages_exact
);
2566 static unsigned int nr_free_zone_pages(int offset
)
2571 /* Just pick one node, since fallback list is circular */
2572 unsigned int sum
= 0;
2574 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2576 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2577 unsigned long size
= zone
->present_pages
;
2578 unsigned long high
= high_wmark_pages(zone
);
2587 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2589 unsigned int nr_free_buffer_pages(void)
2591 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2593 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2596 * Amount of free RAM allocatable within all zones
2598 unsigned int nr_free_pagecache_pages(void)
2600 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2603 static inline void show_node(struct zone
*zone
)
2606 printk("Node %d ", zone_to_nid(zone
));
2609 void si_meminfo(struct sysinfo
*val
)
2611 val
->totalram
= totalram_pages
;
2613 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2614 val
->bufferram
= nr_blockdev_pages();
2615 val
->totalhigh
= totalhigh_pages
;
2616 val
->freehigh
= nr_free_highpages();
2617 val
->mem_unit
= PAGE_SIZE
;
2620 EXPORT_SYMBOL(si_meminfo
);
2623 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2625 pg_data_t
*pgdat
= NODE_DATA(nid
);
2627 val
->totalram
= pgdat
->node_present_pages
;
2628 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2629 #ifdef CONFIG_HIGHMEM
2630 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2631 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2637 val
->mem_unit
= PAGE_SIZE
;
2642 * Determine whether the node should be displayed or not, depending on whether
2643 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2645 bool skip_free_areas_node(unsigned int flags
, int nid
)
2648 unsigned int cpuset_mems_cookie
;
2650 if (!(flags
& SHOW_MEM_FILTER_NODES
))
2654 cpuset_mems_cookie
= get_mems_allowed();
2655 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
2656 } while (!put_mems_allowed(cpuset_mems_cookie
));
2661 #define K(x) ((x) << (PAGE_SHIFT-10))
2664 * Show free area list (used inside shift_scroll-lock stuff)
2665 * We also calculate the percentage fragmentation. We do this by counting the
2666 * memory on each free list with the exception of the first item on the list.
2667 * Suppresses nodes that are not allowed by current's cpuset if
2668 * SHOW_MEM_FILTER_NODES is passed.
2670 void show_free_areas(unsigned int filter
)
2675 for_each_populated_zone(zone
) {
2676 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2679 printk("%s per-cpu:\n", zone
->name
);
2681 for_each_online_cpu(cpu
) {
2682 struct per_cpu_pageset
*pageset
;
2684 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2686 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2687 cpu
, pageset
->pcp
.high
,
2688 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2692 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2693 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2695 " dirty:%lu writeback:%lu unstable:%lu\n"
2696 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2697 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2698 global_page_state(NR_ACTIVE_ANON
),
2699 global_page_state(NR_INACTIVE_ANON
),
2700 global_page_state(NR_ISOLATED_ANON
),
2701 global_page_state(NR_ACTIVE_FILE
),
2702 global_page_state(NR_INACTIVE_FILE
),
2703 global_page_state(NR_ISOLATED_FILE
),
2704 global_page_state(NR_UNEVICTABLE
),
2705 global_page_state(NR_FILE_DIRTY
),
2706 global_page_state(NR_WRITEBACK
),
2707 global_page_state(NR_UNSTABLE_NFS
),
2708 global_page_state(NR_FREE_PAGES
),
2709 global_page_state(NR_SLAB_RECLAIMABLE
),
2710 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2711 global_page_state(NR_FILE_MAPPED
),
2712 global_page_state(NR_SHMEM
),
2713 global_page_state(NR_PAGETABLE
),
2714 global_page_state(NR_BOUNCE
));
2716 for_each_populated_zone(zone
) {
2719 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2727 " active_anon:%lukB"
2728 " inactive_anon:%lukB"
2729 " active_file:%lukB"
2730 " inactive_file:%lukB"
2731 " unevictable:%lukB"
2732 " isolated(anon):%lukB"
2733 " isolated(file):%lukB"
2740 " slab_reclaimable:%lukB"
2741 " slab_unreclaimable:%lukB"
2742 " kernel_stack:%lukB"
2746 " writeback_tmp:%lukB"
2747 " pages_scanned:%lu"
2748 " all_unreclaimable? %s"
2751 K(zone_page_state(zone
, NR_FREE_PAGES
)),
2752 K(min_wmark_pages(zone
)),
2753 K(low_wmark_pages(zone
)),
2754 K(high_wmark_pages(zone
)),
2755 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
2756 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
2757 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
2758 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
2759 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
2760 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
2761 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
2762 K(zone
->present_pages
),
2763 K(zone_page_state(zone
, NR_MLOCK
)),
2764 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
2765 K(zone_page_state(zone
, NR_WRITEBACK
)),
2766 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
2767 K(zone_page_state(zone
, NR_SHMEM
)),
2768 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
2769 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
2770 zone_page_state(zone
, NR_KERNEL_STACK
) *
2772 K(zone_page_state(zone
, NR_PAGETABLE
)),
2773 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
2774 K(zone_page_state(zone
, NR_BOUNCE
)),
2775 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
2776 zone
->pages_scanned
,
2777 (zone
->all_unreclaimable
? "yes" : "no")
2779 printk("lowmem_reserve[]:");
2780 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
2781 printk(" %lu", zone
->lowmem_reserve
[i
]);
2785 for_each_populated_zone(zone
) {
2786 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
2788 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2791 printk("%s: ", zone
->name
);
2793 spin_lock_irqsave(&zone
->lock
, flags
);
2794 for (order
= 0; order
< MAX_ORDER
; order
++) {
2795 nr
[order
] = zone
->free_area
[order
].nr_free
;
2796 total
+= nr
[order
] << order
;
2798 spin_unlock_irqrestore(&zone
->lock
, flags
);
2799 for (order
= 0; order
< MAX_ORDER
; order
++)
2800 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
2801 printk("= %lukB\n", K(total
));
2804 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
2806 show_swap_cache_info();
2809 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
2811 zoneref
->zone
= zone
;
2812 zoneref
->zone_idx
= zone_idx(zone
);
2816 * Builds allocation fallback zone lists.
2818 * Add all populated zones of a node to the zonelist.
2820 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
2821 int nr_zones
, enum zone_type zone_type
)
2825 BUG_ON(zone_type
>= MAX_NR_ZONES
);
2830 zone
= pgdat
->node_zones
+ zone_type
;
2831 if (populated_zone(zone
)) {
2832 zoneref_set_zone(zone
,
2833 &zonelist
->_zonerefs
[nr_zones
++]);
2834 check_highest_zone(zone_type
);
2837 } while (zone_type
);
2844 * 0 = automatic detection of better ordering.
2845 * 1 = order by ([node] distance, -zonetype)
2846 * 2 = order by (-zonetype, [node] distance)
2848 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2849 * the same zonelist. So only NUMA can configure this param.
2851 #define ZONELIST_ORDER_DEFAULT 0
2852 #define ZONELIST_ORDER_NODE 1
2853 #define ZONELIST_ORDER_ZONE 2
2855 /* zonelist order in the kernel.
2856 * set_zonelist_order() will set this to NODE or ZONE.
2858 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2859 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
2863 /* The value user specified ....changed by config */
2864 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2865 /* string for sysctl */
2866 #define NUMA_ZONELIST_ORDER_LEN 16
2867 char numa_zonelist_order
[16] = "default";
2870 * interface for configure zonelist ordering.
2871 * command line option "numa_zonelist_order"
2872 * = "[dD]efault - default, automatic configuration.
2873 * = "[nN]ode - order by node locality, then by zone within node
2874 * = "[zZ]one - order by zone, then by locality within zone
2877 static int __parse_numa_zonelist_order(char *s
)
2879 if (*s
== 'd' || *s
== 'D') {
2880 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
2881 } else if (*s
== 'n' || *s
== 'N') {
2882 user_zonelist_order
= ZONELIST_ORDER_NODE
;
2883 } else if (*s
== 'z' || *s
== 'Z') {
2884 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
2887 "Ignoring invalid numa_zonelist_order value: "
2894 static __init
int setup_numa_zonelist_order(char *s
)
2901 ret
= __parse_numa_zonelist_order(s
);
2903 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
2907 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
2910 * sysctl handler for numa_zonelist_order
2912 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
2913 void __user
*buffer
, size_t *length
,
2916 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
2918 static DEFINE_MUTEX(zl_order_mutex
);
2920 mutex_lock(&zl_order_mutex
);
2922 strcpy(saved_string
, (char*)table
->data
);
2923 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
2927 int oldval
= user_zonelist_order
;
2928 if (__parse_numa_zonelist_order((char*)table
->data
)) {
2930 * bogus value. restore saved string
2932 strncpy((char*)table
->data
, saved_string
,
2933 NUMA_ZONELIST_ORDER_LEN
);
2934 user_zonelist_order
= oldval
;
2935 } else if (oldval
!= user_zonelist_order
) {
2936 mutex_lock(&zonelists_mutex
);
2937 build_all_zonelists(NULL
);
2938 mutex_unlock(&zonelists_mutex
);
2942 mutex_unlock(&zl_order_mutex
);
2947 #define MAX_NODE_LOAD (nr_online_nodes)
2948 static int node_load
[MAX_NUMNODES
];
2951 * find_next_best_node - find the next node that should appear in a given node's fallback list
2952 * @node: node whose fallback list we're appending
2953 * @used_node_mask: nodemask_t of already used nodes
2955 * We use a number of factors to determine which is the next node that should
2956 * appear on a given node's fallback list. The node should not have appeared
2957 * already in @node's fallback list, and it should be the next closest node
2958 * according to the distance array (which contains arbitrary distance values
2959 * from each node to each node in the system), and should also prefer nodes
2960 * with no CPUs, since presumably they'll have very little allocation pressure
2961 * on them otherwise.
