2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.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 #ifdef CONFIG_MOVABLE_NODE
94 [N_MEMORY
] = { { [0] = 1UL } },
96 [N_CPU
] = { { [0] = 1UL } },
99 EXPORT_SYMBOL(node_states
);
101 unsigned long totalram_pages __read_mostly
;
102 unsigned long totalreserve_pages __read_mostly
;
104 * When calculating the number of globally allowed dirty pages, there
105 * is a certain number of per-zone reserves that should not be
106 * considered dirtyable memory. This is the sum of those reserves
107 * over all existing zones that contribute dirtyable memory.
109 unsigned long dirty_balance_reserve __read_mostly
;
111 int percpu_pagelist_fraction
;
112 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
114 #ifdef CONFIG_PM_SLEEP
116 * The following functions are used by the suspend/hibernate code to temporarily
117 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
118 * while devices are suspended. To avoid races with the suspend/hibernate code,
119 * they should always be called with pm_mutex held (gfp_allowed_mask also should
120 * only be modified with pm_mutex held, unless the suspend/hibernate code is
121 * guaranteed not to run in parallel with that modification).
124 static gfp_t saved_gfp_mask
;
126 void pm_restore_gfp_mask(void)
128 WARN_ON(!mutex_is_locked(&pm_mutex
));
129 if (saved_gfp_mask
) {
130 gfp_allowed_mask
= saved_gfp_mask
;
135 void pm_restrict_gfp_mask(void)
137 WARN_ON(!mutex_is_locked(&pm_mutex
));
138 WARN_ON(saved_gfp_mask
);
139 saved_gfp_mask
= gfp_allowed_mask
;
140 gfp_allowed_mask
&= ~GFP_IOFS
;
143 bool pm_suspended_storage(void)
145 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
149 #endif /* CONFIG_PM_SLEEP */
151 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
152 int pageblock_order __read_mostly
;
155 static void __free_pages_ok(struct page
*page
, unsigned int order
);
158 * results with 256, 32 in the lowmem_reserve sysctl:
159 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
160 * 1G machine -> (16M dma, 784M normal, 224M high)
161 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
162 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
163 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
165 * TBD: should special case ZONE_DMA32 machines here - in those we normally
166 * don't need any ZONE_NORMAL reservation
168 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
169 #ifdef CONFIG_ZONE_DMA
172 #ifdef CONFIG_ZONE_DMA32
175 #ifdef CONFIG_HIGHMEM
181 EXPORT_SYMBOL(totalram_pages
);
183 static char * const zone_names
[MAX_NR_ZONES
] = {
184 #ifdef CONFIG_ZONE_DMA
187 #ifdef CONFIG_ZONE_DMA32
191 #ifdef CONFIG_HIGHMEM
197 int min_free_kbytes
= 1024;
199 static unsigned long __meminitdata nr_kernel_pages
;
200 static unsigned long __meminitdata nr_all_pages
;
201 static unsigned long __meminitdata dma_reserve
;
203 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
204 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
205 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
206 static unsigned long __initdata required_kernelcore
;
207 static unsigned long __initdata required_movablecore
;
208 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
210 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
212 EXPORT_SYMBOL(movable_zone
);
213 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
216 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
217 int nr_online_nodes __read_mostly
= 1;
218 EXPORT_SYMBOL(nr_node_ids
);
219 EXPORT_SYMBOL(nr_online_nodes
);
222 int page_group_by_mobility_disabled __read_mostly
;
224 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
227 if (unlikely(page_group_by_mobility_disabled
))
228 migratetype
= MIGRATE_UNMOVABLE
;
230 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
231 PB_migrate
, PB_migrate_end
);
234 bool oom_killer_disabled __read_mostly
;
236 #ifdef CONFIG_DEBUG_VM
237 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
241 unsigned long pfn
= page_to_pfn(page
);
244 seq
= zone_span_seqbegin(zone
);
245 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
247 else if (pfn
< zone
->zone_start_pfn
)
249 } while (zone_span_seqretry(zone
, seq
));
254 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
256 if (!pfn_valid_within(page_to_pfn(page
)))
258 if (zone
!= page_zone(page
))
264 * Temporary debugging check for pages not lying within a given zone.
266 static int bad_range(struct zone
*zone
, struct page
*page
)
268 if (page_outside_zone_boundaries(zone
, page
))
270 if (!page_is_consistent(zone
, page
))
276 static inline int bad_range(struct zone
*zone
, struct page
*page
)
282 static void bad_page(struct page
*page
)
284 static unsigned long resume
;
285 static unsigned long nr_shown
;
286 static unsigned long nr_unshown
;
288 /* Don't complain about poisoned pages */
289 if (PageHWPoison(page
)) {
290 reset_page_mapcount(page
); /* remove PageBuddy */
295 * Allow a burst of 60 reports, then keep quiet for that minute;
296 * or allow a steady drip of one report per second.
298 if (nr_shown
== 60) {
299 if (time_before(jiffies
, resume
)) {
305 "BUG: Bad page state: %lu messages suppressed\n",
312 resume
= jiffies
+ 60 * HZ
;
314 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
315 current
->comm
, page_to_pfn(page
));
321 /* Leave bad fields for debug, except PageBuddy could make trouble */
322 reset_page_mapcount(page
); /* remove PageBuddy */
323 add_taint(TAINT_BAD_PAGE
);
327 * Higher-order pages are called "compound pages". They are structured thusly:
329 * The first PAGE_SIZE page is called the "head page".
331 * The remaining PAGE_SIZE pages are called "tail pages".
333 * All pages have PG_compound set. All tail pages have their ->first_page
334 * pointing at the head page.
336 * The first tail page's ->lru.next holds the address of the compound page's
337 * put_page() function. Its ->lru.prev holds the order of allocation.
338 * This usage means that zero-order pages may not be compound.
341 static void free_compound_page(struct page
*page
)
343 __free_pages_ok(page
, compound_order(page
));
346 void prep_compound_page(struct page
*page
, unsigned long order
)
349 int nr_pages
= 1 << order
;
351 set_compound_page_dtor(page
, free_compound_page
);
352 set_compound_order(page
, order
);
354 for (i
= 1; i
< nr_pages
; i
++) {
355 struct page
*p
= page
+ i
;
357 set_page_count(p
, 0);
358 p
->first_page
= page
;
362 /* update __split_huge_page_refcount if you change this function */
363 static int destroy_compound_page(struct page
*page
, unsigned long order
)
366 int nr_pages
= 1 << order
;
369 if (unlikely(compound_order(page
) != order
)) {
374 __ClearPageHead(page
);
376 for (i
= 1; i
< nr_pages
; i
++) {
377 struct page
*p
= page
+ i
;
379 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
389 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
394 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
395 * and __GFP_HIGHMEM from hard or soft interrupt context.
397 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
398 for (i
= 0; i
< (1 << order
); i
++)
399 clear_highpage(page
+ i
);
402 #ifdef CONFIG_DEBUG_PAGEALLOC
403 unsigned int _debug_guardpage_minorder
;
405 static int __init
debug_guardpage_minorder_setup(char *buf
)
409 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
410 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
413 _debug_guardpage_minorder
= res
;
414 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
417 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
419 static inline void set_page_guard_flag(struct page
*page
)
421 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
424 static inline void clear_page_guard_flag(struct page
*page
)
426 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
429 static inline void set_page_guard_flag(struct page
*page
) { }
430 static inline void clear_page_guard_flag(struct page
*page
) { }
433 static inline void set_page_order(struct page
*page
, int order
)
435 set_page_private(page
, order
);
436 __SetPageBuddy(page
);
439 static inline void rmv_page_order(struct page
*page
)
441 __ClearPageBuddy(page
);
442 set_page_private(page
, 0);
446 * Locate the struct page for both the matching buddy in our
447 * pair (buddy1) and the combined O(n+1) page they form (page).
449 * 1) Any buddy B1 will have an order O twin B2 which satisfies
450 * the following equation:
452 * For example, if the starting buddy (buddy2) is #8 its order
454 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
456 * 2) Any buddy B will have an order O+1 parent P which
457 * satisfies the following equation:
460 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
462 static inline unsigned long
463 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
465 return page_idx
^ (1 << order
);
469 * This function checks whether a page is free && is the buddy
470 * we can do coalesce a page and its buddy if
471 * (a) the buddy is not in a hole &&
472 * (b) the buddy is in the buddy system &&
473 * (c) a page and its buddy have the same order &&
474 * (d) a page and its buddy are in the same zone.
476 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
477 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
479 * For recording page's order, we use page_private(page).
481 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
484 if (!pfn_valid_within(page_to_pfn(buddy
)))
487 if (page_zone_id(page
) != page_zone_id(buddy
))
490 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
491 VM_BUG_ON(page_count(buddy
) != 0);
495 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
496 VM_BUG_ON(page_count(buddy
) != 0);
503 * Freeing function for a buddy system allocator.
505 * The concept of a buddy system is to maintain direct-mapped table
506 * (containing bit values) for memory blocks of various "orders".
507 * The bottom level table contains the map for the smallest allocatable
508 * units of memory (here, pages), and each level above it describes
509 * pairs of units from the levels below, hence, "buddies".
510 * At a high level, all that happens here is marking the table entry
511 * at the bottom level available, and propagating the changes upward
512 * as necessary, plus some accounting needed to play nicely with other
513 * parts of the VM system.
514 * At each level, we keep a list of pages, which are heads of continuous
515 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
516 * order is recorded in page_private(page) field.
517 * So when we are allocating or freeing one, we can derive the state of the
518 * other. That is, if we allocate a small block, and both were
519 * free, the remainder of the region must be split into blocks.
520 * If a block is freed, and its buddy is also free, then this
521 * triggers coalescing into a block of larger size.
526 static inline void __free_one_page(struct page
*page
,
527 struct zone
*zone
, unsigned int order
,
530 unsigned long page_idx
;
531 unsigned long combined_idx
;
532 unsigned long uninitialized_var(buddy_idx
);
535 if (unlikely(PageCompound(page
)))
536 if (unlikely(destroy_compound_page(page
, order
)))
539 VM_BUG_ON(migratetype
== -1);
541 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
543 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
544 VM_BUG_ON(bad_range(zone
, page
));
546 while (order
< MAX_ORDER
-1) {
547 buddy_idx
= __find_buddy_index(page_idx
, order
);
548 buddy
= page
+ (buddy_idx
- page_idx
);
549 if (!page_is_buddy(page
, buddy
, order
))
552 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
553 * merge with it and move up one order.
555 if (page_is_guard(buddy
)) {
556 clear_page_guard_flag(buddy
);
557 set_page_private(page
, 0);
558 __mod_zone_freepage_state(zone
, 1 << order
,
561 list_del(&buddy
->lru
);
562 zone
->free_area
[order
].nr_free
--;
563 rmv_page_order(buddy
);
565 combined_idx
= buddy_idx
& page_idx
;
566 page
= page
+ (combined_idx
- page_idx
);
567 page_idx
= combined_idx
;
570 set_page_order(page
, order
);
573 * If this is not the largest possible page, check if the buddy
574 * of the next-highest order is free. If it is, it's possible
575 * that pages are being freed that will coalesce soon. In case,
576 * that is happening, add the free page to the tail of the list
577 * so it's less likely to be used soon and more likely to be merged
578 * as a higher order page
580 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
581 struct page
*higher_page
, *higher_buddy
;
582 combined_idx
= buddy_idx
& page_idx
;
583 higher_page
= page
+ (combined_idx
- page_idx
);
584 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
585 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
586 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
587 list_add_tail(&page
->lru
,
588 &zone
->free_area
[order
].free_list
[migratetype
]);
593 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
595 zone
->free_area
[order
].nr_free
++;
598 static inline int free_pages_check(struct page
*page
)
600 if (unlikely(page_mapcount(page
) |
601 (page
->mapping
!= NULL
) |
602 (atomic_read(&page
->_count
) != 0) |
603 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
) |
604 (mem_cgroup_bad_page_check(page
)))) {
608 reset_page_last_nid(page
);
609 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
610 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
615 * Frees a number of pages from the PCP lists
616 * Assumes all pages on list are in same zone, and of same order.
617 * count is the number of pages to free.
619 * If the zone was previously in an "all pages pinned" state then look to
620 * see if this freeing clears that state.
622 * And clear the zone's pages_scanned counter, to hold off the "all pages are
623 * pinned" detection logic.
625 static void free_pcppages_bulk(struct zone
*zone
, int count
,
626 struct per_cpu_pages
*pcp
)
632 spin_lock(&zone
->lock
);
633 zone
->all_unreclaimable
= 0;
634 zone
->pages_scanned
= 0;
638 struct list_head
*list
;
641 * Remove pages from lists in a round-robin fashion. A
642 * batch_free count is maintained that is incremented when an
643 * empty list is encountered. This is so more pages are freed
644 * off fuller lists instead of spinning excessively around empty
649 if (++migratetype
== MIGRATE_PCPTYPES
)
651 list
= &pcp
->lists
[migratetype
];
652 } while (list_empty(list
));
654 /* This is the only non-empty list. Free them all. */
655 if (batch_free
== MIGRATE_PCPTYPES
)
656 batch_free
= to_free
;
659 int mt
; /* migratetype of the to-be-freed page */
661 page
= list_entry(list
->prev
, struct page
, lru
);
662 /* must delete as __free_one_page list manipulates */
663 list_del(&page
->lru
);
664 mt
= get_freepage_migratetype(page
);
665 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
666 __free_one_page(page
, zone
, 0, mt
);
667 trace_mm_page_pcpu_drain(page
, 0, mt
);
668 if (likely(get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)) {
669 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1);
670 if (is_migrate_cma(mt
))
671 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
, 1);
673 } while (--to_free
&& --batch_free
&& !list_empty(list
));
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 if (unlikely(migratetype
!= MIGRATE_ISOLATE
))
687 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
688 spin_unlock(&zone
->lock
);
691 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
696 trace_mm_page_free(page
, order
);
697 kmemcheck_free_shadow(page
, order
);
700 page
->mapping
= NULL
;
701 for (i
= 0; i
< (1 << order
); i
++)
702 bad
+= free_pages_check(page
+ i
);
706 if (!PageHighMem(page
)) {
707 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
708 debug_check_no_obj_freed(page_address(page
),
711 arch_free_page(page
, order
);
712 kernel_map_pages(page
, 1 << order
, 0);
717 static void __free_pages_ok(struct page
*page
, unsigned int order
)
722 if (!free_pages_prepare(page
, order
))
725 local_irq_save(flags
);
726 __count_vm_events(PGFREE
, 1 << order
);
727 migratetype
= get_pageblock_migratetype(page
);
728 set_freepage_migratetype(page
, migratetype
);
729 free_one_page(page_zone(page
), page
, order
, migratetype
);
730 local_irq_restore(flags
);
734 * Read access to zone->managed_pages is safe because it's unsigned long,
735 * but we still need to serialize writers. Currently all callers of
736 * __free_pages_bootmem() except put_page_bootmem() should only be used
737 * at boot time. So for shorter boot time, we shift the burden to
738 * put_page_bootmem() to serialize writers.
740 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
742 unsigned int nr_pages
= 1 << order
;
746 for (loop
= 0; loop
< nr_pages
; loop
++) {
747 struct page
*p
= &page
[loop
];
749 if (loop
+ 1 < nr_pages
)
751 __ClearPageReserved(p
);
752 set_page_count(p
, 0);
755 page_zone(page
)->managed_pages
+= 1 << order
;
756 set_page_refcounted(page
);
757 __free_pages(page
, order
);
761 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
762 void __init
init_cma_reserved_pageblock(struct page
*page
)
764 unsigned i
= pageblock_nr_pages
;
765 struct page
*p
= page
;
768 __ClearPageReserved(p
);
769 set_page_count(p
, 0);
772 set_page_refcounted(page
);
773 set_pageblock_migratetype(page
, MIGRATE_CMA
);
774 __free_pages(page
, pageblock_order
);
775 totalram_pages
+= pageblock_nr_pages
;
780 * The order of subdivision here is critical for the IO subsystem.
781 * Please do not alter this order without good reasons and regression
782 * testing. Specifically, as large blocks of memory are subdivided,
783 * the order in which smaller blocks are delivered depends on the order
784 * they're subdivided in this function. This is the primary factor
785 * influencing the order in which pages are delivered to the IO
786 * subsystem according to empirical testing, and this is also justified
787 * by considering the behavior of a buddy system containing a single
788 * large block of memory acted on by a series of small allocations.
789 * This behavior is a critical factor in sglist merging's success.
793 static inline void expand(struct zone
*zone
, struct page
*page
,
794 int low
, int high
, struct free_area
*area
,
797 unsigned long size
= 1 << high
;
803 VM_BUG_ON(bad_range(zone
, &page
[size
]));
805 #ifdef CONFIG_DEBUG_PAGEALLOC
806 if (high
< debug_guardpage_minorder()) {
808 * Mark as guard pages (or page), that will allow to
809 * merge back to allocator when buddy will be freed.
