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/hugetlb.h>
62 #include <linux/sched/rt.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
69 static DEFINE_MUTEX(pcp_batch_high_lock
);
71 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
72 DEFINE_PER_CPU(int, numa_node
);
73 EXPORT_PER_CPU_SYMBOL(numa_node
);
76 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
78 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
79 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
80 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
81 * defined in <linux/topology.h>.
83 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
84 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
88 * Array of node states.
90 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
91 [N_POSSIBLE
] = NODE_MASK_ALL
,
92 [N_ONLINE
] = { { [0] = 1UL } },
94 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
96 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
98 #ifdef CONFIG_MOVABLE_NODE
99 [N_MEMORY
] = { { [0] = 1UL } },
101 [N_CPU
] = { { [0] = 1UL } },
104 EXPORT_SYMBOL(node_states
);
106 unsigned long totalram_pages __read_mostly
;
107 unsigned long totalreserve_pages __read_mostly
;
109 * When calculating the number of globally allowed dirty pages, there
110 * is a certain number of per-zone reserves that should not be
111 * considered dirtyable memory. This is the sum of those reserves
112 * over all existing zones that contribute dirtyable memory.
114 unsigned long dirty_balance_reserve __read_mostly
;
116 int percpu_pagelist_fraction
;
117 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
119 #ifdef CONFIG_PM_SLEEP
121 * The following functions are used by the suspend/hibernate code to temporarily
122 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
123 * while devices are suspended. To avoid races with the suspend/hibernate code,
124 * they should always be called with pm_mutex held (gfp_allowed_mask also should
125 * only be modified with pm_mutex held, unless the suspend/hibernate code is
126 * guaranteed not to run in parallel with that modification).
129 static gfp_t saved_gfp_mask
;
131 void pm_restore_gfp_mask(void)
133 WARN_ON(!mutex_is_locked(&pm_mutex
));
134 if (saved_gfp_mask
) {
135 gfp_allowed_mask
= saved_gfp_mask
;
140 void pm_restrict_gfp_mask(void)
142 WARN_ON(!mutex_is_locked(&pm_mutex
));
143 WARN_ON(saved_gfp_mask
);
144 saved_gfp_mask
= gfp_allowed_mask
;
145 gfp_allowed_mask
&= ~GFP_IOFS
;
148 bool pm_suspended_storage(void)
150 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
154 #endif /* CONFIG_PM_SLEEP */
156 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
157 int pageblock_order __read_mostly
;
160 static void __free_pages_ok(struct page
*page
, unsigned int order
);
163 * results with 256, 32 in the lowmem_reserve sysctl:
164 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
165 * 1G machine -> (16M dma, 784M normal, 224M high)
166 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
167 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
168 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
170 * TBD: should special case ZONE_DMA32 machines here - in those we normally
171 * don't need any ZONE_NORMAL reservation
173 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
174 #ifdef CONFIG_ZONE_DMA
177 #ifdef CONFIG_ZONE_DMA32
180 #ifdef CONFIG_HIGHMEM
186 EXPORT_SYMBOL(totalram_pages
);
188 static char * const zone_names
[MAX_NR_ZONES
] = {
189 #ifdef CONFIG_ZONE_DMA
192 #ifdef CONFIG_ZONE_DMA32
196 #ifdef CONFIG_HIGHMEM
202 int min_free_kbytes
= 1024;
204 static unsigned long __meminitdata nr_kernel_pages
;
205 static unsigned long __meminitdata nr_all_pages
;
206 static unsigned long __meminitdata dma_reserve
;
208 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
209 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
210 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
211 static unsigned long __initdata required_kernelcore
;
212 static unsigned long __initdata required_movablecore
;
213 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
215 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
217 EXPORT_SYMBOL(movable_zone
);
218 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
221 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
222 int nr_online_nodes __read_mostly
= 1;
223 EXPORT_SYMBOL(nr_node_ids
);
224 EXPORT_SYMBOL(nr_online_nodes
);
227 int page_group_by_mobility_disabled __read_mostly
;
229 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
232 if (unlikely(page_group_by_mobility_disabled
))
233 migratetype
= MIGRATE_UNMOVABLE
;
235 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
236 PB_migrate
, PB_migrate_end
);
239 bool oom_killer_disabled __read_mostly
;
241 #ifdef CONFIG_DEBUG_VM
242 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
246 unsigned long pfn
= page_to_pfn(page
);
247 unsigned long sp
, start_pfn
;
250 seq
= zone_span_seqbegin(zone
);
251 start_pfn
= zone
->zone_start_pfn
;
252 sp
= zone
->spanned_pages
;
253 if (!zone_spans_pfn(zone
, pfn
))
255 } while (zone_span_seqretry(zone
, seq
));
258 pr_err("page %lu outside zone [ %lu - %lu ]\n",
259 pfn
, start_pfn
, start_pfn
+ sp
);
264 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
266 if (!pfn_valid_within(page_to_pfn(page
)))
268 if (zone
!= page_zone(page
))
274 * Temporary debugging check for pages not lying within a given zone.
276 static int bad_range(struct zone
*zone
, struct page
*page
)
278 if (page_outside_zone_boundaries(zone
, page
))
280 if (!page_is_consistent(zone
, page
))
286 static inline int bad_range(struct zone
*zone
, struct page
*page
)
292 static void bad_page(struct page
*page
)
294 static unsigned long resume
;
295 static unsigned long nr_shown
;
296 static unsigned long nr_unshown
;
298 /* Don't complain about poisoned pages */
299 if (PageHWPoison(page
)) {
300 page_mapcount_reset(page
); /* remove PageBuddy */
305 * Allow a burst of 60 reports, then keep quiet for that minute;
306 * or allow a steady drip of one report per second.
308 if (nr_shown
== 60) {
309 if (time_before(jiffies
, resume
)) {
315 "BUG: Bad page state: %lu messages suppressed\n",
322 resume
= jiffies
+ 60 * HZ
;
324 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
325 current
->comm
, page_to_pfn(page
));
331 /* Leave bad fields for debug, except PageBuddy could make trouble */
332 page_mapcount_reset(page
); /* remove PageBuddy */
333 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
337 * Higher-order pages are called "compound pages". They are structured thusly:
339 * The first PAGE_SIZE page is called the "head page".
341 * The remaining PAGE_SIZE pages are called "tail pages".
343 * All pages have PG_compound set. All tail pages have their ->first_page
344 * pointing at the head page.
346 * The first tail page's ->lru.next holds the address of the compound page's
347 * put_page() function. Its ->lru.prev holds the order of allocation.
348 * This usage means that zero-order pages may not be compound.
351 static void free_compound_page(struct page
*page
)
353 __free_pages_ok(page
, compound_order(page
));
356 void prep_compound_page(struct page
*page
, unsigned long order
)
359 int nr_pages
= 1 << order
;
361 set_compound_page_dtor(page
, free_compound_page
);
362 set_compound_order(page
, order
);
364 for (i
= 1; i
< nr_pages
; i
++) {
365 struct page
*p
= page
+ i
;
367 set_page_count(p
, 0);
368 p
->first_page
= page
;
372 /* update __split_huge_page_refcount if you change this function */
373 static int destroy_compound_page(struct page
*page
, unsigned long order
)
376 int nr_pages
= 1 << order
;
379 if (unlikely(compound_order(page
) != order
)) {
384 __ClearPageHead(page
);
386 for (i
= 1; i
< nr_pages
; i
++) {
387 struct page
*p
= page
+ i
;
389 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
399 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
404 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
405 * and __GFP_HIGHMEM from hard or soft interrupt context.
407 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
408 for (i
= 0; i
< (1 << order
); i
++)
409 clear_highpage(page
+ i
);
412 #ifdef CONFIG_DEBUG_PAGEALLOC
413 unsigned int _debug_guardpage_minorder
;
415 static int __init
debug_guardpage_minorder_setup(char *buf
)
419 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
420 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
423 _debug_guardpage_minorder
= res
;
424 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
427 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
429 static inline void set_page_guard_flag(struct page
*page
)
431 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
434 static inline void clear_page_guard_flag(struct page
*page
)
436 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
439 static inline void set_page_guard_flag(struct page
*page
) { }
440 static inline void clear_page_guard_flag(struct page
*page
) { }
443 static inline void set_page_order(struct page
*page
, int order
)
445 set_page_private(page
, order
);
446 __SetPageBuddy(page
);
449 static inline void rmv_page_order(struct page
*page
)
451 __ClearPageBuddy(page
);
452 set_page_private(page
, 0);
456 * Locate the struct page for both the matching buddy in our
457 * pair (buddy1) and the combined O(n+1) page they form (page).
459 * 1) Any buddy B1 will have an order O twin B2 which satisfies
460 * the following equation:
462 * For example, if the starting buddy (buddy2) is #8 its order
464 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
466 * 2) Any buddy B will have an order O+1 parent P which
467 * satisfies the following equation:
470 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
472 static inline unsigned long
473 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
475 return page_idx
^ (1 << order
);
479 * This function checks whether a page is free && is the buddy
480 * we can do coalesce a page and its buddy if
481 * (a) the buddy is not in a hole &&
482 * (b) the buddy is in the buddy system &&
483 * (c) a page and its buddy have the same order &&
484 * (d) a page and its buddy are in the same zone.
486 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
487 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
489 * For recording page's order, we use page_private(page).
491 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
494 if (!pfn_valid_within(page_to_pfn(buddy
)))
497 if (page_zone_id(page
) != page_zone_id(buddy
))
500 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
501 VM_BUG_ON(page_count(buddy
) != 0);
505 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
506 VM_BUG_ON(page_count(buddy
) != 0);
513 * Freeing function for a buddy system allocator.
515 * The concept of a buddy system is to maintain direct-mapped table
516 * (containing bit values) for memory blocks of various "orders".