2962 * It returns -1 if no node is found.
2964 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
2967 int min_val
= INT_MAX
;
2969 const struct cpumask
*tmp
= cpumask_of_node(0);
2971 /* Use the local node if we haven't already */
2972 if (!node_isset(node
, *used_node_mask
)) {
2973 node_set(node
, *used_node_mask
);
2977 for_each_node_state(n
, N_HIGH_MEMORY
) {
2979 /* Don't want a node to appear more than once */
2980 if (node_isset(n
, *used_node_mask
))
2983 /* Use the distance array to find the distance */
2984 val
= node_distance(node
, n
);
2986 /* Penalize nodes under us ("prefer the next node") */
2989 /* Give preference to headless and unused nodes */
2990 tmp
= cpumask_of_node(n
);
2991 if (!cpumask_empty(tmp
))
2992 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
2994 /* Slight preference for less loaded node */
2995 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
2996 val
+= node_load
[n
];
2998 if (val
< min_val
) {
3005 node_set(best_node
, *used_node_mask
);
3012 * Build zonelists ordered by node and zones within node.
3013 * This results in maximum locality--normal zone overflows into local
3014 * DMA zone, if any--but risks exhausting DMA zone.
3016 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3019 struct zonelist
*zonelist
;
3021 zonelist
= &pgdat
->node_zonelists
[0];
3022 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3024 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3026 zonelist
->_zonerefs
[j
].zone
= NULL
;
3027 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3031 * Build gfp_thisnode zonelists
3033 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3036 struct zonelist
*zonelist
;
3038 zonelist
= &pgdat
->node_zonelists
[1];
3039 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3040 zonelist
->_zonerefs
[j
].zone
= NULL
;
3041 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3045 * Build zonelists ordered by zone and nodes within zones.
3046 * This results in conserving DMA zone[s] until all Normal memory is
3047 * exhausted, but results in overflowing to remote node while memory
3048 * may still exist in local DMA zone.
3050 static int node_order
[MAX_NUMNODES
];
3052 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3055 int zone_type
; /* needs to be signed */
3057 struct zonelist
*zonelist
;
3059 zonelist
= &pgdat
->node_zonelists
[0];
3061 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3062 for (j
= 0; j
< nr_nodes
; j
++) {
3063 node
= node_order
[j
];
3064 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3065 if (populated_zone(z
)) {
3067 &zonelist
->_zonerefs
[pos
++]);
3068 check_highest_zone(zone_type
);
3072 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3073 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3076 static int default_zonelist_order(void)
3079 unsigned long low_kmem_size
,total_size
;
3083 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3084 * If they are really small and used heavily, the system can fall
3085 * into OOM very easily.
3086 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3088 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3091 for_each_online_node(nid
) {
3092 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3093 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3094 if (populated_zone(z
)) {
3095 if (zone_type
< ZONE_NORMAL
)
3096 low_kmem_size
+= z
->present_pages
;
3097 total_size
+= z
->present_pages
;
3098 } else if (zone_type
== ZONE_NORMAL
) {
3100 * If any node has only lowmem, then node order
3101 * is preferred to allow kernel allocations
3102 * locally; otherwise, they can easily infringe
3103 * on other nodes when there is an abundance of
3104 * lowmem available to allocate from.
3106 return ZONELIST_ORDER_NODE
;
3110 if (!low_kmem_size
|| /* there are no DMA area. */
3111 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
3112 return ZONELIST_ORDER_NODE
;
3114 * look into each node's config.
3115 * If there is a node whose DMA/DMA32 memory is very big area on
3116 * local memory, NODE_ORDER may be suitable.
3118 average_size
= total_size
/
3119 (nodes_weight(node_states
[N_HIGH_MEMORY
]) + 1);
3120 for_each_online_node(nid
) {
3123 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3124 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3125 if (populated_zone(z
)) {
3126 if (zone_type
< ZONE_NORMAL
)
3127 low_kmem_size
+= z
->present_pages
;
3128 total_size
+= z
->present_pages
;
3131 if (low_kmem_size
&&
3132 total_size
> average_size
&& /* ignore small node */
3133 low_kmem_size
> total_size
* 70/100)
3134 return ZONELIST_ORDER_NODE
;
3136 return ZONELIST_ORDER_ZONE
;
3139 static void set_zonelist_order(void)
3141 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3142 current_zonelist_order
= default_zonelist_order();
3144 current_zonelist_order
= user_zonelist_order
;
3147 static void build_zonelists(pg_data_t
*pgdat
)
3151 nodemask_t used_mask
;
3152 int local_node
, prev_node
;
3153 struct zonelist
*zonelist
;
3154 int order
= current_zonelist_order
;
3156 /* initialize zonelists */
3157 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3158 zonelist
= pgdat
->node_zonelists
+ i
;
3159 zonelist
->_zonerefs
[0].zone
= NULL
;
3160 zonelist
->_zonerefs
[0].zone_idx
= 0;
3163 /* NUMA-aware ordering of nodes */
3164 local_node
= pgdat
->node_id
;
3165 load
= nr_online_nodes
;
3166 prev_node
= local_node
;
3167 nodes_clear(used_mask
);
3169 memset(node_order
, 0, sizeof(node_order
));
3172 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3173 int distance
= node_distance(local_node
, node
);
3176 * If another node is sufficiently far away then it is better
3177 * to reclaim pages in a zone before going off node.
3179 if (distance
> RECLAIM_DISTANCE
)
3180 zone_reclaim_mode
= 1;
3183 * We don't want to pressure a particular node.
3184 * So adding penalty to the first node in same
3185 * distance group to make it round-robin.
3187 if (distance
!= node_distance(local_node
, prev_node
))
3188 node_load
[node
] = load
;
3192 if (order
== ZONELIST_ORDER_NODE
)
3193 build_zonelists_in_node_order(pgdat
, node
);
3195 node_order
[j
++] = node
; /* remember order */
3198 if (order
== ZONELIST_ORDER_ZONE
) {
3199 /* calculate node order -- i.e., DMA last! */
3200 build_zonelists_in_zone_order(pgdat
, j
);
3203 build_thisnode_zonelists(pgdat
);
3206 /* Construct the zonelist performance cache - see further mmzone.h */
3207 static void build_zonelist_cache(pg_data_t
*pgdat
)
3209 struct zonelist
*zonelist
;
3210 struct zonelist_cache
*zlc
;
3213 zonelist
= &pgdat
->node_zonelists
[0];
3214 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3215 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3216 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3217 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3220 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3222 * Return node id of node used for "local" allocations.
3223 * I.e., first node id of first zone in arg node's generic zonelist.
3224 * Used for initializing percpu 'numa_mem', which is used primarily
3225 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3227 int local_memory_node(int node
)
3231 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3232 gfp_zone(GFP_KERNEL
),
3239 #else /* CONFIG_NUMA */
3241 static void set_zonelist_order(void)
3243 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3246 static void build_zonelists(pg_data_t
*pgdat
)
3248 int node
, local_node
;
3250 struct zonelist
*zonelist
;
3252 local_node
= pgdat
->node_id
;
3254 zonelist
= &pgdat
->node_zonelists
[0];
3255 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3258 * Now we build the zonelist so that it contains the zones
3259 * of all the other nodes.
3260 * We don't want to pressure a particular node, so when
3261 * building the zones for node N, we make sure that the
3262 * zones coming right after the local ones are those from
3263 * node N+1 (modulo N)
3265 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3266 if (!node_online(node
))
3268 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3271 for (node
= 0; node
< local_node
; node
++) {
3272 if (!node_online(node
))
3274 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3278 zonelist
->_zonerefs
[j
].zone
= NULL
;
3279 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3282 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3283 static void build_zonelist_cache(pg_data_t
*pgdat
)
3285 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3288 #endif /* CONFIG_NUMA */
3291 * Boot pageset table. One per cpu which is going to be used for all
3292 * zones and all nodes. The parameters will be set in such a way
3293 * that an item put on a list will immediately be handed over to
3294 * the buddy list. This is safe since pageset manipulation is done
3295 * with interrupts disabled.
3297 * The boot_pagesets must be kept even after bootup is complete for
3298 * unused processors and/or zones. They do play a role for bootstrapping
3299 * hotplugged processors.
3301 * zoneinfo_show() and maybe other functions do
3302 * not check if the processor is online before following the pageset pointer.
3303 * Other parts of the kernel may not check if the zone is available.
3305 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3306 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3307 static void setup_zone_pageset(struct zone
*zone
);
3310 * Global mutex to protect against size modification of zonelists
3311 * as well as to serialize pageset setup for the new populated zone.
3313 DEFINE_MUTEX(zonelists_mutex
);
3315 /* return values int ....just for stop_machine() */
3316 static __init_refok
int __build_all_zonelists(void *data
)
3322 memset(node_load
, 0, sizeof(node_load
));
3324 for_each_online_node(nid
) {
3325 pg_data_t
*pgdat
= NODE_DATA(nid
);
3327 build_zonelists(pgdat
);
3328 build_zonelist_cache(pgdat
);
3332 * Initialize the boot_pagesets that are going to be used
3333 * for bootstrapping processors. The real pagesets for
3334 * each zone will be allocated later when the per cpu
3335 * allocator is available.
3337 * boot_pagesets are used also for bootstrapping offline
3338 * cpus if the system is already booted because the pagesets
3339 * are needed to initialize allocators on a specific cpu too.
3340 * F.e. the percpu allocator needs the page allocator which
3341 * needs the percpu allocator in order to allocate its pagesets
3342 * (a chicken-egg dilemma).
3344 for_each_possible_cpu(cpu
) {
3345 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3347 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3349 * We now know the "local memory node" for each node--
3350 * i.e., the node of the first zone in the generic zonelist.