810 * Corresponding page table entries will not be touched,
811 * pages will stay not present in virtual address space
813 INIT_LIST_HEAD(&page
[size
].lru
);
814 set_page_guard_flag(&page
[size
]);
815 set_page_private(&page
[size
], high
);
816 /* Guard pages are not available for any usage */
817 __mod_zone_freepage_state(zone
, -(1 << high
),
822 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
824 set_page_order(&page
[size
], high
);
829 * This page is about to be returned from the page allocator
831 static inline int check_new_page(struct page
*page
)
833 if (unlikely(page_mapcount(page
) |
834 (page
->mapping
!= NULL
) |
835 (atomic_read(&page
->_count
) != 0) |
836 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
) |
837 (mem_cgroup_bad_page_check(page
)))) {
844 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
848 for (i
= 0; i
< (1 << order
); i
++) {
849 struct page
*p
= page
+ i
;
850 if (unlikely(check_new_page(p
)))
854 set_page_private(page
, 0);
855 set_page_refcounted(page
);
857 arch_alloc_page(page
, order
);
858 kernel_map_pages(page
, 1 << order
, 1);
860 if (gfp_flags
& __GFP_ZERO
)
861 prep_zero_page(page
, order
, gfp_flags
);
863 if (order
&& (gfp_flags
& __GFP_COMP
))
864 prep_compound_page(page
, order
);
870 * Go through the free lists for the given migratetype and remove
871 * the smallest available page from the freelists
874 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
877 unsigned int current_order
;
878 struct free_area
* area
;
881 /* Find a page of the appropriate size in the preferred list */
882 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
883 area
= &(zone
->free_area
[current_order
]);
884 if (list_empty(&area
->free_list
[migratetype
]))
887 page
= list_entry(area
->free_list
[migratetype
].next
,
889 list_del(&page
->lru
);
890 rmv_page_order(page
);
892 expand(zone
, page
, order
, current_order
, area
, migratetype
);
901 * This array describes the order lists are fallen back to when
902 * the free lists for the desirable migrate type are depleted
904 static int fallbacks
[MIGRATE_TYPES
][4] = {
905 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
906 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
908 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
909 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
911 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
913 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
914 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
918 * Move the free pages in a range to the free lists of the requested type.
919 * Note that start_page and end_pages are not aligned on a pageblock
920 * boundary. If alignment is required, use move_freepages_block()
922 int move_freepages(struct zone
*zone
,
923 struct page
*start_page
, struct page
*end_page
,
930 #ifndef CONFIG_HOLES_IN_ZONE
932 * page_zone is not safe to call in this context when
933 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
934 * anyway as we check zone boundaries in move_freepages_block().
935 * Remove at a later date when no bug reports exist related to
936 * grouping pages by mobility
938 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
941 for (page
= start_page
; page
<= end_page
;) {
942 /* Make sure we are not inadvertently changing nodes */
943 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
945 if (!pfn_valid_within(page_to_pfn(page
))) {
950 if (!PageBuddy(page
)) {
955 order
= page_order(page
);
956 list_move(&page
->lru
,
957 &zone
->free_area
[order
].free_list
[migratetype
]);
958 set_freepage_migratetype(page
, migratetype
);
960 pages_moved
+= 1 << order
;
966 int move_freepages_block(struct zone
*zone
, struct page
*page
,
969 unsigned long start_pfn
, end_pfn
;
970 struct page
*start_page
, *end_page
;
972 start_pfn
= page_to_pfn(page
);
973 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
974 start_page
= pfn_to_page(start_pfn
);
975 end_page
= start_page
+ pageblock_nr_pages
- 1;
976 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
978 /* Do not cross zone boundaries */
979 if (start_pfn
< zone
->zone_start_pfn
)
981 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
984 return move_freepages(zone
, start_page
, end_page
, migratetype
);
987 static void change_pageblock_range(struct page
*pageblock_page
,
988 int start_order
, int migratetype
)
990 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
992 while (nr_pageblocks
--) {
993 set_pageblock_migratetype(pageblock_page
, migratetype
);
994 pageblock_page
+= pageblock_nr_pages
;
998 /* Remove an element from the buddy allocator from the fallback list */
999 static inline struct page
*
1000 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
1002 struct free_area
* area
;
1007 /* Find the largest possible block of pages in the other list */
1008 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
1011 migratetype
= fallbacks
[start_migratetype
][i
];
1013 /* MIGRATE_RESERVE handled later if necessary */
1014 if (migratetype
== MIGRATE_RESERVE
)
1017 area
= &(zone
->free_area
[current_order
]);
1018 if (list_empty(&area
->free_list
[migratetype
]))
1021 page
= list_entry(area
->free_list
[migratetype
].next
,
1026 * If breaking a large block of pages, move all free
1027 * pages to the preferred allocation list. If falling
1028 * back for a reclaimable kernel allocation, be more
1029 * aggressive about taking ownership of free pages
1031 * On the other hand, never change migration
1032 * type of MIGRATE_CMA pageblocks nor move CMA
1033 * pages on different free lists. We don't
1034 * want unmovable pages to be allocated from
1035 * MIGRATE_CMA areas.
1037 if (!is_migrate_cma(migratetype
) &&
1038 (unlikely(current_order
>= pageblock_order
/ 2) ||
1039 start_migratetype
== MIGRATE_RECLAIMABLE
||
1040 page_group_by_mobility_disabled
)) {
1042 pages
= move_freepages_block(zone
, page
,
1045 /* Claim the whole block if over half of it is free */
1046 if (pages
>= (1 << (pageblock_order
-1)) ||
1047 page_group_by_mobility_disabled
)
1048 set_pageblock_migratetype(page
,
1051 migratetype
= start_migratetype
;
1054 /* Remove the page from the freelists */
1055 list_del(&page
->lru
);
1056 rmv_page_order(page
);
1058 /* Take ownership for orders >= pageblock_order */
1059 if (current_order
>= pageblock_order
&&
1060 !is_migrate_cma(migratetype
))
1061 change_pageblock_range(page
, current_order
,
1064 expand(zone
, page
, order
, current_order
, area
,
1065 is_migrate_cma(migratetype
)
1066 ? migratetype
: start_migratetype
);
1068 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1069 start_migratetype
, migratetype
);
1079 * Do the hard work of removing an element from the buddy allocator.
1080 * Call me with the zone->lock already held.
1082 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1088 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1090 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1091 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1094 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1095 * is used because __rmqueue_smallest is an inline function
1096 * and we want just one call site
1099 migratetype
= MIGRATE_RESERVE
;
1104 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1109 * Obtain a specified number of elements from the buddy allocator, all under
1110 * a single hold of the lock, for efficiency. Add them to the supplied list.
1111 * Returns the number of new pages which were placed at *list.
1113 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1114 unsigned long count
, struct list_head
*list
,
1115 int migratetype
, int cold
)
1117 int mt
= migratetype
, i
;
1119 spin_lock(&zone
->lock
);
1120 for (i
= 0; i
< count
; ++i
) {
1121 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1122 if (unlikely(page
== NULL
))
1126 * Split buddy pages returned by expand() are received here
1127 * in physical page order. The page is added to the callers and
1128 * list and the list head then moves forward. From the callers
1129 * perspective, the linked list is ordered by page number in
1130 * some conditions. This is useful for IO devices that can
1131 * merge IO requests if the physical pages are ordered
1134 if (likely(cold
== 0))
1135 list_add(&page
->lru
, list
);
1137 list_add_tail(&page
->lru
, list
);
1138 if (IS_ENABLED(CONFIG_CMA
)) {
1139 mt
= get_pageblock_migratetype(page
);
1140 if (!is_migrate_cma(mt
) && mt
!= MIGRATE_ISOLATE
)
1143 set_freepage_migratetype(page
, mt
);
1145 if (is_migrate_cma(mt
))
1146 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1149 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1150 spin_unlock(&zone
->lock
);
1156 * Called from the vmstat counter updater to drain pagesets of this
1157 * currently executing processor on remote nodes after they have
1160 * Note that this function must be called with the thread pinned to
1161 * a single processor.
1163 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1165 unsigned long flags
;
1168 local_irq_save(flags
);
1169 if (pcp
->count
>= pcp
->batch
)
1170 to_drain
= pcp
->batch
;
1172 to_drain
= pcp
->count
;
1174 free_pcppages_bulk(zone
, to_drain
, pcp
);
1175 pcp
->count
-= to_drain
;
1177 local_irq_restore(flags
);
1182 * Drain pages of the indicated processor.
1184 * The processor must either be the current processor and the
1185 * thread pinned to the current processor or a processor that
1188 static void drain_pages(unsigned int cpu
)
1190 unsigned long flags
;
1193 for_each_populated_zone(zone
) {
1194 struct per_cpu_pageset
*pset
;
1195 struct per_cpu_pages
*pcp
;
1197 local_irq_save(flags
);
1198 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1202 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1205 local_irq_restore(flags
);
1210 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1212 void drain_local_pages(void *arg
)
1214 drain_pages(smp_processor_id());
1218 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1220 * Note that this code is protected against sending an IPI to an offline
1221 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1222 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1223 * nothing keeps CPUs from showing up after we populated the cpumask and
1224 * before the call to on_each_cpu_mask().
1226 void drain_all_pages(void)
1229 struct per_cpu_pageset
*pcp
;
1233 * Allocate in the BSS so we wont require allocation in
1234 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1236 static cpumask_t cpus_with_pcps
;
1239 * We don't care about racing with CPU hotplug event
1240 * as offline notification will cause the notified
1241 * cpu to drain that CPU pcps and on_each_cpu_mask
1242 * disables preemption as part of its processing
1244 for_each_online_cpu(cpu
) {
1245 bool has_pcps
= false;
1246 for_each_populated_zone(zone
) {
1247 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1248 if (pcp
->pcp
.count
) {
1254 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1256 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1258 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1261 #ifdef CONFIG_HIBERNATION
1263 void mark_free_pages(struct zone
*zone
)
1265 unsigned long pfn
, max_zone_pfn
;
1266 unsigned long flags
;
1268 struct list_head
*curr
;
1270 if (!zone
->spanned_pages
)
1273 spin_lock_irqsave(&zone
->lock
, flags
);
1275 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1276 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1277 if (pfn_valid(pfn
)) {
1278 struct page
*page
= pfn_to_page(pfn
);
1280 if (!swsusp_page_is_forbidden(page
))
1281 swsusp_unset_page_free(page
);
1284 for_each_migratetype_order(order
, t
) {
1285 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1288 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1289 for (i
= 0; i
< (1UL << order
); i
++)
1290 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1293 spin_unlock_irqrestore(&zone
->lock
, flags
);
1295 #endif /* CONFIG_PM */
1298 * Free a 0-order page
1299 * cold == 1 ? free a cold page : free a hot page
1301 void free_hot_cold_page(struct page
*page
, int cold
)
1303 struct zone
*zone
= page_zone(page
);
1304 struct per_cpu_pages
*pcp
;
1305 unsigned long flags
;
1308 if (!free_pages_prepare(page
, 0))
1311 migratetype
= get_pageblock_migratetype(page
);
1312 set_freepage_migratetype(page
, migratetype
);
1313 local_irq_save(flags
);
1314 __count_vm_event(PGFREE
);
1317 * We only track unmovable, reclaimable and movable on pcp lists.
1318 * Free ISOLATE pages back to the allocator because they are being
1319 * offlined but treat RESERVE as movable pages so we can get those
1320 * areas back if necessary. Otherwise, we may have to free
1321 * excessively into the page allocator
1323 if (migratetype
>= MIGRATE_PCPTYPES
) {
1324 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1325 free_one_page(zone
, page
, 0, migratetype
);
1328 migratetype
= MIGRATE_MOVABLE
;
1331 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1333 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1335 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1337 if (pcp
->count
>= pcp
->high
) {
1338 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1339 pcp
->count
-= pcp
->batch
;
1343 local_irq_restore(flags
);
1347 * Free a list of 0-order pages
1349 void free_hot_cold_page_list(struct list_head
*list
, int cold
)
1351 struct page
*page
, *next
;
1353 list_for_each_entry_safe(page
, next
, list
, lru
) {
1354 trace_mm_page_free_batched(page
, cold
);
1355 free_hot_cold_page(page
, cold
);
1360 * split_page takes a non-compound higher-order page, and splits it into
1361 * n (1<<order) sub-pages: page[0..n]
1362 * Each sub-page must be freed individually.
1364 * Note: this is probably too low level an operation for use in drivers.
1365 * Please consult with lkml before using this in your driver.
1367 void split_page(struct page
*page
, unsigned int order
)
1371 VM_BUG_ON(PageCompound(page
));
1372 VM_BUG_ON(!page_count(page
));
1374 #ifdef CONFIG_KMEMCHECK
1376 * Split shadow pages too, because free(page[0]) would
1377 * otherwise free the whole shadow.
1379 if (kmemcheck_page_is_tracked(page
))
1380 split_page(virt_to_page(page
[0].shadow
), order
);
1383 for (i
= 1; i
< (1 << order
); i
++)
1384 set_page_refcounted(page
+ i
);
1388 * Similar to the split_page family of functions except that the page
1389 * required at the given order and being isolated now to prevent races
1390 * with parallel allocators
1392 int capture_free_page(struct page
*page
, int alloc_order
, int migratetype
)
1395 unsigned long watermark
;
1399 BUG_ON(!PageBuddy(page
));
1401 zone
= page_zone(page
);
1402 order
= page_order(page
);
1403 mt
= get_pageblock_migratetype(page
);
1405 if (mt
!= MIGRATE_ISOLATE
) {
1406 /* Obey watermarks as if the page was being allocated */
1407 watermark
= low_wmark_pages(zone
) + (1 << order
);
1408 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1411 __mod_zone_freepage_state(zone
, -(1UL << alloc_order
), mt
);
1414 /* Remove page from free list */
1415 list_del(&page
->lru
);
1416 zone
->free_area
[order
].nr_free
--;
1417 rmv_page_order(page
);
1419 if (alloc_order
!= order
)
1420 expand(zone
, page
, alloc_order
, order
,
1421 &zone
->free_area
[order
], migratetype
);
1423 /* Set the pageblock if the captured page is at least a pageblock */
1424 if (order
>= pageblock_order
- 1) {
1425 struct page
*endpage
= page
+ (1 << order
) - 1;
1426 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1427 int mt
= get_pageblock_migratetype(page
);
1428 if (mt
!= MIGRATE_ISOLATE
&& !is_migrate_cma(mt
))
1429 set_pageblock_migratetype(page
,
1434 return 1UL << alloc_order
;
1438 * Similar to split_page except the page is already free. As this is only
1439 * being used for migration, the migratetype of the block also changes.
1440 * As this is called with interrupts disabled, the caller is responsible
1441 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1444 * Note: this is probably too low level an operation for use in drivers.
1445 * Please consult with lkml before using this in your driver.
1447 int split_free_page(struct page
*page
)
1452 BUG_ON(!PageBuddy(page
));
1453 order
= page_order(page
);
1455 nr_pages
= capture_free_page(page
, order
, 0);
1459 /* Split into individual pages */
1460 set_page_refcounted(page
);
1461 split_page(page
, order
);
1466 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1467 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1471 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1472 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1475 unsigned long flags
;
1477 int cold
= !!(gfp_flags
& __GFP_COLD
);
1480 if (likely(order
== 0)) {
1481 struct per_cpu_pages
*pcp
;
1482 struct list_head
*list
;
1484 local_irq_save(flags
);
1485 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1486 list
= &pcp
->lists
[migratetype
];
1487 if (list_empty(list
)) {
1488 pcp
->count
+= rmqueue_bulk(zone
, 0,
1491 if (unlikely(list_empty(list
)))
1496 page
= list_entry(list
->prev
, struct page
, lru
);
1498 page
= list_entry(list
->next
, struct page
, lru
);
1500 list_del(&page
->lru
);
1503 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1505 * __GFP_NOFAIL is not to be used in new code.
1507 * All __GFP_NOFAIL callers should be fixed so that they
1508 * properly detect and handle allocation failures.
1510 * We most definitely don't want callers attempting to
1511 * allocate greater than order-1 page units with
1514 WARN_ON_ONCE(order
> 1);
1516 spin_lock_irqsave(&zone
->lock
, flags
);
1517 page
= __rmqueue(zone
, order
, migratetype
);
1518 spin_unlock(&zone
->lock
);
1521 __mod_zone_freepage_state(zone
, -(1 << order
),
1522 get_pageblock_migratetype(page
));
1525 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1526 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1527 local_irq_restore(flags
);
1529 VM_BUG_ON(bad_range(zone
, page
));
1530 if (prep_new_page(page
, order
, gfp_flags
))
1535 local_irq_restore(flags
);
1539 #ifdef CONFIG_FAIL_PAGE_ALLOC
1542 struct fault_attr attr
;
1544 u32 ignore_gfp_highmem
;
1545 u32 ignore_gfp_wait
;
1547 } fail_page_alloc
= {
1548 .attr
= FAULT_ATTR_INITIALIZER
,
1549 .ignore_gfp_wait
= 1,
1550 .ignore_gfp_highmem
= 1,
1554 static int __init
setup_fail_page_alloc(char *str
)
1556 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1558 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1560 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1562 if (order
< fail_page_alloc
.min_order
)
1564 if (gfp_mask
& __GFP_NOFAIL
)
1566 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1568 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1571 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1574 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1576 static int __init
fail_page_alloc_debugfs(void)
1578 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1581 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1582 &fail_page_alloc
.attr
);
1584 return PTR_ERR(dir
);
1586 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1587 &fail_page_alloc
.ignore_gfp_wait
))
1589 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1590 &fail_page_alloc
.ignore_gfp_highmem
))
1592 if (!debugfs_create_u32("min-order", mode
, dir
,
1593 &fail_page_alloc
.min_order
))
1598 debugfs_remove_recursive(dir
);
1603 late_initcall(fail_page_alloc_debugfs
);
1605 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1607 #else /* CONFIG_FAIL_PAGE_ALLOC */
1609 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1614 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1617 * Return true if free pages are above 'mark'. This takes into account the order
1618 * of the allocation.