517 * The bottom level table contains the map for the smallest allocatable
518 * units of memory (here, pages), and each level above it describes
519 * pairs of units from the levels below, hence, "buddies".
520 * At a high level, all that happens here is marking the table entry
521 * at the bottom level available, and propagating the changes upward
522 * as necessary, plus some accounting needed to play nicely with other
523 * parts of the VM system.
524 * At each level, we keep a list of pages, which are heads of continuous
525 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
526 * order is recorded in page_private(page) field.
527 * So when we are allocating or freeing one, we can derive the state of the
528 * other. That is, if we allocate a small block, and both were
529 * free, the remainder of the region must be split into blocks.
530 * If a block is freed, and its buddy is also free, then this
531 * triggers coalescing into a block of larger size.
536 static inline void __free_one_page(struct page
*page
,
537 struct zone
*zone
, unsigned int order
,
540 unsigned long page_idx
;
541 unsigned long combined_idx
;
542 unsigned long uninitialized_var(buddy_idx
);
545 VM_BUG_ON(!zone_is_initialized(zone
));
547 if (unlikely(PageCompound(page
)))
548 if (unlikely(destroy_compound_page(page
, order
)))
551 VM_BUG_ON(migratetype
== -1);
553 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
555 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
556 VM_BUG_ON(bad_range(zone
, page
));
558 while (order
< MAX_ORDER
-1) {
559 buddy_idx
= __find_buddy_index(page_idx
, order
);
560 buddy
= page
+ (buddy_idx
- page_idx
);
561 if (!page_is_buddy(page
, buddy
, order
))
564 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
565 * merge with it and move up one order.
567 if (page_is_guard(buddy
)) {
568 clear_page_guard_flag(buddy
);
569 set_page_private(page
, 0);
570 __mod_zone_freepage_state(zone
, 1 << order
,
573 list_del(&buddy
->lru
);
574 zone
->free_area
[order
].nr_free
--;
575 rmv_page_order(buddy
);
577 combined_idx
= buddy_idx
& page_idx
;
578 page
= page
+ (combined_idx
- page_idx
);
579 page_idx
= combined_idx
;
582 set_page_order(page
, order
);
585 * If this is not the largest possible page, check if the buddy
586 * of the next-highest order is free. If it is, it's possible
587 * that pages are being freed that will coalesce soon. In case,
588 * that is happening, add the free page to the tail of the list
589 * so it's less likely to be used soon and more likely to be merged
590 * as a higher order page
592 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
593 struct page
*higher_page
, *higher_buddy
;
594 combined_idx
= buddy_idx
& page_idx
;
595 higher_page
= page
+ (combined_idx
- page_idx
);
596 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
597 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
598 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
599 list_add_tail(&page
->lru
,
600 &zone
->free_area
[order
].free_list
[migratetype
]);
605 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
607 zone
->free_area
[order
].nr_free
++;
610 static inline int free_pages_check(struct page
*page
)
612 if (unlikely(page_mapcount(page
) |
613 (page
->mapping
!= NULL
) |
614 (atomic_read(&page
->_count
) != 0) |
615 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
) |
616 (mem_cgroup_bad_page_check(page
)))) {
620 page_nid_reset_last(page
);
621 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
622 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
627 * Frees a number of pages from the PCP lists
628 * Assumes all pages on list are in same zone, and of same order.
629 * count is the number of pages to free.
631 * If the zone was previously in an "all pages pinned" state then look to
632 * see if this freeing clears that state.
634 * And clear the zone's pages_scanned counter, to hold off the "all pages are
635 * pinned" detection logic.
637 static void free_pcppages_bulk(struct zone
*zone
, int count
,
638 struct per_cpu_pages
*pcp
)
644 spin_lock(&zone
->lock
);
645 zone
->all_unreclaimable
= 0;
646 zone
->pages_scanned
= 0;
650 struct list_head
*list
;
653 * Remove pages from lists in a round-robin fashion. A
654 * batch_free count is maintained that is incremented when an
655 * empty list is encountered. This is so more pages are freed
656 * off fuller lists instead of spinning excessively around empty
661 if (++migratetype
== MIGRATE_PCPTYPES
)
663 list
= &pcp
->lists
[migratetype
];
664 } while (list_empty(list
));
666 /* This is the only non-empty list. Free them all. */
667 if (batch_free
== MIGRATE_PCPTYPES
)
668 batch_free
= to_free
;
671 int mt
; /* migratetype of the to-be-freed page */
673 page
= list_entry(list
->prev
, struct page
, lru
);
674 /* must delete as __free_one_page list manipulates */
675 list_del(&page
->lru
);
676 mt
= get_freepage_migratetype(page
);
677 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
678 __free_one_page(page
, zone
, 0, mt
);
679 trace_mm_page_pcpu_drain(page
, 0, mt
);
680 if (likely(!is_migrate_isolate_page(page
))) {
681 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1);
682 if (is_migrate_cma(mt
))
683 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
, 1);
685 } while (--to_free
&& --batch_free
&& !list_empty(list
));
687 spin_unlock(&zone
->lock
);
690 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
693 spin_lock(&zone
->lock
);
694 zone
->all_unreclaimable
= 0;
695 zone
->pages_scanned
= 0;
697 __free_one_page(page
, zone
, order
, migratetype
);
698 if (unlikely(!is_migrate_isolate(migratetype
)))
699 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
700 spin_unlock(&zone
->lock
);
703 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
708 trace_mm_page_free(page
, order
);
709 kmemcheck_free_shadow(page
, order
);
712 page
->mapping
= NULL
;
713 for (i
= 0; i
< (1 << order
); i
++)
714 bad
+= free_pages_check(page
+ i
);
718 if (!PageHighMem(page
)) {
719 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
720 debug_check_no_obj_freed(page_address(page
),
723 arch_free_page(page
, order
);
724 kernel_map_pages(page
, 1 << order
, 0);
729 static void __free_pages_ok(struct page
*page
, unsigned int order
)
734 if (!free_pages_prepare(page
, order
))
737 local_irq_save(flags
);
738 __count_vm_events(PGFREE
, 1 << order
);
739 migratetype
= get_pageblock_migratetype(page
);
740 set_freepage_migratetype(page
, migratetype
);
741 free_one_page(page_zone(page
), page
, order
, migratetype
);
742 local_irq_restore(flags
);
746 * Read access to zone->managed_pages is safe because it's unsigned long,
747 * but we still need to serialize writers. Currently all callers of
748 * __free_pages_bootmem() except put_page_bootmem() should only be used
749 * at boot time. So for shorter boot time, we shift the burden to
750 * put_page_bootmem() to serialize writers.
752 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
754 unsigned int nr_pages
= 1 << order
;
758 for (loop
= 0; loop
< nr_pages
; loop
++) {
759 struct page
*p
= &page
[loop
];
761 if (loop
+ 1 < nr_pages
)
763 __ClearPageReserved(p
);
764 set_page_count(p
, 0);
767 page_zone(page
)->managed_pages
+= 1 << order
;
768 set_page_refcounted(page
);
769 __free_pages(page
, order
);
773 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
774 void __init
init_cma_reserved_pageblock(struct page
*page
)
776 unsigned i
= pageblock_nr_pages
;
777 struct page
*p
= page
;
780 __ClearPageReserved(p
);
781 set_page_count(p
, 0);
784 set_page_refcounted(page
);
785 set_pageblock_migratetype(page
, MIGRATE_CMA
);
786 __free_pages(page
, pageblock_order
);
787 totalram_pages
+= pageblock_nr_pages
;
788 #ifdef CONFIG_HIGHMEM
789 if (PageHighMem(page
))
790 totalhigh_pages
+= pageblock_nr_pages
;
796 * The order of subdivision here is critical for the IO subsystem.
797 * Please do not alter this order without good reasons and regression
798 * testing. Specifically, as large blocks of memory are subdivided,
799 * the order in which smaller blocks are delivered depends on the order
800 * they're subdivided in this function. This is the primary factor
801 * influencing the order in which pages are delivered to the IO
802 * subsystem according to empirical testing, and this is also justified
803 * by considering the behavior of a buddy system containing a single
804 * large block of memory acted on by a series of small allocations.
805 * This behavior is a critical factor in sglist merging's success.
809 static inline void expand(struct zone
*zone
, struct page
*page
,
810 int low
, int high
, struct free_area
*area
,
813 unsigned long size
= 1 << high
;
819 VM_BUG_ON(bad_range(zone
, &page
[size
]));
821 #ifdef CONFIG_DEBUG_PAGEALLOC
822 if (high
< debug_guardpage_minorder()) {
824 * Mark as guard pages (or page), that will allow to
825 * merge back to allocator when buddy will be freed.
826 * Corresponding page table entries will not be touched,
827 * pages will stay not present in virtual address space
829 INIT_LIST_HEAD(&page
[size
].lru
);
830 set_page_guard_flag(&page
[size
]);
831 set_page_private(&page
[size
], high
);
832 /* Guard pages are not available for any usage */
833 __mod_zone_freepage_state(zone
, -(1 << high
),
838 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
840 set_page_order(&page
[size
], high
);
845 * This page is about to be returned from the page allocator
847 static inline int check_new_page(struct page
*page
)
849 if (unlikely(page_mapcount(page
) |
850 (page
->mapping
!= NULL
) |
851 (atomic_read(&page
->_count
) != 0) |
852 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
) |
853 (mem_cgroup_bad_page_check(page
)))) {
860 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
864 for (i
= 0; i
< (1 << order
); i
++) {
865 struct page
*p
= page
+ i
;
866 if (unlikely(check_new_page(p
)))
870 set_page_private(page
, 0);
871 set_page_refcounted(page
);
873 arch_alloc_page(page
, order
);
874 kernel_map_pages(page
, 1 << order
, 1);
876 if (gfp_flags
& __GFP_ZERO
)
877 prep_zero_page(page
, order
, gfp_flags
);
879 if (order
&& (gfp_flags
& __GFP_COMP
))
880 prep_compound_page(page
, order
);
886 * Go through the free lists for the given migratetype and remove
887 * the smallest available page from the freelists
890 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
893 unsigned int current_order
;
894 struct free_area
* area
;
897 /* Find a page of the appropriate size in the preferred list */
898 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
899 area
= &(zone
->free_area
[current_order
]);
900 if (list_empty(&area
->free_list
[migratetype
]))
903 page
= list_entry(area
->free_list
[migratetype
].next
,
905 list_del(&page
->lru
);
906 rmv_page_order(page
);
908 expand(zone
, page
, order
, current_order
, area
, migratetype
);
917 * This array describes the order lists are fallen back to when
918 * the free lists for the desirable migrate type are depleted
920 static int fallbacks
[MIGRATE_TYPES
][4] = {
921 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
922 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
924 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
925 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
927 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
929 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
930 #ifdef CONFIG_MEMORY_ISOLATION
931 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
936 * Move the free pages in a range to the free lists of the requested type.