3351 * Set up numa_mem percpu variable for on-line cpus. During
3352 * boot, only the boot cpu should be on-line; we'll init the
3353 * secondary cpus' numa_mem as they come on-line. During
3354 * node/memory hotplug, we'll fixup all on-line cpus.
3356 if (cpu_online(cpu
))
3357 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3365 * Called with zonelists_mutex held always
3366 * unless system_state == SYSTEM_BOOTING.
3368 void __ref
build_all_zonelists(void *data
)
3370 set_zonelist_order();
3372 if (system_state
== SYSTEM_BOOTING
) {
3373 __build_all_zonelists(NULL
);
3374 mminit_verify_zonelist();
3375 cpuset_init_current_mems_allowed();
3377 /* we have to stop all cpus to guarantee there is no user
3379 #ifdef CONFIG_MEMORY_HOTPLUG
3381 setup_zone_pageset((struct zone
*)data
);
3383 stop_machine(__build_all_zonelists
, NULL
, NULL
);
3384 /* cpuset refresh routine should be here */
3386 vm_total_pages
= nr_free_pagecache_pages();
3388 * Disable grouping by mobility if the number of pages in the
3389 * system is too low to allow the mechanism to work. It would be
3390 * more accurate, but expensive to check per-zone. This check is
3391 * made on memory-hotadd so a system can start with mobility
3392 * disabled and enable it later
3394 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3395 page_group_by_mobility_disabled
= 1;
3397 page_group_by_mobility_disabled
= 0;
3399 printk("Built %i zonelists in %s order, mobility grouping %s. "
3400 "Total pages: %ld\n",
3402 zonelist_order_name
[current_zonelist_order
],
3403 page_group_by_mobility_disabled
? "off" : "on",
3406 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3411 * Helper functions to size the waitqueue hash table.
3412 * Essentially these want to choose hash table sizes sufficiently
3413 * large so that collisions trying to wait on pages are rare.
3414 * But in fact, the number of active page waitqueues on typical
3415 * systems is ridiculously low, less than 200. So this is even
3416 * conservative, even though it seems large.
3418 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3419 * waitqueues, i.e. the size of the waitq table given the number of pages.
3421 #define PAGES_PER_WAITQUEUE 256
3423 #ifndef CONFIG_MEMORY_HOTPLUG
3424 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3426 unsigned long size
= 1;
3428 pages
/= PAGES_PER_WAITQUEUE
;
3430 while (size
< pages
)
3434 * Once we have dozens or even hundreds of threads sleeping
3435 * on IO we've got bigger problems than wait queue collision.
3436 * Limit the size of the wait table to a reasonable size.
3438 size
= min(size
, 4096UL);
3440 return max(size
, 4UL);
3444 * A zone's size might be changed by hot-add, so it is not possible to determine
3445 * a suitable size for its wait_table. So we use the maximum size now.
3447 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3449 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3450 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3451 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3453 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3454 * or more by the traditional way. (See above). It equals:
3456 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3457 * ia64(16K page size) : = ( 8G + 4M)byte.
3458 * powerpc (64K page size) : = (32G +16M)byte.
3460 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3467 * This is an integer logarithm so that shifts can be used later
3468 * to extract the more random high bits from the multiplicative
3469 * hash function before the remainder is taken.
3471 static inline unsigned long wait_table_bits(unsigned long size
)
3476 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3479 * Check if a pageblock contains reserved pages
3481 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3485 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3486 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3493 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3494 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3495 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3496 * higher will lead to a bigger reserve which will get freed as contiguous
3497 * blocks as reclaim kicks in
3499 static void setup_zone_migrate_reserve(struct zone
*zone
)
3501 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3503 unsigned long block_migratetype
;
3507 * Get the start pfn, end pfn and the number of blocks to reserve
3508 * We have to be careful to be aligned to pageblock_nr_pages to
3509 * make sure that we always check pfn_valid for the first page in
3512 start_pfn
= zone
->zone_start_pfn
;
3513 end_pfn
= start_pfn
+ zone
->spanned_pages
;
3514 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
3515 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3519 * Reserve blocks are generally in place to help high-order atomic
3520 * allocations that are short-lived. A min_free_kbytes value that
3521 * would result in more than 2 reserve blocks for atomic allocations
3522 * is assumed to be in place to help anti-fragmentation for the
3523 * future allocation of hugepages at runtime.
3525 reserve
= min(2, reserve
);
3527 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3528 if (!pfn_valid(pfn
))
3530 page
= pfn_to_page(pfn
);
3532 /* Watch out for overlapping nodes */
3533 if (page_to_nid(page
) != zone_to_nid(zone
))
3536 block_migratetype
= get_pageblock_migratetype(page
);
3538 /* Only test what is necessary when the reserves are not met */
3541 * Blocks with reserved pages will never free, skip
3544 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
3545 if (pageblock_is_reserved(pfn
, block_end_pfn
))
3548 /* If this block is reserved, account for it */
3549 if (block_migratetype
== MIGRATE_RESERVE
) {
3554 /* Suitable for reserving if this block is movable */
3555 if (block_migratetype
== MIGRATE_MOVABLE
) {
3556 set_pageblock_migratetype(page
,
3558 move_freepages_block(zone
, page
,
3566 * If the reserve is met and this is a previous reserved block,
3569 if (block_migratetype
== MIGRATE_RESERVE
) {
3570 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3571 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3577 * Initially all pages are reserved - free ones are freed
3578 * up by free_all_bootmem() once the early boot process is
3579 * done. Non-atomic initialization, single-pass.
3581 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3582 unsigned long start_pfn
, enum memmap_context context
)
3585 unsigned long end_pfn
= start_pfn
+ size
;
3589 if (highest_memmap_pfn
< end_pfn
- 1)
3590 highest_memmap_pfn
= end_pfn
- 1;
3592 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3593 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3595 * There can be holes in boot-time mem_map[]s
3596 * handed to this function. They do not
3597 * exist on hotplugged memory.
3599 if (context
== MEMMAP_EARLY
) {
3600 if (!early_pfn_valid(pfn
))
3602 if (!early_pfn_in_nid(pfn
, nid
))
3605 page
= pfn_to_page(pfn
);
3606 set_page_links(page
, zone
, nid
, pfn
);
3607 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3608 init_page_count(page
);
3609 reset_page_mapcount(page
);
3610 SetPageReserved(page
);
3612 * Mark the block movable so that blocks are reserved for
3613 * movable at startup. This will force kernel allocations
3614 * to reserve their blocks rather than leaking throughout
3615 * the address space during boot when many long-lived
3616 * kernel allocations are made. Later some blocks near
3617 * the start are marked MIGRATE_RESERVE by
3618 * setup_zone_migrate_reserve()
3620 * bitmap is created for zone's valid pfn range. but memmap
3621 * can be created for invalid pages (for alignment)
3622 * check here not to call set_pageblock_migratetype() against
3625 if ((z
->zone_start_pfn
<= pfn
)
3626 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
3627 && !(pfn
& (pageblock_nr_pages
- 1)))
3628 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3630 INIT_LIST_HEAD(&page
->lru
);
3631 #ifdef WANT_PAGE_VIRTUAL
3632 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3633 if (!is_highmem_idx(zone
))
3634 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3639 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3642 for_each_migratetype_order(order
, t
) {
3643 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3644 zone
->free_area
[order
].nr_free
= 0;
3648 #ifndef __HAVE_ARCH_MEMMAP_INIT
3649 #define memmap_init(size, nid, zone, start_pfn) \
3650 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3653 static int zone_batchsize(struct zone
*zone
)
3659 * The per-cpu-pages pools are set to around 1000th of the
3660 * size of the zone. But no more than 1/2 of a meg.
3662 * OK, so we don't know how big the cache is. So guess.
3664 batch
= zone
->present_pages
/ 1024;
3665 if (batch
* PAGE_SIZE
> 512 * 1024)
3666 batch
= (512 * 1024) / PAGE_SIZE
;
3667 batch
/= 4; /* We effectively *= 4 below */
3672 * Clamp the batch to a 2^n - 1 value. Having a power
3673 * of 2 value was found to be more likely to have
3674 * suboptimal cache aliasing properties in some cases.
3676 * For example if 2 tasks are alternately allocating
3677 * batches of pages, one task can end up with a lot
3678 * of pages of one half of the possible page colors
3679 * and the other with pages of the other colors.
3681 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3686 /* The deferral and batching of frees should be suppressed under NOMMU
3689 * The problem is that NOMMU needs to be able to allocate large chunks
3690 * of contiguous memory as there's no hardware page translation to
3691 * assemble apparent contiguous memory from discontiguous pages.
3693 * Queueing large contiguous runs of pages for batching, however,
3694 * causes the pages to actually be freed in smaller chunks. As there
3695 * can be a significant delay between the individual batches being
3696 * recycled, this leads to the once large chunks of space being
3697 * fragmented and becoming unavailable for high-order allocations.
3703 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3705 struct per_cpu_pages
*pcp
;
3708 memset(p
, 0, sizeof(*p
));
3712 pcp
->high
= 6 * batch
;
3713 pcp
->batch
= max(1UL, 1 * batch
);
3714 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3715 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3719 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3720 * to the value high for the pageset p.
3723 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
3726 struct per_cpu_pages
*pcp
;
3730 pcp
->batch
= max(1UL, high
/4);
3731 if ((high
/4) > (PAGE_SHIFT
* 8))
3732 pcp
->batch
= PAGE_SHIFT
* 8;
3735 static void setup_zone_pageset(struct zone
*zone
)
3739 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
3741 for_each_possible_cpu(cpu
) {
3742 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
3744 setup_pageset(pcp
, zone_batchsize(zone
));
3746 if (percpu_pagelist_fraction
)
3747 setup_pagelist_highmark(pcp
,
3748 (zone
->present_pages
/
3749 percpu_pagelist_fraction
));
3754 * Allocate per cpu pagesets and initialize them.
