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>
61 #include <linux/sched/rt.h>
63 #include <asm/tlbflush.h>
64 #include <asm/div64.h>
67 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
68 DEFINE_PER_CPU(int, numa_node
);
69 EXPORT_PER_CPU_SYMBOL(numa_node
);
72 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
74 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
75 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
76 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
77 * defined in <linux/topology.h>.
79 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
80 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
84 * Array of node states.
86 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
87 [N_POSSIBLE
] = NODE_MASK_ALL
,
88 [N_ONLINE
] = { { [0] = 1UL } },
90 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
92 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
94 #ifdef CONFIG_MOVABLE_NODE
95 [N_MEMORY
] = { { [0] = 1UL } },
97 [N_CPU
] = { { [0] = 1UL } },
100 EXPORT_SYMBOL(node_states
);
102 unsigned long totalram_pages __read_mostly
;
103 unsigned long totalreserve_pages __read_mostly
;
105 * When calculating the number of globally allowed dirty pages, there
106 * is a certain number of per-zone reserves that should not be
107 * considered dirtyable memory. This is the sum of those reserves
108 * over all existing zones that contribute dirtyable memory.
110 unsigned long dirty_balance_reserve __read_mostly
;
112 int percpu_pagelist_fraction
;
113 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
115 #ifdef CONFIG_PM_SLEEP
117 * The following functions are used by the suspend/hibernate code to temporarily
118 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
119 * while devices are suspended. To avoid races with the suspend/hibernate code,
120 * they should always be called with pm_mutex held (gfp_allowed_mask also should
121 * only be modified with pm_mutex held, unless the suspend/hibernate code is
122 * guaranteed not to run in parallel with that modification).
125 static gfp_t saved_gfp_mask
;
127 void pm_restore_gfp_mask(void)
129 WARN_ON(!mutex_is_locked(&pm_mutex
));
130 if (saved_gfp_mask
) {
131 gfp_allowed_mask
= saved_gfp_mask
;
136 void pm_restrict_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex
));
139 WARN_ON(saved_gfp_mask
);
140 saved_gfp_mask
= gfp_allowed_mask
;
141 gfp_allowed_mask
&= ~GFP_IOFS
;
144 bool pm_suspended_storage(void)
146 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
150 #endif /* CONFIG_PM_SLEEP */
152 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
153 int pageblock_order __read_mostly
;
156 static void __free_pages_ok(struct page
*page
, unsigned int order
);
159 * results with 256, 32 in the lowmem_reserve sysctl:
160 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
161 * 1G machine -> (16M dma, 784M normal, 224M high)
162 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
163 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
164 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
166 * TBD: should special case ZONE_DMA32 machines here - in those we normally
167 * don't need any ZONE_NORMAL reservation
169 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
170 #ifdef CONFIG_ZONE_DMA
173 #ifdef CONFIG_ZONE_DMA32
176 #ifdef CONFIG_HIGHMEM
182 EXPORT_SYMBOL(totalram_pages
);
184 static char * const zone_names
[MAX_NR_ZONES
] = {
185 #ifdef CONFIG_ZONE_DMA
188 #ifdef CONFIG_ZONE_DMA32
192 #ifdef CONFIG_HIGHMEM
198 int min_free_kbytes
= 1024;
200 static unsigned long __meminitdata nr_kernel_pages
;
201 static unsigned long __meminitdata nr_all_pages
;
202 static unsigned long __meminitdata dma_reserve
;
204 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
205 /* Movable memory ranges, will also be used by memblock subsystem. */
206 struct movablemem_map movablemem_map
= {
211 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
212 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
213 static unsigned long __initdata required_kernelcore
;
214 static unsigned long __initdata required_movablecore
;
215 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
216 static unsigned long __meminitdata zone_movable_limit
[MAX_NUMNODES
];
218 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
220 EXPORT_SYMBOL(movable_zone
);
221 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
224 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
225 int nr_online_nodes __read_mostly
= 1;
226 EXPORT_SYMBOL(nr_node_ids
);
227 EXPORT_SYMBOL(nr_online_nodes
);
230 int page_group_by_mobility_disabled __read_mostly
;
232 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
235 if (unlikely(page_group_by_mobility_disabled
))
236 migratetype
= MIGRATE_UNMOVABLE
;
238 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
239 PB_migrate
, PB_migrate_end
);
242 bool oom_killer_disabled __read_mostly
;
244 #ifdef CONFIG_DEBUG_VM
245 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
249 unsigned long pfn
= page_to_pfn(page
);
250 unsigned long sp
, start_pfn
;
253 seq
= zone_span_seqbegin(zone
);
254 start_pfn
= zone
->zone_start_pfn
;
255 sp
= zone
->spanned_pages
;
256 if (!zone_spans_pfn(zone
, pfn
))
258 } while (zone_span_seqretry(zone
, seq
));
261 pr_err("page %lu outside zone [ %lu - %lu ]\n",
262 pfn
, start_pfn
, start_pfn
+ sp
);
267 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
269 if (!pfn_valid_within(page_to_pfn(page
)))
271 if (zone
!= page_zone(page
))
277 * Temporary debugging check for pages not lying within a given zone.
279 static int bad_range(struct zone
*zone
, struct page
*page
)
281 if (page_outside_zone_boundaries(zone
, page
))
283 if (!page_is_consistent(zone
, page
))
289 static inline int bad_range(struct zone
*zone
, struct page
*page
)
295 static void bad_page(struct page
*page
)
297 static unsigned long resume
;
298 static unsigned long nr_shown
;
299 static unsigned long nr_unshown
;
301 /* Don't complain about poisoned pages */
302 if (PageHWPoison(page
)) {
303 page_mapcount_reset(page
); /* remove PageBuddy */
308 * Allow a burst of 60 reports, then keep quiet for that minute;
309 * or allow a steady drip of one report per second.
311 if (nr_shown
== 60) {
312 if (time_before(jiffies
, resume
)) {
318 "BUG: Bad page state: %lu messages suppressed\n",
325 resume
= jiffies
+ 60 * HZ
;
327 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
328 current
->comm
, page_to_pfn(page
));
334 /* Leave bad fields for debug, except PageBuddy could make trouble */
335 page_mapcount_reset(page
); /* remove PageBuddy */
336 add_taint(TAINT_BAD_PAGE
);
340 * Higher-order pages are called "compound pages". They are structured thusly:
342 * The first PAGE_SIZE page is called the "head page".
344 * The remaining PAGE_SIZE pages are called "tail pages".
346 * All pages have PG_compound set. All tail pages have their ->first_page
347 * pointing at the head page.
349 * The first tail page's ->lru.next holds the address of the compound page's
350 * put_page() function. Its ->lru.prev holds the order of allocation.
351 * This usage means that zero-order pages may not be compound.
354 static void free_compound_page(struct page
*page
)
356 __free_pages_ok(page
, compound_order(page
));
359 void prep_compound_page(struct page
*page
, unsigned long order
)
362 int nr_pages
= 1 << order
;
364 set_compound_page_dtor(page
, free_compound_page
);
365 set_compound_order(page
, order
);
367 for (i
= 1; i
< nr_pages
; i
++) {
368 struct page
*p
= page
+ i
;
370 set_page_count(p
, 0);
371 p
->first_page
= page
;
375 /* update __split_huge_page_refcount if you change this function */
376 static int destroy_compound_page(struct page
*page
, unsigned long order
)
379 int nr_pages
= 1 << order
;
382 if (unlikely(compound_order(page
) != order
)) {
387 __ClearPageHead(page
);
389 for (i
= 1; i
< nr_pages
; i
++) {
390 struct page
*p
= page
+ i
;
392 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
402 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
407 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
408 * and __GFP_HIGHMEM from hard or soft interrupt context.
410 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
411 for (i
= 0; i
< (1 << order
); i
++)
412 clear_highpage(page
+ i
);
415 #ifdef CONFIG_DEBUG_PAGEALLOC
416 unsigned int _debug_guardpage_minorder
;
418 static int __init
debug_guardpage_minorder_setup(char *buf
)
422 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
423 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
426 _debug_guardpage_minorder
= res
;
427 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
430 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
432 static inline void set_page_guard_flag(struct page
*page
)
434 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
437 static inline void clear_page_guard_flag(struct page
*page
)
439 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
442 static inline void set_page_guard_flag(struct page
*page
) { }
443 static inline void clear_page_guard_flag(struct page
*page
) { }
446 static inline void set_page_order(struct page
*page
, int order
)
448 set_page_private(page
, order
);
449 __SetPageBuddy(page
);
452 static inline void rmv_page_order(struct page
*page
)
454 __ClearPageBuddy(page
);
455 set_page_private(page
, 0);
459 * Locate the struct page for both the matching buddy in our
460 * pair (buddy1) and the combined O(n+1) page they form (page).
462 * 1) Any buddy B1 will have an order O twin B2 which satisfies
463 * the following equation:
465 * For example, if the starting buddy (buddy2) is #8 its order
467 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
469 * 2) Any buddy B will have an order O+1 parent P which
470 * satisfies the following equation:
473 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
475 static inline unsigned long
476 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
478 return page_idx
^ (1 << order
);
482 * This function checks whether a page is free && is the buddy
483 * we can do coalesce a page and its buddy if
484 * (a) the buddy is not in a hole &&
485 * (b) the buddy is in the buddy system &&
486 * (c) a page and its buddy have the same order &&
487 * (d) a page and its buddy are in the same zone.
489 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
490 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
492 * For recording page's order, we use page_private(page).
494 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
497 if (!pfn_valid_within(page_to_pfn(buddy
)))
500 if (page_zone_id(page
) != page_zone_id(buddy
))
503 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
504 VM_BUG_ON(page_count(buddy
) != 0);
508 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
509 VM_BUG_ON(page_count(buddy
) != 0);
516 * Freeing function for a buddy system allocator.
518 * The concept of a buddy system is to maintain direct-mapped table
519 * (containing bit values) for memory blocks of various "orders".
520 * The bottom level table contains the map for the smallest allocatable
521 * units of memory (here, pages), and each level above it describes
522 * pairs of units from the levels below, hence, "buddies".
523 * At a high level, all that happens here is marking the table entry
524 * at the bottom level available, and propagating the changes upward
525 * as necessary, plus some accounting needed to play nicely with other
526 * parts of the VM system.
527 * At each level, we keep a list of pages, which are heads of continuous
528 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
529 * order is recorded in page_private(page) field.
530 * So when we are allocating or freeing one, we can derive the state of the
531 * other. That is, if we allocate a small block, and both were
532 * free, the remainder of the region must be split into blocks.
533 * If a block is freed, and its buddy is also free, then this
534 * triggers coalescing into a block of larger size.
539 static inline void __free_one_page(struct page
*page
,
540 struct zone
*zone
, unsigned int order
,
543 unsigned long page_idx
;
544 unsigned long combined_idx
;
545 unsigned long uninitialized_var(buddy_idx
);
548 VM_BUG_ON(!zone_is_initialized(zone
));
550 if (unlikely(PageCompound(page
)))
551 if (unlikely(destroy_compound_page(page
, order
)))
554 VM_BUG_ON(migratetype
== -1);
556 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
558 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
559 VM_BUG_ON(bad_range(zone
, page
));
561 while (order
< MAX_ORDER
-1) {
562 buddy_idx
= __find_buddy_index(page_idx
, order
);
563 buddy
= page
+ (buddy_idx
- page_idx
);
564 if (!page_is_buddy(page
, buddy
, order
))
567 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
568 * merge with it and move up one order.
570 if (page_is_guard(buddy
)) {
571 clear_page_guard_flag(buddy
);
572 set_page_private(page
, 0);
573 __mod_zone_freepage_state(zone
, 1 << order
,
576 list_del(&buddy
->lru
);
577 zone
->free_area
[order
].nr_free
--;
578 rmv_page_order(buddy
);
580 combined_idx
= buddy_idx
& page_idx
;
581 page
= page
+ (combined_idx
- page_idx
);
582 page_idx
= combined_idx
;
585 set_page_order(page
, order
);
588 * If this is not the largest possible page, check if the buddy
589 * of the next-highest order is free. If it is, it's possible
590 * that pages are being freed that will coalesce soon. In case,
591 * that is happening, add the free page to the tail of the list
592 * so it's less likely to be used soon and more likely to be merged
593 * as a higher order page
595 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
596 struct page
*higher_page
, *higher_buddy
;
597 combined_idx
= buddy_idx
& page_idx
;
598 higher_page
= page
+ (combined_idx
- page_idx
);
599 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
600 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
601 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
602 list_add_tail(&page
->lru
,
603 &zone
->free_area
[order
].free_list
[migratetype
]);
608 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
610 zone
->free_area
[order
].nr_free
++;
613 static inline int free_pages_check(struct page
*page
)
615 if (unlikely(page_mapcount(page
) |
616 (page
->mapping
!= NULL
) |
617 (atomic_read(&page
->_count
) != 0) |
618 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
) |
619 (mem_cgroup_bad_page_check(page
)))) {
623 page_nid_reset_last(page
);
624 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
625 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
630 * Frees a number of pages from the PCP lists
631 * Assumes all pages on list are in same zone, and of same order.
632 * count is the number of pages to free.
634 * If the zone was previously in an "all pages pinned" state then look to
635 * see if this freeing clears that state.
637 * And clear the zone's pages_scanned counter, to hold off the "all pages are
638 * pinned" detection logic.
640 static void free_pcppages_bulk(struct zone
*zone
, int count
,
641 struct per_cpu_pages
*pcp
)
647 spin_lock(&zone
->lock
);
648 zone
->all_unreclaimable
= 0;
649 zone
->pages_scanned
= 0;
653 struct list_head
*list
;
656 * Remove pages from lists in a round-robin fashion. A
657 * batch_free count is maintained that is incremented when an
658 * empty list is encountered. This is so more pages are freed
659 * off fuller lists instead of spinning excessively around empty
664 if (++migratetype
== MIGRATE_PCPTYPES
)
666 list
= &pcp
->lists
[migratetype
];
667 } while (list_empty(list
));
669 /* This is the only non-empty list. Free them all. */
670 if (batch_free
== MIGRATE_PCPTYPES
)
671 batch_free
= to_free
;
674 int mt
; /* migratetype of the to-be-freed page */
676 page
= list_entry(list
->prev
, struct page
, lru
);
677 /* must delete as __free_one_page list manipulates */
678 list_del(&page
->lru
);
679 mt
= get_freepage_migratetype(page
);
680 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
681 __free_one_page(page
, zone
, 0, mt
);
682 trace_mm_page_pcpu_drain(page
, 0, mt
);
683 if (likely(!is_migrate_isolate_page(page
))) {
684 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1);
685 if (is_migrate_cma(mt
))
686 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
, 1);
688 } while (--to_free
&& --batch_free
&& !list_empty(list
));
690 spin_unlock(&zone
->lock
);
693 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
696 spin_lock(&zone
->lock
);
697 zone
->all_unreclaimable
= 0;
698 zone
->pages_scanned
= 0;
700 __free_one_page(page
, zone
, order
, migratetype
);
701 if (unlikely(!is_migrate_isolate(migratetype
)))
702 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
703 spin_unlock(&zone
->lock
);
706 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
711 trace_mm_page_free(page
, order
);
712 kmemcheck_free_shadow(page
, order
);
715 page
->mapping
= NULL
;
716 for (i
= 0; i
< (1 << order
); i
++)
717 bad
+= free_pages_check(page
+ i
);
721 if (!PageHighMem(page
)) {
722 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
723 debug_check_no_obj_freed(page_address(page
),
726 arch_free_page(page
, order
);
727 kernel_map_pages(page
, 1 << order
, 0);
732 static void __free_pages_ok(struct page
*page
, unsigned int order
)
737 if (!free_pages_prepare(page
, order
))
740 local_irq_save(flags
);
741 __count_vm_events(PGFREE
, 1 << order
);
742 migratetype
= get_pageblock_migratetype(page
);
743 set_freepage_migratetype(page
, migratetype
);
744 free_one_page(page_zone(page
), page
, order
, migratetype
);
745 local_irq_restore(flags
);
749 * Read access to zone->managed_pages is safe because it's unsigned long,
750 * but we still need to serialize writers. Currently all callers of
751 * __free_pages_bootmem() except put_page_bootmem() should only be used
752 * at boot time. So for shorter boot time, we shift the burden to
753 * put_page_bootmem() to serialize writers.
755 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
757 unsigned int nr_pages
= 1 << order
;
761 for (loop
= 0; loop
< nr_pages
; loop
++) {
762 struct page
*p
= &page
[loop
];
764 if (loop
+ 1 < nr_pages
)
766 __ClearPageReserved(p
);
767 set_page_count(p
, 0);
770 page_zone(page
)->managed_pages
+= 1 << order
;
771 set_page_refcounted(page
);
772 __free_pages(page
, order
);
776 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
777 void __init
init_cma_reserved_pageblock(struct page
*page
)
779 unsigned i
= pageblock_nr_pages
;
780 struct page
*p
= page
;
783 __ClearPageReserved(p
);
784 set_page_count(p
, 0);
787 set_page_refcounted(page
);
788 set_pageblock_migratetype(page
, MIGRATE_CMA
);
789 __free_pages(page
, pageblock_order
);
790 totalram_pages
+= pageblock_nr_pages
;
791 #ifdef CONFIG_HIGHMEM
792 if (PageHighMem(page
))
793 totalhigh_pages
+= pageblock_nr_pages
;
799 * The order of subdivision here is critical for the IO subsystem.
800 * Please do not alter this order without good reasons and regression
801 * testing. Specifically, as large blocks of memory are subdivided,
802 * the order in which smaller blocks are delivered depends on the order
803 * they're subdivided in this function. This is the primary factor
804 * influencing the order in which pages are delivered to the IO
805 * subsystem according to empirical testing, and this is also justified
806 * by considering the behavior of a buddy system containing a single
807 * large block of memory acted on by a series of small allocations.
808 * This behavior is a critical factor in sglist merging's success.
812 static inline void expand(struct zone
*zone
, struct page
*page
,
813 int low
, int high
, struct free_area
*area
,
816 unsigned long size
= 1 << high
;
822 VM_BUG_ON(bad_range(zone
, &page
[size
]));
824 #ifdef CONFIG_DEBUG_PAGEALLOC
825 if (high
< debug_guardpage_minorder()) {
827 * Mark as guard pages (or page), that will allow to
828 * merge back to allocator when buddy will be freed.