1620 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1621 int classzone_idx
, int alloc_flags
, long free_pages
)
1623 /* free_pages my go negative - that's OK */
1625 long lowmem_reserve
= z
->lowmem_reserve
[classzone_idx
];
1628 free_pages
-= (1 << order
) - 1;
1629 if (alloc_flags
& ALLOC_HIGH
)
1631 if (alloc_flags
& ALLOC_HARDER
)
1634 /* If allocation can't use CMA areas don't use free CMA pages */
1635 if (!(alloc_flags
& ALLOC_CMA
))
1636 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1638 if (free_pages
<= min
+ lowmem_reserve
)
1640 for (o
= 0; o
< order
; o
++) {
1641 /* At the next order, this order's pages become unavailable */
1642 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1644 /* Require fewer higher order pages to be free */
1647 if (free_pages
<= min
)
1653 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1654 int classzone_idx
, int alloc_flags
)
1656 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1657 zone_page_state(z
, NR_FREE_PAGES
));
1660 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1661 int classzone_idx
, int alloc_flags
)
1663 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1665 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1666 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1668 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1674 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1675 * skip over zones that are not allowed by the cpuset, or that have
1676 * been recently (in last second) found to be nearly full. See further
1677 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1678 * that have to skip over a lot of full or unallowed zones.
1680 * If the zonelist cache is present in the passed in zonelist, then
1681 * returns a pointer to the allowed node mask (either the current
1682 * tasks mems_allowed, or node_states[N_MEMORY].)
1684 * If the zonelist cache is not available for this zonelist, does
1685 * nothing and returns NULL.
1687 * If the fullzones BITMAP in the zonelist cache is stale (more than
1688 * a second since last zap'd) then we zap it out (clear its bits.)
1690 * We hold off even calling zlc_setup, until after we've checked the
1691 * first zone in the zonelist, on the theory that most allocations will
1692 * be satisfied from that first zone, so best to examine that zone as
1693 * quickly as we can.
1695 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1697 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1698 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1700 zlc
= zonelist
->zlcache_ptr
;
1704 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1705 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1706 zlc
->last_full_zap
= jiffies
;
1709 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1710 &cpuset_current_mems_allowed
:
1711 &node_states
[N_MEMORY
];
1712 return allowednodes
;
1716 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1717 * if it is worth looking at further for free memory:
1718 * 1) Check that the zone isn't thought to be full (doesn't have its
1719 * bit set in the zonelist_cache fullzones BITMAP).
1720 * 2) Check that the zones node (obtained from the zonelist_cache
1721 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1722 * Return true (non-zero) if zone is worth looking at further, or
1723 * else return false (zero) if it is not.
1725 * This check -ignores- the distinction between various watermarks,
1726 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1727 * found to be full for any variation of these watermarks, it will
1728 * be considered full for up to one second by all requests, unless
1729 * we are so low on memory on all allowed nodes that we are forced
1730 * into the second scan of the zonelist.
1732 * In the second scan we ignore this zonelist cache and exactly
1733 * apply the watermarks to all zones, even it is slower to do so.
1734 * We are low on memory in the second scan, and should leave no stone
1735 * unturned looking for a free page.
1737 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1738 nodemask_t
*allowednodes
)
1740 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1741 int i
; /* index of *z in zonelist zones */
1742 int n
; /* node that zone *z is on */
1744 zlc
= zonelist
->zlcache_ptr
;
1748 i
= z
- zonelist
->_zonerefs
;
1751 /* This zone is worth trying if it is allowed but not full */
1752 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1756 * Given 'z' scanning a zonelist, set the corresponding bit in
1757 * zlc->fullzones, so that subsequent attempts to allocate a page
1758 * from that zone don't waste time re-examining it.
1760 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1762 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1763 int i
; /* index of *z in zonelist zones */
1765 zlc
= zonelist
->zlcache_ptr
;
1769 i
= z
- zonelist
->_zonerefs
;
1771 set_bit(i
, zlc
->fullzones
);
1775 * clear all zones full, called after direct reclaim makes progress so that
1776 * a zone that was recently full is not skipped over for up to a second
1778 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1780 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1782 zlc
= zonelist
->zlcache_ptr
;
1786 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1789 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1791 return node_isset(local_zone
->node
, zone
->zone_pgdat
->reclaim_nodes
);
1794 static void __paginginit
init_zone_allows_reclaim(int nid
)
1798 for_each_online_node(i
)
1799 if (node_distance(nid
, i
) <= RECLAIM_DISTANCE
)
1800 node_set(i
, NODE_DATA(nid
)->reclaim_nodes
);
1802 zone_reclaim_mode
= 1;
1805 #else /* CONFIG_NUMA */
1807 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1812 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1813 nodemask_t
*allowednodes
)
1818 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1822 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1826 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1831 static inline void init_zone_allows_reclaim(int nid
)
1834 #endif /* CONFIG_NUMA */
1837 * get_page_from_freelist goes through the zonelist trying to allocate
1840 static struct page
*
1841 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1842 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1843 struct zone
*preferred_zone
, int migratetype
)
1846 struct page
*page
= NULL
;
1849 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1850 int zlc_active
= 0; /* set if using zonelist_cache */
1851 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1853 classzone_idx
= zone_idx(preferred_zone
);
1856 * Scan zonelist, looking for a zone with enough free.
1857 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1859 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1860 high_zoneidx
, nodemask
) {
1861 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1862 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1864 if ((alloc_flags
& ALLOC_CPUSET
) &&
1865 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1868 * When allocating a page cache page for writing, we
1869 * want to get it from a zone that is within its dirty
1870 * limit, such that no single zone holds more than its
1871 * proportional share of globally allowed dirty pages.
1872 * The dirty limits take into account the zone's
1873 * lowmem reserves and high watermark so that kswapd
1874 * should be able to balance it without having to
1875 * write pages from its LRU list.
1877 * This may look like it could increase pressure on
1878 * lower zones by failing allocations in higher zones
1879 * before they are full. But the pages that do spill
1880 * over are limited as the lower zones are protected
1881 * by this very same mechanism. It should not become
1882 * a practical burden to them.
1884 * XXX: For now, allow allocations to potentially
1885 * exceed the per-zone dirty limit in the slowpath
1886 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1887 * which is important when on a NUMA setup the allowed
1888 * zones are together not big enough to reach the
1889 * global limit. The proper fix for these situations
1890 * will require awareness of zones in the
1891 * dirty-throttling and the flusher threads.
1893 if ((alloc_flags
& ALLOC_WMARK_LOW
) &&
1894 (gfp_mask
& __GFP_WRITE
) && !zone_dirty_ok(zone
))
1895 goto this_zone_full
;
1897 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1898 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1902 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1903 if (zone_watermark_ok(zone
, order
, mark
,
1904 classzone_idx
, alloc_flags
))
1907 if (IS_ENABLED(CONFIG_NUMA
) &&
1908 !did_zlc_setup
&& nr_online_nodes
> 1) {
1910 * we do zlc_setup if there are multiple nodes
1911 * and before considering the first zone allowed
1914 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1919 if (zone_reclaim_mode
== 0 ||
1920 !zone_allows_reclaim(preferred_zone
, zone
))
1921 goto this_zone_full
;
1924 * As we may have just activated ZLC, check if the first
1925 * eligible zone has failed zone_reclaim recently.
1927 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1928 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1931 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1933 case ZONE_RECLAIM_NOSCAN
:
1936 case ZONE_RECLAIM_FULL
:
1937 /* scanned but unreclaimable */
1940 /* did we reclaim enough */
1941 if (!zone_watermark_ok(zone
, order
, mark
,
1942 classzone_idx
, alloc_flags
))
1943 goto this_zone_full
;
1948 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1949 gfp_mask
, migratetype
);
1953 if (IS_ENABLED(CONFIG_NUMA
))
1954 zlc_mark_zone_full(zonelist
, z
);
1957 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && page
== NULL
&& zlc_active
)) {
1958 /* Disable zlc cache for second zonelist scan */
1965 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1966 * necessary to allocate the page. The expectation is
1967 * that the caller is taking steps that will free more
1968 * memory. The caller should avoid the page being used
1969 * for !PFMEMALLOC purposes.
1971 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
1977 * Large machines with many possible nodes should not always dump per-node
1978 * meminfo in irq context.
1980 static inline bool should_suppress_show_mem(void)
1985 ret
= in_interrupt();
1990 static DEFINE_RATELIMIT_STATE(nopage_rs
,
1991 DEFAULT_RATELIMIT_INTERVAL
,
1992 DEFAULT_RATELIMIT_BURST
);
1994 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
1996 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
1998 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
1999 debug_guardpage_minorder() > 0)
2003 * This documents exceptions given to allocations in certain
2004 * contexts that are allowed to allocate outside current's set
2007 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2008 if (test_thread_flag(TIF_MEMDIE
) ||
2009 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2010 filter
&= ~SHOW_MEM_FILTER_NODES
;
2011 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2012 filter
&= ~SHOW_MEM_FILTER_NODES
;
2015 struct va_format vaf
;
2018 va_start(args
, fmt
);
2023 pr_warn("%pV", &vaf
);
2028 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2029 current
->comm
, order
, gfp_mask
);
2032 if (!should_suppress_show_mem())
2037 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2038 unsigned long did_some_progress
,
2039 unsigned long pages_reclaimed
)
2041 /* Do not loop if specifically requested */
2042 if (gfp_mask
& __GFP_NORETRY
)
2045 /* Always retry if specifically requested */
2046 if (gfp_mask
& __GFP_NOFAIL
)
2050 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2051 * making forward progress without invoking OOM. Suspend also disables
2052 * storage devices so kswapd will not help. Bail if we are suspending.
2054 if (!did_some_progress
&& pm_suspended_storage())
2058 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2059 * means __GFP_NOFAIL, but that may not be true in other
2062 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2066 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2067 * specified, then we retry until we no longer reclaim any pages
2068 * (above), or we've reclaimed an order of pages at least as
2069 * large as the allocation's order. In both cases, if the
2070 * allocation still fails, we stop retrying.
2072 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2078 static inline struct page
*
2079 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2080 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2081 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2086 /* Acquire the OOM killer lock for the zones in zonelist */
2087 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
2088 schedule_timeout_uninterruptible(1);
2093 * Go through the zonelist yet one more time, keep very high watermark
2094 * here, this is only to catch a parallel oom killing, we must fail if
2095 * we're still under heavy pressure.
2097 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2098 order
, zonelist
, high_zoneidx
,
2099 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2100 preferred_zone
, migratetype
);
2104 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2105 /* The OOM killer will not help higher order allocs */
2106 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2108 /* The OOM killer does not needlessly kill tasks for lowmem */
2109 if (high_zoneidx
< ZONE_NORMAL
)
2112 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2113 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2114 * The caller should handle page allocation failure by itself if
2115 * it specifies __GFP_THISNODE.
2116 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2118 if (gfp_mask
& __GFP_THISNODE
)
2121 /* Exhausted what can be done so it's blamo time */
2122 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2125 clear_zonelist_oom(zonelist
, gfp_mask
);
2129 #ifdef CONFIG_COMPACTION
2130 /* Try memory compaction for high-order allocations before reclaim */
2131 static struct page
*
2132 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2133 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2134 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2135 int migratetype
, bool sync_migration
,
2136 bool *contended_compaction
, bool *deferred_compaction
,
2137 unsigned long *did_some_progress
)
2139 struct page
*page
= NULL
;
2144 if (compaction_deferred(preferred_zone
, order
)) {
2145 *deferred_compaction
= true;
2149 current
->flags
|= PF_MEMALLOC
;
2150 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2151 nodemask
, sync_migration
,
2152 contended_compaction
, &page
);
2153 current
->flags
&= ~PF_MEMALLOC
;
2155 /* If compaction captured a page, prep and use it */
2157 prep_new_page(page
, order
, gfp_mask
);
2161 if (*did_some_progress
!= COMPACT_SKIPPED
) {
2162 /* Page migration frees to the PCP lists but we want merging */
2163 drain_pages(get_cpu());
2166 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2167 order
, zonelist
, high_zoneidx
,
2168 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2169 preferred_zone
, migratetype
);
2172 preferred_zone
->compact_blockskip_flush
= false;
2173 preferred_zone
->compact_considered
= 0;
2174 preferred_zone
->compact_defer_shift
= 0;
2175 if (order
>= preferred_zone
->compact_order_failed
)
2176 preferred_zone
->compact_order_failed
= order
+ 1;
2177 count_vm_event(COMPACTSUCCESS
);
2182 * It's bad if compaction run occurs and fails.
2183 * The most likely reason is that pages exist,
2184 * but not enough to satisfy watermarks.
2186 count_vm_event(COMPACTFAIL
);
2189 * As async compaction considers a subset of pageblocks, only
2190 * defer if the failure was a sync compaction failure.
2193 defer_compaction(preferred_zone
, order
);
2201 static inline struct page
*
2202 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2203 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2204 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2205 int migratetype
, bool sync_migration
,
2206 bool *contended_compaction
, bool *deferred_compaction
,
2207 unsigned long *did_some_progress
)
2211 #endif /* CONFIG_COMPACTION */
2213 /* Perform direct synchronous page reclaim */
2215 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2216 nodemask_t
*nodemask
)
2218 struct reclaim_state reclaim_state
;
2223 /* We now go into synchronous reclaim */
2224 cpuset_memory_pressure_bump();
2225 current
->flags
|= PF_MEMALLOC
;
2226 lockdep_set_current_reclaim_state(gfp_mask
);
2227 reclaim_state
.reclaimed_slab
= 0;
2228 current
->reclaim_state
= &reclaim_state
;
2230 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2232 current
->reclaim_state
= NULL
;
2233 lockdep_clear_current_reclaim_state();
2234 current
->flags
&= ~PF_MEMALLOC
;
2241 /* The really slow allocator path where we enter direct reclaim */
2242 static inline struct page
*
2243 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2244 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2245 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2246 int migratetype
, unsigned long *did_some_progress
)
2248 struct page
*page
= NULL
;
2249 bool drained
= false;
2251 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2253 if (unlikely(!(*did_some_progress
)))
2256 /* After successful reclaim, reconsider all zones for allocation */
2257 if (IS_ENABLED(CONFIG_NUMA
))
2258 zlc_clear_zones_full(zonelist
);
2261 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2262 zonelist
, high_zoneidx
,
2263 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2264 preferred_zone
, migratetype
);
2267 * If an allocation failed after direct reclaim, it could be because
2268 * pages are pinned on the per-cpu lists. Drain them and try again
2270 if (!page
&& !drained
) {
2280 * This is called in the allocator slow-path if the allocation request is of
2281 * sufficient urgency to ignore watermarks and take other desperate measures
2283 static inline struct page
*
2284 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2285 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2286 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2292 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2293 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2294 preferred_zone
, migratetype
);
2296 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2297 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2298 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2304 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
2305 enum zone_type high_zoneidx
,
2306 enum zone_type classzone_idx
)
2311 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
2312 wakeup_kswapd(zone
, order
, classzone_idx
);
2316 gfp_to_alloc_flags(gfp_t gfp_mask
)
2318 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2319 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2321 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2322 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2325 * The caller may dip into page reserves a bit more if the caller
2326 * cannot run direct reclaim, or if the caller has realtime scheduling
2327 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2328 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2330 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2334 * Not worth trying to allocate harder for
2335 * __GFP_NOMEMALLOC even if it can't schedule.
2337 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2338 alloc_flags
|= ALLOC_HARDER
;
2340 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2341 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2343 alloc_flags
&= ~ALLOC_CPUSET
;
2344 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2345 alloc_flags
|= ALLOC_HARDER
;
2347 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2348 if (gfp_mask
& __GFP_MEMALLOC
)
2349 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2350 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2351 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2352 else if (!in_interrupt() &&
2353 ((current
->flags
& PF_MEMALLOC
) ||
2354 unlikely(test_thread_flag(TIF_MEMDIE
))))
2355 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2358 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2359 alloc_flags
|= ALLOC_CMA
;
2364 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2366 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2369 static inline struct page
*
2370 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2371 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2372 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2375 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2376 struct page
*page
= NULL
;
2378 unsigned long pages_reclaimed
= 0;
2379 unsigned long did_some_progress
;
2380 bool sync_migration
= false;
2381 bool deferred_compaction
= false;
2382 bool contended_compaction
= false;
2385 * In the slowpath, we sanity check order to avoid ever trying to
2386 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2387 * be using allocators in order of preference for an area that is
2390 if (order
>= MAX_ORDER
) {
2391 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2396 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2397 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2398 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2399 * using a larger set of nodes after it has established that the
2400 * allowed per node queues are empty and that nodes are
2403 if (IS_ENABLED(CONFIG_NUMA
) &&
2404 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2408 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2409 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2410 zone_idx(preferred_zone
));
2413 * OK, we're below the kswapd watermark and have kicked background
2414 * reclaim. Now things get more complex, so set up alloc_flags according
2415 * to how we want to proceed.
2417 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2420 * Find the true preferred zone if the allocation is unconstrained by
2423 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2424 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2428 /* This is the last chance, in general, before the goto nopage. */
2429 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2430 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2431 preferred_zone
, migratetype
);
2435 /* Allocate without watermarks if the context allows */
2436 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2438 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2439 * the allocation is high priority and these type of
2440 * allocations are system rather than user orientated
2442 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2444 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2445 zonelist
, high_zoneidx
, nodemask
,
2446 preferred_zone
, migratetype
);
2452 /* Atomic allocations - we can't balance anything */
2456 /* Avoid recursion of direct reclaim */
2457 if (current
->flags
& PF_MEMALLOC
)
2460 /* Avoid allocations with no watermarks from looping endlessly */
2461 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2465 * Try direct compaction. The first pass is asynchronous. Subsequent
2466 * attempts after direct reclaim are synchronous
2468 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2469 zonelist
, high_zoneidx
,
2471 alloc_flags
, preferred_zone
,
2472 migratetype
, sync_migration
,
2473 &contended_compaction
,
2474 &deferred_compaction
,
2475 &did_some_progress
);
2478 sync_migration
= true;
2481 * If compaction is deferred for high-order allocations, it is because
2482 * sync compaction recently failed. In this is the case and the caller
2483 * requested a movable allocation that does not heavily disrupt the
2484 * system then fail the allocation instead of entering direct reclaim.