937 * Note that start_page and end_pages are not aligned on a pageblock
938 * boundary. If alignment is required, use move_freepages_block()
940 int move_freepages(struct zone
*zone
,
941 struct page
*start_page
, struct page
*end_page
,
948 #ifndef CONFIG_HOLES_IN_ZONE
950 * page_zone is not safe to call in this context when
951 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
952 * anyway as we check zone boundaries in move_freepages_block().
953 * Remove at a later date when no bug reports exist related to
954 * grouping pages by mobility
956 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
959 for (page
= start_page
; page
<= end_page
;) {
960 /* Make sure we are not inadvertently changing nodes */
961 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
963 if (!pfn_valid_within(page_to_pfn(page
))) {
968 if (!PageBuddy(page
)) {
973 order
= page_order(page
);
974 list_move(&page
->lru
,
975 &zone
->free_area
[order
].free_list
[migratetype
]);
976 set_freepage_migratetype(page
, migratetype
);
978 pages_moved
+= 1 << order
;
984 int move_freepages_block(struct zone
*zone
, struct page
*page
,
987 unsigned long start_pfn
, end_pfn
;
988 struct page
*start_page
, *end_page
;
990 start_pfn
= page_to_pfn(page
);
991 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
992 start_page
= pfn_to_page(start_pfn
);
993 end_page
= start_page
+ pageblock_nr_pages
- 1;
994 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
996 /* Do not cross zone boundaries */
997 if (!zone_spans_pfn(zone
, start_pfn
))
999 if (!zone_spans_pfn(zone
, end_pfn
))
1002 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1005 static void change_pageblock_range(struct page
*pageblock_page
,
1006 int start_order
, int migratetype
)
1008 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1010 while (nr_pageblocks
--) {
1011 set_pageblock_migratetype(pageblock_page
, migratetype
);
1012 pageblock_page
+= pageblock_nr_pages
;
1016 /* Remove an element from the buddy allocator from the fallback list */
1017 static inline struct page
*
1018 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
1020 struct free_area
* area
;
1025 /* Find the largest possible block of pages in the other list */
1026 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
1029 migratetype
= fallbacks
[start_migratetype
][i
];
1031 /* MIGRATE_RESERVE handled later if necessary */
1032 if (migratetype
== MIGRATE_RESERVE
)
1035 area
= &(zone
->free_area
[current_order
]);
1036 if (list_empty(&area
->free_list
[migratetype
]))
1039 page
= list_entry(area
->free_list
[migratetype
].next
,
1044 * If breaking a large block of pages, move all free
1045 * pages to the preferred allocation list. If falling
1046 * back for a reclaimable kernel allocation, be more
1047 * aggressive about taking ownership of free pages
1049 * On the other hand, never change migration
1050 * type of MIGRATE_CMA pageblocks nor move CMA
1051 * pages on different free lists. We don't
1052 * want unmovable pages to be allocated from
1053 * MIGRATE_CMA areas.
1055 if (!is_migrate_cma(migratetype
) &&
1056 (unlikely(current_order
>= pageblock_order
/ 2) ||
1057 start_migratetype
== MIGRATE_RECLAIMABLE
||
1058 page_group_by_mobility_disabled
)) {
1060 pages
= move_freepages_block(zone
, page
,
1063 /* Claim the whole block if over half of it is free */
1064 if (pages
>= (1 << (pageblock_order
-1)) ||
1065 page_group_by_mobility_disabled
)
1066 set_pageblock_migratetype(page
,
1069 migratetype
= start_migratetype
;
1072 /* Remove the page from the freelists */
1073 list_del(&page
->lru
);
1074 rmv_page_order(page
);
1076 /* Take ownership for orders >= pageblock_order */
1077 if (current_order
>= pageblock_order
&&
1078 !is_migrate_cma(migratetype
))
1079 change_pageblock_range(page
, current_order
,
1082 expand(zone
, page
, order
, current_order
, area
,
1083 is_migrate_cma(migratetype
)
1084 ? migratetype
: start_migratetype
);
1086 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1087 start_migratetype
, migratetype
);
1097 * Do the hard work of removing an element from the buddy allocator.
1098 * Call me with the zone->lock already held.
1100 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1106 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1108 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1109 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1112 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1113 * is used because __rmqueue_smallest is an inline function
1114 * and we want just one call site
1117 migratetype
= MIGRATE_RESERVE
;
1122 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1127 * Obtain a specified number of elements from the buddy allocator, all under
1128 * a single hold of the lock, for efficiency. Add them to the supplied list.
1129 * Returns the number of new pages which were placed at *list.
1131 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1132 unsigned long count
, struct list_head
*list
,
1133 int migratetype
, int cold
)
1135 int mt
= migratetype
, i
;
1137 spin_lock(&zone
->lock
);
1138 for (i
= 0; i
< count
; ++i
) {
1139 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1140 if (unlikely(page
== NULL
))
1144 * Split buddy pages returned by expand() are received here
1145 * in physical page order. The page is added to the callers and
1146 * list and the list head then moves forward. From the callers
1147 * perspective, the linked list is ordered by page number in
1148 * some conditions. This is useful for IO devices that can
1149 * merge IO requests if the physical pages are ordered
1152 if (likely(cold
== 0))
1153 list_add(&page
->lru
, list
);
1155 list_add_tail(&page
->lru
, list
);
1156 if (IS_ENABLED(CONFIG_CMA
)) {
1157 mt
= get_pageblock_migratetype(page
);
1158 if (!is_migrate_cma(mt
) && !is_migrate_isolate(mt
))
1161 set_freepage_migratetype(page
, mt
);
1163 if (is_migrate_cma(mt
))
1164 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1167 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1168 spin_unlock(&zone
->lock
);
1174 * Called from the vmstat counter updater to drain pagesets of this
1175 * currently executing processor on remote nodes after they have
1178 * Note that this function must be called with the thread pinned to
1179 * a single processor.
1181 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1183 unsigned long flags
;
1185 unsigned long batch
;
1187 local_irq_save(flags
);
1188 batch
= ACCESS_ONCE(pcp
->batch
);
1189 if (pcp
->count
>= batch
)
1192 to_drain
= pcp
->count
;
1194 free_pcppages_bulk(zone
, to_drain
, pcp
);
1195 pcp
->count
-= to_drain
;
1197 local_irq_restore(flags
);
1202 * Drain pages of the indicated processor.
1204 * The processor must either be the current processor and the
1205 * thread pinned to the current processor or a processor that
1208 static void drain_pages(unsigned int cpu
)
1210 unsigned long flags
;
1213 for_each_populated_zone(zone
) {
1214 struct per_cpu_pageset
*pset
;
1215 struct per_cpu_pages
*pcp
;
1217 local_irq_save(flags
);
1218 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1222 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1225 local_irq_restore(flags
);
1230 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1232 void drain_local_pages(void *arg
)
1234 drain_pages(smp_processor_id());
1238 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1240 * Note that this code is protected against sending an IPI to an offline
1241 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1242 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1243 * nothing keeps CPUs from showing up after we populated the cpumask and
1244 * before the call to on_each_cpu_mask().
1246 void drain_all_pages(void)
1249 struct per_cpu_pageset
*pcp
;
1253 * Allocate in the BSS so we wont require allocation in
1254 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1256 static cpumask_t cpus_with_pcps
;
1259 * We don't care about racing with CPU hotplug event
1260 * as offline notification will cause the notified
1261 * cpu to drain that CPU pcps and on_each_cpu_mask
1262 * disables preemption as part of its processing
1264 for_each_online_cpu(cpu
) {
1265 bool has_pcps
= false;
1266 for_each_populated_zone(zone
) {
1267 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1268 if (pcp
->pcp
.count
) {
1274 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1276 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1278 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1281 #ifdef CONFIG_HIBERNATION
1283 void mark_free_pages(struct zone
*zone
)
1285 unsigned long pfn
, max_zone_pfn
;
1286 unsigned long flags
;
1288 struct list_head
*curr
;
1290 if (!zone
->spanned_pages
)
1293 spin_lock_irqsave(&zone
->lock
, flags
);
1295 max_zone_pfn
= zone_end_pfn(zone
);
1296 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1297 if (pfn_valid(pfn
)) {
1298 struct page
*page
= pfn_to_page(pfn
);
1300 if (!swsusp_page_is_forbidden(page
))
1301 swsusp_unset_page_free(page
);
1304 for_each_migratetype_order(order
, t
) {
1305 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1308 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1309 for (i
= 0; i
< (1UL << order
); i
++)
1310 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1313 spin_unlock_irqrestore(&zone
->lock
, flags
);
1315 #endif /* CONFIG_PM */
1318 * Free a 0-order page
1319 * cold == 1 ? free a cold page : free a hot page
1321 void free_hot_cold_page(struct page
*page
, int cold
)
1323 struct zone
*zone
= page_zone(page
);
1324 struct per_cpu_pages
*pcp
;
1325 unsigned long flags
;
1328 if (!free_pages_prepare(page
, 0))
1331 migratetype
= get_pageblock_migratetype(page
);
1332 set_freepage_migratetype(page
, migratetype
);
1333 local_irq_save(flags
);
1334 __count_vm_event(PGFREE
);
1337 * We only track unmovable, reclaimable and movable on pcp lists.