829 * Corresponding page table entries will not be touched,
830 * pages will stay not present in virtual address space
832 INIT_LIST_HEAD(&page
[size
].lru
);
833 set_page_guard_flag(&page
[size
]);
834 set_page_private(&page
[size
], high
);
835 /* Guard pages are not available for any usage */
836 __mod_zone_freepage_state(zone
, -(1 << high
),
841 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
843 set_page_order(&page
[size
], high
);
848 * This page is about to be returned from the page allocator
850 static inline int check_new_page(struct page
*page
)
852 if (unlikely(page_mapcount(page
) |
853 (page
->mapping
!= NULL
) |
854 (atomic_read(&page
->_count
) != 0) |
855 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
) |
856 (mem_cgroup_bad_page_check(page
)))) {
863 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
867 for (i
= 0; i
< (1 << order
); i
++) {
868 struct page
*p
= page
+ i
;
869 if (unlikely(check_new_page(p
)))
873 set_page_private(page
, 0);
874 set_page_refcounted(page
);
876 arch_alloc_page(page
, order
);
877 kernel_map_pages(page
, 1 << order
, 1);
879 if (gfp_flags
& __GFP_ZERO
)
880 prep_zero_page(page
, order
, gfp_flags
);
882 if (order
&& (gfp_flags
& __GFP_COMP
))
883 prep_compound_page(page
, order
);
889 * Go through the free lists for the given migratetype and remove
890 * the smallest available page from the freelists
893 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
896 unsigned int current_order
;
897 struct free_area
* area
;
900 /* Find a page of the appropriate size in the preferred list */
901 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
902 area
= &(zone
->free_area
[current_order
]);
903 if (list_empty(&area
->free_list
[migratetype
]))
906 page
= list_entry(area
->free_list
[migratetype
].next
,
908 list_del(&page
->lru
);
909 rmv_page_order(page
);
911 expand(zone
, page
, order
, current_order
, area
, migratetype
);
920 * This array describes the order lists are fallen back to when
921 * the free lists for the desirable migrate type are depleted
923 static int fallbacks
[MIGRATE_TYPES
][4] = {
924 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
925 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
927 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
928 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
930 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
932 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
933 #ifdef CONFIG_MEMORY_ISOLATION
934 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
939 * Move the free pages in a range to the free lists of the requested type.
940 * Note that start_page and end_pages are not aligned on a pageblock
941 * boundary. If alignment is required, use move_freepages_block()
943 int move_freepages(struct zone
*zone
,
944 struct page
*start_page
, struct page
*end_page
,
951 #ifndef CONFIG_HOLES_IN_ZONE
953 * page_zone is not safe to call in this context when
954 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
955 * anyway as we check zone boundaries in move_freepages_block().
956 * Remove at a later date when no bug reports exist related to
957 * grouping pages by mobility
959 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
962 for (page
= start_page
; page
<= end_page
;) {
963 /* Make sure we are not inadvertently changing nodes */
964 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
966 if (!pfn_valid_within(page_to_pfn(page
))) {
971 if (!PageBuddy(page
)) {
976 order
= page_order(page
);
977 list_move(&page
->lru
,
978 &zone
->free_area
[order
].free_list
[migratetype
]);
979 set_freepage_migratetype(page
, migratetype
);
981 pages_moved
+= 1 << order
;
987 int move_freepages_block(struct zone
*zone
, struct page
*page
,
990 unsigned long start_pfn
, end_pfn
;
991 struct page
*start_page
, *end_page
;
993 start_pfn
= page_to_pfn(page
);
994 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
995 start_page
= pfn_to_page(start_pfn
);
996 end_page
= start_page
+ pageblock_nr_pages
- 1;
997 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
999 /* Do not cross zone boundaries */
1000 if (!zone_spans_pfn(zone
, start_pfn
))
1002 if (!zone_spans_pfn(zone
, end_pfn
))
1005 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1008 static void change_pageblock_range(struct page
*pageblock_page
,
1009 int start_order
, int migratetype
)
1011 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1013 while (nr_pageblocks
--) {
1014 set_pageblock_migratetype(pageblock_page
, migratetype
);
1015 pageblock_page
+= pageblock_nr_pages
;
1019 /* Remove an element from the buddy allocator from the fallback list */
1020 static inline struct page
*
1021 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
1023 struct free_area
* area
;
1028 /* Find the largest possible block of pages in the other list */
1029 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
1032 migratetype
= fallbacks
[start_migratetype
][i
];
1034 /* MIGRATE_RESERVE handled later if necessary */
1035 if (migratetype
== MIGRATE_RESERVE
)
1038 area
= &(zone
->free_area
[current_order
]);
1039 if (list_empty(&area
->free_list
[migratetype
]))
1042 page
= list_entry(area
->free_list
[migratetype
].next
,
1047 * If breaking a large block of pages, move all free
1048 * pages to the preferred allocation list. If falling
1049 * back for a reclaimable kernel allocation, be more
1050 * aggressive about taking ownership of free pages
1052 * On the other hand, never change migration
1053 * type of MIGRATE_CMA pageblocks nor move CMA
1054 * pages on different free lists. We don't
1055 * want unmovable pages to be allocated from
1056 * MIGRATE_CMA areas.
1058 if (!is_migrate_cma(migratetype
) &&
1059 (unlikely(current_order
>= pageblock_order
/ 2) ||
1060 start_migratetype
== MIGRATE_RECLAIMABLE
||
1061 page_group_by_mobility_disabled
)) {
1063 pages
= move_freepages_block(zone
, page
,
1066 /* Claim the whole block if over half of it is free */
1067 if (pages
>= (1 << (pageblock_order
-1)) ||
1068 page_group_by_mobility_disabled
)
1069 set_pageblock_migratetype(page
,
1072 migratetype
= start_migratetype
;
1075 /* Remove the page from the freelists */
1076 list_del(&page
->lru
);
1077 rmv_page_order(page
);
1079 /* Take ownership for orders >= pageblock_order */
1080 if (current_order
>= pageblock_order
&&
1081 !is_migrate_cma(migratetype
))
1082 change_pageblock_range(page
, current_order
,
1085 expand(zone
, page
, order
, current_order
, area
,
1086 is_migrate_cma(migratetype
)
1087 ? migratetype
: start_migratetype
);
1089 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1090 start_migratetype
, migratetype
);
1100 * Do the hard work of removing an element from the buddy allocator.
1101 * Call me with the zone->lock already held.
1103 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1109 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1111 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1112 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1115 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1116 * is used because __rmqueue_smallest is an inline function
1117 * and we want just one call site
1120 migratetype
= MIGRATE_RESERVE
;
1125 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1130 * Obtain a specified number of elements from the buddy allocator, all under
1131 * a single hold of the lock, for efficiency. Add them to the supplied list.
1132 * Returns the number of new pages which were placed at *list.
1134 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1135 unsigned long count
, struct list_head
*list
,
1136 int migratetype
, int cold
)
1138 int mt
= migratetype
, i
;
1140 spin_lock(&zone
->lock
);
1141 for (i
= 0; i
< count
; ++i
) {
1142 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1143 if (unlikely(page
== NULL
))
1147 * Split buddy pages returned by expand() are received here
1148 * in physical page order. The page is added to the callers and
1149 * list and the list head then moves forward. From the callers
1150 * perspective, the linked list is ordered by page number in
1151 * some conditions. This is useful for IO devices that can
1152 * merge IO requests if the physical pages are ordered
1155 if (likely(cold
== 0))
1156 list_add(&page
->lru
, list
);
1158 list_add_tail(&page
->lru
, list
);
1159 if (IS_ENABLED(CONFIG_CMA
)) {
1160 mt
= get_pageblock_migratetype(page
);
1161 if (!is_migrate_cma(mt
) && !is_migrate_isolate(mt
))
1164 set_freepage_migratetype(page
, mt
);
1166 if (is_migrate_cma(mt
))
1167 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1170 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1171 spin_unlock(&zone
->lock
);
1177 * Called from the vmstat counter updater to drain pagesets of this
1178 * currently executing processor on remote nodes after they have
1181 * Note that this function must be called with the thread pinned to
1182 * a single processor.
1184 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1186 unsigned long flags
;
1189 local_irq_save(flags
);
1190 if (pcp
->count
>= pcp
->batch
)
1191 to_drain
= pcp
->batch
;
1193 to_drain
= pcp
->count
;
1195 free_pcppages_bulk(zone
, to_drain
, pcp
);
1196 pcp
->count
-= to_drain
;
1198 local_irq_restore(flags
);
1203 * Drain pages of the indicated processor.
1205 * The processor must either be the current processor and the
1206 * thread pinned to the current processor or a processor that
1209 static void drain_pages(unsigned int cpu
)
1211 unsigned long flags
;
1214 for_each_populated_zone(zone
) {
1215 struct per_cpu_pageset
*pset
;
1216 struct per_cpu_pages
*pcp
;
1218 local_irq_save(flags
);
1219 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1223 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1226 local_irq_restore(flags
);
1231 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1233 void drain_local_pages(void *arg
)
1235 drain_pages(smp_processor_id());
1239 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1241 * Note that this code is protected against sending an IPI to an offline
1242 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1243 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1244 * nothing keeps CPUs from showing up after we populated the cpumask and
1245 * before the call to on_each_cpu_mask().
1247 void drain_all_pages(void)
1250 struct per_cpu_pageset
*pcp
;
1254 * Allocate in the BSS so we wont require allocation in
1255 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1257 static cpumask_t cpus_with_pcps
;
1260 * We don't care about racing with CPU hotplug event
1261 * as offline notification will cause the notified
1262 * cpu to drain that CPU pcps and on_each_cpu_mask
1263 * disables preemption as part of its processing
1265 for_each_online_cpu(cpu
) {
1266 bool has_pcps
= false;
1267 for_each_populated_zone(zone
) {
1268 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1269 if (pcp
->pcp
.count
) {
1275 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1277 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1279 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1282 #ifdef CONFIG_HIBERNATION
1284 void mark_free_pages(struct zone
*zone
)
1286 unsigned long pfn
, max_zone_pfn
;
1287 unsigned long flags
;
1289 struct list_head
*curr
;
1291 if (!zone
->spanned_pages
)
1294 spin_lock_irqsave(&zone
->lock
, flags
);
1296 max_zone_pfn
= zone_end_pfn(zone
);
1297 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1298 if (pfn_valid(pfn
)) {
1299 struct page
*page
= pfn_to_page(pfn
);
1301 if (!swsusp_page_is_forbidden(page
))
1302 swsusp_unset_page_free(page
);
1305 for_each_migratetype_order(order
, t
) {
1306 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1309 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1310 for (i
= 0; i
< (1UL << order
); i
++)
1311 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1314 spin_unlock_irqrestore(&zone
->lock
, flags
);
1316 #endif /* CONFIG_PM */
1319 * Free a 0-order page
1320 * cold == 1 ? free a cold page : free a hot page
1322 void free_hot_cold_page(struct page
*page
, int cold
)
1324 struct zone
*zone
= page_zone(page
);
1325 struct per_cpu_pages
*pcp
;
1326 unsigned long flags
;
1329 if (!free_pages_prepare(page
, 0))
1332 migratetype
= get_pageblock_migratetype(page
);
1333 set_freepage_migratetype(page
, migratetype
);
1334 local_irq_save(flags
);
1335 __count_vm_event(PGFREE
);
1338 * We only track unmovable, reclaimable and movable on pcp lists.
1339 * Free ISOLATE pages back to the allocator because they are being
1340 * offlined but treat RESERVE as movable pages so we can get those
1341 * areas back if necessary. Otherwise, we may have to free
1342 * excessively into the page allocator
1344 if (migratetype
>= MIGRATE_PCPTYPES
) {
1345 if (unlikely(is_migrate_isolate(migratetype
))) {
1346 free_one_page(zone
, page
, 0, migratetype
);
1349 migratetype
= MIGRATE_MOVABLE
;
1352 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1354 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1356 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1358 if (pcp
->count
>= pcp
->high
) {
1359 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1360 pcp
->count
-= pcp
->batch
;
1364 local_irq_restore(flags
);
1368 * Free a list of 0-order pages
1370 void free_hot_cold_page_list(struct list_head
*list
, int cold
)
1372 struct page
*page
, *next
;
1374 list_for_each_entry_safe(page
, next
, list
, lru
) {
1375 trace_mm_page_free_batched(page
, cold
);
1376 free_hot_cold_page(page
, cold
);
1381 * split_page takes a non-compound higher-order page, and splits it into
1382 * n (1<<order) sub-pages: page[0..n]
1383 * Each sub-page must be freed individually.
1385 * Note: this is probably too low level an operation for use in drivers.
1386 * Please consult with lkml before using this in your driver.
1388 void split_page(struct page
*page
, unsigned int order
)
1392 VM_BUG_ON(PageCompound(page
));
1393 VM_BUG_ON(!page_count(page
));
1395 #ifdef CONFIG_KMEMCHECK
1397 * Split shadow pages too, because free(page[0]) would
1398 * otherwise free the whole shadow.
1400 if (kmemcheck_page_is_tracked(page
))
1401 split_page(virt_to_page(page
[0].shadow
), order
);
1404 for (i
= 1; i
< (1 << order
); i
++)
1405 set_page_refcounted(page
+ i
);
1408 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1410 unsigned long watermark
;
1414 BUG_ON(!PageBuddy(page
));
1416 zone
= page_zone(page
);
1417 mt
= get_pageblock_migratetype(page
);
1419 if (!is_migrate_isolate(mt
)) {
1420 /* Obey watermarks as if the page was being allocated */
1421 watermark
= low_wmark_pages(zone
) + (1 << order
);
1422 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1425 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1428 /* Remove page from free list */
1429 list_del(&page
->lru
);
1430 zone
->free_area
[order
].nr_free
--;
1431 rmv_page_order(page
);
1433 /* Set the pageblock if the isolated page is at least a pageblock */
1434 if (order
>= pageblock_order
- 1) {
1435 struct page
*endpage
= page
+ (1 << order
) - 1;
1436 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1437 int mt
= get_pageblock_migratetype(page
);
1438 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1439 set_pageblock_migratetype(page
,
1444 return 1UL << order
;
1448 * Similar to split_page except the page is already free. As this is only
1449 * being used for migration, the migratetype of the block also changes.
1450 * As this is called with interrupts disabled, the caller is responsible
1451 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1454 * Note: this is probably too low level an operation for use in drivers.
1455 * Please consult with lkml before using this in your driver.
1457 int split_free_page(struct page
*page
)
1462 order
= page_order(page
);
1464 nr_pages
= __isolate_free_page(page
, order
);
1468 /* Split into individual pages */
1469 set_page_refcounted(page
);
1470 split_page(page
, order
);
1475 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1476 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1480 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1481 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1484 unsigned long flags
;
1486 int cold
= !!(gfp_flags
& __GFP_COLD
);
1489 if (likely(order
== 0)) {
1490 struct per_cpu_pages
*pcp
;
1491 struct list_head
*list
;
1493 local_irq_save(flags
);
1494 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1495 list
= &pcp
->lists
[migratetype
];
1496 if (list_empty(list
)) {
1497 pcp
->count
+= rmqueue_bulk(zone
, 0,
1500 if (unlikely(list_empty(list
)))
1505 page
= list_entry(list
->prev
, struct page
, lru
);
1507 page
= list_entry(list
->next
, struct page
, lru
);
1509 list_del(&page
->lru
);
1512 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1514 * __GFP_NOFAIL is not to be used in new code.
1516 * All __GFP_NOFAIL callers should be fixed so that they
1517 * properly detect and handle allocation failures.
1519 * We most definitely don't want callers attempting to
1520 * allocate greater than order-1 page units with
1523 WARN_ON_ONCE(order
> 1);
1525 spin_lock_irqsave(&zone
->lock
, flags
);
1526 page
= __rmqueue(zone
, order
, migratetype
);
1527 spin_unlock(&zone
->lock
);
1530 __mod_zone_freepage_state(zone
, -(1 << order
),
1531 get_pageblock_migratetype(page
));
1534 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1535 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1536 local_irq_restore(flags
);
1538 VM_BUG_ON(bad_range(zone
, page
));
1539 if (prep_new_page(page
, order
, gfp_flags
))
1544 local_irq_restore(flags
);
1548 #ifdef CONFIG_FAIL_PAGE_ALLOC
1551 struct fault_attr attr
;
1553 u32 ignore_gfp_highmem
;
1554 u32 ignore_gfp_wait
;
1556 } fail_page_alloc
= {
1557 .attr
= FAULT_ATTR_INITIALIZER
,
1558 .ignore_gfp_wait
= 1,
1559 .ignore_gfp_highmem
= 1,
1563 static int __init
setup_fail_page_alloc(char *str
)
1565 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1567 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1569 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1571 if (order
< fail_page_alloc
.min_order
)
1573 if (gfp_mask
& __GFP_NOFAIL
)
1575 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1577 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1580 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1583 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1585 static int __init
fail_page_alloc_debugfs(void)
1587 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1590 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1591 &fail_page_alloc
.attr
);
1593 return PTR_ERR(dir
);
1595 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1596 &fail_page_alloc
.ignore_gfp_wait
))
1598 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1599 &fail_page_alloc
.ignore_gfp_highmem
))
1601 if (!debugfs_create_u32("min-order", mode
, dir
,
1602 &fail_page_alloc
.min_order
))
1607 debugfs_remove_recursive(dir
);
1612 late_initcall(fail_page_alloc_debugfs
);
1614 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1616 #else /* CONFIG_FAIL_PAGE_ALLOC */
1618 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1623 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1626 * Return true if free pages are above 'mark'. This takes into account the order
1627 * of the allocation.
1629 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1630 int classzone_idx
, int alloc_flags
, long free_pages
)
1632 /* free_pages my go negative - that's OK */
1634 long lowmem_reserve
= z
->lowmem_reserve
[classzone_idx
];
1637 free_pages
-= (1 << order
) - 1;
1638 if (alloc_flags
& ALLOC_HIGH
)
1640 if (alloc_flags
& ALLOC_HARDER
)
1643 /* If allocation can't use CMA areas don't use free CMA pages */
1644 if (!(alloc_flags
& ALLOC_CMA
))
1645 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1647 if (free_pages
<= min
+ lowmem_reserve
)
1649 for (o
= 0; o
< order
; o
++) {
1650 /* At the next order, this order's pages become unavailable */
1651 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1653 /* Require fewer higher order pages to be free */
1656 if (free_pages
<= min
)
1662 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1663 int classzone_idx
, int alloc_flags
)
1665 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1666 zone_page_state(z
, NR_FREE_PAGES
));
1669 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1670 int classzone_idx
, int alloc_flags
)
1672 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1674 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1675 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1677 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1683 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1684 * skip over zones that are not allowed by the cpuset, or that have
1685 * been recently (in last second) found to be nearly full. See further
1686 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1687 * that have to skip over a lot of full or unallowed zones.
1689 * If the zonelist cache is present in the passed in zonelist, then
1690 * returns a pointer to the allowed node mask (either the current
1691 * tasks mems_allowed, or node_states[N_MEMORY].)
1693 * If the zonelist cache is not available for this zonelist, does
1694 * nothing and returns NULL.