2486 if ((deferred_compaction
|| contended_compaction
) &&
2487 (gfp_mask
& __GFP_NO_KSWAPD
))
2490 /* Try direct reclaim and then allocating */
2491 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2492 zonelist
, high_zoneidx
,
2494 alloc_flags
, preferred_zone
,
2495 migratetype
, &did_some_progress
);
2500 * If we failed to make any progress reclaiming, then we are
2501 * running out of options and have to consider going OOM
2503 if (!did_some_progress
) {
2504 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2505 if (oom_killer_disabled
)
2507 /* Coredumps can quickly deplete all memory reserves */
2508 if ((current
->flags
& PF_DUMPCORE
) &&
2509 !(gfp_mask
& __GFP_NOFAIL
))
2511 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2512 zonelist
, high_zoneidx
,
2513 nodemask
, preferred_zone
,
2518 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2520 * The oom killer is not called for high-order
2521 * allocations that may fail, so if no progress
2522 * is being made, there are no other options and
2523 * retrying is unlikely to help.
2525 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2528 * The oom killer is not called for lowmem
2529 * allocations to prevent needlessly killing
2532 if (high_zoneidx
< ZONE_NORMAL
)
2540 /* Check if we should retry the allocation */
2541 pages_reclaimed
+= did_some_progress
;
2542 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2544 /* Wait for some write requests to complete then retry */
2545 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2549 * High-order allocations do not necessarily loop after
2550 * direct reclaim and reclaim/compaction depends on compaction
2551 * being called after reclaim so call directly if necessary
2553 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2554 zonelist
, high_zoneidx
,
2556 alloc_flags
, preferred_zone
,
2557 migratetype
, sync_migration
,
2558 &contended_compaction
,
2559 &deferred_compaction
,
2560 &did_some_progress
);
2566 warn_alloc_failed(gfp_mask
, order
, NULL
);
2569 if (kmemcheck_enabled
)
2570 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2576 * This is the 'heart' of the zoned buddy allocator.
2579 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2580 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2582 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2583 struct zone
*preferred_zone
;
2584 struct page
*page
= NULL
;
2585 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2586 unsigned int cpuset_mems_cookie
;
2587 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
;
2588 struct mem_cgroup
*memcg
= NULL
;
2590 gfp_mask
&= gfp_allowed_mask
;
2592 lockdep_trace_alloc(gfp_mask
);
2594 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2596 if (should_fail_alloc_page(gfp_mask
, order
))
2600 * Check the zones suitable for the gfp_mask contain at least one
2601 * valid zone. It's possible to have an empty zonelist as a result
2602 * of GFP_THISNODE and a memoryless node
2604 if (unlikely(!zonelist
->_zonerefs
->zone
))
2608 * Will only have any effect when __GFP_KMEMCG is set. This is
2609 * verified in the (always inline) callee
2611 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2615 cpuset_mems_cookie
= get_mems_allowed();
2617 /* The preferred zone is used for statistics later */
2618 first_zones_zonelist(zonelist
, high_zoneidx
,
2619 nodemask
? : &cpuset_current_mems_allowed
,
2621 if (!preferred_zone
)
2625 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2626 alloc_flags
|= ALLOC_CMA
;
2628 /* First allocation attempt */
2629 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2630 zonelist
, high_zoneidx
, alloc_flags
,
2631 preferred_zone
, migratetype
);
2632 if (unlikely(!page
))
2633 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2634 zonelist
, high_zoneidx
, nodemask
,
2635 preferred_zone
, migratetype
);
2637 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2641 * When updating a task's mems_allowed, it is possible to race with
2642 * parallel threads in such a way that an allocation can fail while
2643 * the mask is being updated. If a page allocation is about to fail,
2644 * check if the cpuset changed during allocation and if so, retry.
2646 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2649 memcg_kmem_commit_charge(page
, memcg
, order
);
2653 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2656 * Common helper functions.
2658 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2663 * __get_free_pages() returns a 32-bit address, which cannot represent
2666 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2668 page
= alloc_pages(gfp_mask
, order
);
2671 return (unsigned long) page_address(page
);
2673 EXPORT_SYMBOL(__get_free_pages
);
2675 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2677 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2679 EXPORT_SYMBOL(get_zeroed_page
);
2681 void __free_pages(struct page
*page
, unsigned int order
)
2683 if (put_page_testzero(page
)) {
2685 free_hot_cold_page(page
, 0);
2687 __free_pages_ok(page
, order
);
2691 EXPORT_SYMBOL(__free_pages
);
2693 void free_pages(unsigned long addr
, unsigned int order
)
2696 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2697 __free_pages(virt_to_page((void *)addr
), order
);
2701 EXPORT_SYMBOL(free_pages
);
2704 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2705 * pages allocated with __GFP_KMEMCG.
2707 * Those pages are accounted to a particular memcg, embedded in the
2708 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2709 * for that information only to find out that it is NULL for users who have no
2710 * interest in that whatsoever, we provide these functions.
2712 * The caller knows better which flags it relies on.
2714 void __free_memcg_kmem_pages(struct page
*page
, unsigned int order
)
2716 memcg_kmem_uncharge_pages(page
, order
);
2717 __free_pages(page
, order
);
2720 void free_memcg_kmem_pages(unsigned long addr
, unsigned int order
)
2723 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2724 __free_memcg_kmem_pages(virt_to_page((void *)addr
), order
);
2728 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2731 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2732 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2734 split_page(virt_to_page((void *)addr
), order
);
2735 while (used
< alloc_end
) {
2740 return (void *)addr
;
2744 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2745 * @size: the number of bytes to allocate
2746 * @gfp_mask: GFP flags for the allocation
2748 * This function is similar to alloc_pages(), except that it allocates the
2749 * minimum number of pages to satisfy the request. alloc_pages() can only
2750 * allocate memory in power-of-two pages.
2752 * This function is also limited by MAX_ORDER.
2754 * Memory allocated by this function must be released by free_pages_exact().
2756 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2758 unsigned int order
= get_order(size
);
2761 addr
= __get_free_pages(gfp_mask
, order
);
2762 return make_alloc_exact(addr
, order
, size
);
2764 EXPORT_SYMBOL(alloc_pages_exact
);
2767 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2769 * @nid: the preferred node ID where memory should be allocated
2770 * @size: the number of bytes to allocate
2771 * @gfp_mask: GFP flags for the allocation
2773 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2775 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2778 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2780 unsigned order
= get_order(size
);
2781 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2784 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2786 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2789 * free_pages_exact - release memory allocated via alloc_pages_exact()
2790 * @virt: the value returned by alloc_pages_exact.
2791 * @size: size of allocation, same value as passed to alloc_pages_exact().
2793 * Release the memory allocated by a previous call to alloc_pages_exact.
2795 void free_pages_exact(void *virt
, size_t size
)
2797 unsigned long addr
= (unsigned long)virt
;
2798 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2800 while (addr
< end
) {
2805 EXPORT_SYMBOL(free_pages_exact
);
2807 static unsigned int nr_free_zone_pages(int offset
)
2812 /* Just pick one node, since fallback list is circular */
2813 unsigned int sum
= 0;
2815 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2817 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2818 unsigned long size
= zone
->present_pages
;
2819 unsigned long high
= high_wmark_pages(zone
);
2828 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2830 unsigned int nr_free_buffer_pages(void)
2832 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2834 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2837 * Amount of free RAM allocatable within all zones
2839 unsigned int nr_free_pagecache_pages(void)
2841 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2844 static inline void show_node(struct zone
*zone
)
2846 if (IS_ENABLED(CONFIG_NUMA
))
2847 printk("Node %d ", zone_to_nid(zone
));
2850 void si_meminfo(struct sysinfo
*val
)
2852 val
->totalram
= totalram_pages
;
2854 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2855 val
->bufferram
= nr_blockdev_pages();
2856 val
->totalhigh
= totalhigh_pages
;
2857 val
->freehigh
= nr_free_highpages();
2858 val
->mem_unit
= PAGE_SIZE
;
2861 EXPORT_SYMBOL(si_meminfo
);
2864 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2866 pg_data_t
*pgdat
= NODE_DATA(nid
);
2868 val
->totalram
= pgdat
->node_present_pages
;
2869 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2870 #ifdef CONFIG_HIGHMEM
2871 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].present_pages
;
2872 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2878 val
->mem_unit
= PAGE_SIZE
;
2883 * Determine whether the node should be displayed or not, depending on whether
2884 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2886 bool skip_free_areas_node(unsigned int flags
, int nid
)
2889 unsigned int cpuset_mems_cookie
;
2891 if (!(flags
& SHOW_MEM_FILTER_NODES
))
2895 cpuset_mems_cookie
= get_mems_allowed();
2896 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
2897 } while (!put_mems_allowed(cpuset_mems_cookie
));
2902 #define K(x) ((x) << (PAGE_SHIFT-10))
2904 static void show_migration_types(unsigned char type
)
2906 static const char types
[MIGRATE_TYPES
] = {
2907 [MIGRATE_UNMOVABLE
] = 'U',
2908 [MIGRATE_RECLAIMABLE
] = 'E',
2909 [MIGRATE_MOVABLE
] = 'M',
2910 [MIGRATE_RESERVE
] = 'R',
2912 [MIGRATE_CMA
] = 'C',
2914 [MIGRATE_ISOLATE
] = 'I',
2916 char tmp
[MIGRATE_TYPES
+ 1];
2920 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
2921 if (type
& (1 << i
))
2926 printk("(%s) ", tmp
);
2930 * Show free area list (used inside shift_scroll-lock stuff)
2931 * We also calculate the percentage fragmentation. We do this by counting the
2932 * memory on each free list with the exception of the first item on the list.
2933 * Suppresses nodes that are not allowed by current's cpuset if
2934 * SHOW_MEM_FILTER_NODES is passed.
2936 void show_free_areas(unsigned int filter
)
2941 for_each_populated_zone(zone
) {
2942 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2945 printk("%s per-cpu:\n", zone
->name
);
2947 for_each_online_cpu(cpu
) {
2948 struct per_cpu_pageset
*pageset
;
2950 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2952 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2953 cpu
, pageset
->pcp
.high
,
2954 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2958 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2959 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2961 " dirty:%lu writeback:%lu unstable:%lu\n"
2962 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2963 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2965 global_page_state(NR_ACTIVE_ANON
),
2966 global_page_state(NR_INACTIVE_ANON
),
2967 global_page_state(NR_ISOLATED_ANON
),
2968 global_page_state(NR_ACTIVE_FILE
),
2969 global_page_state(NR_INACTIVE_FILE
),
2970 global_page_state(NR_ISOLATED_FILE
),
2971 global_page_state(NR_UNEVICTABLE
),
2972 global_page_state(NR_FILE_DIRTY
),
2973 global_page_state(NR_WRITEBACK
),
2974 global_page_state(NR_UNSTABLE_NFS
),
2975 global_page_state(NR_FREE_PAGES
),
2976 global_page_state(NR_SLAB_RECLAIMABLE
),
2977 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2978 global_page_state(NR_FILE_MAPPED
),
2979 global_page_state(NR_SHMEM
),
2980 global_page_state(NR_PAGETABLE
),
2981 global_page_state(NR_BOUNCE
),
2982 global_page_state(NR_FREE_CMA_PAGES
));
2984 for_each_populated_zone(zone
) {
2987 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2995 " active_anon:%lukB"
2996 " inactive_anon:%lukB"
2997 " active_file:%lukB"
2998 " inactive_file:%lukB"
2999 " unevictable:%lukB"
3000 " isolated(anon):%lukB"
3001 " isolated(file):%lukB"
3009 " slab_reclaimable:%lukB"
3010 " slab_unreclaimable:%lukB"
3011 " kernel_stack:%lukB"
3016 " writeback_tmp:%lukB"
3017 " pages_scanned:%lu"
3018 " all_unreclaimable? %s"
3021 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3022 K(min_wmark_pages(zone
)),
3023 K(low_wmark_pages(zone
)),
3024 K(high_wmark_pages(zone
)),
3025 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3026 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3027 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3028 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3029 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3030 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3031 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3032 K(zone
->present_pages
),
3033 K(zone
->managed_pages
),
3034 K(zone_page_state(zone
, NR_MLOCK
)),
3035 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3036 K(zone_page_state(zone
, NR_WRITEBACK
)),
3037 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3038 K(zone_page_state(zone
, NR_SHMEM
)),
3039 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3040 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3041 zone_page_state(zone
, NR_KERNEL_STACK
) *
3043 K(zone_page_state(zone
, NR_PAGETABLE
)),
3044 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3045 K(zone_page_state(zone
, NR_BOUNCE
)),
3046 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3047 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3048 zone
->pages_scanned
,
3049 (zone
->all_unreclaimable
? "yes" : "no")
3051 printk("lowmem_reserve[]:");
3052 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3053 printk(" %lu", zone
->lowmem_reserve
[i
]);
3057 for_each_populated_zone(zone
) {
3058 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3059 unsigned char types
[MAX_ORDER
];
3061 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3064 printk("%s: ", zone
->name
);
3066 spin_lock_irqsave(&zone
->lock
, flags
);
3067 for (order
= 0; order
< MAX_ORDER
; order
++) {
3068 struct free_area
*area
= &zone
->free_area
[order
];
3071 nr
[order
] = area
->nr_free
;
3072 total
+= nr
[order
] << order
;
3075 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3076 if (!list_empty(&area
->free_list
[type
]))
3077 types
[order
] |= 1 << type
;
3080 spin_unlock_irqrestore(&zone
->lock
, flags
);
3081 for (order
= 0; order
< MAX_ORDER
; order
++) {
3082 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3084 show_migration_types(types
[order
]);
3086 printk("= %lukB\n", K(total
));
3089 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3091 show_swap_cache_info();
3094 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3096 zoneref
->zone
= zone
;
3097 zoneref
->zone_idx
= zone_idx(zone
);
3101 * Builds allocation fallback zone lists.
3103 * Add all populated zones of a node to the zonelist.
3105 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3106 int nr_zones
, enum zone_type zone_type
)
3110 BUG_ON(zone_type
>= MAX_NR_ZONES
);
3115 zone
= pgdat
->node_zones
+ zone_type
;
3116 if (populated_zone(zone
)) {
3117 zoneref_set_zone(zone
,
3118 &zonelist
->_zonerefs
[nr_zones
++]);
3119 check_highest_zone(zone_type
);
3122 } while (zone_type
);
3129 * 0 = automatic detection of better ordering.
3130 * 1 = order by ([node] distance, -zonetype)
3131 * 2 = order by (-zonetype, [node] distance)
3133 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3134 * the same zonelist. So only NUMA can configure this param.
3136 #define ZONELIST_ORDER_DEFAULT 0
3137 #define ZONELIST_ORDER_NODE 1
3138 #define ZONELIST_ORDER_ZONE 2
3140 /* zonelist order in the kernel.
3141 * set_zonelist_order() will set this to NODE or ZONE.
3143 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3144 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3148 /* The value user specified ....changed by config */
3149 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3150 /* string for sysctl */
3151 #define NUMA_ZONELIST_ORDER_LEN 16
3152 char numa_zonelist_order
[16] = "default";
3155 * interface for configure zonelist ordering.
3156 * command line option "numa_zonelist_order"
3157 * = "[dD]efault - default, automatic configuration.
3158 * = "[nN]ode - order by node locality, then by zone within node
3159 * = "[zZ]one - order by zone, then by locality within zone
3162 static int __parse_numa_zonelist_order(char *s
)
3164 if (*s
== 'd' || *s
== 'D') {
3165 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3166 } else if (*s
== 'n' || *s
== 'N') {
3167 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3168 } else if (*s
== 'z' || *s
== 'Z') {
3169 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3172 "Ignoring invalid numa_zonelist_order value: "
3179 static __init
int setup_numa_zonelist_order(char *s
)
3186 ret
= __parse_numa_zonelist_order(s
);
3188 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3192 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3195 * sysctl handler for numa_zonelist_order
3197 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
3198 void __user
*buffer
, size_t *length
,
3201 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3203 static DEFINE_MUTEX(zl_order_mutex
);
3205 mutex_lock(&zl_order_mutex
);
3207 strcpy(saved_string
, (char*)table
->data
);
3208 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3212 int oldval
= user_zonelist_order
;
3213 if (__parse_numa_zonelist_order((char*)table
->data
)) {
3215 * bogus value. restore saved string
3217 strncpy((char*)table
->data
, saved_string
,
3218 NUMA_ZONELIST_ORDER_LEN
);
3219 user_zonelist_order
= oldval
;
3220 } else if (oldval
!= user_zonelist_order
) {
3221 mutex_lock(&zonelists_mutex
);
3222 build_all_zonelists(NULL
, NULL
);
3223 mutex_unlock(&zonelists_mutex
);
3227 mutex_unlock(&zl_order_mutex
);
3232 #define MAX_NODE_LOAD (nr_online_nodes)
3233 static int node_load
[MAX_NUMNODES
];
3236 * find_next_best_node - find the next node that should appear in a given node's fallback list
3237 * @node: node whose fallback list we're appending
3238 * @used_node_mask: nodemask_t of already used nodes
3240 * We use a number of factors to determine which is the next node that should
3241 * appear on a given node's fallback list. The node should not have appeared
3242 * already in @node's fallback list, and it should be the next closest node
3243 * according to the distance array (which contains arbitrary distance values
3244 * from each node to each node in the system), and should also prefer nodes
3245 * with no CPUs, since presumably they'll have very little allocation pressure
3246 * on them otherwise.