1338 * Free ISOLATE pages back to the allocator because they are being
1339 * offlined but treat RESERVE as movable pages so we can get those
1340 * areas back if necessary. Otherwise, we may have to free
1341 * excessively into the page allocator
1343 if (migratetype
>= MIGRATE_PCPTYPES
) {
1344 if (unlikely(is_migrate_isolate(migratetype
))) {
1345 free_one_page(zone
, page
, 0, migratetype
);
1348 migratetype
= MIGRATE_MOVABLE
;
1351 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1353 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1355 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1357 if (pcp
->count
>= pcp
->high
) {
1358 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1359 free_pcppages_bulk(zone
, batch
, pcp
);
1360 pcp
->count
-= 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
);
1407 EXPORT_SYMBOL_GPL(split_page
);
1409 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1411 unsigned long watermark
;
1415 BUG_ON(!PageBuddy(page
));
1417 zone
= page_zone(page
);
1418 mt
= get_pageblock_migratetype(page
);
1420 if (!is_migrate_isolate(mt
)) {
1421 /* Obey watermarks as if the page was being allocated */
1422 watermark
= low_wmark_pages(zone
) + (1 << order
);
1423 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1426 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1429 /* Remove page from free list */
1430 list_del(&page
->lru
);
1431 zone
->free_area
[order
].nr_free
--;
1432 rmv_page_order(page
);
1434 /* Set the pageblock if the isolated page is at least a pageblock */
1435 if (order
>= pageblock_order
- 1) {
1436 struct page
*endpage
= page
+ (1 << order
) - 1;
1437 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1438 int mt
= get_pageblock_migratetype(page
);
1439 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1440 set_pageblock_migratetype(page
,
1445 return 1UL << order
;
1449 * Similar to split_page except the page is already free. As this is only
1450 * being used for migration, the migratetype of the block also changes.
1451 * As this is called with interrupts disabled, the caller is responsible
1452 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1455 * Note: this is probably too low level an operation for use in drivers.
1456 * Please consult with lkml before using this in your driver.
1458 int split_free_page(struct page
*page
)
1463 order
= page_order(page
);
1465 nr_pages
= __isolate_free_page(page
, order
);
1469 /* Split into individual pages */
1470 set_page_refcounted(page
);
1471 split_page(page
, order
);
1476 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1477 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1481 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1482 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1485 unsigned long flags
;
1487 int cold
= !!(gfp_flags
& __GFP_COLD
);
1490 if (likely(order
== 0)) {
1491 struct per_cpu_pages
*pcp
;
1492 struct list_head
*list
;
1494 local_irq_save(flags
);
1495 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1496 list
= &pcp
->lists
[migratetype
];
1497 if (list_empty(list
)) {
1498 pcp
->count
+= rmqueue_bulk(zone
, 0,
1501 if (unlikely(list_empty(list
)))
1506 page
= list_entry(list
->prev
, struct page
, lru
);
1508 page
= list_entry(list
->next
, struct page
, lru
);
1510 list_del(&page
->lru
);
1513 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1515 * __GFP_NOFAIL is not to be used in new code.
1517 * All __GFP_NOFAIL callers should be fixed so that they
1518 * properly detect and handle allocation failures.
1520 * We most definitely don't want callers attempting to
1521 * allocate greater than order-1 page units with
1524 WARN_ON_ONCE(order
> 1);
1526 spin_lock_irqsave(&zone
->lock
, flags
);
1527 page
= __rmqueue(zone
, order
, migratetype
);
1528 spin_unlock(&zone
->lock
);
1531 __mod_zone_freepage_state(zone
, -(1 << order
),
1532 get_pageblock_migratetype(page
));
1535 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1536 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1537 local_irq_restore(flags
);
1539 VM_BUG_ON(bad_range(zone
, page
));
1540 if (prep_new_page(page
, order
, gfp_flags
))
1545 local_irq_restore(flags
);
1549 #ifdef CONFIG_FAIL_PAGE_ALLOC
1552 struct fault_attr attr
;
1554 u32 ignore_gfp_highmem
;
1555 u32 ignore_gfp_wait
;
1557 } fail_page_alloc
= {
1558 .attr
= FAULT_ATTR_INITIALIZER
,
1559 .ignore_gfp_wait
= 1,
1560 .ignore_gfp_highmem
= 1,
1564 static int __init
setup_fail_page_alloc(char *str
)
1566 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1568 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1570 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1572 if (order
< fail_page_alloc
.min_order
)
1574 if (gfp_mask
& __GFP_NOFAIL
)
1576 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1578 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1581 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1584 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1586 static int __init
fail_page_alloc_debugfs(void)
1588 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1591 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1592 &fail_page_alloc
.attr
);
1594 return PTR_ERR(dir
);
1596 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1597 &fail_page_alloc
.ignore_gfp_wait
))
1599 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1600 &fail_page_alloc
.ignore_gfp_highmem
))
1602 if (!debugfs_create_u32("min-order", mode
, dir
,
1603 &fail_page_alloc
.min_order
))
1608 debugfs_remove_recursive(dir
);
1613 late_initcall(fail_page_alloc_debugfs
);
1615 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1617 #else /* CONFIG_FAIL_PAGE_ALLOC */
1619 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1624 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1627 * Return true if free pages are above 'mark'. This takes into account the order
1628 * of the allocation.
1630 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1631 int classzone_idx
, int alloc_flags
, long free_pages
)
1633 /* free_pages my go negative - that's OK */
1635 long lowmem_reserve
= z
->lowmem_reserve
[classzone_idx
];
1639 free_pages
-= (1 << order
) - 1;
1640 if (alloc_flags
& ALLOC_HIGH
)
1642 if (alloc_flags
& ALLOC_HARDER
)
1645 /* If allocation can't use CMA areas don't use free CMA pages */
1646 if (!(alloc_flags
& ALLOC_CMA
))
1647 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1650 if (free_pages
- free_cma
<= min
+ lowmem_reserve
)
1652 for (o
= 0; o
< order
; o
++) {
1653 /* At the next order, this order's pages become unavailable */
1654 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1656 /* Require fewer higher order pages to be free */
1659 if (free_pages
<= min
)
1665 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1666 int classzone_idx
, int alloc_flags
)
1668 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1669 zone_page_state(z
, NR_FREE_PAGES
));
1672 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1673 int classzone_idx
, int alloc_flags
)
1675 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1677 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1678 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1680 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1686 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1687 * skip over zones that are not allowed by the cpuset, or that have
1688 * been recently (in last second) found to be nearly full. See further
1689 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1690 * that have to skip over a lot of full or unallowed zones.
1692 * If the zonelist cache is present in the passed in zonelist, then
1693 * returns a pointer to the allowed node mask (either the current
1694 * tasks mems_allowed, or node_states[N_MEMORY].)
1696 * If the zonelist cache is not available for this zonelist, does
1697 * nothing and returns NULL.
1699 * If the fullzones BITMAP in the zonelist cache is stale (more than
1700 * a second since last zap'd) then we zap it out (clear its bits.)
1702 * We hold off even calling zlc_setup, until after we've checked the
1703 * first zone in the zonelist, on the theory that most allocations will
1704 * be satisfied from that first zone, so best to examine that zone as
1705 * quickly as we can.
1707 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1709 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1710 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1712 zlc
= zonelist
->zlcache_ptr
;
1716 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1717 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1718 zlc
->last_full_zap
= jiffies
;
1721 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1722 &cpuset_current_mems_allowed
:
1723 &node_states
[N_MEMORY
];
1724 return allowednodes
;
1728 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1729 * if it is worth looking at further for free memory:
1730 * 1) Check that the zone isn't thought to be full (doesn't have its
1731 * bit set in the zonelist_cache fullzones BITMAP).
1732 * 2) Check that the zones node (obtained from the zonelist_cache
1733 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1734 * Return true (non-zero) if zone is worth looking at further, or
1735 * else return false (zero) if it is not.
1737 * This check -ignores- the distinction between various watermarks,
1738 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1739 * found to be full for any variation of these watermarks, it will
1740 * be considered full for up to one second by all requests, unless
1741 * we are so low on memory on all allowed nodes that we are forced
1742 * into the second scan of the zonelist.
1744 * In the second scan we ignore this zonelist cache and exactly
1745 * apply the watermarks to all zones, even it is slower to do so.
1746 * We are low on memory in the second scan, and should leave no stone
1747 * unturned looking for a free page.
1749 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1750 nodemask_t
*allowednodes
)
1752 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1753 int i
; /* index of *z in zonelist zones */
1754 int n
; /* node that zone *z is on */
1756 zlc
= zonelist
->zlcache_ptr
;
1760 i
= z
- zonelist
->_zonerefs
;
1763 /* This zone is worth trying if it is allowed but not full */
1764 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1768 * Given 'z' scanning a zonelist, set the corresponding bit in
1769 * zlc->fullzones, so that subsequent attempts to allocate a page
1770 * from that zone don't waste time re-examining it.