1696 * If the fullzones BITMAP in the zonelist cache is stale (more than
1697 * a second since last zap'd) then we zap it out (clear its bits.)
1699 * We hold off even calling zlc_setup, until after we've checked the
1700 * first zone in the zonelist, on the theory that most allocations will
1701 * be satisfied from that first zone, so best to examine that zone as
1702 * quickly as we can.
1704 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1706 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1707 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1709 zlc
= zonelist
->zlcache_ptr
;
1713 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1714 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1715 zlc
->last_full_zap
= jiffies
;
1718 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1719 &cpuset_current_mems_allowed
:
1720 &node_states
[N_MEMORY
];
1721 return allowednodes
;
1725 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1726 * if it is worth looking at further for free memory:
1727 * 1) Check that the zone isn't thought to be full (doesn't have its
1728 * bit set in the zonelist_cache fullzones BITMAP).
1729 * 2) Check that the zones node (obtained from the zonelist_cache
1730 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1731 * Return true (non-zero) if zone is worth looking at further, or
1732 * else return false (zero) if it is not.
1734 * This check -ignores- the distinction between various watermarks,
1735 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1736 * found to be full for any variation of these watermarks, it will
1737 * be considered full for up to one second by all requests, unless
1738 * we are so low on memory on all allowed nodes that we are forced
1739 * into the second scan of the zonelist.
1741 * In the second scan we ignore this zonelist cache and exactly
1742 * apply the watermarks to all zones, even it is slower to do so.
1743 * We are low on memory in the second scan, and should leave no stone
1744 * unturned looking for a free page.
1746 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1747 nodemask_t
*allowednodes
)
1749 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1750 int i
; /* index of *z in zonelist zones */
1751 int n
; /* node that zone *z is on */
1753 zlc
= zonelist
->zlcache_ptr
;
1757 i
= z
- zonelist
->_zonerefs
;
1760 /* This zone is worth trying if it is allowed but not full */
1761 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1765 * Given 'z' scanning a zonelist, set the corresponding bit in
1766 * zlc->fullzones, so that subsequent attempts to allocate a page
1767 * from that zone don't waste time re-examining it.
1769 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1771 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1772 int i
; /* index of *z in zonelist zones */
1774 zlc
= zonelist
->zlcache_ptr
;
1778 i
= z
- zonelist
->_zonerefs
;
1780 set_bit(i
, zlc
->fullzones
);
1784 * clear all zones full, called after direct reclaim makes progress so that
1785 * a zone that was recently full is not skipped over for up to a second
1787 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1789 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1791 zlc
= zonelist
->zlcache_ptr
;
1795 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1798 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1800 return node_isset(local_zone
->node
, zone
->zone_pgdat
->reclaim_nodes
);
1803 static void __paginginit
init_zone_allows_reclaim(int nid
)
1807 for_each_online_node(i
)
1808 if (node_distance(nid
, i
) <= RECLAIM_DISTANCE
)
1809 node_set(i
, NODE_DATA(nid
)->reclaim_nodes
);
1811 zone_reclaim_mode
= 1;
1814 #else /* CONFIG_NUMA */
1816 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1821 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1822 nodemask_t
*allowednodes
)
1827 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1831 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1835 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1840 static inline void init_zone_allows_reclaim(int nid
)
1843 #endif /* CONFIG_NUMA */
1846 * get_page_from_freelist goes through the zonelist trying to allocate
1849 static struct page
*
1850 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1851 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1852 struct zone
*preferred_zone
, int migratetype
)
1855 struct page
*page
= NULL
;
1858 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1859 int zlc_active
= 0; /* set if using zonelist_cache */
1860 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1862 classzone_idx
= zone_idx(preferred_zone
);
1865 * Scan zonelist, looking for a zone with enough free.
1866 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1868 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1869 high_zoneidx
, nodemask
) {
1870 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1871 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1873 if ((alloc_flags
& ALLOC_CPUSET
) &&
1874 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1877 * When allocating a page cache page for writing, we
1878 * want to get it from a zone that is within its dirty
1879 * limit, such that no single zone holds more than its
1880 * proportional share of globally allowed dirty pages.
1881 * The dirty limits take into account the zone's
1882 * lowmem reserves and high watermark so that kswapd
1883 * should be able to balance it without having to
1884 * write pages from its LRU list.
1886 * This may look like it could increase pressure on
1887 * lower zones by failing allocations in higher zones
1888 * before they are full. But the pages that do spill
1889 * over are limited as the lower zones are protected
1890 * by this very same mechanism. It should not become
1891 * a practical burden to them.
1893 * XXX: For now, allow allocations to potentially
1894 * exceed the per-zone dirty limit in the slowpath
1895 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1896 * which is important when on a NUMA setup the allowed
1897 * zones are together not big enough to reach the
1898 * global limit. The proper fix for these situations
1899 * will require awareness of zones in the
1900 * dirty-throttling and the flusher threads.
1902 if ((alloc_flags
& ALLOC_WMARK_LOW
) &&
1903 (gfp_mask
& __GFP_WRITE
) && !zone_dirty_ok(zone
))
1904 goto this_zone_full
;
1906 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1907 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1911 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1912 if (zone_watermark_ok(zone
, order
, mark
,
1913 classzone_idx
, alloc_flags
))
1916 if (IS_ENABLED(CONFIG_NUMA
) &&
1917 !did_zlc_setup
&& nr_online_nodes
> 1) {
1919 * we do zlc_setup if there are multiple nodes
1920 * and before considering the first zone allowed
1923 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1928 if (zone_reclaim_mode
== 0 ||
1929 !zone_allows_reclaim(preferred_zone
, zone
))
1930 goto this_zone_full
;
1933 * As we may have just activated ZLC, check if the first
1934 * eligible zone has failed zone_reclaim recently.
1936 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1937 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1940 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1942 case ZONE_RECLAIM_NOSCAN
:
1945 case ZONE_RECLAIM_FULL
:
1946 /* scanned but unreclaimable */
1949 /* did we reclaim enough */
1950 if (!zone_watermark_ok(zone
, order
, mark
,
1951 classzone_idx
, alloc_flags
))
1952 goto this_zone_full
;
1957 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1958 gfp_mask
, migratetype
);
1962 if (IS_ENABLED(CONFIG_NUMA
))
1963 zlc_mark_zone_full(zonelist
, z
);
1966 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && page
== NULL
&& zlc_active
)) {
1967 /* Disable zlc cache for second zonelist scan */
1974 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1975 * necessary to allocate the page. The expectation is
1976 * that the caller is taking steps that will free more
1977 * memory. The caller should avoid the page being used
1978 * for !PFMEMALLOC purposes.
1980 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
1986 * Large machines with many possible nodes should not always dump per-node
1987 * meminfo in irq context.
1989 static inline bool should_suppress_show_mem(void)
1994 ret
= in_interrupt();
1999 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2000 DEFAULT_RATELIMIT_INTERVAL
,
2001 DEFAULT_RATELIMIT_BURST
);
2003 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2005 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2007 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2008 debug_guardpage_minorder() > 0)
2012 * This documents exceptions given to allocations in certain
2013 * contexts that are allowed to allocate outside current's set
2016 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2017 if (test_thread_flag(TIF_MEMDIE
) ||
2018 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2019 filter
&= ~SHOW_MEM_FILTER_NODES
;
2020 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2021 filter
&= ~SHOW_MEM_FILTER_NODES
;
2024 struct va_format vaf
;
2027 va_start(args
, fmt
);
2032 pr_warn("%pV", &vaf
);
2037 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2038 current
->comm
, order
, gfp_mask
);
2041 if (!should_suppress_show_mem())
2046 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2047 unsigned long did_some_progress
,
2048 unsigned long pages_reclaimed
)
2050 /* Do not loop if specifically requested */
2051 if (gfp_mask
& __GFP_NORETRY
)
2054 /* Always retry if specifically requested */
2055 if (gfp_mask
& __GFP_NOFAIL
)
2059 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2060 * making forward progress without invoking OOM. Suspend also disables
2061 * storage devices so kswapd will not help. Bail if we are suspending.
2063 if (!did_some_progress
&& pm_suspended_storage())
2067 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2068 * means __GFP_NOFAIL, but that may not be true in other
2071 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2075 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2076 * specified, then we retry until we no longer reclaim any pages
2077 * (above), or we've reclaimed an order of pages at least as
2078 * large as the allocation's order. In both cases, if the
2079 * allocation still fails, we stop retrying.
2081 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2087 static inline struct page
*
2088 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2089 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2090 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2095 /* Acquire the OOM killer lock for the zones in zonelist */
2096 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
2097 schedule_timeout_uninterruptible(1);
2102 * Go through the zonelist yet one more time, keep very high watermark
2103 * here, this is only to catch a parallel oom killing, we must fail if
2104 * we're still under heavy pressure.
2106 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2107 order
, zonelist
, high_zoneidx
,
2108 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2109 preferred_zone
, migratetype
);
2113 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2114 /* The OOM killer will not help higher order allocs */
2115 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2117 /* The OOM killer does not needlessly kill tasks for lowmem */
2118 if (high_zoneidx
< ZONE_NORMAL
)
2121 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2122 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2123 * The caller should handle page allocation failure by itself if
2124 * it specifies __GFP_THISNODE.
2125 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2127 if (gfp_mask
& __GFP_THISNODE
)
2130 /* Exhausted what can be done so it's blamo time */
2131 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2134 clear_zonelist_oom(zonelist
, gfp_mask
);
2138 #ifdef CONFIG_COMPACTION
2139 /* Try memory compaction for high-order allocations before reclaim */
2140 static struct page
*
2141 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2142 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2143 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2144 int migratetype
, bool sync_migration
,
2145 bool *contended_compaction
, bool *deferred_compaction
,
2146 unsigned long *did_some_progress
)
2151 if (compaction_deferred(preferred_zone
, order
)) {
2152 *deferred_compaction
= true;
2156 current
->flags
|= PF_MEMALLOC
;
2157 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2158 nodemask
, sync_migration
,
2159 contended_compaction
);
2160 current
->flags
&= ~PF_MEMALLOC
;
2162 if (*did_some_progress
!= COMPACT_SKIPPED
) {
2165 /* Page migration frees to the PCP lists but we want merging */
2166 drain_pages(get_cpu());
2169 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2170 order
, zonelist
, high_zoneidx
,
2171 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2172 preferred_zone
, migratetype
);
2174 preferred_zone
->compact_blockskip_flush
= false;
2175 preferred_zone
->compact_considered
= 0;
2176 preferred_zone
->compact_defer_shift
= 0;
2177 if (order
>= preferred_zone
->compact_order_failed
)
2178 preferred_zone
->compact_order_failed
= order
+ 1;
2179 count_vm_event(COMPACTSUCCESS
);
2184 * It's bad if compaction run occurs and fails.
2185 * The most likely reason is that pages exist,
2186 * but not enough to satisfy watermarks.
2188 count_vm_event(COMPACTFAIL
);
2191 * As async compaction considers a subset of pageblocks, only
2192 * defer if the failure was a sync compaction failure.
2195 defer_compaction(preferred_zone
, order
);
2203 static inline struct page
*
2204 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2205 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2206 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2207 int migratetype
, bool sync_migration
,
2208 bool *contended_compaction
, bool *deferred_compaction
,
2209 unsigned long *did_some_progress
)
2213 #endif /* CONFIG_COMPACTION */
2215 /* Perform direct synchronous page reclaim */
2217 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2218 nodemask_t
*nodemask
)
2220 struct reclaim_state reclaim_state
;
2225 /* We now go into synchronous reclaim */
2226 cpuset_memory_pressure_bump();
2227 current
->flags
|= PF_MEMALLOC
;
2228 lockdep_set_current_reclaim_state(gfp_mask
);
2229 reclaim_state
.reclaimed_slab
= 0;
2230 current
->reclaim_state
= &reclaim_state
;
2232 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2234 current
->reclaim_state
= NULL
;
2235 lockdep_clear_current_reclaim_state();
2236 current
->flags
&= ~PF_MEMALLOC
;
2243 /* The really slow allocator path where we enter direct reclaim */
2244 static inline struct page
*
2245 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2246 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2247 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2248 int migratetype
, unsigned long *did_some_progress
)
2250 struct page
*page
= NULL
;
2251 bool drained
= false;
2253 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2255 if (unlikely(!(*did_some_progress
)))
2258 /* After successful reclaim, reconsider all zones for allocation */
2259 if (IS_ENABLED(CONFIG_NUMA
))
2260 zlc_clear_zones_full(zonelist
);
2263 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2264 zonelist
, high_zoneidx
,
2265 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2266 preferred_zone
, migratetype
);
2269 * If an allocation failed after direct reclaim, it could be because
2270 * pages are pinned on the per-cpu lists. Drain them and try again
2272 if (!page
&& !drained
) {
2282 * This is called in the allocator slow-path if the allocation request is of
2283 * sufficient urgency to ignore watermarks and take other desperate measures
2285 static inline struct page
*
2286 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2287 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2288 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2294 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2295 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2296 preferred_zone
, migratetype
);
2298 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2299 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2300 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2306 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
2307 enum zone_type high_zoneidx
,
2308 enum zone_type classzone_idx
)
2313 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
2314 wakeup_kswapd(zone
, order
, classzone_idx
);
2318 gfp_to_alloc_flags(gfp_t gfp_mask
)
2320 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2321 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2323 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2324 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2327 * The caller may dip into page reserves a bit more if the caller
2328 * cannot run direct reclaim, or if the caller has realtime scheduling
2329 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2330 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2332 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2336 * Not worth trying to allocate harder for
2337 * __GFP_NOMEMALLOC even if it can't schedule.
2339 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2340 alloc_flags
|= ALLOC_HARDER
;
2342 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2343 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2345 alloc_flags
&= ~ALLOC_CPUSET
;
2346 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2347 alloc_flags
|= ALLOC_HARDER
;
2349 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2350 if (gfp_mask
& __GFP_MEMALLOC
)
2351 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2352 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2353 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2354 else if (!in_interrupt() &&
2355 ((current
->flags
& PF_MEMALLOC
) ||
2356 unlikely(test_thread_flag(TIF_MEMDIE
))))
2357 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2360 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2361 alloc_flags
|= ALLOC_CMA
;
2366 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2368 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2371 static inline struct page
*
2372 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2373 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2374 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2377 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2378 struct page
*page
= NULL
;
2380 unsigned long pages_reclaimed
= 0;
2381 unsigned long did_some_progress
;
2382 bool sync_migration
= false;
2383 bool deferred_compaction
= false;
2384 bool contended_compaction
= false;
2387 * In the slowpath, we sanity check order to avoid ever trying to
2388 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2389 * be using allocators in order of preference for an area that is
2392 if (order
>= MAX_ORDER
) {
2393 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2398 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2399 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2400 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2401 * using a larger set of nodes after it has established that the
2402 * allowed per node queues are empty and that nodes are
2405 if (IS_ENABLED(CONFIG_NUMA
) &&
2406 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2410 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2411 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2412 zone_idx(preferred_zone
));
2415 * OK, we're below the kswapd watermark and have kicked background
2416 * reclaim. Now things get more complex, so set up alloc_flags according
2417 * to how we want to proceed.
2419 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2422 * Find the true preferred zone if the allocation is unconstrained by
2425 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2426 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2430 /* This is the last chance, in general, before the goto nopage. */
2431 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2432 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2433 preferred_zone
, migratetype
);
2437 /* Allocate without watermarks if the context allows */
2438 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2440 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2441 * the allocation is high priority and these type of
2442 * allocations are system rather than user orientated
2444 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2446 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2447 zonelist
, high_zoneidx
, nodemask
,
2448 preferred_zone
, migratetype
);
2454 /* Atomic allocations - we can't balance anything */
2458 /* Avoid recursion of direct reclaim */
2459 if (current
->flags
& PF_MEMALLOC
)
2462 /* Avoid allocations with no watermarks from looping endlessly */
2463 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2467 * Try direct compaction. The first pass is asynchronous. Subsequent
2468 * attempts after direct reclaim are synchronous
2470 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2471 zonelist
, high_zoneidx
,
2473 alloc_flags
, preferred_zone
,
2474 migratetype
, sync_migration
,
2475 &contended_compaction
,
2476 &deferred_compaction
,
2477 &did_some_progress
);
2480 sync_migration
= true;
2483 * If compaction is deferred for high-order allocations, it is because
2484 * sync compaction recently failed. In this is the case and the caller
2485 * requested a movable allocation that does not heavily disrupt the
2486 * system then fail the allocation instead of entering direct reclaim.
2488 if ((deferred_compaction
|| contended_compaction
) &&
2489 (gfp_mask
& __GFP_NO_KSWAPD
))
2492 /* Try direct reclaim and then allocating */
2493 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2494 zonelist
, high_zoneidx
,
2496 alloc_flags
, preferred_zone
,
2497 migratetype
, &did_some_progress
);
2502 * If we failed to make any progress reclaiming, then we are
2503 * running out of options and have to consider going OOM
2505 if (!did_some_progress
) {
2506 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2507 if (oom_killer_disabled
)
2509 /* Coredumps can quickly deplete all memory reserves */
2510 if ((current
->flags
& PF_DUMPCORE
) &&
2511 !(gfp_mask
& __GFP_NOFAIL
))
2513 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2514 zonelist
, high_zoneidx
,
2515 nodemask
, preferred_zone
,
2520 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2522 * The oom killer is not called for high-order
2523 * allocations that may fail, so if no progress
2524 * is being made, there are no other options and
2525 * retrying is unlikely to help.
2527 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2530 * The oom killer is not called for lowmem
2531 * allocations to prevent needlessly killing
2534 if (high_zoneidx
< ZONE_NORMAL
)
2542 /* Check if we should retry the allocation */
2543 pages_reclaimed
+= did_some_progress
;
2544 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2546 /* Wait for some write requests to complete then retry */
2547 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2551 * High-order allocations do not necessarily loop after
2552 * direct reclaim and reclaim/compaction depends on compaction
2553 * being called after reclaim so call directly if necessary
2555 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2556 zonelist
, high_zoneidx
,
2558 alloc_flags
, preferred_zone
,
2559 migratetype
, sync_migration
,
2560 &contended_compaction
,
2561 &deferred_compaction
,
2562 &did_some_progress
);
2568 warn_alloc_failed(gfp_mask
, order
, NULL
);
2571 if (kmemcheck_enabled
)
2572 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2578 * This is the 'heart' of the zoned buddy allocator.