3247 * It returns -1 if no node is found.
3249 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3252 int min_val
= INT_MAX
;
3254 const struct cpumask
*tmp
= cpumask_of_node(0);
3256 /* Use the local node if we haven't already */
3257 if (!node_isset(node
, *used_node_mask
)) {
3258 node_set(node
, *used_node_mask
);
3262 for_each_node_state(n
, N_MEMORY
) {
3264 /* Don't want a node to appear more than once */
3265 if (node_isset(n
, *used_node_mask
))
3268 /* Use the distance array to find the distance */
3269 val
= node_distance(node
, n
);
3271 /* Penalize nodes under us ("prefer the next node") */
3274 /* Give preference to headless and unused nodes */
3275 tmp
= cpumask_of_node(n
);
3276 if (!cpumask_empty(tmp
))
3277 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3279 /* Slight preference for less loaded node */
3280 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3281 val
+= node_load
[n
];
3283 if (val
< min_val
) {
3290 node_set(best_node
, *used_node_mask
);
3297 * Build zonelists ordered by node and zones within node.
3298 * This results in maximum locality--normal zone overflows into local
3299 * DMA zone, if any--but risks exhausting DMA zone.
3301 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3304 struct zonelist
*zonelist
;
3306 zonelist
= &pgdat
->node_zonelists
[0];
3307 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3309 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3311 zonelist
->_zonerefs
[j
].zone
= NULL
;
3312 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3316 * Build gfp_thisnode zonelists
3318 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3321 struct zonelist
*zonelist
;
3323 zonelist
= &pgdat
->node_zonelists
[1];
3324 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3325 zonelist
->_zonerefs
[j
].zone
= NULL
;
3326 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3330 * Build zonelists ordered by zone and nodes within zones.
3331 * This results in conserving DMA zone[s] until all Normal memory is
3332 * exhausted, but results in overflowing to remote node while memory
3333 * may still exist in local DMA zone.
3335 static int node_order
[MAX_NUMNODES
];
3337 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3340 int zone_type
; /* needs to be signed */
3342 struct zonelist
*zonelist
;
3344 zonelist
= &pgdat
->node_zonelists
[0];
3346 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3347 for (j
= 0; j
< nr_nodes
; j
++) {
3348 node
= node_order
[j
];
3349 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3350 if (populated_zone(z
)) {
3352 &zonelist
->_zonerefs
[pos
++]);
3353 check_highest_zone(zone_type
);
3357 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3358 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3361 static int default_zonelist_order(void)
3364 unsigned long low_kmem_size
,total_size
;
3368 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3369 * If they are really small and used heavily, the system can fall
3370 * into OOM very easily.
3371 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3373 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3376 for_each_online_node(nid
) {
3377 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3378 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3379 if (populated_zone(z
)) {
3380 if (zone_type
< ZONE_NORMAL
)
3381 low_kmem_size
+= z
->present_pages
;
3382 total_size
+= z
->present_pages
;
3383 } else if (zone_type
== ZONE_NORMAL
) {
3385 * If any node has only lowmem, then node order
3386 * is preferred to allow kernel allocations
3387 * locally; otherwise, they can easily infringe
3388 * on other nodes when there is an abundance of
3389 * lowmem available to allocate from.
3391 return ZONELIST_ORDER_NODE
;
3395 if (!low_kmem_size
|| /* there are no DMA area. */
3396 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
3397 return ZONELIST_ORDER_NODE
;
3399 * look into each node's config.
3400 * If there is a node whose DMA/DMA32 memory is very big area on
3401 * local memory, NODE_ORDER may be suitable.
3403 average_size
= total_size
/
3404 (nodes_weight(node_states
[N_MEMORY
]) + 1);
3405 for_each_online_node(nid
) {
3408 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3409 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3410 if (populated_zone(z
)) {
3411 if (zone_type
< ZONE_NORMAL
)
3412 low_kmem_size
+= z
->present_pages
;
3413 total_size
+= z
->present_pages
;
3416 if (low_kmem_size
&&
3417 total_size
> average_size
&& /* ignore small node */
3418 low_kmem_size
> total_size
* 70/100)
3419 return ZONELIST_ORDER_NODE
;
3421 return ZONELIST_ORDER_ZONE
;
3424 static void set_zonelist_order(void)
3426 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3427 current_zonelist_order
= default_zonelist_order();
3429 current_zonelist_order
= user_zonelist_order
;
3432 static void build_zonelists(pg_data_t
*pgdat
)
3436 nodemask_t used_mask
;
3437 int local_node
, prev_node
;
3438 struct zonelist
*zonelist
;
3439 int order
= current_zonelist_order
;
3441 /* initialize zonelists */
3442 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3443 zonelist
= pgdat
->node_zonelists
+ i
;
3444 zonelist
->_zonerefs
[0].zone
= NULL
;
3445 zonelist
->_zonerefs
[0].zone_idx
= 0;
3448 /* NUMA-aware ordering of nodes */
3449 local_node
= pgdat
->node_id
;
3450 load
= nr_online_nodes
;
3451 prev_node
= local_node
;
3452 nodes_clear(used_mask
);
3454 memset(node_order
, 0, sizeof(node_order
));
3457 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3459 * We don't want to pressure a particular node.
3460 * So adding penalty to the first node in same
3461 * distance group to make it round-robin.
3463 if (node_distance(local_node
, node
) !=
3464 node_distance(local_node
, prev_node
))
3465 node_load
[node
] = load
;
3469 if (order
== ZONELIST_ORDER_NODE
)
3470 build_zonelists_in_node_order(pgdat
, node
);
3472 node_order
[j
++] = node
; /* remember order */
3475 if (order
== ZONELIST_ORDER_ZONE
) {
3476 /* calculate node order -- i.e., DMA last! */
3477 build_zonelists_in_zone_order(pgdat
, j
);
3480 build_thisnode_zonelists(pgdat
);
3483 /* Construct the zonelist performance cache - see further mmzone.h */
3484 static void build_zonelist_cache(pg_data_t
*pgdat
)
3486 struct zonelist
*zonelist
;
3487 struct zonelist_cache
*zlc
;
3490 zonelist
= &pgdat
->node_zonelists
[0];
3491 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3492 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3493 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3494 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3497 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3499 * Return node id of node used for "local" allocations.
3500 * I.e., first node id of first zone in arg node's generic zonelist.
3501 * Used for initializing percpu 'numa_mem', which is used primarily
3502 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3504 int local_memory_node(int node
)
3508 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3509 gfp_zone(GFP_KERNEL
),
3516 #else /* CONFIG_NUMA */
3518 static void set_zonelist_order(void)
3520 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3523 static void build_zonelists(pg_data_t
*pgdat
)
3525 int node
, local_node
;
3527 struct zonelist
*zonelist
;
3529 local_node
= pgdat
->node_id
;
3531 zonelist
= &pgdat
->node_zonelists
[0];
3532 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3535 * Now we build the zonelist so that it contains the zones
3536 * of all the other nodes.
3537 * We don't want to pressure a particular node, so when
3538 * building the zones for node N, we make sure that the
3539 * zones coming right after the local ones are those from
3540 * node N+1 (modulo N)
3542 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3543 if (!node_online(node
))
3545 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3548 for (node
= 0; node
< local_node
; node
++) {
3549 if (!node_online(node
))
3551 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3555 zonelist
->_zonerefs
[j
].zone
= NULL
;
3556 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3559 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3560 static void build_zonelist_cache(pg_data_t
*pgdat
)
3562 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3565 #endif /* CONFIG_NUMA */
3568 * Boot pageset table. One per cpu which is going to be used for all
3569 * zones and all nodes. The parameters will be set in such a way
3570 * that an item put on a list will immediately be handed over to
3571 * the buddy list. This is safe since pageset manipulation is done
3572 * with interrupts disabled.
3574 * The boot_pagesets must be kept even after bootup is complete for
3575 * unused processors and/or zones. They do play a role for bootstrapping
3576 * hotplugged processors.
3578 * zoneinfo_show() and maybe other functions do
3579 * not check if the processor is online before following the pageset pointer.
3580 * Other parts of the kernel may not check if the zone is available.
3582 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3583 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3584 static void setup_zone_pageset(struct zone
*zone
);
3587 * Global mutex to protect against size modification of zonelists
3588 * as well as to serialize pageset setup for the new populated zone.
3590 DEFINE_MUTEX(zonelists_mutex
);
3592 /* return values int ....just for stop_machine() */
3593 static int __build_all_zonelists(void *data
)
3597 pg_data_t
*self
= data
;
3600 memset(node_load
, 0, sizeof(node_load
));
3603 if (self
&& !node_online(self
->node_id
)) {
3604 build_zonelists(self
);
3605 build_zonelist_cache(self
);
3608 for_each_online_node(nid
) {
3609 pg_data_t
*pgdat
= NODE_DATA(nid
);
3611 build_zonelists(pgdat
);
3612 build_zonelist_cache(pgdat
);
3616 * Initialize the boot_pagesets that are going to be used
3617 * for bootstrapping processors. The real pagesets for
3618 * each zone will be allocated later when the per cpu
3619 * allocator is available.
3621 * boot_pagesets are used also for bootstrapping offline
3622 * cpus if the system is already booted because the pagesets
3623 * are needed to initialize allocators on a specific cpu too.
3624 * F.e. the percpu allocator needs the page allocator which
3625 * needs the percpu allocator in order to allocate its pagesets
3626 * (a chicken-egg dilemma).
3628 for_each_possible_cpu(cpu
) {
3629 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3631 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3633 * We now know the "local memory node" for each node--
3634 * i.e., the node of the first zone in the generic zonelist.
3635 * Set up numa_mem percpu variable for on-line cpus. During
3636 * boot, only the boot cpu should be on-line; we'll init the
3637 * secondary cpus' numa_mem as they come on-line. During
3638 * node/memory hotplug, we'll fixup all on-line cpus.
3640 if (cpu_online(cpu
))
3641 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3649 * Called with zonelists_mutex held always
3650 * unless system_state == SYSTEM_BOOTING.
3652 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3654 set_zonelist_order();
3656 if (system_state
== SYSTEM_BOOTING
) {
3657 __build_all_zonelists(NULL
);
3658 mminit_verify_zonelist();
3659 cpuset_init_current_mems_allowed();
3661 /* we have to stop all cpus to guarantee there is no user
3663 #ifdef CONFIG_MEMORY_HOTPLUG
3665 setup_zone_pageset(zone
);
3667 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3668 /* cpuset refresh routine should be here */
3670 vm_total_pages
= nr_free_pagecache_pages();
3672 * Disable grouping by mobility if the number of pages in the
3673 * system is too low to allow the mechanism to work. It would be
3674 * more accurate, but expensive to check per-zone. This check is
3675 * made on memory-hotadd so a system can start with mobility
3676 * disabled and enable it later
3678 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3679 page_group_by_mobility_disabled
= 1;
3681 page_group_by_mobility_disabled
= 0;
3683 printk("Built %i zonelists in %s order, mobility grouping %s. "
3684 "Total pages: %ld\n",
3686 zonelist_order_name
[current_zonelist_order
],
3687 page_group_by_mobility_disabled
? "off" : "on",
3690 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3695 * Helper functions to size the waitqueue hash table.
3696 * Essentially these want to choose hash table sizes sufficiently
3697 * large so that collisions trying to wait on pages are rare.
3698 * But in fact, the number of active page waitqueues on typical
3699 * systems is ridiculously low, less than 200. So this is even
3700 * conservative, even though it seems large.
3702 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3703 * waitqueues, i.e. the size of the waitq table given the number of pages.
3705 #define PAGES_PER_WAITQUEUE 256
3707 #ifndef CONFIG_MEMORY_HOTPLUG
3708 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3710 unsigned long size
= 1;
3712 pages
/= PAGES_PER_WAITQUEUE
;
3714 while (size
< pages
)
3718 * Once we have dozens or even hundreds of threads sleeping
3719 * on IO we've got bigger problems than wait queue collision.
3720 * Limit the size of the wait table to a reasonable size.
3722 size
= min(size
, 4096UL);
3724 return max(size
, 4UL);
3728 * A zone's size might be changed by hot-add, so it is not possible to determine
3729 * a suitable size for its wait_table. So we use the maximum size now.
3731 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3733 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3734 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3735 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3737 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3738 * or more by the traditional way. (See above). It equals:
3740 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3741 * ia64(16K page size) : = ( 8G + 4M)byte.
3742 * powerpc (64K page size) : = (32G +16M)byte.
3744 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3751 * This is an integer logarithm so that shifts can be used later
3752 * to extract the more random high bits from the multiplicative
3753 * hash function before the remainder is taken.
3755 static inline unsigned long wait_table_bits(unsigned long size
)
3760 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3763 * Check if a pageblock contains reserved pages
3765 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3769 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3770 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3777 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3778 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3779 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3780 * higher will lead to a bigger reserve which will get freed as contiguous
3781 * blocks as reclaim kicks in
3783 static void setup_zone_migrate_reserve(struct zone
*zone
)
3785 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3787 unsigned long block_migratetype
;
3791 * Get the start pfn, end pfn and the number of blocks to reserve
3792 * We have to be careful to be aligned to pageblock_nr_pages to
3793 * make sure that we always check pfn_valid for the first page in
3796 start_pfn
= zone
->zone_start_pfn
;
3797 end_pfn
= start_pfn
+ zone
->spanned_pages
;
3798 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
3799 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3803 * Reserve blocks are generally in place to help high-order atomic
3804 * allocations that are short-lived. A min_free_kbytes value that
3805 * would result in more than 2 reserve blocks for atomic allocations
3806 * is assumed to be in place to help anti-fragmentation for the
3807 * future allocation of hugepages at runtime.
3809 reserve
= min(2, reserve
);
3811 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3812 if (!pfn_valid(pfn
))
3814 page
= pfn_to_page(pfn
);
3816 /* Watch out for overlapping nodes */
3817 if (page_to_nid(page
) != zone_to_nid(zone
))
3820 block_migratetype
= get_pageblock_migratetype(page
);
3822 /* Only test what is necessary when the reserves are not met */
3825 * Blocks with reserved pages will never free, skip
3828 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
3829 if (pageblock_is_reserved(pfn
, block_end_pfn
))
3832 /* If this block is reserved, account for it */
3833 if (block_migratetype
== MIGRATE_RESERVE
) {
3838 /* Suitable for reserving if this block is movable */
3839 if (block_migratetype
== MIGRATE_MOVABLE
) {
3840 set_pageblock_migratetype(page
,
3842 move_freepages_block(zone
, page
,
3850 * If the reserve is met and this is a previous reserved block,
3853 if (block_migratetype
== MIGRATE_RESERVE
) {
3854 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3855 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3861 * Initially all pages are reserved - free ones are freed
3862 * up by free_all_bootmem() once the early boot process is
3863 * done. Non-atomic initialization, single-pass.
3865 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3866 unsigned long start_pfn
, enum memmap_context context
)
3869 unsigned long end_pfn
= start_pfn
+ size
;
3873 if (highest_memmap_pfn
< end_pfn
- 1)
3874 highest_memmap_pfn
= end_pfn
- 1;
3876 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3877 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3879 * There can be holes in boot-time mem_map[]s
3880 * handed to this function. They do not
3881 * exist on hotplugged memory.
3883 if (context
== MEMMAP_EARLY
) {
3884 if (!early_pfn_valid(pfn
))
3886 if (!early_pfn_in_nid(pfn
, nid
))
3889 page
= pfn_to_page(pfn
);
3890 set_page_links(page
, zone
, nid
, pfn
);
3891 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3892 init_page_count(page
);
3893 reset_page_mapcount(page
);
3894 reset_page_last_nid(page
);
3895 SetPageReserved(page
);
3897 * Mark the block movable so that blocks are reserved for
3898 * movable at startup. This will force kernel allocations
3899 * to reserve their blocks rather than leaking throughout
3900 * the address space during boot when many long-lived
3901 * kernel allocations are made. Later some blocks near
3902 * the start are marked MIGRATE_RESERVE by
3903 * setup_zone_migrate_reserve()
3905 * bitmap is created for zone's valid pfn range. but memmap
3906 * can be created for invalid pages (for alignment)
3907 * check here not to call set_pageblock_migratetype() against
3910 if ((z
->zone_start_pfn
<= pfn
)
3911 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
3912 && !(pfn
& (pageblock_nr_pages
- 1)))
3913 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3915 INIT_LIST_HEAD(&page
->lru
);
3916 #ifdef WANT_PAGE_VIRTUAL
3917 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3918 if (!is_highmem_idx(zone
))
3919 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3924 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3927 for_each_migratetype_order(order
, t
) {
3928 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3929 zone
->free_area
[order
].nr_free
= 0;
3933 #ifndef __HAVE_ARCH_MEMMAP_INIT
3934 #define memmap_init(size, nid, zone, start_pfn) \
3935 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3938 static int __meminit
zone_batchsize(struct zone
*zone
)
3944 * The per-cpu-pages pools are set to around 1000th of the
3945 * size of the zone. But no more than 1/2 of a meg.
3947 * OK, so we don't know how big the cache is. So guess.
3949 batch
= zone
->present_pages
/ 1024;
3950 if (batch
* PAGE_SIZE
> 512 * 1024)
3951 batch
= (512 * 1024) / PAGE_SIZE
;
3952 batch
/= 4; /* We effectively *= 4 below */
3957 * Clamp the batch to a 2^n - 1 value. Having a power
3958 * of 2 value was found to be more likely to have
3959 * suboptimal cache aliasing properties in some cases.