1772 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1774 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1775 int i
; /* index of *z in zonelist zones */
1777 zlc
= zonelist
->zlcache_ptr
;
1781 i
= z
- zonelist
->_zonerefs
;
1783 set_bit(i
, zlc
->fullzones
);
1787 * clear all zones full, called after direct reclaim makes progress so that
1788 * a zone that was recently full is not skipped over for up to a second
1790 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1792 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1794 zlc
= zonelist
->zlcache_ptr
;
1798 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1801 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1803 return node_isset(local_zone
->node
, zone
->zone_pgdat
->reclaim_nodes
);
1806 static void __paginginit
init_zone_allows_reclaim(int nid
)
1810 for_each_online_node(i
)
1811 if (node_distance(nid
, i
) <= RECLAIM_DISTANCE
)
1812 node_set(i
, NODE_DATA(nid
)->reclaim_nodes
);
1814 zone_reclaim_mode
= 1;
1817 #else /* CONFIG_NUMA */
1819 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1824 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1825 nodemask_t
*allowednodes
)
1830 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1834 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1838 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1843 static inline void init_zone_allows_reclaim(int nid
)
1846 #endif /* CONFIG_NUMA */
1849 * get_page_from_freelist goes through the zonelist trying to allocate
1852 static struct page
*
1853 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1854 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1855 struct zone
*preferred_zone
, int migratetype
)
1858 struct page
*page
= NULL
;
1861 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1862 int zlc_active
= 0; /* set if using zonelist_cache */
1863 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1865 classzone_idx
= zone_idx(preferred_zone
);
1868 * Scan zonelist, looking for a zone with enough free.
1869 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1871 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1872 high_zoneidx
, nodemask
) {
1873 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1874 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1876 if ((alloc_flags
& ALLOC_CPUSET
) &&
1877 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1880 * When allocating a page cache page for writing, we
1881 * want to get it from a zone that is within its dirty
1882 * limit, such that no single zone holds more than its
1883 * proportional share of globally allowed dirty pages.
1884 * The dirty limits take into account the zone's
1885 * lowmem reserves and high watermark so that kswapd
1886 * should be able to balance it without having to
1887 * write pages from its LRU list.
1889 * This may look like it could increase pressure on
1890 * lower zones by failing allocations in higher zones
1891 * before they are full. But the pages that do spill
1892 * over are limited as the lower zones are protected
1893 * by this very same mechanism. It should not become
1894 * a practical burden to them.
1896 * XXX: For now, allow allocations to potentially
1897 * exceed the per-zone dirty limit in the slowpath
1898 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1899 * which is important when on a NUMA setup the allowed
1900 * zones are together not big enough to reach the
1901 * global limit. The proper fix for these situations
1902 * will require awareness of zones in the
1903 * dirty-throttling and the flusher threads.
1905 if ((alloc_flags
& ALLOC_WMARK_LOW
) &&
1906 (gfp_mask
& __GFP_WRITE
) && !zone_dirty_ok(zone
))
1907 goto this_zone_full
;
1909 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1910 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1914 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1915 if (zone_watermark_ok(zone
, order
, mark
,
1916 classzone_idx
, alloc_flags
))
1919 if (IS_ENABLED(CONFIG_NUMA
) &&
1920 !did_zlc_setup
&& nr_online_nodes
> 1) {
1922 * we do zlc_setup if there are multiple nodes
1923 * and before considering the first zone allowed
1926 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1931 if (zone_reclaim_mode
== 0 ||
1932 !zone_allows_reclaim(preferred_zone
, zone
))
1933 goto this_zone_full
;
1936 * As we may have just activated ZLC, check if the first
1937 * eligible zone has failed zone_reclaim recently.
1939 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1940 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1943 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1945 case ZONE_RECLAIM_NOSCAN
:
1948 case ZONE_RECLAIM_FULL
:
1949 /* scanned but unreclaimable */
1952 /* did we reclaim enough */
1953 if (zone_watermark_ok(zone
, order
, mark
,
1954 classzone_idx
, alloc_flags
))
1958 * Failed to reclaim enough to meet watermark.
1959 * Only mark the zone full if checking the min
1960 * watermark or if we failed to reclaim just
1961 * 1<<order pages or else the page allocator
1962 * fastpath will prematurely mark zones full
1963 * when the watermark is between the low and
1966 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
1967 ret
== ZONE_RECLAIM_SOME
)
1968 goto this_zone_full
;
1975 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1976 gfp_mask
, migratetype
);
1980 if (IS_ENABLED(CONFIG_NUMA
))
1981 zlc_mark_zone_full(zonelist
, z
);
1984 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && page
== NULL
&& zlc_active
)) {
1985 /* Disable zlc cache for second zonelist scan */
1992 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1993 * necessary to allocate the page. The expectation is
1994 * that the caller is taking steps that will free more
1995 * memory. The caller should avoid the page being used
1996 * for !PFMEMALLOC purposes.
1998 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2004 * Large machines with many possible nodes should not always dump per-node
2005 * meminfo in irq context.
2007 static inline bool should_suppress_show_mem(void)
2012 ret
= in_interrupt();
2017 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2018 DEFAULT_RATELIMIT_INTERVAL
,
2019 DEFAULT_RATELIMIT_BURST
);
2021 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2023 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2025 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2026 debug_guardpage_minorder() > 0)
2030 * Walking all memory to count page types is very expensive and should
2031 * be inhibited in non-blockable contexts.
2033 if (!(gfp_mask
& __GFP_WAIT
))
2034 filter
|= SHOW_MEM_FILTER_PAGE_COUNT
;
2037 * This documents exceptions given to allocations in certain
2038 * contexts that are allowed to allocate outside current's set
2041 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2042 if (test_thread_flag(TIF_MEMDIE
) ||
2043 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2044 filter
&= ~SHOW_MEM_FILTER_NODES
;
2045 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2046 filter
&= ~SHOW_MEM_FILTER_NODES
;
2049 struct va_format vaf
;
2052 va_start(args
, fmt
);
2057 pr_warn("%pV", &vaf
);
2062 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2063 current
->comm
, order
, gfp_mask
);
2066 if (!should_suppress_show_mem())
2071 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2072 unsigned long did_some_progress
,
2073 unsigned long pages_reclaimed
)
2075 /* Do not loop if specifically requested */
2076 if (gfp_mask
& __GFP_NORETRY
)
2079 /* Always retry if specifically requested */
2080 if (gfp_mask
& __GFP_NOFAIL
)
2084 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2085 * making forward progress without invoking OOM. Suspend also disables
2086 * storage devices so kswapd will not help. Bail if we are suspending.
2088 if (!did_some_progress
&& pm_suspended_storage())
2092 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2093 * means __GFP_NOFAIL, but that may not be true in other
2096 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2100 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2101 * specified, then we retry until we no longer reclaim any pages
2102 * (above), or we've reclaimed an order of pages at least as
2103 * large as the allocation's order. In both cases, if the
2104 * allocation still fails, we stop retrying.
2106 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2112 static inline struct page
*
2113 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2114 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2115 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2120 /* Acquire the OOM killer lock for the zones in zonelist */
2121 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
2122 schedule_timeout_uninterruptible(1);
2127 * Go through the zonelist yet one more time, keep very high watermark
2128 * here, this is only to catch a parallel oom killing, we must fail if
2129 * we're still under heavy pressure.
2131 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2132 order
, zonelist
, high_zoneidx
,
2133 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2134 preferred_zone
, migratetype
);
2138 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2139 /* The OOM killer will not help higher order allocs */
2140 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2142 /* The OOM killer does not needlessly kill tasks for lowmem */
2143 if (high_zoneidx
< ZONE_NORMAL
)
2146 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2147 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2148 * The caller should handle page allocation failure by itself if
2149 * it specifies __GFP_THISNODE.
2150 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2152 if (gfp_mask
& __GFP_THISNODE
)
2155 /* Exhausted what can be done so it's blamo time */
2156 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2159 clear_zonelist_oom(zonelist
, gfp_mask
);
2163 #ifdef CONFIG_COMPACTION
2164 /* Try memory compaction for high-order allocations before reclaim */
2165 static struct page
*
2166 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2167 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2168 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2169 int migratetype
, bool sync_migration
,
2170 bool *contended_compaction
, bool *deferred_compaction
,
2171 unsigned long *did_some_progress
)
2176 if (compaction_deferred(preferred_zone
, order
)) {
2177 *deferred_compaction
= true;
2181 current
->flags
|= PF_MEMALLOC
;
2182 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2183 nodemask
, sync_migration
,
2184 contended_compaction
);
2185 current
->flags
&= ~PF_MEMALLOC
;
2187 if (*did_some_progress
!= COMPACT_SKIPPED
) {
2190 /* Page migration frees to the PCP lists but we want merging */
2191 drain_pages(get_cpu());
2194 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2195 order
, zonelist
, high_zoneidx
,
2196 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2197 preferred_zone
, migratetype
);
2199 preferred_zone
->compact_blockskip_flush
= false;
2200 preferred_zone
->compact_considered
= 0;
2201 preferred_zone
->compact_defer_shift
= 0;
2202 if (order
>= preferred_zone
->compact_order_failed
)
2203 preferred_zone
->compact_order_failed
= order
+ 1;
2204 count_vm_event(COMPACTSUCCESS
);
2209 * It's bad if compaction run occurs and fails.
2210 * The most likely reason is that pages exist,
2211 * but not enough to satisfy watermarks.
2213 count_vm_event(COMPACTFAIL
);
2216 * As async compaction considers a subset of pageblocks, only
2217 * defer if the failure was a sync compaction failure.