3755 * Before this call only boot pagesets were available.
3757 void __init
setup_per_cpu_pageset(void)
3761 for_each_populated_zone(zone
)
3762 setup_zone_pageset(zone
);
3765 static noinline __init_refok
3766 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
3769 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3773 * The per-page waitqueue mechanism uses hashed waitqueues
3776 zone
->wait_table_hash_nr_entries
=
3777 wait_table_hash_nr_entries(zone_size_pages
);
3778 zone
->wait_table_bits
=
3779 wait_table_bits(zone
->wait_table_hash_nr_entries
);
3780 alloc_size
= zone
->wait_table_hash_nr_entries
3781 * sizeof(wait_queue_head_t
);
3783 if (!slab_is_available()) {
3784 zone
->wait_table
= (wait_queue_head_t
*)
3785 alloc_bootmem_node_nopanic(pgdat
, alloc_size
);
3788 * This case means that a zone whose size was 0 gets new memory
3789 * via memory hot-add.
3790 * But it may be the case that a new node was hot-added. In
3791 * this case vmalloc() will not be able to use this new node's
3792 * memory - this wait_table must be initialized to use this new
3793 * node itself as well.
3794 * To use this new node's memory, further consideration will be
3797 zone
->wait_table
= vmalloc(alloc_size
);
3799 if (!zone
->wait_table
)
3802 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
3803 init_waitqueue_head(zone
->wait_table
+ i
);
3808 static int __zone_pcp_update(void *data
)
3810 struct zone
*zone
= data
;
3812 unsigned long batch
= zone_batchsize(zone
), flags
;
3814 for_each_possible_cpu(cpu
) {
3815 struct per_cpu_pageset
*pset
;
3816 struct per_cpu_pages
*pcp
;
3818 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
3821 local_irq_save(flags
);
3822 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
3823 setup_pageset(pset
, batch
);
3824 local_irq_restore(flags
);
3829 void zone_pcp_update(struct zone
*zone
)
3831 stop_machine(__zone_pcp_update
, zone
, NULL
);
3834 static __meminit
void zone_pcp_init(struct zone
*zone
)
3837 * per cpu subsystem is not up at this point. The following code
3838 * relies on the ability of the linker to provide the
3839 * offset of a (static) per cpu variable into the per cpu area.
3841 zone
->pageset
= &boot_pageset
;
3843 if (zone
->present_pages
)
3844 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
3845 zone
->name
, zone
->present_pages
,
3846 zone_batchsize(zone
));
3849 __meminit
int init_currently_empty_zone(struct zone
*zone
,
3850 unsigned long zone_start_pfn
,
3852 enum memmap_context context
)
3854 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
3856 ret
= zone_wait_table_init(zone
, size
);
3859 pgdat
->nr_zones
= zone_idx(zone
) + 1;
3861 zone
->zone_start_pfn
= zone_start_pfn
;
3863 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
3864 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3866 (unsigned long)zone_idx(zone
),
3867 zone_start_pfn
, (zone_start_pfn
+ size
));
3869 zone_init_free_lists(zone
);
3874 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
3875 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3877 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3878 * Architectures may implement their own version but if add_active_range()
3879 * was used and there are no special requirements, this is a convenient
3882 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
3884 unsigned long start_pfn
, end_pfn
;
3887 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
3888 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
3890 /* This is a memory hole */
3893 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3895 int __meminit
early_pfn_to_nid(unsigned long pfn
)
3899 nid
= __early_pfn_to_nid(pfn
);
3902 /* just returns 0 */
3906 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
3907 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
3911 nid
= __early_pfn_to_nid(pfn
);
3912 if (nid
>= 0 && nid
!= node
)
3919 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3920 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3921 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3923 * If an architecture guarantees that all ranges registered with
3924 * add_active_ranges() contain no holes and may be freed, this
3925 * this function may be used instead of calling free_bootmem() manually.
3927 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
3929 unsigned long start_pfn
, end_pfn
;
3932 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
3933 start_pfn
= min(start_pfn
, max_low_pfn
);
3934 end_pfn
= min(end_pfn
, max_low_pfn
);
3936 if (start_pfn
< end_pfn
)
3937 free_bootmem_node(NODE_DATA(this_nid
),
3938 PFN_PHYS(start_pfn
),
3939 (end_pfn
- start_pfn
) << PAGE_SHIFT
);
3943 int __init
add_from_early_node_map(struct range
*range
, int az
,
3944 int nr_range
, int nid
)
3946 unsigned long start_pfn
, end_pfn
;
3949 /* need to go over early_node_map to find out good range for node */
3950 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
)
3951 nr_range
= add_range(range
, az
, nr_range
, start_pfn
, end_pfn
);
3956 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3957 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3959 * If an architecture guarantees that all ranges registered with
3960 * add_active_ranges() contain no holes and may be freed, this
3961 * function may be used instead of calling memory_present() manually.
3963 void __init
sparse_memory_present_with_active_regions(int nid
)
3965 unsigned long start_pfn
, end_pfn
;
3968 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
3969 memory_present(this_nid
, start_pfn
, end_pfn
);
3973 * get_pfn_range_for_nid - Return the start and end page frames for a node
3974 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3975 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3976 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3978 * It returns the start and end page frame of a node based on information
3979 * provided by an arch calling add_active_range(). If called for a node
3980 * with no available memory, a warning is printed and the start and end
3983 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
3984 unsigned long *start_pfn
, unsigned long *end_pfn
)
3986 unsigned long this_start_pfn
, this_end_pfn
;
3992 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
3993 *start_pfn
= min(*start_pfn
, this_start_pfn
);
3994 *end_pfn
= max(*end_pfn
, this_end_pfn
);
3997 if (*start_pfn
== -1UL)
4002 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4003 * assumption is made that zones within a node are ordered in monotonic
4004 * increasing memory addresses so that the "highest" populated zone is used
4006 static void __init
find_usable_zone_for_movable(void)
4009 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4010 if (zone_index
== ZONE_MOVABLE
)
4013 if (arch_zone_highest_possible_pfn
[zone_index
] >
4014 arch_zone_lowest_possible_pfn
[zone_index
])
4018 VM_BUG_ON(zone_index
== -1);
4019 movable_zone
= zone_index
;
4023 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4024 * because it is sized independent of architecture. Unlike the other zones,
4025 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4026 * in each node depending on the size of each node and how evenly kernelcore
4027 * is distributed. This helper function adjusts the zone ranges
4028 * provided by the architecture for a given node by using the end of the
4029 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4030 * zones within a node are in order of monotonic increases memory addresses
4032 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4033 unsigned long zone_type
,
4034 unsigned long node_start_pfn
,
4035 unsigned long node_end_pfn
,
4036 unsigned long *zone_start_pfn
,
4037 unsigned long *zone_end_pfn
)
4039 /* Only adjust if ZONE_MOVABLE is on this node */
4040 if (zone_movable_pfn
[nid
]) {
4041 /* Size ZONE_MOVABLE */
4042 if (zone_type
== ZONE_MOVABLE
) {
4043 *zone_start_pfn
= zone_movable_pfn
[nid
];
4044 *zone_end_pfn
= min(node_end_pfn
,
4045 arch_zone_highest_possible_pfn
[movable_zone
]);
4047 /* Adjust for ZONE_MOVABLE starting within this range */
4048 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4049 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4050 *zone_end_pfn
= zone_movable_pfn
[nid
];
4052 /* Check if this whole range is within ZONE_MOVABLE */
4053 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4054 *zone_start_pfn
= *zone_end_pfn
;
4059 * Return the number of pages a zone spans in a node, including holes
4060 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4062 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4063 unsigned long zone_type
,
4064 unsigned long *ignored
)
4066 unsigned long node_start_pfn
, node_end_pfn
;
4067 unsigned long zone_start_pfn
, zone_end_pfn
;
4069 /* Get the start and end of the node and zone */
4070 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4071 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4072 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4073 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4074 node_start_pfn
, node_end_pfn
,
4075 &zone_start_pfn
, &zone_end_pfn
);
4077 /* Check that this node has pages within the zone's required range */
4078 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4081 /* Move the zone boundaries inside the node if necessary */
4082 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4083 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4085 /* Return the spanned pages */
4086 return zone_end_pfn
- zone_start_pfn
;
4090 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4091 * then all holes in the requested range will be accounted for.
4093 unsigned long __meminit
__absent_pages_in_range(int nid
,
4094 unsigned long range_start_pfn
,
4095 unsigned long range_end_pfn
)
4097 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4098 unsigned long start_pfn
, end_pfn
;
4101 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4102 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4103 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4104 nr_absent
-= end_pfn
- start_pfn
;
4110 * absent_pages_in_range - Return number of page frames in holes within a range
4111 * @start_pfn: The start PFN to start searching for holes
4112 * @end_pfn: The end PFN to stop searching for holes
4114 * It returns the number of pages frames in memory holes within a range.