2581 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2582 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2584 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2585 struct zone
*preferred_zone
;
2586 struct page
*page
= NULL
;
2587 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2588 unsigned int cpuset_mems_cookie
;
2589 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
;
2590 struct mem_cgroup
*memcg
= NULL
;
2592 gfp_mask
&= gfp_allowed_mask
;
2594 lockdep_trace_alloc(gfp_mask
);
2596 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2598 if (should_fail_alloc_page(gfp_mask
, order
))
2602 * Check the zones suitable for the gfp_mask contain at least one
2603 * valid zone. It's possible to have an empty zonelist as a result
2604 * of GFP_THISNODE and a memoryless node
2606 if (unlikely(!zonelist
->_zonerefs
->zone
))
2610 * Will only have any effect when __GFP_KMEMCG is set. This is
2611 * verified in the (always inline) callee
2613 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2617 cpuset_mems_cookie
= get_mems_allowed();
2619 /* The preferred zone is used for statistics later */
2620 first_zones_zonelist(zonelist
, high_zoneidx
,
2621 nodemask
? : &cpuset_current_mems_allowed
,
2623 if (!preferred_zone
)
2627 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2628 alloc_flags
|= ALLOC_CMA
;
2630 /* First allocation attempt */
2631 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2632 zonelist
, high_zoneidx
, alloc_flags
,
2633 preferred_zone
, migratetype
);
2634 if (unlikely(!page
)) {
2636 * Runtime PM, block IO and its error handling path
2637 * can deadlock because I/O on the device might not
2640 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2641 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2642 zonelist
, high_zoneidx
, nodemask
,
2643 preferred_zone
, migratetype
);
2646 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2650 * When updating a task's mems_allowed, it is possible to race with
2651 * parallel threads in such a way that an allocation can fail while
2652 * the mask is being updated. If a page allocation is about to fail,
2653 * check if the cpuset changed during allocation and if so, retry.
2655 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2658 memcg_kmem_commit_charge(page
, memcg
, order
);
2662 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2665 * Common helper functions.
2667 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2672 * __get_free_pages() returns a 32-bit address, which cannot represent
2675 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2677 page
= alloc_pages(gfp_mask
, order
);
2680 return (unsigned long) page_address(page
);
2682 EXPORT_SYMBOL(__get_free_pages
);
2684 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2686 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2688 EXPORT_SYMBOL(get_zeroed_page
);
2690 void __free_pages(struct page
*page
, unsigned int order
)
2692 if (put_page_testzero(page
)) {
2694 free_hot_cold_page(page
, 0);
2696 __free_pages_ok(page
, order
);
2700 EXPORT_SYMBOL(__free_pages
);
2702 void free_pages(unsigned long addr
, unsigned int order
)
2705 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2706 __free_pages(virt_to_page((void *)addr
), order
);
2710 EXPORT_SYMBOL(free_pages
);
2713 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2714 * pages allocated with __GFP_KMEMCG.
2716 * Those pages are accounted to a particular memcg, embedded in the
2717 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2718 * for that information only to find out that it is NULL for users who have no
2719 * interest in that whatsoever, we provide these functions.
2721 * The caller knows better which flags it relies on.
2723 void __free_memcg_kmem_pages(struct page
*page
, unsigned int order
)
2725 memcg_kmem_uncharge_pages(page
, order
);
2726 __free_pages(page
, order
);
2729 void free_memcg_kmem_pages(unsigned long addr
, unsigned int order
)
2732 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2733 __free_memcg_kmem_pages(virt_to_page((void *)addr
), order
);
2737 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2740 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2741 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2743 split_page(virt_to_page((void *)addr
), order
);
2744 while (used
< alloc_end
) {
2749 return (void *)addr
;
2753 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2754 * @size: the number of bytes to allocate
2755 * @gfp_mask: GFP flags for the allocation
2757 * This function is similar to alloc_pages(), except that it allocates the
2758 * minimum number of pages to satisfy the request. alloc_pages() can only
2759 * allocate memory in power-of-two pages.
2761 * This function is also limited by MAX_ORDER.
2763 * Memory allocated by this function must be released by free_pages_exact().
2765 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2767 unsigned int order
= get_order(size
);
2770 addr
= __get_free_pages(gfp_mask
, order
);
2771 return make_alloc_exact(addr
, order
, size
);
2773 EXPORT_SYMBOL(alloc_pages_exact
);
2776 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2778 * @nid: the preferred node ID where memory should be allocated
2779 * @size: the number of bytes to allocate
2780 * @gfp_mask: GFP flags for the allocation
2782 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2784 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2787 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2789 unsigned order
= get_order(size
);
2790 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2793 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2795 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2798 * free_pages_exact - release memory allocated via alloc_pages_exact()
2799 * @virt: the value returned by alloc_pages_exact.
2800 * @size: size of allocation, same value as passed to alloc_pages_exact().
2802 * Release the memory allocated by a previous call to alloc_pages_exact.
2804 void free_pages_exact(void *virt
, size_t size
)
2806 unsigned long addr
= (unsigned long)virt
;
2807 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2809 while (addr
< end
) {
2814 EXPORT_SYMBOL(free_pages_exact
);
2816 static unsigned int nr_free_zone_pages(int offset
)
2821 /* Just pick one node, since fallback list is circular */
2822 unsigned int sum
= 0;
2824 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2826 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2827 unsigned long size
= zone
->managed_pages
;
2828 unsigned long high
= high_wmark_pages(zone
);
2837 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2839 unsigned int nr_free_buffer_pages(void)
2841 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2843 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2846 * Amount of free RAM allocatable within all zones
2848 unsigned int nr_free_pagecache_pages(void)
2850 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2853 static inline void show_node(struct zone
*zone
)
2855 if (IS_ENABLED(CONFIG_NUMA
))
2856 printk("Node %d ", zone_to_nid(zone
));
2859 void si_meminfo(struct sysinfo
*val
)
2861 val
->totalram
= totalram_pages
;
2863 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2864 val
->bufferram
= nr_blockdev_pages();
2865 val
->totalhigh
= totalhigh_pages
;
2866 val
->freehigh
= nr_free_highpages();
2867 val
->mem_unit
= PAGE_SIZE
;
2870 EXPORT_SYMBOL(si_meminfo
);
2873 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2875 pg_data_t
*pgdat
= NODE_DATA(nid
);
2877 val
->totalram
= pgdat
->node_present_pages
;
2878 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2879 #ifdef CONFIG_HIGHMEM
2880 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
2881 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2887 val
->mem_unit
= PAGE_SIZE
;
2892 * Determine whether the node should be displayed or not, depending on whether
2893 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2895 bool skip_free_areas_node(unsigned int flags
, int nid
)
2898 unsigned int cpuset_mems_cookie
;
2900 if (!(flags
& SHOW_MEM_FILTER_NODES
))
2904 cpuset_mems_cookie
= get_mems_allowed();
2905 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
2906 } while (!put_mems_allowed(cpuset_mems_cookie
));
2911 #define K(x) ((x) << (PAGE_SHIFT-10))
2913 static void show_migration_types(unsigned char type
)
2915 static const char types
[MIGRATE_TYPES
] = {
2916 [MIGRATE_UNMOVABLE
] = 'U',
2917 [MIGRATE_RECLAIMABLE
] = 'E',
2918 [MIGRATE_MOVABLE
] = 'M',
2919 [MIGRATE_RESERVE
] = 'R',
2921 [MIGRATE_CMA
] = 'C',
2923 #ifdef CONFIG_MEMORY_ISOLATION
2924 [MIGRATE_ISOLATE
] = 'I',
2927 char tmp
[MIGRATE_TYPES
+ 1];
2931 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
2932 if (type
& (1 << i
))
2937 printk("(%s) ", tmp
);
2941 * Show free area list (used inside shift_scroll-lock stuff)
2942 * We also calculate the percentage fragmentation. We do this by counting the
2943 * memory on each free list with the exception of the first item on the list.
2944 * Suppresses nodes that are not allowed by current's cpuset if
2945 * SHOW_MEM_FILTER_NODES is passed.
2947 void show_free_areas(unsigned int filter
)
2952 for_each_populated_zone(zone
) {
2953 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2956 printk("%s per-cpu:\n", zone
->name
);
2958 for_each_online_cpu(cpu
) {
2959 struct per_cpu_pageset
*pageset
;
2961 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2963 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2964 cpu
, pageset
->pcp
.high
,
2965 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2969 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2970 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2972 " dirty:%lu writeback:%lu unstable:%lu\n"
2973 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2974 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2976 global_page_state(NR_ACTIVE_ANON
),
2977 global_page_state(NR_INACTIVE_ANON
),
2978 global_page_state(NR_ISOLATED_ANON
),
2979 global_page_state(NR_ACTIVE_FILE
),
2980 global_page_state(NR_INACTIVE_FILE
),
2981 global_page_state(NR_ISOLATED_FILE
),
2982 global_page_state(NR_UNEVICTABLE
),
2983 global_page_state(NR_FILE_DIRTY
),
2984 global_page_state(NR_WRITEBACK
),
2985 global_page_state(NR_UNSTABLE_NFS
),
2986 global_page_state(NR_FREE_PAGES
),
2987 global_page_state(NR_SLAB_RECLAIMABLE
),
2988 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2989 global_page_state(NR_FILE_MAPPED
),
2990 global_page_state(NR_SHMEM
),
2991 global_page_state(NR_PAGETABLE
),
2992 global_page_state(NR_BOUNCE
),
2993 global_page_state(NR_FREE_CMA_PAGES
));
2995 for_each_populated_zone(zone
) {
2998 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3006 " active_anon:%lukB"
3007 " inactive_anon:%lukB"
3008 " active_file:%lukB"
3009 " inactive_file:%lukB"
3010 " unevictable:%lukB"
3011 " isolated(anon):%lukB"
3012 " isolated(file):%lukB"
3020 " slab_reclaimable:%lukB"
3021 " slab_unreclaimable:%lukB"
3022 " kernel_stack:%lukB"
3027 " writeback_tmp:%lukB"
3028 " pages_scanned:%lu"
3029 " all_unreclaimable? %s"
3032 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3033 K(min_wmark_pages(zone
)),
3034 K(low_wmark_pages(zone
)),
3035 K(high_wmark_pages(zone
)),
3036 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3037 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3038 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3039 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3040 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3041 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3042 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3043 K(zone
->present_pages
),
3044 K(zone
->managed_pages
),
3045 K(zone_page_state(zone
, NR_MLOCK
)),
3046 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3047 K(zone_page_state(zone
, NR_WRITEBACK
)),
3048 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3049 K(zone_page_state(zone
, NR_SHMEM
)),
3050 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3051 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3052 zone_page_state(zone
, NR_KERNEL_STACK
) *
3054 K(zone_page_state(zone
, NR_PAGETABLE
)),
3055 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3056 K(zone_page_state(zone
, NR_BOUNCE
)),
3057 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3058 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3059 zone
->pages_scanned
,
3060 (zone
->all_unreclaimable
? "yes" : "no")
3062 printk("lowmem_reserve[]:");
3063 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3064 printk(" %lu", zone
->lowmem_reserve
[i
]);
3068 for_each_populated_zone(zone
) {
3069 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3070 unsigned char types
[MAX_ORDER
];
3072 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3075 printk("%s: ", zone
->name
);
3077 spin_lock_irqsave(&zone
->lock
, flags
);
3078 for (order
= 0; order
< MAX_ORDER
; order
++) {
3079 struct free_area
*area
= &zone
->free_area
[order
];
3082 nr
[order
] = area
->nr_free
;
3083 total
+= nr
[order
] << order
;
3086 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3087 if (!list_empty(&area
->free_list
[type
]))
3088 types
[order
] |= 1 << type
;
3091 spin_unlock_irqrestore(&zone
->lock
, flags
);
3092 for (order
= 0; order
< MAX_ORDER
; order
++) {
3093 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3095 show_migration_types(types
[order
]);
3097 printk("= %lukB\n", K(total
));
3100 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3102 show_swap_cache_info();
3105 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3107 zoneref
->zone
= zone
;
3108 zoneref
->zone_idx
= zone_idx(zone
);
3112 * Builds allocation fallback zone lists.
3114 * Add all populated zones of a node to the zonelist.
3116 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3117 int nr_zones
, enum zone_type zone_type
)
3121 BUG_ON(zone_type
>= MAX_NR_ZONES
);
3126 zone
= pgdat
->node_zones
+ zone_type
;
3127 if (populated_zone(zone
)) {
3128 zoneref_set_zone(zone
,
3129 &zonelist
->_zonerefs
[nr_zones
++]);
3130 check_highest_zone(zone_type
);
3133 } while (zone_type
);
3140 * 0 = automatic detection of better ordering.
3141 * 1 = order by ([node] distance, -zonetype)
3142 * 2 = order by (-zonetype, [node] distance)
3144 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3145 * the same zonelist. So only NUMA can configure this param.
3147 #define ZONELIST_ORDER_DEFAULT 0
3148 #define ZONELIST_ORDER_NODE 1
3149 #define ZONELIST_ORDER_ZONE 2
3151 /* zonelist order in the kernel.
3152 * set_zonelist_order() will set this to NODE or ZONE.
3154 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3155 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3159 /* The value user specified ....changed by config */
3160 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3161 /* string for sysctl */
3162 #define NUMA_ZONELIST_ORDER_LEN 16
3163 char numa_zonelist_order
[16] = "default";
3166 * interface for configure zonelist ordering.
3167 * command line option "numa_zonelist_order"
3168 * = "[dD]efault - default, automatic configuration.
3169 * = "[nN]ode - order by node locality, then by zone within node
3170 * = "[zZ]one - order by zone, then by locality within zone
3173 static int __parse_numa_zonelist_order(char *s
)
3175 if (*s
== 'd' || *s
== 'D') {
3176 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3177 } else if (*s
== 'n' || *s
== 'N') {
3178 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3179 } else if (*s
== 'z' || *s
== 'Z') {
3180 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3183 "Ignoring invalid numa_zonelist_order value: "
3190 static __init
int setup_numa_zonelist_order(char *s
)
3197 ret
= __parse_numa_zonelist_order(s
);
3199 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3203 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3206 * sysctl handler for numa_zonelist_order
3208 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
3209 void __user
*buffer
, size_t *length
,
3212 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3214 static DEFINE_MUTEX(zl_order_mutex
);
3216 mutex_lock(&zl_order_mutex
);
3218 strcpy(saved_string
, (char*)table
->data
);
3219 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3223 int oldval
= user_zonelist_order
;
3224 if (__parse_numa_zonelist_order((char*)table
->data
)) {
3226 * bogus value. restore saved string
3228 strncpy((char*)table
->data
, saved_string
,
3229 NUMA_ZONELIST_ORDER_LEN
);
3230 user_zonelist_order
= oldval
;
3231 } else if (oldval
!= user_zonelist_order
) {
3232 mutex_lock(&zonelists_mutex
);
3233 build_all_zonelists(NULL
, NULL
);
3234 mutex_unlock(&zonelists_mutex
);
3238 mutex_unlock(&zl_order_mutex
);
3243 #define MAX_NODE_LOAD (nr_online_nodes)
3244 static int node_load
[MAX_NUMNODES
];
3247 * find_next_best_node - find the next node that should appear in a given node's fallback list
3248 * @node: node whose fallback list we're appending
3249 * @used_node_mask: nodemask_t of already used nodes
3251 * We use a number of factors to determine which is the next node that should
3252 * appear on a given node's fallback list. The node should not have appeared
3253 * already in @node's fallback list, and it should be the next closest node
3254 * according to the distance array (which contains arbitrary distance values
3255 * from each node to each node in the system), and should also prefer nodes
3256 * with no CPUs, since presumably they'll have very little allocation pressure
3257 * on them otherwise.
3258 * It returns -1 if no node is found.
3260 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3263 int min_val
= INT_MAX
;
3264 int best_node
= NUMA_NO_NODE
;
3265 const struct cpumask
*tmp
= cpumask_of_node(0);
3267 /* Use the local node if we haven't already */
3268 if (!node_isset(node
, *used_node_mask
)) {
3269 node_set(node
, *used_node_mask
);
3273 for_each_node_state(n
, N_MEMORY
) {
3275 /* Don't want a node to appear more than once */
3276 if (node_isset(n
, *used_node_mask
))
3279 /* Use the distance array to find the distance */
3280 val
= node_distance(node
, n
);
3282 /* Penalize nodes under us ("prefer the next node") */
3285 /* Give preference to headless and unused nodes */
3286 tmp
= cpumask_of_node(n
);
3287 if (!cpumask_empty(tmp
))
3288 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3290 /* Slight preference for less loaded node */
3291 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3292 val
+= node_load
[n
];
3294 if (val
< min_val
) {
3301 node_set(best_node
, *used_node_mask
);
3308 * Build zonelists ordered by node and zones within node.
3309 * This results in maximum locality--normal zone overflows into local
3310 * DMA zone, if any--but risks exhausting DMA zone.
3312 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3315 struct zonelist
*zonelist
;
3317 zonelist
= &pgdat
->node_zonelists
[0];
3318 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3320 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3322 zonelist
->_zonerefs
[j
].zone
= NULL
;
3323 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3327 * Build gfp_thisnode zonelists
3329 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3332 struct zonelist
*zonelist
;
3334 zonelist
= &pgdat
->node_zonelists
[1];
3335 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3336 zonelist
->_zonerefs
[j
].zone
= NULL
;
3337 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3341 * Build zonelists ordered by zone and nodes within zones.
3342 * This results in conserving DMA zone[s] until all Normal memory is
3343 * exhausted, but results in overflowing to remote node while memory
3344 * may still exist in local DMA zone.
3346 static int node_order
[MAX_NUMNODES
];
3348 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3351 int zone_type
; /* needs to be signed */
3353 struct zonelist
*zonelist
;
3355 zonelist
= &pgdat
->node_zonelists
[0];
3357 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3358 for (j
= 0; j
< nr_nodes
; j
++) {
3359 node
= node_order
[j
];
3360 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3361 if (populated_zone(z
)) {
3363 &zonelist
->_zonerefs
[pos
++]);
3364 check_highest_zone(zone_type
);
3368 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3369 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3372 static int default_zonelist_order(void)
3375 unsigned long low_kmem_size
,total_size
;
3379 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3380 * If they are really small and used heavily, the system can fall
3381 * into OOM very easily.
3382 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3384 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3387 for_each_online_node(nid
) {
3388 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3389 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3390 if (populated_zone(z
)) {
3391 if (zone_type
< ZONE_NORMAL
)
3392 low_kmem_size
+= z
->present_pages
;
3393 total_size
+= z
->present_pages
;
3394 } else if (zone_type
== ZONE_NORMAL
) {
3396 * If any node has only lowmem, then node order
3397 * is preferred to allow kernel allocations
3398 * locally; otherwise, they can easily infringe
3399 * on other nodes when there is an abundance of
3400 * lowmem available to allocate from.
3402 return ZONELIST_ORDER_NODE
;
3406 if (!low_kmem_size
|| /* there are no DMA area. */
3407 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
3408 return ZONELIST_ORDER_NODE
;
3410 * look into each node's config.