2220 defer_compaction(preferred_zone
, order
);
2228 static inline struct page
*
2229 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2230 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2231 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2232 int migratetype
, bool sync_migration
,
2233 bool *contended_compaction
, bool *deferred_compaction
,
2234 unsigned long *did_some_progress
)
2238 #endif /* CONFIG_COMPACTION */
2240 /* Perform direct synchronous page reclaim */
2242 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2243 nodemask_t
*nodemask
)
2245 struct reclaim_state reclaim_state
;
2250 /* We now go into synchronous reclaim */
2251 cpuset_memory_pressure_bump();
2252 current
->flags
|= PF_MEMALLOC
;
2253 lockdep_set_current_reclaim_state(gfp_mask
);
2254 reclaim_state
.reclaimed_slab
= 0;
2255 current
->reclaim_state
= &reclaim_state
;
2257 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2259 current
->reclaim_state
= NULL
;
2260 lockdep_clear_current_reclaim_state();
2261 current
->flags
&= ~PF_MEMALLOC
;
2268 /* The really slow allocator path where we enter direct reclaim */
2269 static inline struct page
*
2270 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2271 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2272 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2273 int migratetype
, unsigned long *did_some_progress
)
2275 struct page
*page
= NULL
;
2276 bool drained
= false;
2278 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2280 if (unlikely(!(*did_some_progress
)))
2283 /* After successful reclaim, reconsider all zones for allocation */
2284 if (IS_ENABLED(CONFIG_NUMA
))
2285 zlc_clear_zones_full(zonelist
);
2288 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2289 zonelist
, high_zoneidx
,
2290 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2291 preferred_zone
, migratetype
);
2294 * If an allocation failed after direct reclaim, it could be because
2295 * pages are pinned on the per-cpu lists. Drain them and try again
2297 if (!page
&& !drained
) {
2307 * This is called in the allocator slow-path if the allocation request is of
2308 * sufficient urgency to ignore watermarks and take other desperate measures
2310 static inline struct page
*
2311 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2312 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2313 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2319 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2320 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2321 preferred_zone
, migratetype
);
2323 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2324 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2325 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2331 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
2332 enum zone_type high_zoneidx
,
2333 enum zone_type classzone_idx
)
2338 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
2339 wakeup_kswapd(zone
, order
, classzone_idx
);
2343 gfp_to_alloc_flags(gfp_t gfp_mask
)
2345 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2346 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2348 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2349 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2352 * The caller may dip into page reserves a bit more if the caller
2353 * cannot run direct reclaim, or if the caller has realtime scheduling
2354 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2355 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2357 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2361 * Not worth trying to allocate harder for
2362 * __GFP_NOMEMALLOC even if it can't schedule.
2364 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2365 alloc_flags
|= ALLOC_HARDER
;
2367 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2368 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2370 alloc_flags
&= ~ALLOC_CPUSET
;
2371 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2372 alloc_flags
|= ALLOC_HARDER
;
2374 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2375 if (gfp_mask
& __GFP_MEMALLOC
)
2376 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2377 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2378 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2379 else if (!in_interrupt() &&
2380 ((current
->flags
& PF_MEMALLOC
) ||
2381 unlikely(test_thread_flag(TIF_MEMDIE
))))
2382 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2385 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2386 alloc_flags
|= ALLOC_CMA
;
2391 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2393 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2396 static inline struct page
*
2397 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2398 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2399 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2402 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2403 struct page
*page
= NULL
;
2405 unsigned long pages_reclaimed
= 0;
2406 unsigned long did_some_progress
;
2407 bool sync_migration
= false;
2408 bool deferred_compaction
= false;
2409 bool contended_compaction
= false;
2412 * In the slowpath, we sanity check order to avoid ever trying to
2413 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2414 * be using allocators in order of preference for an area that is
2417 if (order
>= MAX_ORDER
) {
2418 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2423 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2424 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2425 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2426 * using a larger set of nodes after it has established that the
2427 * allowed per node queues are empty and that nodes are
2430 if (IS_ENABLED(CONFIG_NUMA
) &&
2431 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2435 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2436 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2437 zone_idx(preferred_zone
));
2440 * OK, we're below the kswapd watermark and have kicked background
2441 * reclaim. Now things get more complex, so set up alloc_flags according
2442 * to how we want to proceed.
2444 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2447 * Find the true preferred zone if the allocation is unconstrained by
2450 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2451 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2455 /* This is the last chance, in general, before the goto nopage. */
2456 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2457 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2458 preferred_zone
, migratetype
);
2462 /* Allocate without watermarks if the context allows */
2463 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2465 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2466 * the allocation is high priority and these type of
2467 * allocations are system rather than user orientated
2469 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2471 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2472 zonelist
, high_zoneidx
, nodemask
,
2473 preferred_zone
, migratetype
);
2479 /* Atomic allocations - we can't balance anything */
2483 /* Avoid recursion of direct reclaim */
2484 if (current
->flags
& PF_MEMALLOC
)
2487 /* Avoid allocations with no watermarks from looping endlessly */
2488 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2492 * Try direct compaction. The first pass is asynchronous. Subsequent
2493 * attempts after direct reclaim are synchronous
2495 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2496 zonelist
, high_zoneidx
,
2498 alloc_flags
, preferred_zone
,
2499 migratetype
, sync_migration
,
2500 &contended_compaction
,
2501 &deferred_compaction
,
2502 &did_some_progress
);
2505 sync_migration
= true;
2508 * If compaction is deferred for high-order allocations, it is because
2509 * sync compaction recently failed. In this is the case and the caller
2510 * requested a movable allocation that does not heavily disrupt the
2511 * system then fail the allocation instead of entering direct reclaim.
2513 if ((deferred_compaction
|| contended_compaction
) &&
2514 (gfp_mask
& __GFP_NO_KSWAPD
))
2517 /* Try direct reclaim and then allocating */
2518 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2519 zonelist
, high_zoneidx
,
2521 alloc_flags
, preferred_zone
,
2522 migratetype
, &did_some_progress
);
2527 * If we failed to make any progress reclaiming, then we are
2528 * running out of options and have to consider going OOM
2530 if (!did_some_progress
) {
2531 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2532 if (oom_killer_disabled
)
2534 /* Coredumps can quickly deplete all memory reserves */
2535 if ((current
->flags
& PF_DUMPCORE
) &&
2536 !(gfp_mask
& __GFP_NOFAIL
))
2538 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2539 zonelist
, high_zoneidx
,
2540 nodemask
, preferred_zone
,
2545 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2547 * The oom killer is not called for high-order
2548 * allocations that may fail, so if no progress
2549 * is being made, there are no other options and
2550 * retrying is unlikely to help.
2552 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2555 * The oom killer is not called for lowmem
2556 * allocations to prevent needlessly killing
2559 if (high_zoneidx
< ZONE_NORMAL
)
2567 /* Check if we should retry the allocation */
2568 pages_reclaimed
+= did_some_progress
;
2569 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2571 /* Wait for some write requests to complete then retry */
2572 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2576 * High-order allocations do not necessarily loop after
2577 * direct reclaim and reclaim/compaction depends on compaction
2578 * being called after reclaim so call directly if necessary
2580 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2581 zonelist
, high_zoneidx
,
2583 alloc_flags
, preferred_zone
,
2584 migratetype
, sync_migration
,
2585 &contended_compaction
,
2586 &deferred_compaction
,
2587 &did_some_progress
);
2593 warn_alloc_failed(gfp_mask
, order
, NULL
);
2596 if (kmemcheck_enabled
)
2597 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2603 * This is the 'heart' of the zoned buddy allocator.
2606 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2607 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2609 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2610 struct zone
*preferred_zone
;
2611 struct page
*page
= NULL
;
2612 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2613 unsigned int cpuset_mems_cookie
;
2614 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
;
2615 struct mem_cgroup
*memcg
= NULL
;
2617 gfp_mask
&= gfp_allowed_mask
;
2619 lockdep_trace_alloc(gfp_mask
);
2621 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2623 if (should_fail_alloc_page(gfp_mask
, order
))
2627 * Check the zones suitable for the gfp_mask contain at least one
2628 * valid zone. It's possible to have an empty zonelist as a result
2629 * of GFP_THISNODE and a memoryless node
2631 if (unlikely(!zonelist
->_zonerefs
->zone
))
2635 * Will only have any effect when __GFP_KMEMCG is set. This is
2636 * verified in the (always inline) callee
2638 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2642 cpuset_mems_cookie
= get_mems_allowed();
2644 /* The preferred zone is used for statistics later */
2645 first_zones_zonelist(zonelist
, high_zoneidx
,
2646 nodemask
? : &cpuset_current_mems_allowed
,
2648 if (!preferred_zone
)
2652 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2653 alloc_flags
|= ALLOC_CMA
;
2655 /* First allocation attempt */
2656 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2657 zonelist
, high_zoneidx
, alloc_flags
,
2658 preferred_zone
, migratetype
);
2659 if (unlikely(!page
)) {
2661 * Runtime PM, block IO and its error handling path
2662 * can deadlock because I/O on the device might not
2665 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2666 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2667 zonelist
, high_zoneidx
, nodemask
,
2668 preferred_zone
, migratetype
);
2671 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2675 * When updating a task's mems_allowed, it is possible to race with
2676 * parallel threads in such a way that an allocation can fail while
2677 * the mask is being updated. If a page allocation is about to fail,
2678 * check if the cpuset changed during allocation and if so, retry.
2680 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2683 memcg_kmem_commit_charge(page
, memcg
, order
);
2687 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2690 * Common helper functions.
2692 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2697 * __get_free_pages() returns a 32-bit address, which cannot represent
2700 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2702 page
= alloc_pages(gfp_mask
, order
);
2705 return (unsigned long) page_address(page
);
2707 EXPORT_SYMBOL(__get_free_pages
);
2709 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2711 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2713 EXPORT_SYMBOL(get_zeroed_page
);
2715 void __free_pages(struct page
*page
, unsigned int order
)
2717 if (put_page_testzero(page
)) {
2719 free_hot_cold_page(page
, 0);
2721 __free_pages_ok(page
, order
);
2725 EXPORT_SYMBOL(__free_pages
);
2727 void free_pages(unsigned long addr
, unsigned int order
)
2730 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2731 __free_pages(virt_to_page((void *)addr
), order
);
2735 EXPORT_SYMBOL(free_pages
);
2738 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2739 * pages allocated with __GFP_KMEMCG.