3961 * For example if 2 tasks are alternately allocating
3962 * batches of pages, one task can end up with a lot
3963 * of pages of one half of the possible page colors
3964 * and the other with pages of the other colors.
3966 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3971 /* The deferral and batching of frees should be suppressed under NOMMU
3974 * The problem is that NOMMU needs to be able to allocate large chunks
3975 * of contiguous memory as there's no hardware page translation to
3976 * assemble apparent contiguous memory from discontiguous pages.
3978 * Queueing large contiguous runs of pages for batching, however,
3979 * causes the pages to actually be freed in smaller chunks. As there
3980 * can be a significant delay between the individual batches being
3981 * recycled, this leads to the once large chunks of space being
3982 * fragmented and becoming unavailable for high-order allocations.
3988 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3990 struct per_cpu_pages
*pcp
;
3993 memset(p
, 0, sizeof(*p
));
3997 pcp
->high
= 6 * batch
;
3998 pcp
->batch
= max(1UL, 1 * batch
);
3999 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4000 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4004 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4005 * to the value high for the pageset p.
4008 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
4011 struct per_cpu_pages
*pcp
;
4015 pcp
->batch
= max(1UL, high
/4);
4016 if ((high
/4) > (PAGE_SHIFT
* 8))
4017 pcp
->batch
= PAGE_SHIFT
* 8;
4020 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4024 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4026 for_each_possible_cpu(cpu
) {
4027 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4029 setup_pageset(pcp
, zone_batchsize(zone
));
4031 if (percpu_pagelist_fraction
)
4032 setup_pagelist_highmark(pcp
,
4033 (zone
->present_pages
/
4034 percpu_pagelist_fraction
));
4039 * Allocate per cpu pagesets and initialize them.
4040 * Before this call only boot pagesets were available.
4042 void __init
setup_per_cpu_pageset(void)
4046 for_each_populated_zone(zone
)
4047 setup_zone_pageset(zone
);
4050 static noinline __init_refok
4051 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4054 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4058 * The per-page waitqueue mechanism uses hashed waitqueues
4061 zone
->wait_table_hash_nr_entries
=
4062 wait_table_hash_nr_entries(zone_size_pages
);
4063 zone
->wait_table_bits
=
4064 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4065 alloc_size
= zone
->wait_table_hash_nr_entries
4066 * sizeof(wait_queue_head_t
);
4068 if (!slab_is_available()) {
4069 zone
->wait_table
= (wait_queue_head_t
*)
4070 alloc_bootmem_node_nopanic(pgdat
, alloc_size
);
4073 * This case means that a zone whose size was 0 gets new memory
4074 * via memory hot-add.
4075 * But it may be the case that a new node was hot-added. In
4076 * this case vmalloc() will not be able to use this new node's
4077 * memory - this wait_table must be initialized to use this new
4078 * node itself as well.
4079 * To use this new node's memory, further consideration will be
4082 zone
->wait_table
= vmalloc(alloc_size
);
4084 if (!zone
->wait_table
)
4087 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4088 init_waitqueue_head(zone
->wait_table
+ i
);
4093 static __meminit
void zone_pcp_init(struct zone
*zone
)
4096 * per cpu subsystem is not up at this point. The following code
4097 * relies on the ability of the linker to provide the
4098 * offset of a (static) per cpu variable into the per cpu area.
4100 zone
->pageset
= &boot_pageset
;
4102 if (zone
->present_pages
)
4103 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4104 zone
->name
, zone
->present_pages
,
4105 zone_batchsize(zone
));
4108 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4109 unsigned long zone_start_pfn
,
4111 enum memmap_context context
)
4113 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4115 ret
= zone_wait_table_init(zone
, size
);
4118 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4120 zone
->zone_start_pfn
= zone_start_pfn
;
4122 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4123 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4125 (unsigned long)zone_idx(zone
),
4126 zone_start_pfn
, (zone_start_pfn
+ size
));
4128 zone_init_free_lists(zone
);
4133 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4134 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4136 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4137 * Architectures may implement their own version but if add_active_range()
4138 * was used and there are no special requirements, this is a convenient
4141 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4143 unsigned long start_pfn
, end_pfn
;
4146 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
4147 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
4149 /* This is a memory hole */
4152 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4154 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4158 nid
= __early_pfn_to_nid(pfn
);
4161 /* just returns 0 */
4165 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4166 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4170 nid
= __early_pfn_to_nid(pfn
);
4171 if (nid
>= 0 && nid
!= node
)
4178 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4179 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4180 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4182 * If an architecture guarantees that all ranges registered with
4183 * add_active_ranges() contain no holes and may be freed, this
4184 * this function may be used instead of calling free_bootmem() manually.
4186 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4188 unsigned long start_pfn
, end_pfn
;
4191 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4192 start_pfn
= min(start_pfn
, max_low_pfn
);
4193 end_pfn
= min(end_pfn
, max_low_pfn
);
4195 if (start_pfn
< end_pfn
)
4196 free_bootmem_node(NODE_DATA(this_nid
),
4197 PFN_PHYS(start_pfn
),
4198 (end_pfn
- start_pfn
) << PAGE_SHIFT
);
4203 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4204 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4206 * If an architecture guarantees that all ranges registered with
4207 * add_active_ranges() contain no holes and may be freed, this
4208 * function may be used instead of calling memory_present() manually.
4210 void __init
sparse_memory_present_with_active_regions(int nid
)
4212 unsigned long start_pfn
, end_pfn
;
4215 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4216 memory_present(this_nid
, start_pfn
, end_pfn
);
4220 * get_pfn_range_for_nid - Return the start and end page frames for a node
4221 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4222 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4223 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4225 * It returns the start and end page frame of a node based on information
4226 * provided by an arch calling add_active_range(). If called for a node
4227 * with no available memory, a warning is printed and the start and end
4230 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4231 unsigned long *start_pfn
, unsigned long *end_pfn
)
4233 unsigned long this_start_pfn
, this_end_pfn
;
4239 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4240 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4241 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4244 if (*start_pfn
== -1UL)
4249 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4250 * assumption is made that zones within a node are ordered in monotonic
4251 * increasing memory addresses so that the "highest" populated zone is used
4253 static void __init
find_usable_zone_for_movable(void)
4256 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4257 if (zone_index
== ZONE_MOVABLE
)
4260 if (arch_zone_highest_possible_pfn
[zone_index
] >
4261 arch_zone_lowest_possible_pfn
[zone_index
])
4265 VM_BUG_ON(zone_index
== -1);
4266 movable_zone
= zone_index
;
4270 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4271 * because it is sized independent of architecture. Unlike the other zones,
4272 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4273 * in each node depending on the size of each node and how evenly kernelcore
4274 * is distributed. This helper function adjusts the zone ranges
4275 * provided by the architecture for a given node by using the end of the
4276 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4277 * zones within a node are in order of monotonic increases memory addresses
4279 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4280 unsigned long zone_type
,
4281 unsigned long node_start_pfn
,
4282 unsigned long node_end_pfn
,
4283 unsigned long *zone_start_pfn
,
4284 unsigned long *zone_end_pfn
)
4286 /* Only adjust if ZONE_MOVABLE is on this node */
4287 if (zone_movable_pfn
[nid
]) {
4288 /* Size ZONE_MOVABLE */
4289 if (zone_type
== ZONE_MOVABLE
) {
4290 *zone_start_pfn
= zone_movable_pfn
[nid
];
4291 *zone_end_pfn
= min(node_end_pfn
,
4292 arch_zone_highest_possible_pfn
[movable_zone
]);
4294 /* Adjust for ZONE_MOVABLE starting within this range */
4295 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4296 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4297 *zone_end_pfn
= zone_movable_pfn
[nid
];
4299 /* Check if this whole range is within ZONE_MOVABLE */
4300 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4301 *zone_start_pfn
= *zone_end_pfn
;
4306 * Return the number of pages a zone spans in a node, including holes
4307 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4309 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4310 unsigned long zone_type
,
4311 unsigned long *ignored
)
4313 unsigned long node_start_pfn
, node_end_pfn
;
4314 unsigned long zone_start_pfn
, zone_end_pfn
;
4316 /* Get the start and end of the node and zone */
4317 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4318 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4319 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4320 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4321 node_start_pfn
, node_end_pfn
,
4322 &zone_start_pfn
, &zone_end_pfn
);
4324 /* Check that this node has pages within the zone's required range */
4325 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4328 /* Move the zone boundaries inside the node if necessary */
4329 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4330 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4332 /* Return the spanned pages */
4333 return zone_end_pfn
- zone_start_pfn
;
4337 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4338 * then all holes in the requested range will be accounted for.
4340 unsigned long __meminit
__absent_pages_in_range(int nid
,
4341 unsigned long range_start_pfn
,
4342 unsigned long range_end_pfn
)
4344 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4345 unsigned long start_pfn
, end_pfn
;
4348 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4349 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4350 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4351 nr_absent
-= end_pfn
- start_pfn
;
4357 * absent_pages_in_range - Return number of page frames in holes within a range
4358 * @start_pfn: The start PFN to start searching for holes
4359 * @end_pfn: The end PFN to stop searching for holes
4361 * It returns the number of pages frames in memory holes within a range.
4363 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4364 unsigned long end_pfn
)
4366 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4369 /* Return the number of page frames in holes in a zone on a node */
4370 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4371 unsigned long zone_type
,
4372 unsigned long *ignored
)
4374 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4375 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4376 unsigned long node_start_pfn
, node_end_pfn
;
4377 unsigned long zone_start_pfn
, zone_end_pfn
;
4379 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4380 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4381 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4383 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4384 node_start_pfn
, node_end_pfn
,
4385 &zone_start_pfn
, &zone_end_pfn
);
4386 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4389 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4390 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4391 unsigned long zone_type
,
4392 unsigned long *zones_size
)
4394 return zones_size
[zone_type
];
4397 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4398 unsigned long zone_type
,
4399 unsigned long *zholes_size
)
4404 return zholes_size
[zone_type
];
4407 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4409 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4410 unsigned long *zones_size
, unsigned long *zholes_size
)
4412 unsigned long realtotalpages
, totalpages
= 0;
4415 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4416 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4418 pgdat
->node_spanned_pages
= totalpages
;
4420 realtotalpages
= totalpages
;
4421 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4423 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4425 pgdat
->node_present_pages
= realtotalpages
;
4426 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4430 #ifndef CONFIG_SPARSEMEM
4432 * Calculate the size of the zone->blockflags rounded to an unsigned long
4433 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4434 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4435 * round what is now in bits to nearest long in bits, then return it in
4438 static unsigned long __init
usemap_size(unsigned long zonesize
)
4440 unsigned long usemapsize
;
4442 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4443 usemapsize
= usemapsize
>> pageblock_order
;
4444 usemapsize
*= NR_PAGEBLOCK_BITS
;
4445 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4447 return usemapsize
/ 8;
4450 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4451 struct zone
*zone
, unsigned long zonesize
)
4453 unsigned long usemapsize
= usemap_size(zonesize
);
4454 zone
->pageblock_flags
= NULL
;
4456 zone
->pageblock_flags
= alloc_bootmem_node_nopanic(pgdat
,
4460 static inline void setup_usemap(struct pglist_data
*pgdat
,
4461 struct zone
*zone
, unsigned long zonesize
) {}
4462 #endif /* CONFIG_SPARSEMEM */
4464 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4466 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4467 void __init
set_pageblock_order(void)
4471 /* Check that pageblock_nr_pages has not already been setup */
4472 if (pageblock_order
)
4475 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4476 order
= HUGETLB_PAGE_ORDER
;
4478 order
= MAX_ORDER
- 1;
4481 * Assume the largest contiguous order of interest is a huge page.
4482 * This value may be variable depending on boot parameters on IA64 and
4485 pageblock_order
= order
;
4487 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4490 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4491 * is unused as pageblock_order is set at compile-time. See
4492 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4495 void __init
set_pageblock_order(void)
4499 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4501 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4502 unsigned long present_pages
)
4504 unsigned long pages
= spanned_pages
;
4507 * Provide a more accurate estimation if there are holes within
4508 * the zone and SPARSEMEM is in use. If there are holes within the
4509 * zone, each populated memory region may cost us one or two extra
4510 * memmap pages due to alignment because memmap pages for each
4511 * populated regions may not naturally algined on page boundary.
4512 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4514 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4515 IS_ENABLED(CONFIG_SPARSEMEM
))
4516 pages
= present_pages
;
4518 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4522 * Set up the zone data structures:
4523 * - mark all pages reserved
4524 * - mark all memory queues empty
4525 * - clear the memory bitmaps
4527 * NOTE: pgdat should get zeroed by caller.
4529 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4530 unsigned long *zones_size
, unsigned long *zholes_size
)
4533 int nid
= pgdat
->node_id
;
4534 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4537 pgdat_resize_init(pgdat
);
4538 #ifdef CONFIG_NUMA_BALANCING
4539 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4540 pgdat
->numabalancing_migrate_nr_pages
= 0;
4541 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4543 init_waitqueue_head(&pgdat
->kswapd_wait
);
4544 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4545 pgdat_page_cgroup_init(pgdat
);
4547 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4548 struct zone
*zone
= pgdat
->node_zones
+ j
;
4549 unsigned long size
, realsize
, freesize
, memmap_pages
;
4551 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4552 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4556 * Adjust freesize so that it accounts for how much memory
4557 * is used by this zone for memmap. This affects the watermark
4558 * and per-cpu initialisations
4560 memmap_pages
= calc_memmap_size(size
, realsize
);
4561 if (freesize
>= memmap_pages
) {
4562 freesize
-= memmap_pages
;
4565 " %s zone: %lu pages used for memmap\n",
4566 zone_names
[j
], memmap_pages
);
4569 " %s zone: %lu pages exceeds freesize %lu\n",
4570 zone_names
[j
], memmap_pages
, freesize
);
4572 /* Account for reserved pages */
4573 if (j
== 0 && freesize
> dma_reserve
) {
4574 freesize
-= dma_reserve
;
4575 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4576 zone_names
[0], dma_reserve
);
4579 if (!is_highmem_idx(j
))
4580 nr_kernel_pages
+= freesize
;
4581 /* Charge for highmem memmap if there are enough kernel pages */
4582 else if (nr_kernel_pages
> memmap_pages
* 2)
4583 nr_kernel_pages
-= memmap_pages
;
4584 nr_all_pages
+= freesize
;
4586 zone
->spanned_pages
= size
;
4587 zone
->present_pages
= freesize
;
4589 * Set an approximate value for lowmem here, it will be adjusted
4590 * when the bootmem allocator frees pages into the buddy system.
4591 * And all highmem pages will be managed by the buddy system.
4593 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4596 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4598 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4600 zone
->name
= zone_names
[j
];
4601 spin_lock_init(&zone
->lock
);
4602 spin_lock_init(&zone
->lru_lock
);
4603 zone_seqlock_init(zone
);
4604 zone
->zone_pgdat
= pgdat
;
4606 zone_pcp_init(zone
);
4607 lruvec_init(&zone
->lruvec
);
4611 set_pageblock_order();
4612 setup_usemap(pgdat
, zone
, size
);
4613 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4614 size
, MEMMAP_EARLY
);
4616 memmap_init(size
, nid
, j
, zone_start_pfn
);
4617 zone_start_pfn
+= size
;
4621 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4623 /* Skip empty nodes */
4624 if (!pgdat
->node_spanned_pages
)
4627 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4628 /* ia64 gets its own node_mem_map, before this, without bootmem */
4629 if (!pgdat
->node_mem_map
) {
4630 unsigned long size
, start
, end
;
4634 * The zone's endpoints aren't required to be MAX_ORDER
4635 * aligned but the node_mem_map endpoints must be in order
4636 * for the buddy allocator to function correctly.
4638 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4639 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
4640 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4641 size
= (end
- start
) * sizeof(struct page
);
4642 map
= alloc_remap(pgdat
->node_id
, size
);
4644 map
= alloc_bootmem_node_nopanic(pgdat
, size
);
4645 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4647 #ifndef CONFIG_NEED_MULTIPLE_NODES
4649 * With no DISCONTIG, the global mem_map is just set as node 0's
4651 if (pgdat
== NODE_DATA(0)) {
4652 mem_map
= NODE_DATA(0)->node_mem_map
;
4653 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4654 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4655 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4656 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4659 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4662 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4663 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4665 pg_data_t
*pgdat
= NODE_DATA(nid
);
4667 /* pg_data_t should be reset to zero when it's allocated */
4668 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4670 pgdat
->node_id
= nid
;
4671 pgdat
->node_start_pfn
= node_start_pfn
;
4672 init_zone_allows_reclaim(nid
);
4673 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4675 alloc_node_mem_map(pgdat
);
4676 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4677 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4678 nid
, (unsigned long)pgdat
,
4679 (unsigned long)pgdat
->node_mem_map
);
4682 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4685 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4687 #if MAX_NUMNODES > 1
4689 * Figure out the number of possible node ids.
4691 static void __init
setup_nr_node_ids(void)
4694 unsigned int highest
= 0;
4696 for_each_node_mask(node
, node_possible_map
)
4698 nr_node_ids
= highest
+ 1;
4701 static inline void setup_nr_node_ids(void)
4707 * node_map_pfn_alignment - determine the maximum internode alignment
4709 * This function should be called after node map is populated and sorted.
4710 * It calculates the maximum power of two alignment which can distinguish
4713 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4714 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4715 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4716 * shifted, 1GiB is enough and this function will indicate so.
4718 * This is used to test whether pfn -> nid mapping of the chosen memory
4719 * model has fine enough granularity to avoid incorrect mapping for the
4720 * populated node map.