4116 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4117 unsigned long end_pfn
)
4119 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4122 /* Return the number of page frames in holes in a zone on a node */
4123 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4124 unsigned long zone_type
,
4125 unsigned long *ignored
)
4127 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4128 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4129 unsigned long node_start_pfn
, node_end_pfn
;
4130 unsigned long zone_start_pfn
, zone_end_pfn
;
4132 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4133 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4134 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4136 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4137 node_start_pfn
, node_end_pfn
,
4138 &zone_start_pfn
, &zone_end_pfn
);
4139 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4142 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4143 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4144 unsigned long zone_type
,
4145 unsigned long *zones_size
)
4147 return zones_size
[zone_type
];
4150 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4151 unsigned long zone_type
,
4152 unsigned long *zholes_size
)
4157 return zholes_size
[zone_type
];
4160 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4162 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4163 unsigned long *zones_size
, unsigned long *zholes_size
)
4165 unsigned long realtotalpages
, totalpages
= 0;
4168 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4169 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4171 pgdat
->node_spanned_pages
= totalpages
;
4173 realtotalpages
= totalpages
;
4174 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4176 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4178 pgdat
->node_present_pages
= realtotalpages
;
4179 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4183 #ifndef CONFIG_SPARSEMEM
4185 * Calculate the size of the zone->blockflags rounded to an unsigned long
4186 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4187 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4188 * round what is now in bits to nearest long in bits, then return it in
4191 static unsigned long __init
usemap_size(unsigned long zonesize
)
4193 unsigned long usemapsize
;
4195 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4196 usemapsize
= usemapsize
>> pageblock_order
;
4197 usemapsize
*= NR_PAGEBLOCK_BITS
;
4198 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4200 return usemapsize
/ 8;
4203 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4204 struct zone
*zone
, unsigned long zonesize
)
4206 unsigned long usemapsize
= usemap_size(zonesize
);
4207 zone
->pageblock_flags
= NULL
;
4209 zone
->pageblock_flags
= alloc_bootmem_node_nopanic(pgdat
,
4213 static inline void setup_usemap(struct pglist_data
*pgdat
,
4214 struct zone
*zone
, unsigned long zonesize
) {}
4215 #endif /* CONFIG_SPARSEMEM */
4217 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4219 /* Return a sensible default order for the pageblock size. */
4220 static inline int pageblock_default_order(void)
4222 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4223 return HUGETLB_PAGE_ORDER
;
4228 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4229 static inline void __init
set_pageblock_order(unsigned int order
)
4231 /* Check that pageblock_nr_pages has not already been setup */
4232 if (pageblock_order
)
4236 * Assume the largest contiguous order of interest is a huge page.
4237 * This value may be variable depending on boot parameters on IA64
4239 pageblock_order
= order
;
4241 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4244 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4245 * and pageblock_default_order() are unused as pageblock_order is set
4246 * at compile-time. See include/linux/pageblock-flags.h for the values of
4247 * pageblock_order based on the kernel config
4249 static inline int pageblock_default_order(unsigned int order
)
4253 #define set_pageblock_order(x) do {} while (0)
4255 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4258 * Set up the zone data structures:
4259 * - mark all pages reserved
4260 * - mark all memory queues empty
4261 * - clear the memory bitmaps
4263 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4264 unsigned long *zones_size
, unsigned long *zholes_size
)
4267 int nid
= pgdat
->node_id
;
4268 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4271 pgdat_resize_init(pgdat
);
4272 pgdat
->nr_zones
= 0;
4273 init_waitqueue_head(&pgdat
->kswapd_wait
);
4274 pgdat
->kswapd_max_order
= 0;
4275 pgdat_page_cgroup_init(pgdat
);
4277 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4278 struct zone
*zone
= pgdat
->node_zones
+ j
;
4279 unsigned long size
, realsize
, memmap_pages
;
4282 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4283 realsize
= size
- zone_absent_pages_in_node(nid
, j
,
4287 * Adjust realsize so that it accounts for how much memory
4288 * is used by this zone for memmap. This affects the watermark
4289 * and per-cpu initialisations
4292 PAGE_ALIGN(size
* sizeof(struct page
)) >> PAGE_SHIFT
;
4293 if (realsize
>= memmap_pages
) {
4294 realsize
-= memmap_pages
;
4297 " %s zone: %lu pages used for memmap\n",
4298 zone_names
[j
], memmap_pages
);
4301 " %s zone: %lu pages exceeds realsize %lu\n",
4302 zone_names
[j
], memmap_pages
, realsize
);
4304 /* Account for reserved pages */
4305 if (j
== 0 && realsize
> dma_reserve
) {
4306 realsize
-= dma_reserve
;
4307 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4308 zone_names
[0], dma_reserve
);
4311 if (!is_highmem_idx(j
))
4312 nr_kernel_pages
+= realsize
;
4313 nr_all_pages
+= realsize
;
4315 zone
->spanned_pages
= size
;
4316 zone
->present_pages
= realsize
;
4319 zone
->min_unmapped_pages
= (realsize
*sysctl_min_unmapped_ratio
)
4321 zone
->min_slab_pages
= (realsize
* sysctl_min_slab_ratio
) / 100;
4323 zone
->name
= zone_names
[j
];
4324 spin_lock_init(&zone
->lock
);
4325 spin_lock_init(&zone
->lru_lock
);
4326 zone_seqlock_init(zone
);
4327 zone
->zone_pgdat
= pgdat
;
4329 zone_pcp_init(zone
);
4331 INIT_LIST_HEAD(&zone
->lruvec
.lists
[lru
]);
4332 zone
->reclaim_stat
.recent_rotated
[0] = 0;
4333 zone
->reclaim_stat
.recent_rotated
[1] = 0;
4334 zone
->reclaim_stat
.recent_scanned
[0] = 0;
4335 zone
->reclaim_stat
.recent_scanned
[1] = 0;
4336 zap_zone_vm_stats(zone
);
4341 set_pageblock_order(pageblock_default_order());
4342 setup_usemap(pgdat
, zone
, size
);
4343 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4344 size
, MEMMAP_EARLY
);
4346 memmap_init(size
, nid
, j
, zone_start_pfn
);
4347 zone_start_pfn
+= size
;
4351 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4353 /* Skip empty nodes */
4354 if (!pgdat
->node_spanned_pages
)
4357 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4358 /* ia64 gets its own node_mem_map, before this, without bootmem */
4359 if (!pgdat
->node_mem_map
) {
4360 unsigned long size
, start
, end
;
4364 * The zone's endpoints aren't required to be MAX_ORDER
4365 * aligned but the node_mem_map endpoints must be in order
4366 * for the buddy allocator to function correctly.
4368 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4369 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
4370 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4371 size
= (end
- start
) * sizeof(struct page
);
4372 map
= alloc_remap(pgdat
->node_id
, size
);
4374 map
= alloc_bootmem_node_nopanic(pgdat
, size
);
4375 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4377 #ifndef CONFIG_NEED_MULTIPLE_NODES
4379 * With no DISCONTIG, the global mem_map is just set as node 0's
4381 if (pgdat
== NODE_DATA(0)) {
4382 mem_map
= NODE_DATA(0)->node_mem_map
;
4383 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4384 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4385 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4386 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4389 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4392 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4393 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4395 pg_data_t
*pgdat
= NODE_DATA(nid
);
4397 pgdat
->node_id
= nid
;
4398 pgdat
->node_start_pfn
= node_start_pfn
;
4399 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4401 alloc_node_mem_map(pgdat
);
4402 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4403 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4404 nid
, (unsigned long)pgdat
,
4405 (unsigned long)pgdat
->node_mem_map
);
4408 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4411 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4413 #if MAX_NUMNODES > 1
4415 * Figure out the number of possible node ids.
4417 static void __init
setup_nr_node_ids(void)
4420 unsigned int highest
= 0;
4422 for_each_node_mask(node
, node_possible_map
)
4424 nr_node_ids
= highest
+ 1;
4427 static inline void setup_nr_node_ids(void)
4433 * node_map_pfn_alignment - determine the maximum internode alignment
4435 * This function should be called after node map is populated and sorted.
4436 * It calculates the maximum power of two alignment which can distinguish
4439 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4440 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4441 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4442 * shifted, 1GiB is enough and this function will indicate so.
4444 * This is used to test whether pfn -> nid mapping of the chosen memory
4445 * model has fine enough granularity to avoid incorrect mapping for the
4446 * populated node map.
4448 * Returns the determined alignment in pfn's. 0 if there is no alignment
4449 * requirement (single node).
4451 unsigned long __init
node_map_pfn_alignment(void)
4453 unsigned long accl_mask
= 0, last_end
= 0;
4454 unsigned long start
, end
, mask
;
4458 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
4459 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
4466 * Start with a mask granular enough to pin-point to the
4467 * start pfn and tick off bits one-by-one until it becomes
4468 * too coarse to separate the current node from the last.
4470 mask
= ~((1 << __ffs(start
)) - 1);
4471 while (mask
&& last_end
<= (start
& (mask
<< 1)))
4474 /* accumulate all internode masks */
4478 /* convert mask to number of pages */
4479 return ~accl_mask
+ 1;
4482 /* Find the lowest pfn for a node */
4483 static unsigned long __init
find_min_pfn_for_node(int nid
)
4485 unsigned long min_pfn
= ULONG_MAX
;
4486 unsigned long start_pfn
;
4489 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
4490 min_pfn
= min(min_pfn
, start_pfn
);
4492 if (min_pfn
== ULONG_MAX
) {
4494 "Could not find start_pfn for node %d\n", nid
);
4502 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4504 * It returns the minimum PFN based on information provided via
4505 * add_active_range().
4507 unsigned long __init
find_min_pfn_with_active_regions(void)
4509 return find_min_pfn_for_node(MAX_NUMNODES
);
4513 * early_calculate_totalpages()
4514 * Sum pages in active regions for movable zone.