2741 * Those pages are accounted to a particular memcg, embedded in the
2742 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2743 * for that information only to find out that it is NULL for users who have no
2744 * interest in that whatsoever, we provide these functions.
2746 * The caller knows better which flags it relies on.
2748 void __free_memcg_kmem_pages(struct page
*page
, unsigned int order
)
2750 memcg_kmem_uncharge_pages(page
, order
);
2751 __free_pages(page
, order
);
2754 void free_memcg_kmem_pages(unsigned long addr
, unsigned int order
)
2757 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2758 __free_memcg_kmem_pages(virt_to_page((void *)addr
), order
);
2762 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2765 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2766 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2768 split_page(virt_to_page((void *)addr
), order
);
2769 while (used
< alloc_end
) {
2774 return (void *)addr
;
2778 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2779 * @size: the number of bytes to allocate
2780 * @gfp_mask: GFP flags for the allocation
2782 * This function is similar to alloc_pages(), except that it allocates the
2783 * minimum number of pages to satisfy the request. alloc_pages() can only
2784 * allocate memory in power-of-two pages.
2786 * This function is also limited by MAX_ORDER.
2788 * Memory allocated by this function must be released by free_pages_exact().
2790 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2792 unsigned int order
= get_order(size
);
2795 addr
= __get_free_pages(gfp_mask
, order
);
2796 return make_alloc_exact(addr
, order
, size
);
2798 EXPORT_SYMBOL(alloc_pages_exact
);
2801 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2803 * @nid: the preferred node ID where memory should be allocated
2804 * @size: the number of bytes to allocate
2805 * @gfp_mask: GFP flags for the allocation
2807 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2809 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2812 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2814 unsigned order
= get_order(size
);
2815 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2818 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2820 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2823 * free_pages_exact - release memory allocated via alloc_pages_exact()
2824 * @virt: the value returned by alloc_pages_exact.
2825 * @size: size of allocation, same value as passed to alloc_pages_exact().
2827 * Release the memory allocated by a previous call to alloc_pages_exact.
2829 void free_pages_exact(void *virt
, size_t size
)
2831 unsigned long addr
= (unsigned long)virt
;
2832 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2834 while (addr
< end
) {
2839 EXPORT_SYMBOL(free_pages_exact
);
2842 * nr_free_zone_pages - count number of pages beyond high watermark
2843 * @offset: The zone index of the highest zone
2845 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2846 * high watermark within all zones at or below a given zone index. For each
2847 * zone, the number of pages is calculated as:
2848 * present_pages - high_pages
2850 static unsigned long nr_free_zone_pages(int offset
)
2855 /* Just pick one node, since fallback list is circular */
2856 unsigned long sum
= 0;
2858 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2860 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2861 unsigned long size
= zone
->managed_pages
;
2862 unsigned long high
= high_wmark_pages(zone
);
2871 * nr_free_buffer_pages - count number of pages beyond high watermark
2873 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2874 * watermark within ZONE_DMA and ZONE_NORMAL.
2876 unsigned long nr_free_buffer_pages(void)
2878 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2880 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2883 * nr_free_pagecache_pages - count number of pages beyond high watermark
2885 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2886 * high watermark within all zones.
2888 unsigned long nr_free_pagecache_pages(void)
2890 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2893 static inline void show_node(struct zone
*zone
)
2895 if (IS_ENABLED(CONFIG_NUMA
))
2896 printk("Node %d ", zone_to_nid(zone
));
2899 void si_meminfo(struct sysinfo
*val
)
2901 val
->totalram
= totalram_pages
;
2903 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2904 val
->bufferram
= nr_blockdev_pages();
2905 val
->totalhigh
= totalhigh_pages
;
2906 val
->freehigh
= nr_free_highpages();
2907 val
->mem_unit
= PAGE_SIZE
;
2910 EXPORT_SYMBOL(si_meminfo
);
2913 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2915 pg_data_t
*pgdat
= NODE_DATA(nid
);
2917 val
->totalram
= pgdat
->node_present_pages
;
2918 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2919 #ifdef CONFIG_HIGHMEM
2920 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
2921 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2927 val
->mem_unit
= PAGE_SIZE
;
2932 * Determine whether the node should be displayed or not, depending on whether
2933 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2935 bool skip_free_areas_node(unsigned int flags
, int nid
)
2938 unsigned int cpuset_mems_cookie
;
2940 if (!(flags
& SHOW_MEM_FILTER_NODES
))
2944 cpuset_mems_cookie
= get_mems_allowed();
2945 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
2946 } while (!put_mems_allowed(cpuset_mems_cookie
));
2951 #define K(x) ((x) << (PAGE_SHIFT-10))
2953 static void show_migration_types(unsigned char type
)
2955 static const char types
[MIGRATE_TYPES
] = {
2956 [MIGRATE_UNMOVABLE
] = 'U',
2957 [MIGRATE_RECLAIMABLE
] = 'E',
2958 [MIGRATE_MOVABLE
] = 'M',
2959 [MIGRATE_RESERVE
] = 'R',
2961 [MIGRATE_CMA
] = 'C',
2963 #ifdef CONFIG_MEMORY_ISOLATION
2964 [MIGRATE_ISOLATE
] = 'I',
2967 char tmp
[MIGRATE_TYPES
+ 1];
2971 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
2972 if (type
& (1 << i
))
2977 printk("(%s) ", tmp
);
2981 * Show free area list (used inside shift_scroll-lock stuff)
2982 * We also calculate the percentage fragmentation. We do this by counting the
2983 * memory on each free list with the exception of the first item on the list.
2984 * Suppresses nodes that are not allowed by current's cpuset if
2985 * SHOW_MEM_FILTER_NODES is passed.
2987 void show_free_areas(unsigned int filter
)
2992 for_each_populated_zone(zone
) {
2993 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2996 printk("%s per-cpu:\n", zone
->name
);
2998 for_each_online_cpu(cpu
) {
2999 struct per_cpu_pageset
*pageset
;
3001 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3003 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3004 cpu
, pageset
->pcp
.high
,
3005 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3009 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3010 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3012 " dirty:%lu writeback:%lu unstable:%lu\n"
3013 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3014 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3016 global_page_state(NR_ACTIVE_ANON
),
3017 global_page_state(NR_INACTIVE_ANON
),
3018 global_page_state(NR_ISOLATED_ANON
),
3019 global_page_state(NR_ACTIVE_FILE
),
3020 global_page_state(NR_INACTIVE_FILE
),
3021 global_page_state(NR_ISOLATED_FILE
),
3022 global_page_state(NR_UNEVICTABLE
),
3023 global_page_state(NR_FILE_DIRTY
),
3024 global_page_state(NR_WRITEBACK
),
3025 global_page_state(NR_UNSTABLE_NFS
),
3026 global_page_state(NR_FREE_PAGES
),
3027 global_page_state(NR_SLAB_RECLAIMABLE
),
3028 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3029 global_page_state(NR_FILE_MAPPED
),
3030 global_page_state(NR_SHMEM
),
3031 global_page_state(NR_PAGETABLE
),
3032 global_page_state(NR_BOUNCE
),
3033 global_page_state(NR_FREE_CMA_PAGES
));
3035 for_each_populated_zone(zone
) {
3038 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3046 " active_anon:%lukB"
3047 " inactive_anon:%lukB"
3048 " active_file:%lukB"
3049 " inactive_file:%lukB"
3050 " unevictable:%lukB"
3051 " isolated(anon):%lukB"
3052 " isolated(file):%lukB"
3060 " slab_reclaimable:%lukB"
3061 " slab_unreclaimable:%lukB"
3062 " kernel_stack:%lukB"
3067 " writeback_tmp:%lukB"
3068 " pages_scanned:%lu"
3069 " all_unreclaimable? %s"
3072 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3073 K(min_wmark_pages(zone
)),
3074 K(low_wmark_pages(zone
)),
3075 K(high_wmark_pages(zone
)),
3076 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3077 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3078 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3079 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3080 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3081 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3082 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3083 K(zone
->present_pages
),
3084 K(zone
->managed_pages
),
3085 K(zone_page_state(zone
, NR_MLOCK
)),
3086 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3087 K(zone_page_state(zone
, NR_WRITEBACK
)),
3088 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3089 K(zone_page_state(zone
, NR_SHMEM
)),
3090 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3091 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3092 zone_page_state(zone
, NR_KERNEL_STACK
) *
3094 K(zone_page_state(zone
, NR_PAGETABLE
)),
3095 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3096 K(zone_page_state(zone
, NR_BOUNCE
)),
3097 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3098 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3099 zone
->pages_scanned
,
3100 (zone
->all_unreclaimable
? "yes" : "no")
3102 printk("lowmem_reserve[]:");
3103 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3104 printk(" %lu", zone
->lowmem_reserve
[i
]);
3108 for_each_populated_zone(zone
) {
3109 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3110 unsigned char types
[MAX_ORDER
];
3112 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3115 printk("%s: ", zone
->name
);
3117 spin_lock_irqsave(&zone
->lock
, flags
);
3118 for (order
= 0; order
< MAX_ORDER
; order
++) {
3119 struct free_area
*area
= &zone
->free_area
[order
];
3122 nr
[order
] = area
->nr_free
;
3123 total
+= nr
[order
] << order
;
3126 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3127 if (!list_empty(&area
->free_list
[type
]))
3128 types
[order
] |= 1 << type
;
3131 spin_unlock_irqrestore(&zone
->lock
, flags
);
3132 for (order
= 0; order
< MAX_ORDER
; order
++) {
3133 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3135 show_migration_types(types
[order
]);
3137 printk("= %lukB\n", K(total
));
3140 hugetlb_show_meminfo();
3142 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3144 show_swap_cache_info();
3147 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3149 zoneref
->zone
= zone
;
3150 zoneref
->zone_idx
= zone_idx(zone
);
3154 * Builds allocation fallback zone lists.