3411 * If there is a node whose DMA/DMA32 memory is very big area on
3412 * local memory, NODE_ORDER may be suitable.
3414 average_size
= total_size
/
3415 (nodes_weight(node_states
[N_MEMORY
]) + 1);
3416 for_each_online_node(nid
) {
3419 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3420 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3421 if (populated_zone(z
)) {
3422 if (zone_type
< ZONE_NORMAL
)
3423 low_kmem_size
+= z
->present_pages
;
3424 total_size
+= z
->present_pages
;
3427 if (low_kmem_size
&&
3428 total_size
> average_size
&& /* ignore small node */
3429 low_kmem_size
> total_size
* 70/100)
3430 return ZONELIST_ORDER_NODE
;
3432 return ZONELIST_ORDER_ZONE
;
3435 static void set_zonelist_order(void)
3437 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3438 current_zonelist_order
= default_zonelist_order();
3440 current_zonelist_order
= user_zonelist_order
;
3443 static void build_zonelists(pg_data_t
*pgdat
)
3447 nodemask_t used_mask
;
3448 int local_node
, prev_node
;
3449 struct zonelist
*zonelist
;
3450 int order
= current_zonelist_order
;
3452 /* initialize zonelists */
3453 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3454 zonelist
= pgdat
->node_zonelists
+ i
;
3455 zonelist
->_zonerefs
[0].zone
= NULL
;
3456 zonelist
->_zonerefs
[0].zone_idx
= 0;
3459 /* NUMA-aware ordering of nodes */
3460 local_node
= pgdat
->node_id
;
3461 load
= nr_online_nodes
;
3462 prev_node
= local_node
;
3463 nodes_clear(used_mask
);
3465 memset(node_order
, 0, sizeof(node_order
));
3468 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3470 * We don't want to pressure a particular node.
3471 * So adding penalty to the first node in same
3472 * distance group to make it round-robin.
3474 if (node_distance(local_node
, node
) !=
3475 node_distance(local_node
, prev_node
))
3476 node_load
[node
] = load
;
3480 if (order
== ZONELIST_ORDER_NODE
)
3481 build_zonelists_in_node_order(pgdat
, node
);
3483 node_order
[j
++] = node
; /* remember order */
3486 if (order
== ZONELIST_ORDER_ZONE
) {
3487 /* calculate node order -- i.e., DMA last! */
3488 build_zonelists_in_zone_order(pgdat
, j
);
3491 build_thisnode_zonelists(pgdat
);
3494 /* Construct the zonelist performance cache - see further mmzone.h */
3495 static void build_zonelist_cache(pg_data_t
*pgdat
)
3497 struct zonelist
*zonelist
;
3498 struct zonelist_cache
*zlc
;
3501 zonelist
= &pgdat
->node_zonelists
[0];
3502 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3503 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3504 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3505 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3508 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3510 * Return node id of node used for "local" allocations.
3511 * I.e., first node id of first zone in arg node's generic zonelist.
3512 * Used for initializing percpu 'numa_mem', which is used primarily
3513 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3515 int local_memory_node(int node
)
3519 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3520 gfp_zone(GFP_KERNEL
),
3527 #else /* CONFIG_NUMA */
3529 static void set_zonelist_order(void)
3531 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3534 static void build_zonelists(pg_data_t
*pgdat
)
3536 int node
, local_node
;
3538 struct zonelist
*zonelist
;
3540 local_node
= pgdat
->node_id
;
3542 zonelist
= &pgdat
->node_zonelists
[0];
3543 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3546 * Now we build the zonelist so that it contains the zones
3547 * of all the other nodes.
3548 * We don't want to pressure a particular node, so when
3549 * building the zones for node N, we make sure that the
3550 * zones coming right after the local ones are those from
3551 * node N+1 (modulo N)
3553 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3554 if (!node_online(node
))
3556 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3559 for (node
= 0; node
< local_node
; node
++) {
3560 if (!node_online(node
))
3562 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3566 zonelist
->_zonerefs
[j
].zone
= NULL
;
3567 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3570 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3571 static void build_zonelist_cache(pg_data_t
*pgdat
)
3573 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3576 #endif /* CONFIG_NUMA */
3579 * Boot pageset table. One per cpu which is going to be used for all
3580 * zones and all nodes. The parameters will be set in such a way
3581 * that an item put on a list will immediately be handed over to
3582 * the buddy list. This is safe since pageset manipulation is done
3583 * with interrupts disabled.
3585 * The boot_pagesets must be kept even after bootup is complete for
3586 * unused processors and/or zones. They do play a role for bootstrapping
3587 * hotplugged processors.
3589 * zoneinfo_show() and maybe other functions do
3590 * not check if the processor is online before following the pageset pointer.
3591 * Other parts of the kernel may not check if the zone is available.
3593 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3594 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3595 static void setup_zone_pageset(struct zone
*zone
);
3598 * Global mutex to protect against size modification of zonelists
3599 * as well as to serialize pageset setup for the new populated zone.
3601 DEFINE_MUTEX(zonelists_mutex
);
3603 /* return values int ....just for stop_machine() */
3604 static int __build_all_zonelists(void *data
)
3608 pg_data_t
*self
= data
;
3611 memset(node_load
, 0, sizeof(node_load
));
3614 if (self
&& !node_online(self
->node_id
)) {
3615 build_zonelists(self
);
3616 build_zonelist_cache(self
);
3619 for_each_online_node(nid
) {
3620 pg_data_t
*pgdat
= NODE_DATA(nid
);
3622 build_zonelists(pgdat
);
3623 build_zonelist_cache(pgdat
);
3627 * Initialize the boot_pagesets that are going to be used
3628 * for bootstrapping processors. The real pagesets for
3629 * each zone will be allocated later when the per cpu
3630 * allocator is available.
3632 * boot_pagesets are used also for bootstrapping offline
3633 * cpus if the system is already booted because the pagesets
3634 * are needed to initialize allocators on a specific cpu too.
3635 * F.e. the percpu allocator needs the page allocator which
3636 * needs the percpu allocator in order to allocate its pagesets
3637 * (a chicken-egg dilemma).
3639 for_each_possible_cpu(cpu
) {
3640 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3642 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3644 * We now know the "local memory node" for each node--
3645 * i.e., the node of the first zone in the generic zonelist.
3646 * Set up numa_mem percpu variable for on-line cpus. During
3647 * boot, only the boot cpu should be on-line; we'll init the
3648 * secondary cpus' numa_mem as they come on-line. During
3649 * node/memory hotplug, we'll fixup all on-line cpus.
3651 if (cpu_online(cpu
))
3652 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3660 * Called with zonelists_mutex held always
3661 * unless system_state == SYSTEM_BOOTING.
3663 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3665 set_zonelist_order();
3667 if (system_state
== SYSTEM_BOOTING
) {
3668 __build_all_zonelists(NULL
);
3669 mminit_verify_zonelist();
3670 cpuset_init_current_mems_allowed();
3672 /* we have to stop all cpus to guarantee there is no user
3674 #ifdef CONFIG_MEMORY_HOTPLUG
3676 setup_zone_pageset(zone
);
3678 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3679 /* cpuset refresh routine should be here */
3681 vm_total_pages
= nr_free_pagecache_pages();
3683 * Disable grouping by mobility if the number of pages in the
3684 * system is too low to allow the mechanism to work. It would be
3685 * more accurate, but expensive to check per-zone. This check is
3686 * made on memory-hotadd so a system can start with mobility
3687 * disabled and enable it later
3689 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3690 page_group_by_mobility_disabled
= 1;
3692 page_group_by_mobility_disabled
= 0;
3694 printk("Built %i zonelists in %s order, mobility grouping %s. "
3695 "Total pages: %ld\n",
3697 zonelist_order_name
[current_zonelist_order
],
3698 page_group_by_mobility_disabled
? "off" : "on",
3701 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3706 * Helper functions to size the waitqueue hash table.
3707 * Essentially these want to choose hash table sizes sufficiently
3708 * large so that collisions trying to wait on pages are rare.
3709 * But in fact, the number of active page waitqueues on typical
3710 * systems is ridiculously low, less than 200. So this is even
3711 * conservative, even though it seems large.
3713 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3714 * waitqueues, i.e. the size of the waitq table given the number of pages.
3716 #define PAGES_PER_WAITQUEUE 256
3718 #ifndef CONFIG_MEMORY_HOTPLUG
3719 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3721 unsigned long size
= 1;
3723 pages
/= PAGES_PER_WAITQUEUE
;
3725 while (size
< pages
)
3729 * Once we have dozens or even hundreds of threads sleeping
3730 * on IO we've got bigger problems than wait queue collision.
3731 * Limit the size of the wait table to a reasonable size.
3733 size
= min(size
, 4096UL);
3735 return max(size
, 4UL);
3739 * A zone's size might be changed by hot-add, so it is not possible to determine
3740 * a suitable size for its wait_table. So we use the maximum size now.
3742 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3744 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3745 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3746 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3748 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3749 * or more by the traditional way. (See above). It equals:
3751 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3752 * ia64(16K page size) : = ( 8G + 4M)byte.
3753 * powerpc (64K page size) : = (32G +16M)byte.
3755 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3762 * This is an integer logarithm so that shifts can be used later
3763 * to extract the more random high bits from the multiplicative
3764 * hash function before the remainder is taken.
3766 static inline unsigned long wait_table_bits(unsigned long size
)
3771 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3774 * Check if a pageblock contains reserved pages
3776 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3780 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3781 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3788 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3789 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3790 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3791 * higher will lead to a bigger reserve which will get freed as contiguous
3792 * blocks as reclaim kicks in
3794 static void setup_zone_migrate_reserve(struct zone
*zone
)
3796 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3798 unsigned long block_migratetype
;
3802 * Get the start pfn, end pfn and the number of blocks to reserve
3803 * We have to be careful to be aligned to pageblock_nr_pages to
3804 * make sure that we always check pfn_valid for the first page in
3807 start_pfn
= zone
->zone_start_pfn
;
3808 end_pfn
= zone_end_pfn(zone
);
3809 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
3810 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3814 * Reserve blocks are generally in place to help high-order atomic
3815 * allocations that are short-lived. A min_free_kbytes value that
3816 * would result in more than 2 reserve blocks for atomic allocations
3817 * is assumed to be in place to help anti-fragmentation for the
3818 * future allocation of hugepages at runtime.
3820 reserve
= min(2, reserve
);
3822 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3823 if (!pfn_valid(pfn
))
3825 page
= pfn_to_page(pfn
);
3827 /* Watch out for overlapping nodes */
3828 if (page_to_nid(page
) != zone_to_nid(zone
))
3831 block_migratetype
= get_pageblock_migratetype(page
);
3833 /* Only test what is necessary when the reserves are not met */
3836 * Blocks with reserved pages will never free, skip
3839 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
3840 if (pageblock_is_reserved(pfn
, block_end_pfn
))
3843 /* If this block is reserved, account for it */
3844 if (block_migratetype
== MIGRATE_RESERVE
) {
3849 /* Suitable for reserving if this block is movable */
3850 if (block_migratetype
== MIGRATE_MOVABLE
) {
3851 set_pageblock_migratetype(page
,
3853 move_freepages_block(zone
, page
,
3861 * If the reserve is met and this is a previous reserved block,
3864 if (block_migratetype
== MIGRATE_RESERVE
) {
3865 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3866 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3872 * Initially all pages are reserved - free ones are freed
3873 * up by free_all_bootmem() once the early boot process is
3874 * done. Non-atomic initialization, single-pass.
3876 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3877 unsigned long start_pfn
, enum memmap_context context
)
3880 unsigned long end_pfn
= start_pfn
+ size
;
3884 if (highest_memmap_pfn
< end_pfn
- 1)
3885 highest_memmap_pfn
= end_pfn
- 1;
3887 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3888 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3890 * There can be holes in boot-time mem_map[]s
3891 * handed to this function. They do not
3892 * exist on hotplugged memory.
3894 if (context
== MEMMAP_EARLY
) {
3895 if (!early_pfn_valid(pfn
))
3897 if (!early_pfn_in_nid(pfn
, nid
))
3900 page
= pfn_to_page(pfn
);
3901 set_page_links(page
, zone
, nid
, pfn
);
3902 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3903 init_page_count(page
);
3904 page_mapcount_reset(page
);
3905 page_nid_reset_last(page
);
3906 SetPageReserved(page
);
3908 * Mark the block movable so that blocks are reserved for
3909 * movable at startup. This will force kernel allocations
3910 * to reserve their blocks rather than leaking throughout
3911 * the address space during boot when many long-lived
3912 * kernel allocations are made. Later some blocks near
3913 * the start are marked MIGRATE_RESERVE by
3914 * setup_zone_migrate_reserve()
3916 * bitmap is created for zone's valid pfn range. but memmap
3917 * can be created for invalid pages (for alignment)
3918 * check here not to call set_pageblock_migratetype() against
3921 if ((z
->zone_start_pfn
<= pfn
)
3922 && (pfn
< zone_end_pfn(z
))
3923 && !(pfn
& (pageblock_nr_pages
- 1)))
3924 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3926 INIT_LIST_HEAD(&page
->lru
);
3927 #ifdef WANT_PAGE_VIRTUAL
3928 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3929 if (!is_highmem_idx(zone
))
3930 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3935 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3938 for_each_migratetype_order(order
, t
) {
3939 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3940 zone
->free_area
[order
].nr_free
= 0;
3944 #ifndef __HAVE_ARCH_MEMMAP_INIT
3945 #define memmap_init(size, nid, zone, start_pfn) \
3946 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3949 static int __meminit
zone_batchsize(struct zone
*zone
)
3955 * The per-cpu-pages pools are set to around 1000th of the
3956 * size of the zone. But no more than 1/2 of a meg.
3958 * OK, so we don't know how big the cache is. So guess.
3960 batch
= zone
->managed_pages
/ 1024;
3961 if (batch
* PAGE_SIZE
> 512 * 1024)
3962 batch
= (512 * 1024) / PAGE_SIZE
;
3963 batch
/= 4; /* We effectively *= 4 below */
3968 * Clamp the batch to a 2^n - 1 value. Having a power
3969 * of 2 value was found to be more likely to have
3970 * suboptimal cache aliasing properties in some cases.
3972 * For example if 2 tasks are alternately allocating
3973 * batches of pages, one task can end up with a lot
3974 * of pages of one half of the possible page colors
3975 * and the other with pages of the other colors.
3977 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3982 /* The deferral and batching of frees should be suppressed under NOMMU
3985 * The problem is that NOMMU needs to be able to allocate large chunks
3986 * of contiguous memory as there's no hardware page translation to
3987 * assemble apparent contiguous memory from discontiguous pages.
3989 * Queueing large contiguous runs of pages for batching, however,
3990 * causes the pages to actually be freed in smaller chunks. As there
3991 * can be a significant delay between the individual batches being
3992 * recycled, this leads to the once large chunks of space being
3993 * fragmented and becoming unavailable for high-order allocations.
3999 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4001 struct per_cpu_pages
*pcp
;
4004 memset(p
, 0, sizeof(*p
));
4008 pcp
->high
= 6 * batch
;
4009 pcp
->batch
= max(1UL, 1 * batch
);
4010 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4011 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4015 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4016 * to the value high for the pageset p.
4019 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
4022 struct per_cpu_pages
*pcp
;
4026 pcp
->batch
= max(1UL, high
/4);
4027 if ((high
/4) > (PAGE_SHIFT
* 8))
4028 pcp
->batch
= PAGE_SHIFT
* 8;
4031 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4035 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4037 for_each_possible_cpu(cpu
) {
4038 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4040 setup_pageset(pcp
, zone_batchsize(zone
));
4042 if (percpu_pagelist_fraction
)
4043 setup_pagelist_highmark(pcp
,
4044 (zone
->managed_pages
/
4045 percpu_pagelist_fraction
));
4050 * Allocate per cpu pagesets and initialize them.
4051 * Before this call only boot pagesets were available.
4053 void __init
setup_per_cpu_pageset(void)
4057 for_each_populated_zone(zone
)
4058 setup_zone_pageset(zone
);
4061 static noinline __init_refok
4062 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4065 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4069 * The per-page waitqueue mechanism uses hashed waitqueues
4072 zone
->wait_table_hash_nr_entries
=
4073 wait_table_hash_nr_entries(zone_size_pages
);
4074 zone
->wait_table_bits
=
4075 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4076 alloc_size
= zone
->wait_table_hash_nr_entries
4077 * sizeof(wait_queue_head_t
);
4079 if (!slab_is_available()) {
4080 zone
->wait_table
= (wait_queue_head_t
*)
4081 alloc_bootmem_node_nopanic(pgdat
, alloc_size
);
4084 * This case means that a zone whose size was 0 gets new memory
4085 * via memory hot-add.
4086 * But it may be the case that a new node was hot-added. In
4087 * this case vmalloc() will not be able to use this new node's
4088 * memory - this wait_table must be initialized to use this new
4089 * node itself as well.
4090 * To use this new node's memory, further consideration will be
4093 zone
->wait_table
= vmalloc(alloc_size
);
4095 if (!zone
->wait_table
)
4098 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4099 init_waitqueue_head(zone
->wait_table
+ i
);
4104 static __meminit
void zone_pcp_init(struct zone
*zone
)
4107 * per cpu subsystem is not up at this point. The following code
4108 * relies on the ability of the linker to provide the
4109 * offset of a (static) per cpu variable into the per cpu area.
4111 zone
->pageset
= &boot_pageset
;
4113 if (zone
->present_pages
)
4114 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4115 zone
->name
, zone
->present_pages
,
4116 zone_batchsize(zone
));
4119 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4120 unsigned long zone_start_pfn
,
4122 enum memmap_context context
)
4124 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4126 ret
= zone_wait_table_init(zone
, size
);
4129 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4131 zone
->zone_start_pfn
= zone_start_pfn
;
4133 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4134 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4136 (unsigned long)zone_idx(zone
),
4137 zone_start_pfn
, (zone_start_pfn
+ size
));
4139 zone_init_free_lists(zone
);
4144 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4145 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4147 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4148 * Architectures may implement their own version but if add_active_range()
4149 * was used and there are no special requirements, this is a convenient
4152 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4154 unsigned long start_pfn
, end_pfn
;
4157 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
4158 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
4160 /* This is a memory hole */
4163 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4165 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4169 nid
= __early_pfn_to_nid(pfn
);
4172 /* just returns 0 */
4176 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4177 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4181 nid
= __early_pfn_to_nid(pfn
);
4182 if (nid
>= 0 && nid
!= node
)
4189 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4190 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4191 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4193 * If an architecture guarantees that all ranges registered with
4194 * add_active_ranges() contain no holes and may be freed, this
4195 * this function may be used instead of calling free_bootmem() manually.