4515 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4517 static unsigned long __init
early_calculate_totalpages(void)
4519 unsigned long totalpages
= 0;
4520 unsigned long start_pfn
, end_pfn
;
4523 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4524 unsigned long pages
= end_pfn
- start_pfn
;
4526 totalpages
+= pages
;
4528 node_set_state(nid
, N_HIGH_MEMORY
);
4534 * Find the PFN the Movable zone begins in each node. Kernel memory
4535 * is spread evenly between nodes as long as the nodes have enough
4536 * memory. When they don't, some nodes will have more kernelcore than
4539 static void __init
find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn
)
4542 unsigned long usable_startpfn
;
4543 unsigned long kernelcore_node
, kernelcore_remaining
;
4544 /* save the state before borrow the nodemask */
4545 nodemask_t saved_node_state
= node_states
[N_HIGH_MEMORY
];
4546 unsigned long totalpages
= early_calculate_totalpages();
4547 int usable_nodes
= nodes_weight(node_states
[N_HIGH_MEMORY
]);
4550 * If movablecore was specified, calculate what size of
4551 * kernelcore that corresponds so that memory usable for
4552 * any allocation type is evenly spread. If both kernelcore
4553 * and movablecore are specified, then the value of kernelcore
4554 * will be used for required_kernelcore if it's greater than
4555 * what movablecore would have allowed.
4557 if (required_movablecore
) {
4558 unsigned long corepages
;
4561 * Round-up so that ZONE_MOVABLE is at least as large as what
4562 * was requested by the user
4564 required_movablecore
=
4565 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4566 corepages
= totalpages
- required_movablecore
;
4568 required_kernelcore
= max(required_kernelcore
, corepages
);
4571 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4572 if (!required_kernelcore
)
4575 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4576 find_usable_zone_for_movable();
4577 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4580 /* Spread kernelcore memory as evenly as possible throughout nodes */
4581 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4582 for_each_node_state(nid
, N_HIGH_MEMORY
) {
4583 unsigned long start_pfn
, end_pfn
;
4586 * Recalculate kernelcore_node if the division per node
4587 * now exceeds what is necessary to satisfy the requested
4588 * amount of memory for the kernel
4590 if (required_kernelcore
< kernelcore_node
)
4591 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4594 * As the map is walked, we track how much memory is usable
4595 * by the kernel using kernelcore_remaining. When it is
4596 * 0, the rest of the node is usable by ZONE_MOVABLE
4598 kernelcore_remaining
= kernelcore_node
;
4600 /* Go through each range of PFNs within this node */
4601 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4602 unsigned long size_pages
;
4604 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
4605 if (start_pfn
>= end_pfn
)
4608 /* Account for what is only usable for kernelcore */
4609 if (start_pfn
< usable_startpfn
) {
4610 unsigned long kernel_pages
;
4611 kernel_pages
= min(end_pfn
, usable_startpfn
)
4614 kernelcore_remaining
-= min(kernel_pages
,
4615 kernelcore_remaining
);
4616 required_kernelcore
-= min(kernel_pages
,
4617 required_kernelcore
);
4619 /* Continue if range is now fully accounted */
4620 if (end_pfn
<= usable_startpfn
) {
4623 * Push zone_movable_pfn to the end so
4624 * that if we have to rebalance
4625 * kernelcore across nodes, we will
4626 * not double account here
4628 zone_movable_pfn
[nid
] = end_pfn
;
4631 start_pfn
= usable_startpfn
;
4635 * The usable PFN range for ZONE_MOVABLE is from
4636 * start_pfn->end_pfn. Calculate size_pages as the
4637 * number of pages used as kernelcore
4639 size_pages
= end_pfn
- start_pfn
;
4640 if (size_pages
> kernelcore_remaining
)
4641 size_pages
= kernelcore_remaining
;
4642 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4645 * Some kernelcore has been met, update counts and
4646 * break if the kernelcore for this node has been
4649 required_kernelcore
-= min(required_kernelcore
,
4651 kernelcore_remaining
-= size_pages
;
4652 if (!kernelcore_remaining
)
4658 * If there is still required_kernelcore, we do another pass with one
4659 * less node in the count. This will push zone_movable_pfn[nid] further
4660 * along on the nodes that still have memory until kernelcore is
4664 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4667 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4668 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4669 zone_movable_pfn
[nid
] =
4670 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4673 /* restore the node_state */
4674 node_states
[N_HIGH_MEMORY
] = saved_node_state
;
4677 /* Any regular memory on that node ? */
4678 static void check_for_regular_memory(pg_data_t
*pgdat
)
4680 #ifdef CONFIG_HIGHMEM
4681 enum zone_type zone_type
;
4683 for (zone_type
= 0; zone_type
<= ZONE_NORMAL
; zone_type
++) {
4684 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4685 if (zone
->present_pages
) {
4686 node_set_state(zone_to_nid(zone
), N_NORMAL_MEMORY
);
4694 * free_area_init_nodes - Initialise all pg_data_t and zone data
4695 * @max_zone_pfn: an array of max PFNs for each zone
4697 * This will call free_area_init_node() for each active node in the system.
4698 * Using the page ranges provided by add_active_range(), the size of each
4699 * zone in each node and their holes is calculated. If the maximum PFN
4700 * between two adjacent zones match, it is assumed that the zone is empty.
4701 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4702 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4703 * starts where the previous one ended. For example, ZONE_DMA32 starts
4704 * at arch_max_dma_pfn.
4706 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4708 unsigned long start_pfn
, end_pfn
;
4711 /* Record where the zone boundaries are */
4712 memset(arch_zone_lowest_possible_pfn
, 0,
4713 sizeof(arch_zone_lowest_possible_pfn
));
4714 memset(arch_zone_highest_possible_pfn
, 0,
4715 sizeof(arch_zone_highest_possible_pfn
));
4716 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4717 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4718 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4719 if (i
== ZONE_MOVABLE
)
4721 arch_zone_lowest_possible_pfn
[i
] =
4722 arch_zone_highest_possible_pfn
[i
-1];
4723 arch_zone_highest_possible_pfn
[i
] =
4724 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
4726 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
4727 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
4729 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4730 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
4731 find_zone_movable_pfns_for_nodes(zone_movable_pfn
);
4733 /* Print out the zone ranges */
4734 printk("Zone PFN ranges:\n");
4735 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4736 if (i
== ZONE_MOVABLE
)
4738 printk(" %-8s ", zone_names
[i
]);
4739 if (arch_zone_lowest_possible_pfn
[i
] ==
4740 arch_zone_highest_possible_pfn
[i
])
4743 printk("%0#10lx -> %0#10lx\n",
4744 arch_zone_lowest_possible_pfn
[i
],
4745 arch_zone_highest_possible_pfn
[i
]);
4748 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4749 printk("Movable zone start PFN for each node\n");
4750 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
4751 if (zone_movable_pfn
[i
])
4752 printk(" Node %d: %lu\n", i
, zone_movable_pfn
[i
]);
4755 /* Print out the early_node_map[] */
4756 printk("Early memory PFN ranges\n");
4757 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
4758 printk(" %3d: %0#10lx -> %0#10lx\n", nid
, start_pfn
, end_pfn
);
4760 /* Initialise every node */
4761 mminit_verify_pageflags_layout();
4762 setup_nr_node_ids();
4763 for_each_online_node(nid
) {
4764 pg_data_t
*pgdat
= NODE_DATA(nid
);
4765 free_area_init_node(nid
, NULL
,
4766 find_min_pfn_for_node(nid
), NULL
);
4768 /* Any memory on that node */
4769 if (pgdat
->node_present_pages
)
4770 node_set_state(nid
, N_HIGH_MEMORY
);
4771 check_for_regular_memory(pgdat
);
4775 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
4777 unsigned long long coremem
;
4781 coremem
= memparse(p
, &p
);
4782 *core
= coremem
>> PAGE_SHIFT
;
4784 /* Paranoid check that UL is enough for the coremem value */
4785 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
4791 * kernelcore=size sets the amount of memory for use for allocations that
4792 * cannot be reclaimed or migrated.
4794 static int __init
cmdline_parse_kernelcore(char *p
)
4796 return cmdline_parse_core(p
, &required_kernelcore
);
4800 * movablecore=size sets the amount of memory for use for allocations that
4801 * can be reclaimed or migrated.
4803 static int __init
cmdline_parse_movablecore(char *p
)
4805 return cmdline_parse_core(p
, &required_movablecore
);
4808 early_param("kernelcore", cmdline_parse_kernelcore
);
4809 early_param("movablecore", cmdline_parse_movablecore
);
4811 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4814 * set_dma_reserve - set the specified number of pages reserved in the first zone
4815 * @new_dma_reserve: The number of pages to mark reserved
4817 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4818 * In the DMA zone, a significant percentage may be consumed by kernel image
4819 * and other unfreeable allocations which can skew the watermarks badly. This
4820 * function may optionally be used to account for unfreeable pages in the
4821 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4822 * smaller per-cpu batchsize.
4824 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
4826 dma_reserve
= new_dma_reserve
;
4829 void __init
free_area_init(unsigned long *zones_size
)
4831 free_area_init_node(0, zones_size
,
4832 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
4835 static int page_alloc_cpu_notify(struct notifier_block
*self
,
4836 unsigned long action
, void *hcpu
)
4838 int cpu
= (unsigned long)hcpu
;
4840 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
4841 lru_add_drain_cpu(cpu
);
4845 * Spill the event counters of the dead processor
4846 * into the current processors event counters.
4847 * This artificially elevates the count of the current
4850 vm_events_fold_cpu(cpu
);
4853 * Zero the differential counters of the dead processor
4854 * so that the vm statistics are consistent.
4856 * This is only okay since the processor is dead and cannot
4857 * race with what we are doing.
4859 refresh_cpu_vm_stats(cpu
);
4864 void __init
page_alloc_init(void)
4866 hotcpu_notifier(page_alloc_cpu_notify
, 0);
4870 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
4871 * or min_free_kbytes changes.