4722 * Returns the determined alignment in pfn's. 0 if there is no alignment
4723 * requirement (single node).
4725 unsigned long __init
node_map_pfn_alignment(void)
4727 unsigned long accl_mask
= 0, last_end
= 0;
4728 unsigned long start
, end
, mask
;
4732 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
4733 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
4740 * Start with a mask granular enough to pin-point to the
4741 * start pfn and tick off bits one-by-one until it becomes
4742 * too coarse to separate the current node from the last.
4744 mask
= ~((1 << __ffs(start
)) - 1);
4745 while (mask
&& last_end
<= (start
& (mask
<< 1)))
4748 /* accumulate all internode masks */
4752 /* convert mask to number of pages */
4753 return ~accl_mask
+ 1;
4756 /* Find the lowest pfn for a node */
4757 static unsigned long __init
find_min_pfn_for_node(int nid
)
4759 unsigned long min_pfn
= ULONG_MAX
;
4760 unsigned long start_pfn
;
4763 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
4764 min_pfn
= min(min_pfn
, start_pfn
);
4766 if (min_pfn
== ULONG_MAX
) {
4768 "Could not find start_pfn for node %d\n", nid
);
4776 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4778 * It returns the minimum PFN based on information provided via
4779 * add_active_range().
4781 unsigned long __init
find_min_pfn_with_active_regions(void)
4783 return find_min_pfn_for_node(MAX_NUMNODES
);
4787 * early_calculate_totalpages()
4788 * Sum pages in active regions for movable zone.
4789 * Populate N_MEMORY for calculating usable_nodes.
4791 static unsigned long __init
early_calculate_totalpages(void)
4793 unsigned long totalpages
= 0;
4794 unsigned long start_pfn
, end_pfn
;
4797 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4798 unsigned long pages
= end_pfn
- start_pfn
;
4800 totalpages
+= pages
;
4802 node_set_state(nid
, N_MEMORY
);
4808 * Find the PFN the Movable zone begins in each node. Kernel memory
4809 * is spread evenly between nodes as long as the nodes have enough
4810 * memory. When they don't, some nodes will have more kernelcore than
4813 static void __init
find_zone_movable_pfns_for_nodes(void)
4816 unsigned long usable_startpfn
;
4817 unsigned long kernelcore_node
, kernelcore_remaining
;
4818 /* save the state before borrow the nodemask */
4819 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
4820 unsigned long totalpages
= early_calculate_totalpages();
4821 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
4824 * If movablecore was specified, calculate what size of
4825 * kernelcore that corresponds so that memory usable for
4826 * any allocation type is evenly spread. If both kernelcore
4827 * and movablecore are specified, then the value of kernelcore
4828 * will be used for required_kernelcore if it's greater than
4829 * what movablecore would have allowed.
4831 if (required_movablecore
) {
4832 unsigned long corepages
;
4835 * Round-up so that ZONE_MOVABLE is at least as large as what
4836 * was requested by the user
4838 required_movablecore
=
4839 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4840 corepages
= totalpages
- required_movablecore
;
4842 required_kernelcore
= max(required_kernelcore
, corepages
);
4845 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4846 if (!required_kernelcore
)
4849 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4850 find_usable_zone_for_movable();
4851 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4854 /* Spread kernelcore memory as evenly as possible throughout nodes */
4855 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4856 for_each_node_state(nid
, N_MEMORY
) {
4857 unsigned long start_pfn
, end_pfn
;
4860 * Recalculate kernelcore_node if the division per node
4861 * now exceeds what is necessary to satisfy the requested
4862 * amount of memory for the kernel
4864 if (required_kernelcore
< kernelcore_node
)
4865 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4868 * As the map is walked, we track how much memory is usable
4869 * by the kernel using kernelcore_remaining. When it is
4870 * 0, the rest of the node is usable by ZONE_MOVABLE
4872 kernelcore_remaining
= kernelcore_node
;
4874 /* Go through each range of PFNs within this node */
4875 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4876 unsigned long size_pages
;
4878 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
4879 if (start_pfn
>= end_pfn
)
4882 /* Account for what is only usable for kernelcore */
4883 if (start_pfn
< usable_startpfn
) {
4884 unsigned long kernel_pages
;
4885 kernel_pages
= min(end_pfn
, usable_startpfn
)
4888 kernelcore_remaining
-= min(kernel_pages
,
4889 kernelcore_remaining
);
4890 required_kernelcore
-= min(kernel_pages
,
4891 required_kernelcore
);
4893 /* Continue if range is now fully accounted */
4894 if (end_pfn
<= usable_startpfn
) {
4897 * Push zone_movable_pfn to the end so
4898 * that if we have to rebalance
4899 * kernelcore across nodes, we will
4900 * not double account here
4902 zone_movable_pfn
[nid
] = end_pfn
;
4905 start_pfn
= usable_startpfn
;
4909 * The usable PFN range for ZONE_MOVABLE is from
4910 * start_pfn->end_pfn. Calculate size_pages as the
4911 * number of pages used as kernelcore
4913 size_pages
= end_pfn
- start_pfn
;
4914 if (size_pages
> kernelcore_remaining
)
4915 size_pages
= kernelcore_remaining
;
4916 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4919 * Some kernelcore has been met, update counts and
4920 * break if the kernelcore for this node has been
4923 required_kernelcore
-= min(required_kernelcore
,
4925 kernelcore_remaining
-= size_pages
;
4926 if (!kernelcore_remaining
)
4932 * If there is still required_kernelcore, we do another pass with one
4933 * less node in the count. This will push zone_movable_pfn[nid] further
4934 * along on the nodes that still have memory until kernelcore is
4938 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
4941 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4942 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
4943 zone_movable_pfn
[nid
] =
4944 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
4947 /* restore the node_state */
4948 node_states
[N_MEMORY
] = saved_node_state
;
4951 /* Any regular or high memory on that node ? */
4952 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
4954 enum zone_type zone_type
;
4956 if (N_MEMORY
== N_NORMAL_MEMORY
)
4959 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
4960 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
4961 if (zone
->present_pages
) {
4962 node_set_state(nid
, N_HIGH_MEMORY
);
4963 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
4964 zone_type
<= ZONE_NORMAL
)
4965 node_set_state(nid
, N_NORMAL_MEMORY
);
4972 * free_area_init_nodes - Initialise all pg_data_t and zone data
4973 * @max_zone_pfn: an array of max PFNs for each zone
4975 * This will call free_area_init_node() for each active node in the system.
4976 * Using the page ranges provided by add_active_range(), the size of each
4977 * zone in each node and their holes is calculated. If the maximum PFN
4978 * between two adjacent zones match, it is assumed that the zone is empty.
4979 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4980 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4981 * starts where the previous one ended. For example, ZONE_DMA32 starts
4982 * at arch_max_dma_pfn.
4984 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
4986 unsigned long start_pfn
, end_pfn
;
4989 /* Record where the zone boundaries are */
4990 memset(arch_zone_lowest_possible_pfn
, 0,
4991 sizeof(arch_zone_lowest_possible_pfn
));
4992 memset(arch_zone_highest_possible_pfn
, 0,
4993 sizeof(arch_zone_highest_possible_pfn
));
4994 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
4995 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
4996 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
4997 if (i
== ZONE_MOVABLE
)
4999 arch_zone_lowest_possible_pfn
[i
] =
5000 arch_zone_highest_possible_pfn
[i
-1];
5001 arch_zone_highest_possible_pfn
[i
] =
5002 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5004 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5005 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5007 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5008 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5009 find_zone_movable_pfns_for_nodes();
5011 /* Print out the zone ranges */
5012 printk("Zone ranges:\n");
5013 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5014 if (i
== ZONE_MOVABLE
)
5016 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5017 if (arch_zone_lowest_possible_pfn
[i
] ==
5018 arch_zone_highest_possible_pfn
[i
])
5019 printk(KERN_CONT
"empty\n");
5021 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5022 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5023 (arch_zone_highest_possible_pfn
[i
]
5024 << PAGE_SHIFT
) - 1);
5027 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5028 printk("Movable zone start for each node\n");
5029 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5030 if (zone_movable_pfn
[i
])
5031 printk(" Node %d: %#010lx\n", i
,
5032 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5035 /* Print out the early node map */
5036 printk("Early memory node ranges\n");
5037 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5038 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5039 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5041 /* Initialise every node */
5042 mminit_verify_pageflags_layout();
5043 setup_nr_node_ids();
5044 for_each_online_node(nid
) {
5045 pg_data_t
*pgdat
= NODE_DATA(nid
);
5046 free_area_init_node(nid
, NULL
,
5047 find_min_pfn_for_node(nid
), NULL
);
5049 /* Any memory on that node */
5050 if (pgdat
->node_present_pages
)
5051 node_set_state(nid
, N_MEMORY
);
5052 check_for_memory(pgdat
, nid
);
5056 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5058 unsigned long long coremem
;
5062 coremem
= memparse(p
, &p
);
5063 *core
= coremem
>> PAGE_SHIFT
;
5065 /* Paranoid check that UL is enough for the coremem value */
5066 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5072 * kernelcore=size sets the amount of memory for use for allocations that
5073 * cannot be reclaimed or migrated.
5075 static int __init
cmdline_parse_kernelcore(char *p
)
5077 return cmdline_parse_core(p
, &required_kernelcore
);
5081 * movablecore=size sets the amount of memory for use for allocations that
5082 * can be reclaimed or migrated.
5084 static int __init
cmdline_parse_movablecore(char *p
)
5086 return cmdline_parse_core(p
, &required_movablecore
);
5089 early_param("kernelcore", cmdline_parse_kernelcore
);
5090 early_param("movablecore", cmdline_parse_movablecore
);
5092 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5095 * set_dma_reserve - set the specified number of pages reserved in the first zone
5096 * @new_dma_reserve: The number of pages to mark reserved
5098 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5099 * In the DMA zone, a significant percentage may be consumed by kernel image
5100 * and other unfreeable allocations which can skew the watermarks badly. This
5101 * function may optionally be used to account for unfreeable pages in the
5102 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5103 * smaller per-cpu batchsize.
5105 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5107 dma_reserve
= new_dma_reserve
;
5110 void __init
free_area_init(unsigned long *zones_size
)
5112 free_area_init_node(0, zones_size
,
5113 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5116 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5117 unsigned long action
, void *hcpu
)
5119 int cpu
= (unsigned long)hcpu
;
5121 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5122 lru_add_drain_cpu(cpu
);
5126 * Spill the event counters of the dead processor
5127 * into the current processors event counters.
5128 * This artificially elevates the count of the current
5131 vm_events_fold_cpu(cpu
);
5134 * Zero the differential counters of the dead processor
5135 * so that the vm statistics are consistent.
5137 * This is only okay since the processor is dead and cannot
5138 * race with what we are doing.
5140 refresh_cpu_vm_stats(cpu
);
5145 void __init
page_alloc_init(void)
5147 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5151 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5152 * or min_free_kbytes changes.
5154 static void calculate_totalreserve_pages(void)
5156 struct pglist_data
*pgdat
;
5157 unsigned long reserve_pages
= 0;
5158 enum zone_type i
, j
;
5160 for_each_online_pgdat(pgdat
) {
5161 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5162 struct zone
*zone
= pgdat
->node_zones
+ i
;
5163 unsigned long max
= 0;
5165 /* Find valid and maximum lowmem_reserve in the zone */
5166 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5167 if (zone
->lowmem_reserve
[j
] > max
)
5168 max
= zone
->lowmem_reserve
[j
];
5171 /* we treat the high watermark as reserved pages. */
5172 max
+= high_wmark_pages(zone
);
5174 if (max
> zone
->present_pages
)
5175 max
= zone
->present_pages
;
5176 reserve_pages
+= max
;
5178 * Lowmem reserves are not available to
5179 * GFP_HIGHUSER page cache allocations and
5180 * kswapd tries to balance zones to their high
5181 * watermark. As a result, neither should be
5182 * regarded as dirtyable memory, to prevent a
5183 * situation where reclaim has to clean pages
5184 * in order to balance the zones.
5186 zone
->dirty_balance_reserve
= max
;
5189 dirty_balance_reserve
= reserve_pages
;
5190 totalreserve_pages
= reserve_pages
;
5194 * setup_per_zone_lowmem_reserve - called whenever
5195 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5196 * has a correct pages reserved value, so an adequate number of
5197 * pages are left in the zone after a successful __alloc_pages().
5199 static void setup_per_zone_lowmem_reserve(void)
5201 struct pglist_data
*pgdat
;
5202 enum zone_type j
, idx
;
5204 for_each_online_pgdat(pgdat
) {
5205 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5206 struct zone
*zone
= pgdat
->node_zones
+ j
;
5207 unsigned long present_pages
= zone
->present_pages
;
5209 zone
->lowmem_reserve
[j
] = 0;
5213 struct zone
*lower_zone
;
5217 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5218 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5220 lower_zone
= pgdat
->node_zones
+ idx
;
5221 lower_zone
->lowmem_reserve
[j
] = present_pages
/
5222 sysctl_lowmem_reserve_ratio
[idx
];
5223 present_pages
+= lower_zone
->present_pages
;
5228 /* update totalreserve_pages */
5229 calculate_totalreserve_pages();
5232 static void __setup_per_zone_wmarks(void)
5234 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5235 unsigned long lowmem_pages
= 0;
5237 unsigned long flags
;
5239 /* Calculate total number of !ZONE_HIGHMEM pages */
5240 for_each_zone(zone
) {
5241 if (!is_highmem(zone
))
5242 lowmem_pages
+= zone
->present_pages
;
5245 for_each_zone(zone
) {
5248 spin_lock_irqsave(&zone
->lock
, flags
);
5249 tmp
= (u64
)pages_min
* zone
->present_pages
;
5250 do_div(tmp
, lowmem_pages
);
5251 if (is_highmem(zone
)) {
5253 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5254 * need highmem pages, so cap pages_min to a small
5257 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5258 * deltas controls asynch page reclaim, and so should
5259 * not be capped for highmem.
5263 min_pages
= zone
->present_pages
/ 1024;
5264 if (min_pages
< SWAP_CLUSTER_MAX
)
5265 min_pages
= SWAP_CLUSTER_MAX
;
5266 if (min_pages
> 128)
5268 zone
->watermark
[WMARK_MIN
] = min_pages
;
5271 * If it's a lowmem zone, reserve a number of pages
5272 * proportionate to the zone's size.
5274 zone
->watermark
[WMARK_MIN
] = tmp
;
5277 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5278 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5280 setup_zone_migrate_reserve(zone
);
5281 spin_unlock_irqrestore(&zone
->lock
, flags
);
5284 /* update totalreserve_pages */
5285 calculate_totalreserve_pages();
5289 * setup_per_zone_wmarks - called when min_free_kbytes changes
5290 * or when memory is hot-{added|removed}
5292 * Ensures that the watermark[min,low,high] values for each zone are set
5293 * correctly with respect to min_free_kbytes.
5295 void setup_per_zone_wmarks(void)
5297 mutex_lock(&zonelists_mutex
);
5298 __setup_per_zone_wmarks();
5299 mutex_unlock(&zonelists_mutex
);
5303 * The inactive anon list should be small enough that the VM never has to
5304 * do too much work, but large enough that each inactive page has a chance
5305 * to be referenced again before it is swapped out.
5307 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5308 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5309 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5310 * the anonymous pages are kept on the inactive list.
5313 * memory ratio inactive anon
5314 * -------------------------------------
5323 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5325 unsigned int gb
, ratio
;
5327 /* Zone size in gigabytes */
5328 gb
= zone
->present_pages
>> (30 - PAGE_SHIFT
);
5330 ratio
= int_sqrt(10 * gb
);
5334 zone
->inactive_ratio
= ratio
;
5337 static void __meminit
setup_per_zone_inactive_ratio(void)
5342 calculate_zone_inactive_ratio(zone
);
5346 * Initialise min_free_kbytes.
5348 * For small machines we want it small (128k min). For large machines
5349 * we want it large (64MB max). But it is not linear, because network
5350 * bandwidth does not increase linearly with machine size. We use
5352 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5353 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5369 int __meminit
init_per_zone_wmark_min(void)
5371 unsigned long lowmem_kbytes
;
5373 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5375 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5376 if (min_free_kbytes
< 128)
5377 min_free_kbytes
= 128;
5378 if (min_free_kbytes
> 65536)
5379 min_free_kbytes
= 65536;
5380 setup_per_zone_wmarks();
5381 refresh_zone_stat_thresholds();
5382 setup_per_zone_lowmem_reserve();
5383 setup_per_zone_inactive_ratio();
5386 module_init(init_per_zone_wmark_min
)
5389 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5390 * that we can call two helper functions whenever min_free_kbytes
5393 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5394 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5396 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5398 setup_per_zone_wmarks();
5403 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5404 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5409 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5414 zone
->min_unmapped_pages
= (zone
->present_pages
*
5415 sysctl_min_unmapped_ratio
) / 100;
5419 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5420 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5425 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5430 zone
->min_slab_pages
= (zone
->present_pages
*
5431 sysctl_min_slab_ratio
) / 100;
5437 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5438 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5439 * whenever sysctl_lowmem_reserve_ratio changes.
5441 * The reserve ratio obviously has absolutely no relation with the
5442 * minimum watermarks. The lowmem reserve ratio can only make sense
5443 * if in function of the boot time zone sizes.
5445 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5446 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5448 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5449 setup_per_zone_lowmem_reserve();
5454 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5455 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5456 * can have before it gets flushed back to buddy allocator.