4197 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4199 unsigned long start_pfn
, end_pfn
;
4202 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4203 start_pfn
= min(start_pfn
, max_low_pfn
);
4204 end_pfn
= min(end_pfn
, max_low_pfn
);
4206 if (start_pfn
< end_pfn
)
4207 free_bootmem_node(NODE_DATA(this_nid
),
4208 PFN_PHYS(start_pfn
),
4209 (end_pfn
- start_pfn
) << PAGE_SHIFT
);
4214 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4215 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4217 * If an architecture guarantees that all ranges registered with
4218 * add_active_ranges() contain no holes and may be freed, this
4219 * function may be used instead of calling memory_present() manually.
4221 void __init
sparse_memory_present_with_active_regions(int nid
)
4223 unsigned long start_pfn
, end_pfn
;
4226 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4227 memory_present(this_nid
, start_pfn
, end_pfn
);
4231 * get_pfn_range_for_nid - Return the start and end page frames for a node
4232 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4233 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4234 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4236 * It returns the start and end page frame of a node based on information
4237 * provided by an arch calling add_active_range(). If called for a node
4238 * with no available memory, a warning is printed and the start and end
4241 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4242 unsigned long *start_pfn
, unsigned long *end_pfn
)
4244 unsigned long this_start_pfn
, this_end_pfn
;
4250 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4251 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4252 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4255 if (*start_pfn
== -1UL)
4260 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4261 * assumption is made that zones within a node are ordered in monotonic
4262 * increasing memory addresses so that the "highest" populated zone is used
4264 static void __init
find_usable_zone_for_movable(void)
4267 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4268 if (zone_index
== ZONE_MOVABLE
)
4271 if (arch_zone_highest_possible_pfn
[zone_index
] >
4272 arch_zone_lowest_possible_pfn
[zone_index
])
4276 VM_BUG_ON(zone_index
== -1);
4277 movable_zone
= zone_index
;
4281 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4282 * because it is sized independent of architecture. Unlike the other zones,
4283 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4284 * in each node depending on the size of each node and how evenly kernelcore
4285 * is distributed. This helper function adjusts the zone ranges
4286 * provided by the architecture for a given node by using the end of the
4287 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4288 * zones within a node are in order of monotonic increases memory addresses
4290 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4291 unsigned long zone_type
,
4292 unsigned long node_start_pfn
,
4293 unsigned long node_end_pfn
,
4294 unsigned long *zone_start_pfn
,
4295 unsigned long *zone_end_pfn
)
4297 /* Only adjust if ZONE_MOVABLE is on this node */
4298 if (zone_movable_pfn
[nid
]) {
4299 /* Size ZONE_MOVABLE */
4300 if (zone_type
== ZONE_MOVABLE
) {
4301 *zone_start_pfn
= zone_movable_pfn
[nid
];
4302 *zone_end_pfn
= min(node_end_pfn
,
4303 arch_zone_highest_possible_pfn
[movable_zone
]);
4305 /* Adjust for ZONE_MOVABLE starting within this range */
4306 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4307 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4308 *zone_end_pfn
= zone_movable_pfn
[nid
];
4310 /* Check if this whole range is within ZONE_MOVABLE */
4311 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4312 *zone_start_pfn
= *zone_end_pfn
;
4317 * Return the number of pages a zone spans in a node, including holes
4318 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4320 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4321 unsigned long zone_type
,
4322 unsigned long *ignored
)
4324 unsigned long node_start_pfn
, node_end_pfn
;
4325 unsigned long zone_start_pfn
, zone_end_pfn
;
4327 /* Get the start and end of the node and zone */
4328 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4329 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4330 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4331 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4332 node_start_pfn
, node_end_pfn
,
4333 &zone_start_pfn
, &zone_end_pfn
);
4335 /* Check that this node has pages within the zone's required range */
4336 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4339 /* Move the zone boundaries inside the node if necessary */
4340 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4341 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4343 /* Return the spanned pages */
4344 return zone_end_pfn
- zone_start_pfn
;
4348 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4349 * then all holes in the requested range will be accounted for.
4351 unsigned long __meminit
__absent_pages_in_range(int nid
,
4352 unsigned long range_start_pfn
,
4353 unsigned long range_end_pfn
)
4355 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4356 unsigned long start_pfn
, end_pfn
;
4359 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4360 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4361 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4362 nr_absent
-= end_pfn
- start_pfn
;
4368 * absent_pages_in_range - Return number of page frames in holes within a range
4369 * @start_pfn: The start PFN to start searching for holes
4370 * @end_pfn: The end PFN to stop searching for holes
4372 * It returns the number of pages frames in memory holes within a range.
4374 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4375 unsigned long end_pfn
)
4377 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4380 /* Return the number of page frames in holes in a zone on a node */
4381 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4382 unsigned long zone_type
,
4383 unsigned long *ignored
)
4385 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4386 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4387 unsigned long node_start_pfn
, node_end_pfn
;
4388 unsigned long zone_start_pfn
, zone_end_pfn
;
4390 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4391 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4392 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4394 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4395 node_start_pfn
, node_end_pfn
,
4396 &zone_start_pfn
, &zone_end_pfn
);
4397 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4401 * sanitize_zone_movable_limit - Sanitize the zone_movable_limit array.
4403 * zone_movable_limit is initialized as 0. This function will try to get
4404 * the first ZONE_MOVABLE pfn of each node from movablemem_map, and
4405 * assigne them to zone_movable_limit.
4406 * zone_movable_limit[nid] == 0 means no limit for the node.
4408 * Note: Each range is represented as [start_pfn, end_pfn)
4410 static void __meminit
sanitize_zone_movable_limit(void)
4412 int map_pos
= 0, i
, nid
;
4413 unsigned long start_pfn
, end_pfn
;
4415 if (!movablemem_map
.nr_map
)
4418 /* Iterate all ranges from minimum to maximum */
4419 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4421 * If we have found lowest pfn of ZONE_MOVABLE of the node
4422 * specified by user, just go on to check next range.
4424 if (zone_movable_limit
[nid
])
4427 #ifdef CONFIG_ZONE_DMA
4428 /* Skip DMA memory. */
4429 if (start_pfn
< arch_zone_highest_possible_pfn
[ZONE_DMA
])
4430 start_pfn
= arch_zone_highest_possible_pfn
[ZONE_DMA
];
4433 #ifdef CONFIG_ZONE_DMA32
4434 /* Skip DMA32 memory. */
4435 if (start_pfn
< arch_zone_highest_possible_pfn
[ZONE_DMA32
])
4436 start_pfn
= arch_zone_highest_possible_pfn
[ZONE_DMA32
];
4439 #ifdef CONFIG_HIGHMEM
4440 /* Skip lowmem if ZONE_MOVABLE is highmem. */
4441 if (zone_movable_is_highmem() &&
4442 start_pfn
< arch_zone_lowest_possible_pfn
[ZONE_HIGHMEM
])
4443 start_pfn
= arch_zone_lowest_possible_pfn
[ZONE_HIGHMEM
];
4446 if (start_pfn
>= end_pfn
)
4449 while (map_pos
< movablemem_map
.nr_map
) {
4450 if (end_pfn
<= movablemem_map
.map
[map_pos
].start_pfn
)
4453 if (start_pfn
>= movablemem_map
.map
[map_pos
].end_pfn
) {
4459 * The start_pfn of ZONE_MOVABLE is either the minimum
4460 * pfn specified by movablemem_map, or 0, which means
4461 * the node has no ZONE_MOVABLE.
4463 zone_movable_limit
[nid
] = max(start_pfn
,
4464 movablemem_map
.map
[map_pos
].start_pfn
);
4471 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4472 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4473 unsigned long zone_type
,
4474 unsigned long *zones_size
)
4476 return zones_size
[zone_type
];
4479 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4480 unsigned long zone_type
,
4481 unsigned long *zholes_size
)
4486 return zholes_size
[zone_type
];
4488 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4490 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4491 unsigned long *zones_size
, unsigned long *zholes_size
)
4493 unsigned long realtotalpages
, totalpages
= 0;
4496 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4497 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4499 pgdat
->node_spanned_pages
= totalpages
;
4501 realtotalpages
= totalpages
;
4502 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4504 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4506 pgdat
->node_present_pages
= realtotalpages
;
4507 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4511 #ifndef CONFIG_SPARSEMEM
4513 * Calculate the size of the zone->blockflags rounded to an unsigned long
4514 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4515 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4516 * round what is now in bits to nearest long in bits, then return it in
4519 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4521 unsigned long usemapsize
;
4523 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4524 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4525 usemapsize
= usemapsize
>> pageblock_order
;
4526 usemapsize
*= NR_PAGEBLOCK_BITS
;
4527 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4529 return usemapsize
/ 8;
4532 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4534 unsigned long zone_start_pfn
,
4535 unsigned long zonesize
)
4537 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4538 zone
->pageblock_flags
= NULL
;
4540 zone
->pageblock_flags
= alloc_bootmem_node_nopanic(pgdat
,
4544 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4545 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4546 #endif /* CONFIG_SPARSEMEM */
4548 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4550 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4551 void __init
set_pageblock_order(void)
4555 /* Check that pageblock_nr_pages has not already been setup */
4556 if (pageblock_order
)
4559 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4560 order
= HUGETLB_PAGE_ORDER
;
4562 order
= MAX_ORDER
- 1;
4565 * Assume the largest contiguous order of interest is a huge page.
4566 * This value may be variable depending on boot parameters on IA64 and
4569 pageblock_order
= order
;
4571 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4574 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4575 * is unused as pageblock_order is set at compile-time. See
4576 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4579 void __init
set_pageblock_order(void)
4583 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4585 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4586 unsigned long present_pages
)
4588 unsigned long pages
= spanned_pages
;
4591 * Provide a more accurate estimation if there are holes within
4592 * the zone and SPARSEMEM is in use. If there are holes within the
4593 * zone, each populated memory region may cost us one or two extra
4594 * memmap pages due to alignment because memmap pages for each
4595 * populated regions may not naturally algined on page boundary.
4596 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4598 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4599 IS_ENABLED(CONFIG_SPARSEMEM
))
4600 pages
= present_pages
;
4602 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4606 * Set up the zone data structures:
4607 * - mark all pages reserved
4608 * - mark all memory queues empty
4609 * - clear the memory bitmaps
4611 * NOTE: pgdat should get zeroed by caller.
4613 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4614 unsigned long *zones_size
, unsigned long *zholes_size
)
4617 int nid
= pgdat
->node_id
;
4618 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4621 pgdat_resize_init(pgdat
);
4622 #ifdef CONFIG_NUMA_BALANCING
4623 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4624 pgdat
->numabalancing_migrate_nr_pages
= 0;
4625 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4627 init_waitqueue_head(&pgdat
->kswapd_wait
);
4628 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4629 pgdat_page_cgroup_init(pgdat
);
4631 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4632 struct zone
*zone
= pgdat
->node_zones
+ j
;
4633 unsigned long size
, realsize
, freesize
, memmap_pages
;
4635 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4636 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4640 * Adjust freesize so that it accounts for how much memory
4641 * is used by this zone for memmap. This affects the watermark
4642 * and per-cpu initialisations
4644 memmap_pages
= calc_memmap_size(size
, realsize
);
4645 if (freesize
>= memmap_pages
) {
4646 freesize
-= memmap_pages
;
4649 " %s zone: %lu pages used for memmap\n",
4650 zone_names
[j
], memmap_pages
);
4653 " %s zone: %lu pages exceeds freesize %lu\n",
4654 zone_names
[j
], memmap_pages
, freesize
);
4656 /* Account for reserved pages */
4657 if (j
== 0 && freesize
> dma_reserve
) {
4658 freesize
-= dma_reserve
;
4659 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4660 zone_names
[0], dma_reserve
);
4663 if (!is_highmem_idx(j
))
4664 nr_kernel_pages
+= freesize
;
4665 /* Charge for highmem memmap if there are enough kernel pages */
4666 else if (nr_kernel_pages
> memmap_pages
* 2)
4667 nr_kernel_pages
-= memmap_pages
;
4668 nr_all_pages
+= freesize
;
4670 zone
->spanned_pages
= size
;
4671 zone
->present_pages
= realsize
;
4673 * Set an approximate value for lowmem here, it will be adjusted
4674 * when the bootmem allocator frees pages into the buddy system.
4675 * And all highmem pages will be managed by the buddy system.
4677 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4680 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4682 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4684 zone
->name
= zone_names
[j
];
4685 spin_lock_init(&zone
->lock
);
4686 spin_lock_init(&zone
->lru_lock
);
4687 zone_seqlock_init(zone
);
4688 zone
->zone_pgdat
= pgdat
;
4690 zone_pcp_init(zone
);
4691 lruvec_init(&zone
->lruvec
);
4695 set_pageblock_order();
4696 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4697 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4698 size
, MEMMAP_EARLY
);
4700 memmap_init(size
, nid
, j
, zone_start_pfn
);
4701 zone_start_pfn
+= size
;
4705 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4707 /* Skip empty nodes */
4708 if (!pgdat
->node_spanned_pages
)
4711 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4712 /* ia64 gets its own node_mem_map, before this, without bootmem */
4713 if (!pgdat
->node_mem_map
) {
4714 unsigned long size
, start
, end
;
4718 * The zone's endpoints aren't required to be MAX_ORDER
4719 * aligned but the node_mem_map endpoints must be in order
4720 * for the buddy allocator to function correctly.
4722 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4723 end
= pgdat_end_pfn(pgdat
);
4724 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4725 size
= (end
- start
) * sizeof(struct page
);
4726 map
= alloc_remap(pgdat
->node_id
, size
);
4728 map
= alloc_bootmem_node_nopanic(pgdat
, size
);
4729 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4731 #ifndef CONFIG_NEED_MULTIPLE_NODES
4733 * With no DISCONTIG, the global mem_map is just set as node 0's
4735 if (pgdat
== NODE_DATA(0)) {
4736 mem_map
= NODE_DATA(0)->node_mem_map
;
4737 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4738 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4739 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4740 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4743 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4746 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4747 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4749 pg_data_t
*pgdat
= NODE_DATA(nid
);
4751 /* pg_data_t should be reset to zero when it's allocated */
4752 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4754 pgdat
->node_id
= nid
;
4755 pgdat
->node_start_pfn
= node_start_pfn
;
4756 init_zone_allows_reclaim(nid
);
4757 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4759 alloc_node_mem_map(pgdat
);
4760 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4761 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4762 nid
, (unsigned long)pgdat
,
4763 (unsigned long)pgdat
->node_mem_map
);
4766 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4769 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4771 #if MAX_NUMNODES > 1
4773 * Figure out the number of possible node ids.
4775 static void __init
setup_nr_node_ids(void)
4778 unsigned int highest
= 0;
4780 for_each_node_mask(node
, node_possible_map
)
4782 nr_node_ids
= highest
+ 1;
4785 static inline void setup_nr_node_ids(void)
4791 * node_map_pfn_alignment - determine the maximum internode alignment
4793 * This function should be called after node map is populated and sorted.
4794 * It calculates the maximum power of two alignment which can distinguish
4797 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4798 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4799 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4800 * shifted, 1GiB is enough and this function will indicate so.
4802 * This is used to test whether pfn -> nid mapping of the chosen memory
4803 * model has fine enough granularity to avoid incorrect mapping for the
4804 * populated node map.
4806 * Returns the determined alignment in pfn's. 0 if there is no alignment
4807 * requirement (single node).
4809 unsigned long __init
node_map_pfn_alignment(void)
4811 unsigned long accl_mask
= 0, last_end
= 0;
4812 unsigned long start
, end
, mask
;
4816 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
4817 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
4824 * Start with a mask granular enough to pin-point to the
4825 * start pfn and tick off bits one-by-one until it becomes
4826 * too coarse to separate the current node from the last.
4828 mask
= ~((1 << __ffs(start
)) - 1);
4829 while (mask
&& last_end
<= (start
& (mask
<< 1)))
4832 /* accumulate all internode masks */
4836 /* convert mask to number of pages */
4837 return ~accl_mask
+ 1;
4840 /* Find the lowest pfn for a node */
4841 static unsigned long __init
find_min_pfn_for_node(int nid
)
4843 unsigned long min_pfn
= ULONG_MAX
;
4844 unsigned long start_pfn
;
4847 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
4848 min_pfn
= min(min_pfn
, start_pfn
);
4850 if (min_pfn
== ULONG_MAX
) {
4852 "Could not find start_pfn for node %d\n", nid
);
4860 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4862 * It returns the minimum PFN based on information provided via
4863 * add_active_range().
4865 unsigned long __init
find_min_pfn_with_active_regions(void)
4867 return find_min_pfn_for_node(MAX_NUMNODES
);
4871 * early_calculate_totalpages()
4872 * Sum pages in active regions for movable zone.
4873 * Populate N_MEMORY for calculating usable_nodes.
4875 static unsigned long __init
early_calculate_totalpages(void)
4877 unsigned long totalpages
= 0;
4878 unsigned long start_pfn
, end_pfn
;
4881 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4882 unsigned long pages
= end_pfn
- start_pfn
;
4884 totalpages
+= pages
;
4886 node_set_state(nid
, N_MEMORY
);
4892 * Find the PFN the Movable zone begins in each node. Kernel memory
4893 * is spread evenly between nodes as long as the nodes have enough
4894 * memory. When they don't, some nodes will have more kernelcore than
4897 static void __init
find_zone_movable_pfns_for_nodes(void)
4900 unsigned long usable_startpfn
;
4901 unsigned long kernelcore_node
, kernelcore_remaining
;
4902 /* save the state before borrow the nodemask */
4903 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
4904 unsigned long totalpages
= early_calculate_totalpages();
4905 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
4908 * If movablecore was specified, calculate what size of
4909 * kernelcore that corresponds so that memory usable for
4910 * any allocation type is evenly spread. If both kernelcore
4911 * and movablecore are specified, then the value of kernelcore
4912 * will be used for required_kernelcore if it's greater than
4913 * what movablecore would have allowed.
4915 if (required_movablecore
) {
4916 unsigned long corepages
;
4919 * Round-up so that ZONE_MOVABLE is at least as large as what
4920 * was requested by the user
4922 required_movablecore
=
4923 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4924 corepages
= totalpages
- required_movablecore
;
4926 required_kernelcore
= max(required_kernelcore
, corepages
);
4930 * If neither kernelcore/movablecore nor movablemem_map is specified,
4931 * there is no ZONE_MOVABLE. But if movablemem_map is specified, the
4932 * start pfn of ZONE_MOVABLE has been stored in zone_movable_limit[].