4873 static void calculate_totalreserve_pages(void)
4875 struct pglist_data
*pgdat
;
4876 unsigned long reserve_pages
= 0;
4877 enum zone_type i
, j
;
4879 for_each_online_pgdat(pgdat
) {
4880 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
4881 struct zone
*zone
= pgdat
->node_zones
+ i
;
4882 unsigned long max
= 0;
4884 /* Find valid and maximum lowmem_reserve in the zone */
4885 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
4886 if (zone
->lowmem_reserve
[j
] > max
)
4887 max
= zone
->lowmem_reserve
[j
];
4890 /* we treat the high watermark as reserved pages. */
4891 max
+= high_wmark_pages(zone
);
4893 if (max
> zone
->present_pages
)
4894 max
= zone
->present_pages
;
4895 reserve_pages
+= max
;
4897 * Lowmem reserves are not available to
4898 * GFP_HIGHUSER page cache allocations and
4899 * kswapd tries to balance zones to their high
4900 * watermark. As a result, neither should be
4901 * regarded as dirtyable memory, to prevent a
4902 * situation where reclaim has to clean pages
4903 * in order to balance the zones.
4905 zone
->dirty_balance_reserve
= max
;
4908 dirty_balance_reserve
= reserve_pages
;
4909 totalreserve_pages
= reserve_pages
;
4913 * setup_per_zone_lowmem_reserve - called whenever
4914 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
4915 * has a correct pages reserved value, so an adequate number of
4916 * pages are left in the zone after a successful __alloc_pages().
4918 static void setup_per_zone_lowmem_reserve(void)
4920 struct pglist_data
*pgdat
;
4921 enum zone_type j
, idx
;
4923 for_each_online_pgdat(pgdat
) {
4924 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4925 struct zone
*zone
= pgdat
->node_zones
+ j
;
4926 unsigned long present_pages
= zone
->present_pages
;
4928 zone
->lowmem_reserve
[j
] = 0;
4932 struct zone
*lower_zone
;
4936 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
4937 sysctl_lowmem_reserve_ratio
[idx
] = 1;
4939 lower_zone
= pgdat
->node_zones
+ idx
;
4940 lower_zone
->lowmem_reserve
[j
] = present_pages
/
4941 sysctl_lowmem_reserve_ratio
[idx
];
4942 present_pages
+= lower_zone
->present_pages
;
4947 /* update totalreserve_pages */
4948 calculate_totalreserve_pages();
4952 * setup_per_zone_wmarks - called when min_free_kbytes changes
4953 * or when memory is hot-{added|removed}
4955 * Ensures that the watermark[min,low,high] values for each zone are set
4956 * correctly with respect to min_free_kbytes.
4958 void setup_per_zone_wmarks(void)
4960 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
4961 unsigned long lowmem_pages
= 0;
4963 unsigned long flags
;
4965 /* Calculate total number of !ZONE_HIGHMEM pages */
4966 for_each_zone(zone
) {
4967 if (!is_highmem(zone
))
4968 lowmem_pages
+= zone
->present_pages
;
4971 for_each_zone(zone
) {
4974 spin_lock_irqsave(&zone
->lock
, flags
);
4975 tmp
= (u64
)pages_min
* zone
->present_pages
;
4976 do_div(tmp
, lowmem_pages
);
4977 if (is_highmem(zone
)) {
4979 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
4980 * need highmem pages, so cap pages_min to a small
4983 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
4984 * deltas controls asynch page reclaim, and so should
4985 * not be capped for highmem.
4989 min_pages
= zone
->present_pages
/ 1024;
4990 if (min_pages
< SWAP_CLUSTER_MAX
)
4991 min_pages
= SWAP_CLUSTER_MAX
;
4992 if (min_pages
> 128)
4994 zone
->watermark
[WMARK_MIN
] = min_pages
;
4997 * If it's a lowmem zone, reserve a number of pages
4998 * proportionate to the zone's size.
5000 zone
->watermark
[WMARK_MIN
] = tmp
;
5003 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5004 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5005 setup_zone_migrate_reserve(zone
);
5006 spin_unlock_irqrestore(&zone
->lock
, flags
);
5009 /* update totalreserve_pages */
5010 calculate_totalreserve_pages();
5014 * The inactive anon list should be small enough that the VM never has to
5015 * do too much work, but large enough that each inactive page has a chance
5016 * to be referenced again before it is swapped out.
5018 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5019 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5020 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5021 * the anonymous pages are kept on the inactive list.
5024 * memory ratio inactive anon
5025 * -------------------------------------
5034 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5036 unsigned int gb
, ratio
;
5038 /* Zone size in gigabytes */
5039 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
5041 ratio
= int_sqrt(10 * gb
);
5045 zone
->inactive_ratio
= ratio
;
5048 static void __meminit
setup_per_zone_inactive_ratio(void)
5053 calculate_zone_inactive_ratio(zone
);
5057 * Initialise min_free_kbytes.
5059 * For small machines we want it small (128k min). For large machines
5060 * we want it large (64MB max). But it is not linear, because network
5061 * bandwidth does not increase linearly with machine size. We use
5063 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5064 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5080 int __meminit
init_per_zone_wmark_min(void)
5082 unsigned long lowmem_kbytes
;
5084 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5086 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5087 if (min_free_kbytes
< 128)
5088 min_free_kbytes
= 128;
5089 if (min_free_kbytes
> 65536)
5090 min_free_kbytes
= 65536;
5091 setup_per_zone_wmarks();
5092 refresh_zone_stat_thresholds();
5093 setup_per_zone_lowmem_reserve();
5094 setup_per_zone_inactive_ratio();
5097 module_init(init_per_zone_wmark_min
)
5100 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5101 * that we can call two helper functions whenever min_free_kbytes
5104 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5105 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5107 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5109 setup_per_zone_wmarks();
5114 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5115 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5120 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5125 zone
->min_unmapped_pages
= (zone
->present_pages
*
5126 sysctl_min_unmapped_ratio
) / 100;
5130 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5131 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5136 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5141 zone
->min_slab_pages
= (zone
->present_pages
*
5142 sysctl_min_slab_ratio
) / 100;
5148 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5149 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5150 * whenever sysctl_lowmem_reserve_ratio changes.
5152 * The reserve ratio obviously has absolutely no relation with the
5153 * minimum watermarks. The lowmem reserve ratio can only make sense
5154 * if in function of the boot time zone sizes.
5156 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5157 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5159 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5160 setup_per_zone_lowmem_reserve();
5165 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5166 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5167 * can have before it gets flushed back to buddy allocator.
5170 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5171 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5177 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5178 if (!write
|| (ret
== -EINVAL
))
5180 for_each_populated_zone(zone
) {
5181 for_each_possible_cpu(cpu
) {
5183 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
5184 setup_pagelist_highmark(
5185 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5191 int hashdist
= HASHDIST_DEFAULT
;
5194 static int __init
set_hashdist(char *str
)
5198 hashdist
= simple_strtoul(str
, &str
, 0);
5201 __setup("hashdist=", set_hashdist
);
5205 * allocate a large system hash table from bootmem
5206 * - it is assumed that the hash table must contain an exact power-of-2
5207 * quantity of entries
5208 * - limit is the number of hash buckets, not the total allocation size
5210 void *__init
alloc_large_system_hash(const char *tablename
,
5211 unsigned long bucketsize
,
5212 unsigned long numentries
,
5215 unsigned int *_hash_shift
,
5216 unsigned int *_hash_mask
,
5217 unsigned long limit
)
5219 unsigned long long max
= limit
;
5220 unsigned long log2qty
, size
;
5223 /* allow the kernel cmdline to have a say */
5225 /* round applicable memory size up to nearest megabyte */
5226 numentries
= nr_kernel_pages
;
5227 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5228 numentries
>>= 20 - PAGE_SHIFT
;
5229 numentries
<<= 20 - PAGE_SHIFT
;
5231 /* limit to 1 bucket per 2^scale bytes of low memory */
5232 if (scale
> PAGE_SHIFT
)
5233 numentries
>>= (scale
- PAGE_SHIFT
);
5235 numentries
<<= (PAGE_SHIFT
- scale
);
5237 /* Make sure we've got at least a 0-order allocation.. */
5238 if (unlikely(flags
& HASH_SMALL
)) {
5239 /* Makes no sense without HASH_EARLY */
5240 WARN_ON(!(flags
& HASH_EARLY
));
5241 if (!(numentries
>> *_hash_shift
)) {
5242 numentries
= 1UL << *_hash_shift
;
5243 BUG_ON(!numentries
);
5245 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5246 numentries
= PAGE_SIZE
/ bucketsize
;
5248 numentries
= roundup_pow_of_two(numentries
);
5250 /* limit allocation size to 1/16 total memory by default */
5252 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5253 do_div(max
, bucketsize
);
5255 max
= min(max
, 0x80000000ULL
);
5257 if (numentries
> max
)
5260 log2qty
= ilog2(numentries
);
5263 size
= bucketsize
<< log2qty
;
5264 if (flags
& HASH_EARLY
)
5265 table
= alloc_bootmem_nopanic(size
);
5267 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5270 * If bucketsize is not a power-of-two, we may free
5271 * some pages at the end of hash table which
5272 * alloc_pages_exact() automatically does
5274 if (get_order(size
) < MAX_ORDER
) {
5275 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5276 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5279 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5282 panic("Failed to allocate %s hash table\n", tablename
);
5284 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5287 ilog2(size
) - PAGE_SHIFT
,
5291 *_hash_shift
= log2qty
;
5293 *_hash_mask
= (1 << log2qty
) - 1;
5298 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5299 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5302 #ifdef CONFIG_SPARSEMEM
5303 return __pfn_to_section(pfn
)->pageblock_flags
;
5305 return zone
->pageblock_flags
;
5306 #endif /* CONFIG_SPARSEMEM */
5309 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5311 #ifdef CONFIG_SPARSEMEM
5312 pfn
&= (PAGES_PER_SECTION
-1);
5313 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5315 pfn
= pfn
- zone
->zone_start_pfn
;
5316 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5317 #endif /* CONFIG_SPARSEMEM */
5321 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5322 * @page: The page within the block of interest
5323 * @start_bitidx: The first bit of interest to retrieve
5324 * @end_bitidx: The last bit of interest
5325 * returns pageblock_bits flags
5327 unsigned long get_pageblock_flags_group(struct page
*page
,
5328 int start_bitidx
, int end_bitidx
)
5331 unsigned long *bitmap
;
5332 unsigned long pfn
, bitidx
;
5333 unsigned long flags
= 0;
5334 unsigned long value
= 1;
5336 zone
= page_zone(page
);
5337 pfn
= page_to_pfn(page
);
5338 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5339 bitidx
= pfn_to_bitidx(zone
, pfn
);
5341 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5342 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5349 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5350 * @page: The page within the block of interest
5351 * @start_bitidx: The first bit of interest
5352 * @end_bitidx: The last bit of interest
5353 * @flags: The flags to set
5355 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5356 int start_bitidx
, int end_bitidx
)
5359 unsigned long *bitmap
;
5360 unsigned long pfn
, bitidx
;
5361 unsigned long value
= 1;
5363 zone
= page_zone(page
);
5364 pfn
= page_to_pfn(page
);
5365 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5366 bitidx
= pfn_to_bitidx(zone
, pfn
);
5367 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
5368 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
5370 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5372 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5374 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5378 * This is designed as sub function...plz see page_isolation.c also.