5459 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5460 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5466 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5467 if (!write
|| (ret
< 0))
5469 for_each_populated_zone(zone
) {
5470 for_each_possible_cpu(cpu
) {
5472 high
= zone
->present_pages
/ percpu_pagelist_fraction
;
5473 setup_pagelist_highmark(
5474 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5480 int hashdist
= HASHDIST_DEFAULT
;
5483 static int __init
set_hashdist(char *str
)
5487 hashdist
= simple_strtoul(str
, &str
, 0);
5490 __setup("hashdist=", set_hashdist
);
5494 * allocate a large system hash table from bootmem
5495 * - it is assumed that the hash table must contain an exact power-of-2
5496 * quantity of entries
5497 * - limit is the number of hash buckets, not the total allocation size
5499 void *__init
alloc_large_system_hash(const char *tablename
,
5500 unsigned long bucketsize
,
5501 unsigned long numentries
,
5504 unsigned int *_hash_shift
,
5505 unsigned int *_hash_mask
,
5506 unsigned long low_limit
,
5507 unsigned long high_limit
)
5509 unsigned long long max
= high_limit
;
5510 unsigned long log2qty
, size
;
5513 /* allow the kernel cmdline to have a say */
5515 /* round applicable memory size up to nearest megabyte */
5516 numentries
= nr_kernel_pages
;
5517 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5518 numentries
>>= 20 - PAGE_SHIFT
;
5519 numentries
<<= 20 - PAGE_SHIFT
;
5521 /* limit to 1 bucket per 2^scale bytes of low memory */
5522 if (scale
> PAGE_SHIFT
)
5523 numentries
>>= (scale
- PAGE_SHIFT
);
5525 numentries
<<= (PAGE_SHIFT
- scale
);
5527 /* Make sure we've got at least a 0-order allocation.. */
5528 if (unlikely(flags
& HASH_SMALL
)) {
5529 /* Makes no sense without HASH_EARLY */
5530 WARN_ON(!(flags
& HASH_EARLY
));
5531 if (!(numentries
>> *_hash_shift
)) {
5532 numentries
= 1UL << *_hash_shift
;
5533 BUG_ON(!numentries
);
5535 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5536 numentries
= PAGE_SIZE
/ bucketsize
;
5538 numentries
= roundup_pow_of_two(numentries
);
5540 /* limit allocation size to 1/16 total memory by default */
5542 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5543 do_div(max
, bucketsize
);
5545 max
= min(max
, 0x80000000ULL
);
5547 if (numentries
< low_limit
)
5548 numentries
= low_limit
;
5549 if (numentries
> max
)
5552 log2qty
= ilog2(numentries
);
5555 size
= bucketsize
<< log2qty
;
5556 if (flags
& HASH_EARLY
)
5557 table
= alloc_bootmem_nopanic(size
);
5559 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5562 * If bucketsize is not a power-of-two, we may free
5563 * some pages at the end of hash table which
5564 * alloc_pages_exact() automatically does
5566 if (get_order(size
) < MAX_ORDER
) {
5567 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5568 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5571 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5574 panic("Failed to allocate %s hash table\n", tablename
);
5576 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5579 ilog2(size
) - PAGE_SHIFT
,
5583 *_hash_shift
= log2qty
;
5585 *_hash_mask
= (1 << log2qty
) - 1;
5590 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5591 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5594 #ifdef CONFIG_SPARSEMEM
5595 return __pfn_to_section(pfn
)->pageblock_flags
;
5597 return zone
->pageblock_flags
;
5598 #endif /* CONFIG_SPARSEMEM */
5601 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5603 #ifdef CONFIG_SPARSEMEM
5604 pfn
&= (PAGES_PER_SECTION
-1);
5605 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5607 pfn
= pfn
- zone
->zone_start_pfn
;
5608 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5609 #endif /* CONFIG_SPARSEMEM */
5613 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5614 * @page: The page within the block of interest
5615 * @start_bitidx: The first bit of interest to retrieve
5616 * @end_bitidx: The last bit of interest
5617 * returns pageblock_bits flags
5619 unsigned long get_pageblock_flags_group(struct page
*page
,
5620 int start_bitidx
, int end_bitidx
)
5623 unsigned long *bitmap
;
5624 unsigned long pfn
, bitidx
;
5625 unsigned long flags
= 0;
5626 unsigned long value
= 1;
5628 zone
= page_zone(page
);
5629 pfn
= page_to_pfn(page
);
5630 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5631 bitidx
= pfn_to_bitidx(zone
, pfn
);
5633 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5634 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5641 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5642 * @page: The page within the block of interest
5643 * @start_bitidx: The first bit of interest
5644 * @end_bitidx: The last bit of interest
5645 * @flags: The flags to set
5647 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5648 int start_bitidx
, int end_bitidx
)
5651 unsigned long *bitmap
;
5652 unsigned long pfn
, bitidx
;
5653 unsigned long value
= 1;
5655 zone
= page_zone(page
);
5656 pfn
= page_to_pfn(page
);
5657 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5658 bitidx
= pfn_to_bitidx(zone
, pfn
);
5659 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
5660 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
5662 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5664 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5666 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5670 * This function checks whether pageblock includes unmovable pages or not.
5671 * If @count is not zero, it is okay to include less @count unmovable pages
5673 * PageLRU check wihtout isolation or lru_lock could race so that
5674 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5675 * expect this function should be exact.
5677 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
5678 bool skip_hwpoisoned_pages
)
5680 unsigned long pfn
, iter
, found
;
5684 * For avoiding noise data, lru_add_drain_all() should be called
5685 * If ZONE_MOVABLE, the zone never contains unmovable pages
5687 if (zone_idx(zone
) == ZONE_MOVABLE
)
5689 mt
= get_pageblock_migratetype(page
);
5690 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
5693 pfn
= page_to_pfn(page
);
5694 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5695 unsigned long check
= pfn
+ iter
;
5697 if (!pfn_valid_within(check
))
5700 page
= pfn_to_page(check
);
5702 * We can't use page_count without pin a page
5703 * because another CPU can free compound page.
5704 * This check already skips compound tails of THP
5705 * because their page->_count is zero at all time.
5707 if (!atomic_read(&page
->_count
)) {
5708 if (PageBuddy(page
))
5709 iter
+= (1 << page_order(page
)) - 1;
5714 * The HWPoisoned page may be not in buddy system, and
5715 * page_count() is not 0.
5717 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
5723 * If there are RECLAIMABLE pages, we need to check it.
5724 * But now, memory offline itself doesn't call shrink_slab()
5725 * and it still to be fixed.
5728 * If the page is not RAM, page_count()should be 0.
5729 * we don't need more check. This is an _used_ not-movable page.
5731 * The problematic thing here is PG_reserved pages. PG_reserved
5732 * is set to both of a memory hole page and a _used_ kernel
5741 bool is_pageblock_removable_nolock(struct page
*page
)
5747 * We have to be careful here because we are iterating over memory
5748 * sections which are not zone aware so we might end up outside of
5749 * the zone but still within the section.
5750 * We have to take care about the node as well. If the node is offline
5751 * its NODE_DATA will be NULL - see page_zone.
5753 if (!node_online(page_to_nid(page
)))
5756 zone
= page_zone(page
);
5757 pfn
= page_to_pfn(page
);
5758 if (zone
->zone_start_pfn
> pfn
||
5759 zone
->zone_start_pfn
+ zone
->spanned_pages
<= pfn
)
5762 return !has_unmovable_pages(zone
, page
, 0, true);
5767 static unsigned long pfn_max_align_down(unsigned long pfn
)
5769 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
5770 pageblock_nr_pages
) - 1);
5773 static unsigned long pfn_max_align_up(unsigned long pfn
)
5775 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
5776 pageblock_nr_pages
));
5779 /* [start, end) must belong to a single zone. */
5780 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
5781 unsigned long start
, unsigned long end
)
5783 /* This function is based on compact_zone() from compaction.c. */
5784 unsigned long nr_reclaimed
;
5785 unsigned long pfn
= start
;
5786 unsigned int tries
= 0;
5791 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
5792 if (fatal_signal_pending(current
)) {
5797 if (list_empty(&cc
->migratepages
)) {
5798 cc
->nr_migratepages
= 0;
5799 pfn
= isolate_migratepages_range(cc
->zone
, cc
,
5806 } else if (++tries
== 5) {
5807 ret
= ret
< 0 ? ret
: -EBUSY
;
5811 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
5813 cc
->nr_migratepages
-= nr_reclaimed
;
5815 ret
= migrate_pages(&cc
->migratepages
,
5816 alloc_migrate_target
,
5817 0, false, MIGRATE_SYNC
,
5821 putback_movable_pages(&cc
->migratepages
);
5822 return ret
> 0 ? 0 : ret
;
5826 * alloc_contig_range() -- tries to allocate given range of pages
5827 * @start: start PFN to allocate
5828 * @end: one-past-the-last PFN to allocate
5829 * @migratetype: migratetype of the underlaying pageblocks (either
5830 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5831 * in range must have the same migratetype and it must
5832 * be either of the two.
5834 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5835 * aligned, however it's the caller's responsibility to guarantee that
5836 * we are the only thread that changes migrate type of pageblocks the
5839 * The PFN range must belong to a single zone.
5841 * Returns zero on success or negative error code. On success all
5842 * pages which PFN is in [start, end) are allocated for the caller and
5843 * need to be freed with free_contig_range().
5845 int alloc_contig_range(unsigned long start
, unsigned long end
,
5846 unsigned migratetype
)
5848 unsigned long outer_start
, outer_end
;
5851 struct compact_control cc
= {
5852 .nr_migratepages
= 0,
5854 .zone
= page_zone(pfn_to_page(start
)),
5856 .ignore_skip_hint
= true,
5858 INIT_LIST_HEAD(&cc
.migratepages
);
5861 * What we do here is we mark all pageblocks in range as
5862 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5863 * have different sizes, and due to the way page allocator
5864 * work, we align the range to biggest of the two pages so
5865 * that page allocator won't try to merge buddies from
5866 * different pageblocks and change MIGRATE_ISOLATE to some
5867 * other migration type.
5869 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5870 * migrate the pages from an unaligned range (ie. pages that
5871 * we are interested in). This will put all the pages in
5872 * range back to page allocator as MIGRATE_ISOLATE.
5874 * When this is done, we take the pages in range from page
5875 * allocator removing them from the buddy system. This way
5876 * page allocator will never consider using them.
5878 * This lets us mark the pageblocks back as
5879 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5880 * aligned range but not in the unaligned, original range are
5881 * put back to page allocator so that buddy can use them.
5884 ret
= start_isolate_page_range(pfn_max_align_down(start
),
5885 pfn_max_align_up(end
), migratetype
,
5890 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
5895 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5896 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5897 * more, all pages in [start, end) are free in page allocator.
5898 * What we are going to do is to allocate all pages from
5899 * [start, end) (that is remove them from page allocator).
5901 * The only problem is that pages at the beginning and at the
5902 * end of interesting range may be not aligned with pages that
5903 * page allocator holds, ie. they can be part of higher order
5904 * pages. Because of this, we reserve the bigger range and
5905 * once this is done free the pages we are not interested in.
5907 * We don't have to hold zone->lock here because the pages are
5908 * isolated thus they won't get removed from buddy.
5911 lru_add_drain_all();
5915 outer_start
= start
;
5916 while (!PageBuddy(pfn_to_page(outer_start
))) {
5917 if (++order
>= MAX_ORDER
) {
5921 outer_start
&= ~0UL << order
;
5924 /* Make sure the range is really isolated. */
5925 if (test_pages_isolated(outer_start
, end
, false)) {
5926 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
5933 /* Grab isolated pages from freelists. */
5934 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
5940 /* Free head and tail (if any) */
5941 if (start
!= outer_start
)
5942 free_contig_range(outer_start
, start
- outer_start
);
5943 if (end
!= outer_end
)
5944 free_contig_range(end
, outer_end
- end
);
5947 undo_isolate_page_range(pfn_max_align_down(start
),
5948 pfn_max_align_up(end
), migratetype
);
5952 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
5954 unsigned int count
= 0;
5956 for (; nr_pages
--; pfn
++) {
5957 struct page
*page
= pfn_to_page(pfn
);
5959 count
+= page_count(page
) != 1;
5962 WARN(count
!= 0, "%d pages are still in use!\n", count
);
5966 #ifdef CONFIG_MEMORY_HOTPLUG
5967 static int __meminit
__zone_pcp_update(void *data
)
5969 struct zone
*zone
= data
;
5971 unsigned long batch
= zone_batchsize(zone
), flags
;
5973 for_each_possible_cpu(cpu
) {
5974 struct per_cpu_pageset
*pset
;
5975 struct per_cpu_pages
*pcp
;
5977 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
5980 local_irq_save(flags
);
5982 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
5983 drain_zonestat(zone
, pset
);
5984 setup_pageset(pset
, batch
);
5985 local_irq_restore(flags
);
5990 void __meminit
zone_pcp_update(struct zone
*zone
)
5992 stop_machine(__zone_pcp_update
, zone
, NULL
);
5996 void zone_pcp_reset(struct zone
*zone
)
5998 unsigned long flags
;
6000 struct per_cpu_pageset
*pset
;
6002 /* avoid races with drain_pages() */
6003 local_irq_save(flags
);
6004 if (zone
->pageset
!= &boot_pageset
) {
6005 for_each_online_cpu(cpu
) {
6006 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6007 drain_zonestat(zone
, pset
);
6009 free_percpu(zone
->pageset
);
6010 zone
->pageset
= &boot_pageset
;
6012 local_irq_restore(flags
);
6015 #ifdef CONFIG_MEMORY_HOTREMOVE
6017 * All pages in the range must be isolated before calling this.
6020 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6026 unsigned long flags
;
6027 /* find the first valid pfn */
6028 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6033 zone
= page_zone(pfn_to_page(pfn
));
6034 spin_lock_irqsave(&zone
->lock
, flags
);
6036 while (pfn
< end_pfn
) {
6037 if (!pfn_valid(pfn
)) {
6041 page
= pfn_to_page(pfn
);
6043 * The HWPoisoned page may be not in buddy system, and
6044 * page_count() is not 0.
6046 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6048 SetPageReserved(page
);
6052 BUG_ON(page_count(page
));
6053 BUG_ON(!PageBuddy(page
));
6054 order
= page_order(page
);
6055 #ifdef CONFIG_DEBUG_VM
6056 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6057 pfn
, 1 << order
, end_pfn
);
6059 list_del(&page
->lru
);
6060 rmv_page_order(page
);
6061 zone
->free_area
[order
].nr_free
--;
6062 for (i
= 0; i
< (1 << order
); i
++)
6063 SetPageReserved((page
+i
));
6064 pfn
+= (1 << order
);
6066 spin_unlock_irqrestore(&zone
->lock
, flags
);
6070 #ifdef CONFIG_MEMORY_FAILURE
6071 bool is_free_buddy_page(struct page
*page
)
6073 struct zone
*zone
= page_zone(page
);
6074 unsigned long pfn
= page_to_pfn(page
);
6075 unsigned long flags
;
6078 spin_lock_irqsave(&zone
->lock
, flags
);
6079 for (order
= 0; order
< MAX_ORDER
; order
++) {
6080 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6082 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6085 spin_unlock_irqrestore(&zone
->lock
, flags
);
6087 return order
< MAX_ORDER
;
6091 static const struct trace_print_flags pageflag_names
[] = {
6092 {1UL << PG_locked
, "locked" },
6093 {1UL << PG_error
, "error" },
6094 {1UL << PG_referenced
, "referenced" },
6095 {1UL << PG_uptodate
, "uptodate" },
6096 {1UL << PG_dirty
, "dirty" },
6097 {1UL << PG_lru
, "lru" },
6098 {1UL << PG_active
, "active" },
6099 {1UL << PG_slab
, "slab" },
6100 {1UL << PG_owner_priv_1
, "owner_priv_1" },
6101 {1UL << PG_arch_1
, "arch_1" },
6102 {1UL << PG_reserved
, "reserved" },
6103 {1UL << PG_private
, "private" },
6104 {1UL << PG_private_2
, "private_2" },
6105 {1UL << PG_writeback
, "writeback" },
6106 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6107 {1UL << PG_head
, "head" },
6108 {1UL << PG_tail
, "tail" },
6110 {1UL << PG_compound
, "compound" },
6112 {1UL << PG_swapcache
, "swapcache" },
6113 {1UL << PG_mappedtodisk
, "mappedtodisk" },
6114 {1UL << PG_reclaim
, "reclaim" },
6115 {1UL << PG_swapbacked
, "swapbacked" },
6116 {1UL << PG_unevictable
, "unevictable" },
6118 {1UL << PG_mlocked
, "mlocked" },
6120 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6121 {1UL << PG_uncached
, "uncached" },
6123 #ifdef CONFIG_MEMORY_FAILURE
6124 {1UL << PG_hwpoison
, "hwpoison" },
6126 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6127 {1UL << PG_compound_lock
, "compound_lock" },
6131 static void dump_page_flags(unsigned long flags
)
6133 const char *delim
= "";
6137 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names
) != __NR_PAGEFLAGS
);
6139 printk(KERN_ALERT
"page flags: %#lx(", flags
);
6141 /* remove zone id */
6142 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
6144 for (i
= 0; i
< ARRAY_SIZE(pageflag_names
) && flags
; i
++) {
6146 mask
= pageflag_names
[i
].mask
;
6147 if ((flags
& mask
) != mask
)
6151 printk("%s%s", delim
, pageflag_names
[i
].name
);
6155 /* check for left over flags */
6157 printk("%s%#lx", delim
, flags
);
6162 void dump_page(struct page
*page
)
6165 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6166 page
, atomic_read(&page
->_count
), page_mapcount(page
),
6167 page
->mapping
, page
->index
);
6168 dump_page_flags(page
->flags
);
6169 mem_cgroup_print_bad_page(page
);