4934 if (!required_kernelcore
) {
4935 if (movablemem_map
.nr_map
)
4936 memcpy(zone_movable_pfn
, zone_movable_limit
,
4937 sizeof(zone_movable_pfn
));
4941 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4942 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4945 /* Spread kernelcore memory as evenly as possible throughout nodes */
4946 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4947 for_each_node_state(nid
, N_MEMORY
) {
4948 unsigned long start_pfn
, end_pfn
;
4951 * Recalculate kernelcore_node if the division per node
4952 * now exceeds what is necessary to satisfy the requested
4953 * amount of memory for the kernel
4955 if (required_kernelcore
< kernelcore_node
)
4956 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4959 * As the map is walked, we track how much memory is usable
4960 * by the kernel using kernelcore_remaining. When it is
4961 * 0, the rest of the node is usable by ZONE_MOVABLE
4963 kernelcore_remaining
= kernelcore_node
;
4965 /* Go through each range of PFNs within this node */
4966 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4967 unsigned long size_pages
;
4970 * Find more memory for kernelcore in
4971 * [zone_movable_pfn[nid], zone_movable_limit[nid]).
4973 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
4974 if (start_pfn
>= end_pfn
)
4977 if (zone_movable_limit
[nid
]) {
4978 end_pfn
= min(end_pfn
, zone_movable_limit
[nid
]);
4979 /* No range left for kernelcore in this node */
4980 if (start_pfn
>= end_pfn
) {
4981 zone_movable_pfn
[nid
] =
4982 zone_movable_limit
[nid
];
4987 /* Account for what is only usable for kernelcore */
4988 if (start_pfn
< usable_startpfn
) {
4989 unsigned long kernel_pages
;
4990 kernel_pages
= min(end_pfn
, usable_startpfn
)
4993 kernelcore_remaining
-= min(kernel_pages
,
4994 kernelcore_remaining
);
4995 required_kernelcore
-= min(kernel_pages
,
4996 required_kernelcore
);
4998 /* Continue if range is now fully accounted */
4999 if (end_pfn
<= usable_startpfn
) {
5002 * Push zone_movable_pfn to the end so
5003 * that if we have to rebalance
5004 * kernelcore across nodes, we will
5005 * not double account here
5007 zone_movable_pfn
[nid
] = end_pfn
;
5010 start_pfn
= usable_startpfn
;
5014 * The usable PFN range for ZONE_MOVABLE is from
5015 * start_pfn->end_pfn. Calculate size_pages as the
5016 * number of pages used as kernelcore
5018 size_pages
= end_pfn
- start_pfn
;
5019 if (size_pages
> kernelcore_remaining
)
5020 size_pages
= kernelcore_remaining
;
5021 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5024 * Some kernelcore has been met, update counts and
5025 * break if the kernelcore for this node has been
5028 required_kernelcore
-= min(required_kernelcore
,
5030 kernelcore_remaining
-= size_pages
;
5031 if (!kernelcore_remaining
)
5037 * If there is still required_kernelcore, we do another pass with one
5038 * less node in the count. This will push zone_movable_pfn[nid] further
5039 * along on the nodes that still have memory until kernelcore is
5043 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5047 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5048 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5049 zone_movable_pfn
[nid
] =
5050 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5052 /* restore the node_state */
5053 node_states
[N_MEMORY
] = saved_node_state
;
5056 /* Any regular or high memory on that node ? */
5057 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5059 enum zone_type zone_type
;
5061 if (N_MEMORY
== N_NORMAL_MEMORY
)
5064 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5065 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5066 if (zone
->present_pages
) {
5067 node_set_state(nid
, N_HIGH_MEMORY
);
5068 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5069 zone_type
<= ZONE_NORMAL
)
5070 node_set_state(nid
, N_NORMAL_MEMORY
);
5077 * free_area_init_nodes - Initialise all pg_data_t and zone data
5078 * @max_zone_pfn: an array of max PFNs for each zone
5080 * This will call free_area_init_node() for each active node in the system.
5081 * Using the page ranges provided by add_active_range(), the size of each
5082 * zone in each node and their holes is calculated. If the maximum PFN
5083 * between two adjacent zones match, it is assumed that the zone is empty.
5084 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5085 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5086 * starts where the previous one ended. For example, ZONE_DMA32 starts
5087 * at arch_max_dma_pfn.
5089 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5091 unsigned long start_pfn
, end_pfn
;
5094 /* Record where the zone boundaries are */
5095 memset(arch_zone_lowest_possible_pfn
, 0,
5096 sizeof(arch_zone_lowest_possible_pfn
));
5097 memset(arch_zone_highest_possible_pfn
, 0,
5098 sizeof(arch_zone_highest_possible_pfn
));
5099 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5100 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5101 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5102 if (i
== ZONE_MOVABLE
)
5104 arch_zone_lowest_possible_pfn
[i
] =
5105 arch_zone_highest_possible_pfn
[i
-1];
5106 arch_zone_highest_possible_pfn
[i
] =
5107 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5109 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5110 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5112 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5113 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5114 find_usable_zone_for_movable();
5115 sanitize_zone_movable_limit();
5116 find_zone_movable_pfns_for_nodes();
5118 /* Print out the zone ranges */
5119 printk("Zone ranges:\n");
5120 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5121 if (i
== ZONE_MOVABLE
)
5123 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5124 if (arch_zone_lowest_possible_pfn
[i
] ==
5125 arch_zone_highest_possible_pfn
[i
])
5126 printk(KERN_CONT
"empty\n");
5128 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5129 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5130 (arch_zone_highest_possible_pfn
[i
]
5131 << PAGE_SHIFT
) - 1);
5134 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5135 printk("Movable zone start for each node\n");
5136 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5137 if (zone_movable_pfn
[i
])
5138 printk(" Node %d: %#010lx\n", i
,
5139 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5142 /* Print out the early node map */
5143 printk("Early memory node ranges\n");
5144 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5145 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5146 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5148 /* Initialise every node */
5149 mminit_verify_pageflags_layout();
5150 setup_nr_node_ids();
5151 for_each_online_node(nid
) {
5152 pg_data_t
*pgdat
= NODE_DATA(nid
);
5153 free_area_init_node(nid
, NULL
,
5154 find_min_pfn_for_node(nid
), NULL
);
5156 /* Any memory on that node */
5157 if (pgdat
->node_present_pages
)
5158 node_set_state(nid
, N_MEMORY
);
5159 check_for_memory(pgdat
, nid
);
5163 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5165 unsigned long long coremem
;
5169 coremem
= memparse(p
, &p
);
5170 *core
= coremem
>> PAGE_SHIFT
;
5172 /* Paranoid check that UL is enough for the coremem value */
5173 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5179 * kernelcore=size sets the amount of memory for use for allocations that
5180 * cannot be reclaimed or migrated.
5182 static int __init
cmdline_parse_kernelcore(char *p
)
5184 return cmdline_parse_core(p
, &required_kernelcore
);
5188 * movablecore=size sets the amount of memory for use for allocations that
5189 * can be reclaimed or migrated.
5191 static int __init
cmdline_parse_movablecore(char *p
)
5193 return cmdline_parse_core(p
, &required_movablecore
);
5196 early_param("kernelcore", cmdline_parse_kernelcore
);
5197 early_param("movablecore", cmdline_parse_movablecore
);
5200 * movablemem_map_overlap() - Check if a range overlaps movablemem_map.map[].
5201 * @start_pfn: start pfn of the range to be checked
5202 * @end_pfn: end pfn of the range to be checked (exclusive)
5204 * This function checks if a given memory range [start_pfn, end_pfn) overlaps
5205 * the movablemem_map.map[] array.
5207 * Return: index of the first overlapped element in movablemem_map.map[]
5208 * or -1 if they don't overlap each other.
5210 int __init
movablemem_map_overlap(unsigned long start_pfn
,
5211 unsigned long end_pfn
)
5215 if (!movablemem_map
.nr_map
)
5218 for (overlap
= 0; overlap
< movablemem_map
.nr_map
; overlap
++)
5219 if (start_pfn
< movablemem_map
.map
[overlap
].end_pfn
)
5222 if (overlap
== movablemem_map
.nr_map
||
5223 end_pfn
<= movablemem_map
.map
[overlap
].start_pfn
)
5230 * insert_movablemem_map - Insert a memory range in to movablemem_map.map.
5231 * @start_pfn: start pfn of the range
5232 * @end_pfn: end pfn of the range
5234 * This function will also merge the overlapped ranges, and sort the array
5235 * by start_pfn in monotonic increasing order.
5237 void __init
insert_movablemem_map(unsigned long start_pfn
,
5238 unsigned long end_pfn
)
5243 * pos will be at the 1st overlapped range, or the position
5244 * where the element should be inserted.
5246 for (pos
= 0; pos
< movablemem_map
.nr_map
; pos
++)
5247 if (start_pfn
<= movablemem_map
.map
[pos
].end_pfn
)
5250 /* If there is no overlapped range, just insert the element. */
5251 if (pos
== movablemem_map
.nr_map
||
5252 end_pfn
< movablemem_map
.map
[pos
].start_pfn
) {
5254 * If pos is not the end of array, we need to move all
5255 * the rest elements backward.
5257 if (pos
< movablemem_map
.nr_map
)
5258 memmove(&movablemem_map
.map
[pos
+1],
5259 &movablemem_map
.map
[pos
],
5260 sizeof(struct movablemem_entry
) *
5261 (movablemem_map
.nr_map
- pos
));
5262 movablemem_map
.map
[pos
].start_pfn
= start_pfn
;
5263 movablemem_map
.map
[pos
].end_pfn
= end_pfn
;
5264 movablemem_map
.nr_map
++;
5268 /* overlap will be at the last overlapped range */
5269 for (overlap
= pos
+ 1; overlap
< movablemem_map
.nr_map
; overlap
++)
5270 if (end_pfn
< movablemem_map
.map
[overlap
].start_pfn
)
5274 * If there are more ranges overlapped, we need to merge them,
5275 * and move the rest elements forward.
5278 movablemem_map
.map
[pos
].start_pfn
= min(start_pfn
,
5279 movablemem_map
.map
[pos
].start_pfn
);
5280 movablemem_map
.map
[pos
].end_pfn
= max(end_pfn
,
5281 movablemem_map
.map
[overlap
].end_pfn
);
5283 if (pos
!= overlap
&& overlap
+ 1 != movablemem_map
.nr_map
)
5284 memmove(&movablemem_map
.map
[pos
+1],
5285 &movablemem_map
.map
[overlap
+1],
5286 sizeof(struct movablemem_entry
) *
5287 (movablemem_map
.nr_map
- overlap
- 1));
5289 movablemem_map
.nr_map
-= overlap
- pos
;
5293 * movablemem_map_add_region - Add a memory range into movablemem_map.
5294 * @start: physical start address of range
5295 * @end: physical end address of range
5297 * This function transform the physical address into pfn, and then add the
5298 * range into movablemem_map by calling insert_movablemem_map().
5300 static void __init
movablemem_map_add_region(u64 start
, u64 size
)
5302 unsigned long start_pfn
, end_pfn
;
5304 /* In case size == 0 or start + size overflows */
5305 if (start
+ size
<= start
)
5308 if (movablemem_map
.nr_map
>= ARRAY_SIZE(movablemem_map
.map
)) {
5309 pr_err("movablemem_map: too many entries;"
5310 " ignoring [mem %#010llx-%#010llx]\n",
5311 (unsigned long long) start
,
5312 (unsigned long long) (start
+ size
- 1));
5316 start_pfn
= PFN_DOWN(start
);
5317 end_pfn
= PFN_UP(start
+ size
);
5318 insert_movablemem_map(start_pfn
, end_pfn
);
5322 * cmdline_parse_movablemem_map - Parse boot option movablemem_map.
5323 * @p: The boot option of the following format:
5324 * movablemem_map=nn[KMG]@ss[KMG]
5326 * This option sets the memory range [ss, ss+nn) to be used as movable memory.
5328 * Return: 0 on success or -EINVAL on failure.
5330 static int __init
cmdline_parse_movablemem_map(char *p
)
5333 u64 start_at
, mem_size
;
5338 if (!strcmp(p
, "acpi"))
5339 movablemem_map
.acpi
= true;
5342 * If user decide to use info from BIOS, all the other user specified
5343 * ranges will be ingored.
5345 if (movablemem_map
.acpi
) {
5346 if (movablemem_map
.nr_map
) {
5347 memset(movablemem_map
.map
, 0,
5348 sizeof(struct movablemem_entry
)
5349 * movablemem_map
.nr_map
);
5350 movablemem_map
.nr_map
= 0;
5356 mem_size
= memparse(p
, &p
);
5362 start_at
= memparse(p
, &p
);
5363 if (p
== oldp
|| *p
!= '\0')
5366 movablemem_map_add_region(start_at
, mem_size
);
5372 early_param("movablemem_map", cmdline_parse_movablemem_map
);
5374 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5377 * set_dma_reserve - set the specified number of pages reserved in the first zone
5378 * @new_dma_reserve: The number of pages to mark reserved
5380 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5381 * In the DMA zone, a significant percentage may be consumed by kernel image
5382 * and other unfreeable allocations which can skew the watermarks badly. This
5383 * function may optionally be used to account for unfreeable pages in the
5384 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5385 * smaller per-cpu batchsize.
5387 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5389 dma_reserve
= new_dma_reserve
;
5392 void __init
free_area_init(unsigned long *zones_size
)
5394 free_area_init_node(0, zones_size
,
5395 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5398 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5399 unsigned long action
, void *hcpu
)
5401 int cpu
= (unsigned long)hcpu
;
5403 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5404 lru_add_drain_cpu(cpu
);
5408 * Spill the event counters of the dead processor
5409 * into the current processors event counters.
5410 * This artificially elevates the count of the current
5413 vm_events_fold_cpu(cpu
);
5416 * Zero the differential counters of the dead processor
5417 * so that the vm statistics are consistent.
5419 * This is only okay since the processor is dead and cannot
5420 * race with what we are doing.
5422 refresh_cpu_vm_stats(cpu
);
5427 void __init
page_alloc_init(void)
5429 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5433 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5434 * or min_free_kbytes changes.
5436 static void calculate_totalreserve_pages(void)
5438 struct pglist_data
*pgdat
;
5439 unsigned long reserve_pages
= 0;
5440 enum zone_type i
, j
;
5442 for_each_online_pgdat(pgdat
) {
5443 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5444 struct zone
*zone
= pgdat
->node_zones
+ i
;
5445 unsigned long max
= 0;
5447 /* Find valid and maximum lowmem_reserve in the zone */
5448 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5449 if (zone
->lowmem_reserve
[j
] > max
)
5450 max
= zone
->lowmem_reserve
[j
];
5453 /* we treat the high watermark as reserved pages. */
5454 max
+= high_wmark_pages(zone
);
5456 if (max
> zone
->managed_pages
)
5457 max
= zone
->managed_pages
;
5458 reserve_pages
+= max
;
5460 * Lowmem reserves are not available to
5461 * GFP_HIGHUSER page cache allocations and
5462 * kswapd tries to balance zones to their high
5463 * watermark. As a result, neither should be
5464 * regarded as dirtyable memory, to prevent a
5465 * situation where reclaim has to clean pages
5466 * in order to balance the zones.
5468 zone
->dirty_balance_reserve
= max
;
5471 dirty_balance_reserve
= reserve_pages
;
5472 totalreserve_pages
= reserve_pages
;
5476 * setup_per_zone_lowmem_reserve - called whenever
5477 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5478 * has a correct pages reserved value, so an adequate number of
5479 * pages are left in the zone after a successful __alloc_pages().
5481 static void setup_per_zone_lowmem_reserve(void)
5483 struct pglist_data
*pgdat
;
5484 enum zone_type j
, idx
;
5486 for_each_online_pgdat(pgdat
) {
5487 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5488 struct zone
*zone
= pgdat
->node_zones
+ j
;
5489 unsigned long managed_pages
= zone
->managed_pages
;
5491 zone
->lowmem_reserve
[j
] = 0;
5495 struct zone
*lower_zone
;
5499 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5500 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5502 lower_zone
= pgdat
->node_zones
+ idx
;
5503 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5504 sysctl_lowmem_reserve_ratio
[idx
];
5505 managed_pages
+= lower_zone
->managed_pages
;
5510 /* update totalreserve_pages */
5511 calculate_totalreserve_pages();
5514 static void __setup_per_zone_wmarks(void)
5516 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5517 unsigned long lowmem_pages
= 0;
5519 unsigned long flags
;
5521 /* Calculate total number of !ZONE_HIGHMEM pages */
5522 for_each_zone(zone
) {
5523 if (!is_highmem(zone
))
5524 lowmem_pages
+= zone
->managed_pages
;
5527 for_each_zone(zone
) {
5530 spin_lock_irqsave(&zone
->lock
, flags
);
5531 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5532 do_div(tmp
, lowmem_pages
);
5533 if (is_highmem(zone
)) {
5535 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5536 * need highmem pages, so cap pages_min to a small
5539 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5540 * deltas controls asynch page reclaim, and so should
5541 * not be capped for highmem.
5543 unsigned long min_pages
;
5545 min_pages
= zone
->managed_pages
/ 1024;
5546 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5547 zone
->watermark
[WMARK_MIN
] = min_pages
;
5550 * If it's a lowmem zone, reserve a number of pages
5551 * proportionate to the zone's size.
5553 zone
->watermark
[WMARK_MIN
] = tmp
;
5556 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5557 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5559 setup_zone_migrate_reserve(zone
);
5560 spin_unlock_irqrestore(&zone
->lock
, flags
);
5563 /* update totalreserve_pages */
5564 calculate_totalreserve_pages();
5568 * setup_per_zone_wmarks - called when min_free_kbytes changes
5569 * or when memory is hot-{added|removed}
5571 * Ensures that the watermark[min,low,high] values for each zone are set
5572 * correctly with respect to min_free_kbytes.
5574 void setup_per_zone_wmarks(void)
5576 mutex_lock(&zonelists_mutex
);
5577 __setup_per_zone_wmarks();
5578 mutex_unlock(&zonelists_mutex
);
5582 * The inactive anon list should be small enough that the VM never has to
5583 * do too much work, but large enough that each inactive page has a chance
5584 * to be referenced again before it is swapped out.
5586 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5587 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5588 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5589 * the anonymous pages are kept on the inactive list.
5592 * memory ratio inactive anon
5593 * -------------------------------------
5602 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5604 unsigned int gb
, ratio
;
5606 /* Zone size in gigabytes */
5607 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5609 ratio
= int_sqrt(10 * gb
);
5613 zone
->inactive_ratio
= ratio
;
5616 static void __meminit
setup_per_zone_inactive_ratio(void)
5621 calculate_zone_inactive_ratio(zone
);
5625 * Initialise min_free_kbytes.