5379 * set/clear page block's type to be ISOLATE.
5380 * page allocater never alloc memory from ISOLATE block.
5384 __count_immobile_pages(struct zone
*zone
, struct page
*page
, int count
)
5386 unsigned long pfn
, iter
, found
;
5388 * For avoiding noise data, lru_add_drain_all() should be called
5389 * If ZONE_MOVABLE, the zone never contains immobile pages
5391 if (zone_idx(zone
) == ZONE_MOVABLE
)
5394 if (get_pageblock_migratetype(page
) == MIGRATE_MOVABLE
)
5397 pfn
= page_to_pfn(page
);
5398 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5399 unsigned long check
= pfn
+ iter
;
5401 if (!pfn_valid_within(check
))
5404 page
= pfn_to_page(check
);
5405 if (!page_count(page
)) {
5406 if (PageBuddy(page
))
5407 iter
+= (1 << page_order(page
)) - 1;
5413 * If there are RECLAIMABLE pages, we need to check it.
5414 * But now, memory offline itself doesn't call shrink_slab()
5415 * and it still to be fixed.
5418 * If the page is not RAM, page_count()should be 0.
5419 * we don't need more check. This is an _used_ not-movable page.
5421 * The problematic thing here is PG_reserved pages. PG_reserved
5422 * is set to both of a memory hole page and a _used_ kernel
5431 bool is_pageblock_removable_nolock(struct page
*page
)
5437 * We have to be careful here because we are iterating over memory
5438 * sections which are not zone aware so we might end up outside of
5439 * the zone but still within the section.
5440 * We have to take care about the node as well. If the node is offline
5441 * its NODE_DATA will be NULL - see page_zone.
5443 if (!node_online(page_to_nid(page
)))
5446 zone
= page_zone(page
);
5447 pfn
= page_to_pfn(page
);
5448 if (zone
->zone_start_pfn
> pfn
||
5449 zone
->zone_start_pfn
+ zone
->spanned_pages
<= pfn
)
5452 return __count_immobile_pages(zone
, page
, 0);
5455 int set_migratetype_isolate(struct page
*page
)
5458 unsigned long flags
, pfn
;
5459 struct memory_isolate_notify arg
;
5463 zone
= page_zone(page
);
5465 spin_lock_irqsave(&zone
->lock
, flags
);
5467 pfn
= page_to_pfn(page
);
5468 arg
.start_pfn
= pfn
;
5469 arg
.nr_pages
= pageblock_nr_pages
;
5470 arg
.pages_found
= 0;
5473 * It may be possible to isolate a pageblock even if the
5474 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5475 * notifier chain is used by balloon drivers to return the
5476 * number of pages in a range that are held by the balloon
5477 * driver to shrink memory. If all the pages are accounted for
5478 * by balloons, are free, or on the LRU, isolation can continue.
5479 * Later, for example, when memory hotplug notifier runs, these
5480 * pages reported as "can be isolated" should be isolated(freed)
5481 * by the balloon driver through the memory notifier chain.
5483 notifier_ret
= memory_isolate_notify(MEM_ISOLATE_COUNT
, &arg
);
5484 notifier_ret
= notifier_to_errno(notifier_ret
);
5488 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5489 * We just check MOVABLE pages.
5491 if (__count_immobile_pages(zone
, page
, arg
.pages_found
))
5495 * immobile means "not-on-lru" paes. If immobile is larger than
5496 * removable-by-driver pages reported by notifier, we'll fail.
5501 set_pageblock_migratetype(page
, MIGRATE_ISOLATE
);
5502 move_freepages_block(zone
, page
, MIGRATE_ISOLATE
);
5505 spin_unlock_irqrestore(&zone
->lock
, flags
);
5511 void unset_migratetype_isolate(struct page
*page
)
5514 unsigned long flags
;
5515 zone
= page_zone(page
);
5516 spin_lock_irqsave(&zone
->lock
, flags
);
5517 if (get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)
5519 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
5520 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
5522 spin_unlock_irqrestore(&zone
->lock
, flags
);
5525 #ifdef CONFIG_MEMORY_HOTREMOVE
5527 * All pages in the range must be isolated before calling this.
5530 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
5536 unsigned long flags
;
5537 /* find the first valid pfn */
5538 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
5543 zone
= page_zone(pfn_to_page(pfn
));
5544 spin_lock_irqsave(&zone
->lock
, flags
);
5546 while (pfn
< end_pfn
) {
5547 if (!pfn_valid(pfn
)) {
5551 page
= pfn_to_page(pfn
);
5552 BUG_ON(page_count(page
));
5553 BUG_ON(!PageBuddy(page
));
5554 order
= page_order(page
);
5555 #ifdef CONFIG_DEBUG_VM
5556 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
5557 pfn
, 1 << order
, end_pfn
);
5559 list_del(&page
->lru
);
5560 rmv_page_order(page
);
5561 zone
->free_area
[order
].nr_free
--;
5562 __mod_zone_page_state(zone
, NR_FREE_PAGES
,
5564 for (i
= 0; i
< (1 << order
); i
++)
5565 SetPageReserved((page
+i
));
5566 pfn
+= (1 << order
);
5568 spin_unlock_irqrestore(&zone
->lock
, flags
);
5572 #ifdef CONFIG_MEMORY_FAILURE
5573 bool is_free_buddy_page(struct page
*page
)
5575 struct zone
*zone
= page_zone(page
);
5576 unsigned long pfn
= page_to_pfn(page
);
5577 unsigned long flags
;
5580 spin_lock_irqsave(&zone
->lock
, flags
);
5581 for (order
= 0; order
< MAX_ORDER
; order
++) {
5582 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
5584 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
5587 spin_unlock_irqrestore(&zone
->lock
, flags
);
5589 return order
< MAX_ORDER
;
5593 static struct trace_print_flags pageflag_names
[] = {
5594 {1UL << PG_locked
, "locked" },
5595 {1UL << PG_error
, "error" },
5596 {1UL << PG_referenced
, "referenced" },
5597 {1UL << PG_uptodate
, "uptodate" },
5598 {1UL << PG_dirty
, "dirty" },
5599 {1UL << PG_lru
, "lru" },
5600 {1UL << PG_active
, "active" },
5601 {1UL << PG_slab
, "slab" },
5602 {1UL << PG_owner_priv_1
, "owner_priv_1" },
5603 {1UL << PG_arch_1
, "arch_1" },
5604 {1UL << PG_reserved
, "reserved" },
5605 {1UL << PG_private
, "private" },
5606 {1UL << PG_private_2
, "private_2" },
5607 {1UL << PG_writeback
, "writeback" },
5608 #ifdef CONFIG_PAGEFLAGS_EXTENDED
5609 {1UL << PG_head
, "head" },
5610 {1UL << PG_tail
, "tail" },
5612 {1UL << PG_compound
, "compound" },
5614 {1UL << PG_swapcache
, "swapcache" },
5615 {1UL << PG_mappedtodisk
, "mappedtodisk" },
5616 {1UL << PG_reclaim
, "reclaim" },
5617 {1UL << PG_swapbacked
, "swapbacked" },
5618 {1UL << PG_unevictable
, "unevictable" },
5620 {1UL << PG_mlocked
, "mlocked" },
5622 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
5623 {1UL << PG_uncached
, "uncached" },
5625 #ifdef CONFIG_MEMORY_FAILURE
5626 {1UL << PG_hwpoison
, "hwpoison" },
5631 static void dump_page_flags(unsigned long flags
)
5633 const char *delim
= "";
5637 printk(KERN_ALERT
"page flags: %#lx(", flags
);
5639 /* remove zone id */
5640 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
5642 for (i
= 0; pageflag_names
[i
].name
&& flags
; i
++) {
5644 mask
= pageflag_names
[i
].mask
;
5645 if ((flags
& mask
) != mask
)
5649 printk("%s%s", delim
, pageflag_names
[i
].name
);
5653 /* check for left over flags */
5655 printk("%s%#lx", delim
, flags
);
5660 void dump_page(struct page
*page
)
5663 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5664 page
, atomic_read(&page
->_count
), page_mapcount(page
),
5665 page
->mapping
, page
->index
);
5666 dump_page_flags(page
->flags
);
5667 mem_cgroup_print_bad_page(page
);