5627 * For small machines we want it small (128k min). For large machines
5628 * we want it large (64MB max). But it is not linear, because network
5629 * bandwidth does not increase linearly with machine size. We use
5631 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5632 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5648 int __meminit
init_per_zone_wmark_min(void)
5650 unsigned long lowmem_kbytes
;
5652 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5654 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5655 if (min_free_kbytes
< 128)
5656 min_free_kbytes
= 128;
5657 if (min_free_kbytes
> 65536)
5658 min_free_kbytes
= 65536;
5659 setup_per_zone_wmarks();
5660 refresh_zone_stat_thresholds();
5661 setup_per_zone_lowmem_reserve();
5662 setup_per_zone_inactive_ratio();
5665 module_init(init_per_zone_wmark_min
)
5668 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5669 * that we can call two helper functions whenever min_free_kbytes
5672 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5673 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5675 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5677 setup_per_zone_wmarks();
5682 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5683 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5688 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5693 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5694 sysctl_min_unmapped_ratio
) / 100;
5698 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5699 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5704 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5709 zone
->min_slab_pages
= (zone
->managed_pages
*
5710 sysctl_min_slab_ratio
) / 100;
5716 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5717 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5718 * whenever sysctl_lowmem_reserve_ratio changes.
5720 * The reserve ratio obviously has absolutely no relation with the
5721 * minimum watermarks. The lowmem reserve ratio can only make sense
5722 * if in function of the boot time zone sizes.
5724 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5725 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5727 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5728 setup_per_zone_lowmem_reserve();
5733 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5734 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5735 * can have before it gets flushed back to buddy allocator.
5738 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5739 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5745 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5746 if (!write
|| (ret
< 0))
5748 for_each_populated_zone(zone
) {
5749 for_each_possible_cpu(cpu
) {
5751 high
= zone
->managed_pages
/ percpu_pagelist_fraction
;
5752 setup_pagelist_highmark(
5753 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5759 int hashdist
= HASHDIST_DEFAULT
;
5762 static int __init
set_hashdist(char *str
)
5766 hashdist
= simple_strtoul(str
, &str
, 0);
5769 __setup("hashdist=", set_hashdist
);
5773 * allocate a large system hash table from bootmem
5774 * - it is assumed that the hash table must contain an exact power-of-2
5775 * quantity of entries
5776 * - limit is the number of hash buckets, not the total allocation size
5778 void *__init
alloc_large_system_hash(const char *tablename
,
5779 unsigned long bucketsize
,
5780 unsigned long numentries
,
5783 unsigned int *_hash_shift
,
5784 unsigned int *_hash_mask
,
5785 unsigned long low_limit
,
5786 unsigned long high_limit
)
5788 unsigned long long max
= high_limit
;
5789 unsigned long log2qty
, size
;
5792 /* allow the kernel cmdline to have a say */
5794 /* round applicable memory size up to nearest megabyte */
5795 numentries
= nr_kernel_pages
;
5796 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5797 numentries
>>= 20 - PAGE_SHIFT
;
5798 numentries
<<= 20 - PAGE_SHIFT
;
5800 /* limit to 1 bucket per 2^scale bytes of low memory */
5801 if (scale
> PAGE_SHIFT
)
5802 numentries
>>= (scale
- PAGE_SHIFT
);
5804 numentries
<<= (PAGE_SHIFT
- scale
);
5806 /* Make sure we've got at least a 0-order allocation.. */
5807 if (unlikely(flags
& HASH_SMALL
)) {
5808 /* Makes no sense without HASH_EARLY */
5809 WARN_ON(!(flags
& HASH_EARLY
));
5810 if (!(numentries
>> *_hash_shift
)) {
5811 numentries
= 1UL << *_hash_shift
;
5812 BUG_ON(!numentries
);
5814 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5815 numentries
= PAGE_SIZE
/ bucketsize
;
5817 numentries
= roundup_pow_of_two(numentries
);
5819 /* limit allocation size to 1/16 total memory by default */
5821 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5822 do_div(max
, bucketsize
);
5824 max
= min(max
, 0x80000000ULL
);
5826 if (numentries
< low_limit
)
5827 numentries
= low_limit
;
5828 if (numentries
> max
)
5831 log2qty
= ilog2(numentries
);
5834 size
= bucketsize
<< log2qty
;
5835 if (flags
& HASH_EARLY
)
5836 table
= alloc_bootmem_nopanic(size
);
5838 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5841 * If bucketsize is not a power-of-two, we may free
5842 * some pages at the end of hash table which
5843 * alloc_pages_exact() automatically does
5845 if (get_order(size
) < MAX_ORDER
) {
5846 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5847 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5850 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5853 panic("Failed to allocate %s hash table\n", tablename
);
5855 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5858 ilog2(size
) - PAGE_SHIFT
,
5862 *_hash_shift
= log2qty
;
5864 *_hash_mask
= (1 << log2qty
) - 1;
5869 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5870 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5873 #ifdef CONFIG_SPARSEMEM
5874 return __pfn_to_section(pfn
)->pageblock_flags
;
5876 return zone
->pageblock_flags
;
5877 #endif /* CONFIG_SPARSEMEM */
5880 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5882 #ifdef CONFIG_SPARSEMEM
5883 pfn
&= (PAGES_PER_SECTION
-1);
5884 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5886 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
5887 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5888 #endif /* CONFIG_SPARSEMEM */
5892 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5893 * @page: The page within the block of interest
5894 * @start_bitidx: The first bit of interest to retrieve
5895 * @end_bitidx: The last bit of interest
5896 * returns pageblock_bits flags
5898 unsigned long get_pageblock_flags_group(struct page
*page
,
5899 int start_bitidx
, int end_bitidx
)
5902 unsigned long *bitmap
;
5903 unsigned long pfn
, bitidx
;
5904 unsigned long flags
= 0;
5905 unsigned long value
= 1;
5907 zone
= page_zone(page
);
5908 pfn
= page_to_pfn(page
);
5909 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5910 bitidx
= pfn_to_bitidx(zone
, pfn
);
5912 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5913 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5920 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5921 * @page: The page within the block of interest
5922 * @start_bitidx: The first bit of interest
5923 * @end_bitidx: The last bit of interest
5924 * @flags: The flags to set
5926 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5927 int start_bitidx
, int end_bitidx
)
5930 unsigned long *bitmap
;
5931 unsigned long pfn
, bitidx
;
5932 unsigned long value
= 1;
5934 zone
= page_zone(page
);
5935 pfn
= page_to_pfn(page
);
5936 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5937 bitidx
= pfn_to_bitidx(zone
, pfn
);
5938 VM_BUG_ON(!zone_spans_pfn(zone
, pfn
));
5940 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5942 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5944 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5948 * This function checks whether pageblock includes unmovable pages or not.
5949 * If @count is not zero, it is okay to include less @count unmovable pages
5951 * PageLRU check wihtout isolation or lru_lock could race so that
5952 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5953 * expect this function should be exact.
5955 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
5956 bool skip_hwpoisoned_pages
)
5958 unsigned long pfn
, iter
, found
;
5962 * For avoiding noise data, lru_add_drain_all() should be called
5963 * If ZONE_MOVABLE, the zone never contains unmovable pages
5965 if (zone_idx(zone
) == ZONE_MOVABLE
)
5967 mt
= get_pageblock_migratetype(page
);
5968 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
5971 pfn
= page_to_pfn(page
);
5972 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5973 unsigned long check
= pfn
+ iter
;
5975 if (!pfn_valid_within(check
))
5978 page
= pfn_to_page(check
);
5980 * We can't use page_count without pin a page
5981 * because another CPU can free compound page.
5982 * This check already skips compound tails of THP
5983 * because their page->_count is zero at all time.
5985 if (!atomic_read(&page
->_count
)) {
5986 if (PageBuddy(page
))
5987 iter
+= (1 << page_order(page
)) - 1;
5992 * The HWPoisoned page may be not in buddy system, and
5993 * page_count() is not 0.
5995 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6001 * If there are RECLAIMABLE pages, we need to check it.
6002 * But now, memory offline itself doesn't call shrink_slab()
6003 * and it still to be fixed.
6006 * If the page is not RAM, page_count()should be 0.
6007 * we don't need more check. This is an _used_ not-movable page.
6009 * The problematic thing here is PG_reserved pages. PG_reserved
6010 * is set to both of a memory hole page and a _used_ kernel
6019 bool is_pageblock_removable_nolock(struct page
*page
)
6025 * We have to be careful here because we are iterating over memory
6026 * sections which are not zone aware so we might end up outside of
6027 * the zone but still within the section.
6028 * We have to take care about the node as well. If the node is offline
6029 * its NODE_DATA will be NULL - see page_zone.
6031 if (!node_online(page_to_nid(page
)))
6034 zone
= page_zone(page
);
6035 pfn
= page_to_pfn(page
);
6036 if (!zone_spans_pfn(zone
, pfn
))
6039 return !has_unmovable_pages(zone
, page
, 0, true);
6044 static unsigned long pfn_max_align_down(unsigned long pfn
)
6046 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6047 pageblock_nr_pages
) - 1);
6050 static unsigned long pfn_max_align_up(unsigned long pfn
)
6052 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6053 pageblock_nr_pages
));
6056 /* [start, end) must belong to a single zone. */
6057 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6058 unsigned long start
, unsigned long end
)
6060 /* This function is based on compact_zone() from compaction.c. */
6061 unsigned long nr_reclaimed
;
6062 unsigned long pfn
= start
;
6063 unsigned int tries
= 0;
6068 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6069 if (fatal_signal_pending(current
)) {
6074 if (list_empty(&cc
->migratepages
)) {
6075 cc
->nr_migratepages
= 0;
6076 pfn
= isolate_migratepages_range(cc
->zone
, cc
,
6083 } else if (++tries
== 5) {
6084 ret
= ret
< 0 ? ret
: -EBUSY
;
6088 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6090 cc
->nr_migratepages
-= nr_reclaimed
;
6092 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6093 0, MIGRATE_SYNC
, MR_CMA
);
6096 putback_movable_pages(&cc
->migratepages
);
6103 * alloc_contig_range() -- tries to allocate given range of pages
6104 * @start: start PFN to allocate
6105 * @end: one-past-the-last PFN to allocate
6106 * @migratetype: migratetype of the underlaying pageblocks (either
6107 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6108 * in range must have the same migratetype and it must
6109 * be either of the two.
6111 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6112 * aligned, however it's the caller's responsibility to guarantee that
6113 * we are the only thread that changes migrate type of pageblocks the
6116 * The PFN range must belong to a single zone.
6118 * Returns zero on success or negative error code. On success all
6119 * pages which PFN is in [start, end) are allocated for the caller and
6120 * need to be freed with free_contig_range().
6122 int alloc_contig_range(unsigned long start
, unsigned long end
,
6123 unsigned migratetype
)
6125 unsigned long outer_start
, outer_end
;
6128 struct compact_control cc
= {
6129 .nr_migratepages
= 0,
6131 .zone
= page_zone(pfn_to_page(start
)),
6133 .ignore_skip_hint
= true,
6135 INIT_LIST_HEAD(&cc
.migratepages
);
6138 * What we do here is we mark all pageblocks in range as
6139 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6140 * have different sizes, and due to the way page allocator
6141 * work, we align the range to biggest of the two pages so
6142 * that page allocator won't try to merge buddies from
6143 * different pageblocks and change MIGRATE_ISOLATE to some
6144 * other migration type.
6146 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6147 * migrate the pages from an unaligned range (ie. pages that
6148 * we are interested in). This will put all the pages in
6149 * range back to page allocator as MIGRATE_ISOLATE.
6151 * When this is done, we take the pages in range from page
6152 * allocator removing them from the buddy system. This way
6153 * page allocator will never consider using them.
6155 * This lets us mark the pageblocks back as
6156 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6157 * aligned range but not in the unaligned, original range are
6158 * put back to page allocator so that buddy can use them.
6161 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6162 pfn_max_align_up(end
), migratetype
,
6167 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6172 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6173 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6174 * more, all pages in [start, end) are free in page allocator.
6175 * What we are going to do is to allocate all pages from
6176 * [start, end) (that is remove them from page allocator).
6178 * The only problem is that pages at the beginning and at the
6179 * end of interesting range may be not aligned with pages that
6180 * page allocator holds, ie. they can be part of higher order
6181 * pages. Because of this, we reserve the bigger range and
6182 * once this is done free the pages we are not interested in.
6184 * We don't have to hold zone->lock here because the pages are
6185 * isolated thus they won't get removed from buddy.
6188 lru_add_drain_all();
6192 outer_start
= start
;
6193 while (!PageBuddy(pfn_to_page(outer_start
))) {
6194 if (++order
>= MAX_ORDER
) {
6198 outer_start
&= ~0UL << order
;
6201 /* Make sure the range is really isolated. */
6202 if (test_pages_isolated(outer_start
, end
, false)) {
6203 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6210 /* Grab isolated pages from freelists. */
6211 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6217 /* Free head and tail (if any) */
6218 if (start
!= outer_start
)
6219 free_contig_range(outer_start
, start
- outer_start
);
6220 if (end
!= outer_end
)
6221 free_contig_range(end
, outer_end
- end
);
6224 undo_isolate_page_range(pfn_max_align_down(start
),
6225 pfn_max_align_up(end
), migratetype
);
6229 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6231 unsigned int count
= 0;
6233 for (; nr_pages
--; pfn
++) {
6234 struct page
*page
= pfn_to_page(pfn
);
6236 count
+= page_count(page
) != 1;
6239 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6243 #ifdef CONFIG_MEMORY_HOTPLUG
6244 static int __meminit
__zone_pcp_update(void *data
)
6246 struct zone
*zone
= data
;
6248 unsigned long batch
= zone_batchsize(zone
), flags
;
6250 for_each_possible_cpu(cpu
) {
6251 struct per_cpu_pageset
*pset
;
6252 struct per_cpu_pages
*pcp
;
6254 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6257 local_irq_save(flags
);
6259 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
6260 drain_zonestat(zone
, pset
);
6261 setup_pageset(pset
, batch
);
6262 local_irq_restore(flags
);
6267 void __meminit
zone_pcp_update(struct zone
*zone
)
6269 stop_machine(__zone_pcp_update
, zone
, NULL
);
6273 void zone_pcp_reset(struct zone
*zone
)
6275 unsigned long flags
;
6277 struct per_cpu_pageset
*pset
;
6279 /* avoid races with drain_pages() */
6280 local_irq_save(flags
);
6281 if (zone
->pageset
!= &boot_pageset
) {
6282 for_each_online_cpu(cpu
) {
6283 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6284 drain_zonestat(zone
, pset
);
6286 free_percpu(zone
->pageset
);
6287 zone
->pageset
= &boot_pageset
;
6289 local_irq_restore(flags
);
6292 #ifdef CONFIG_MEMORY_HOTREMOVE
6294 * All pages in the range must be isolated before calling this.
6297 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6303 unsigned long flags
;
6304 /* find the first valid pfn */
6305 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6310 zone
= page_zone(pfn_to_page(pfn
));
6311 spin_lock_irqsave(&zone
->lock
, flags
);
6313 while (pfn
< end_pfn
) {
6314 if (!pfn_valid(pfn
)) {
6318 page
= pfn_to_page(pfn
);
6320 * The HWPoisoned page may be not in buddy system, and
6321 * page_count() is not 0.
6323 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6325 SetPageReserved(page
);
6329 BUG_ON(page_count(page
));
6330 BUG_ON(!PageBuddy(page
));
6331 order
= page_order(page
);
6332 #ifdef CONFIG_DEBUG_VM
6333 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6334 pfn
, 1 << order
, end_pfn
);
6336 list_del(&page
->lru
);
6337 rmv_page_order(page
);
6338 zone
->free_area
[order
].nr_free
--;
6339 for (i
= 0; i
< (1 << order
); i
++)
6340 SetPageReserved((page
+i
));
6341 pfn
+= (1 << order
);
6343 spin_unlock_irqrestore(&zone
->lock
, flags
);
6347 #ifdef CONFIG_MEMORY_FAILURE
6348 bool is_free_buddy_page(struct page
*page
)
6350 struct zone
*zone
= page_zone(page
);
6351 unsigned long pfn
= page_to_pfn(page
);
6352 unsigned long flags
;
6355 spin_lock_irqsave(&zone
->lock
, flags
);
6356 for (order
= 0; order
< MAX_ORDER
; order
++) {
6357 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6359 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6362 spin_unlock_irqrestore(&zone
->lock
, flags
);
6364 return order
< MAX_ORDER
;
6368 static const struct trace_print_flags pageflag_names
[] = {
6369 {1UL << PG_locked
, "locked" },
6370 {1UL << PG_error
, "error" },
6371 {1UL << PG_referenced
, "referenced" },
6372 {1UL << PG_uptodate
, "uptodate" },
6373 {1UL << PG_dirty
, "dirty" },
6374 {1UL << PG_lru
, "lru" },
6375 {1UL << PG_active
, "active" },
6376 {1UL << PG_slab
, "slab" },
6377 {1UL << PG_owner_priv_1
, "owner_priv_1" },
6378 {1UL << PG_arch_1
, "arch_1" },
6379 {1UL << PG_reserved
, "reserved" },
6380 {1UL << PG_private
, "private" },
6381 {1UL << PG_private_2
, "private_2" },
6382 {1UL << PG_writeback
, "writeback" },
6383 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6384 {1UL << PG_head
, "head" },
6385 {1UL << PG_tail
, "tail" },
6387 {1UL << PG_compound
, "compound" },
6389 {1UL << PG_swapcache
, "swapcache" },
6390 {1UL << PG_mappedtodisk
, "mappedtodisk" },
6391 {1UL << PG_reclaim
, "reclaim" },
6392 {1UL << PG_swapbacked
, "swapbacked" },
6393 {1UL << PG_unevictable
, "unevictable" },
6395 {1UL << PG_mlocked
, "mlocked" },
6397 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6398 {1UL << PG_uncached
, "uncached" },
6400 #ifdef CONFIG_MEMORY_FAILURE
6401 {1UL << PG_hwpoison
, "hwpoison" },
6403 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6404 {1UL << PG_compound_lock
, "compound_lock" },
6408 static void dump_page_flags(unsigned long flags
)
6410 const char *delim
= "";
6414 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names
) != __NR_PAGEFLAGS
);
6416 printk(KERN_ALERT
"page flags: %#lx(", flags
);
6418 /* remove zone id */
6419 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
6421 for (i
= 0; i
< ARRAY_SIZE(pageflag_names
) && flags
; i
++) {
6423 mask
= pageflag_names
[i
].mask
;
6424 if ((flags
& mask
) != mask
)
6428 printk("%s%s", delim
, pageflag_names
[i
].name
);
6432 /* check for left over flags */
6434 printk("%s%#lx", delim
, flags
);
6439 void dump_page(struct page
*page
)
6442 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6443 page
, atomic_read(&page
->_count
), page_mapcount(page
),
6444 page
->mapping
, page
->index
);
6445 dump_page_flags(page
->flags
);
6446 mem_cgroup_print_bad_page(page
);