3156 * Add all populated zones of a node to the zonelist.
3158 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3159 int nr_zones
, enum zone_type zone_type
)
3163 BUG_ON(zone_type
>= MAX_NR_ZONES
);
3168 zone
= pgdat
->node_zones
+ zone_type
;
3169 if (populated_zone(zone
)) {
3170 zoneref_set_zone(zone
,
3171 &zonelist
->_zonerefs
[nr_zones
++]);
3172 check_highest_zone(zone_type
);
3175 } while (zone_type
);
3182 * 0 = automatic detection of better ordering.
3183 * 1 = order by ([node] distance, -zonetype)
3184 * 2 = order by (-zonetype, [node] distance)
3186 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3187 * the same zonelist. So only NUMA can configure this param.
3189 #define ZONELIST_ORDER_DEFAULT 0
3190 #define ZONELIST_ORDER_NODE 1
3191 #define ZONELIST_ORDER_ZONE 2
3193 /* zonelist order in the kernel.
3194 * set_zonelist_order() will set this to NODE or ZONE.
3196 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3197 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3201 /* The value user specified ....changed by config */
3202 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3203 /* string for sysctl */
3204 #define NUMA_ZONELIST_ORDER_LEN 16
3205 char numa_zonelist_order
[16] = "default";
3208 * interface for configure zonelist ordering.
3209 * command line option "numa_zonelist_order"
3210 * = "[dD]efault - default, automatic configuration.
3211 * = "[nN]ode - order by node locality, then by zone within node
3212 * = "[zZ]one - order by zone, then by locality within zone
3215 static int __parse_numa_zonelist_order(char *s
)
3217 if (*s
== 'd' || *s
== 'D') {
3218 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3219 } else if (*s
== 'n' || *s
== 'N') {
3220 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3221 } else if (*s
== 'z' || *s
== 'Z') {
3222 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3225 "Ignoring invalid numa_zonelist_order value: "
3232 static __init
int setup_numa_zonelist_order(char *s
)
3239 ret
= __parse_numa_zonelist_order(s
);
3241 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3245 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3248 * sysctl handler for numa_zonelist_order
3250 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
3251 void __user
*buffer
, size_t *length
,
3254 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3256 static DEFINE_MUTEX(zl_order_mutex
);
3258 mutex_lock(&zl_order_mutex
);
3260 strcpy(saved_string
, (char*)table
->data
);
3261 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3265 int oldval
= user_zonelist_order
;
3266 if (__parse_numa_zonelist_order((char*)table
->data
)) {
3268 * bogus value. restore saved string
3270 strncpy((char*)table
->data
, saved_string
,
3271 NUMA_ZONELIST_ORDER_LEN
);
3272 user_zonelist_order
= oldval
;
3273 } else if (oldval
!= user_zonelist_order
) {
3274 mutex_lock(&zonelists_mutex
);
3275 build_all_zonelists(NULL
, NULL
);
3276 mutex_unlock(&zonelists_mutex
);
3280 mutex_unlock(&zl_order_mutex
);
3285 #define MAX_NODE_LOAD (nr_online_nodes)
3286 static int node_load
[MAX_NUMNODES
];
3289 * find_next_best_node - find the next node that should appear in a given node's fallback list
3290 * @node: node whose fallback list we're appending
3291 * @used_node_mask: nodemask_t of already used nodes
3293 * We use a number of factors to determine which is the next node that should
3294 * appear on a given node's fallback list. The node should not have appeared
3295 * already in @node's fallback list, and it should be the next closest node
3296 * according to the distance array (which contains arbitrary distance values
3297 * from each node to each node in the system), and should also prefer nodes
3298 * with no CPUs, since presumably they'll have very little allocation pressure
3299 * on them otherwise.
3300 * It returns -1 if no node is found.
3302 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3305 int min_val
= INT_MAX
;
3306 int best_node
= NUMA_NO_NODE
;
3307 const struct cpumask
*tmp
= cpumask_of_node(0);
3309 /* Use the local node if we haven't already */
3310 if (!node_isset(node
, *used_node_mask
)) {
3311 node_set(node
, *used_node_mask
);
3315 for_each_node_state(n
, N_MEMORY
) {
3317 /* Don't want a node to appear more than once */
3318 if (node_isset(n
, *used_node_mask
))
3321 /* Use the distance array to find the distance */
3322 val
= node_distance(node
, n
);
3324 /* Penalize nodes under us ("prefer the next node") */
3327 /* Give preference to headless and unused nodes */
3328 tmp
= cpumask_of_node(n
);
3329 if (!cpumask_empty(tmp
))
3330 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3332 /* Slight preference for less loaded node */
3333 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3334 val
+= node_load
[n
];
3336 if (val
< min_val
) {
3343 node_set(best_node
, *used_node_mask
);
3350 * Build zonelists ordered by node and zones within node.
3351 * This results in maximum locality--normal zone overflows into local
3352 * DMA zone, if any--but risks exhausting DMA zone.
3354 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3357 struct zonelist
*zonelist
;
3359 zonelist
= &pgdat
->node_zonelists
[0];
3360 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3362 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3364 zonelist
->_zonerefs
[j
].zone
= NULL
;
3365 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3369 * Build gfp_thisnode zonelists
3371 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3374 struct zonelist
*zonelist
;
3376 zonelist
= &pgdat
->node_zonelists
[1];
3377 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3378 zonelist
->_zonerefs
[j
].zone
= NULL
;
3379 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3383 * Build zonelists ordered by zone and nodes within zones.
3384 * This results in conserving DMA zone[s] until all Normal memory is
3385 * exhausted, but results in overflowing to remote node while memory
3386 * may still exist in local DMA zone.
3388 static int node_order
[MAX_NUMNODES
];
3390 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3393 int zone_type
; /* needs to be signed */
3395 struct zonelist
*zonelist
;
3397 zonelist
= &pgdat
->node_zonelists
[0];
3399 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3400 for (j
= 0; j
< nr_nodes
; j
++) {
3401 node
= node_order
[j
];
3402 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3403 if (populated_zone(z
)) {
3405 &zonelist
->_zonerefs
[pos
++]);
3406 check_highest_zone(zone_type
);
3410 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3411 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3414 static int default_zonelist_order(void)
3417 unsigned long low_kmem_size
,total_size
;
3421 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3422 * If they are really small and used heavily, the system can fall
3423 * into OOM very easily.
3424 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3426 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3429 for_each_online_node(nid
) {
3430 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3431 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3432 if (populated_zone(z
)) {
3433 if (zone_type
< ZONE_NORMAL
)
3434 low_kmem_size
+= z
->present_pages
;
3435 total_size
+= z
->present_pages
;
3436 } else if (zone_type
== ZONE_NORMAL
) {
3438 * If any node has only lowmem, then node order
3439 * is preferred to allow kernel allocations
3440 * locally; otherwise, they can easily infringe
3441 * on other nodes when there is an abundance of
3442 * lowmem available to allocate from.
3444 return ZONELIST_ORDER_NODE
;
3448 if (!low_kmem_size
|| /* there are no DMA area. */
3449 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
3450 return ZONELIST_ORDER_NODE
;
3452 * look into each node's config.
3453 * If there is a node whose DMA/DMA32 memory is very big area on
3454 * local memory, NODE_ORDER may be suitable.
3456 average_size
= total_size
/
3457 (nodes_weight(node_states
[N_MEMORY
]) + 1);
3458 for_each_online_node(nid
) {
3461 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3462 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3463 if (populated_zone(z
)) {
3464 if (zone_type
< ZONE_NORMAL
)
3465 low_kmem_size
+= z
->present_pages
;
3466 total_size
+= z
->present_pages
;
3469 if (low_kmem_size
&&
3470 total_size
> average_size
&& /* ignore small node */
3471 low_kmem_size
> total_size
* 70/100)
3472 return ZONELIST_ORDER_NODE
;
3474 return ZONELIST_ORDER_ZONE
;
3477 static void set_zonelist_order(void)
3479 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3480 current_zonelist_order
= default_zonelist_order();
3482 current_zonelist_order
= user_zonelist_order
;
3485 static void build_zonelists(pg_data_t
*pgdat
)
3489 nodemask_t used_mask
;
3490 int local_node
, prev_node
;
3491 struct zonelist
*zonelist
;
3492 int order
= current_zonelist_order
;
3494 /* initialize zonelists */
3495 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3496 zonelist
= pgdat
->node_zonelists
+ i
;
3497 zonelist
->_zonerefs
[0].zone
= NULL
;
3498 zonelist
->_zonerefs
[0].zone_idx
= 0;
3501 /* NUMA-aware ordering of nodes */
3502 local_node
= pgdat
->node_id
;
3503 load
= nr_online_nodes
;
3504 prev_node
= local_node
;
3505 nodes_clear(used_mask
);
3507 memset(node_order
, 0, sizeof(node_order
));
3510 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3512 * We don't want to pressure a particular node.
3513 * So adding penalty to the first node in same
3514 * distance group to make it round-robin.
3516 if (node_distance(local_node
, node
) !=
3517 node_distance(local_node
, prev_node
))
3518 node_load
[node
] = load
;
3522 if (order
== ZONELIST_ORDER_NODE
)
3523 build_zonelists_in_node_order(pgdat
, node
);
3525 node_order
[j
++] = node
; /* remember order */
3528 if (order
== ZONELIST_ORDER_ZONE
) {
3529 /* calculate node order -- i.e., DMA last! */
3530 build_zonelists_in_zone_order(pgdat
, j
);
3533 build_thisnode_zonelists(pgdat
);
3536 /* Construct the zonelist performance cache - see further mmzone.h */
3537 static void build_zonelist_cache(pg_data_t
*pgdat
)
3539 struct zonelist
*zonelist
;
3540 struct zonelist_cache
*zlc
;
3543 zonelist
= &pgdat
->node_zonelists
[0];
3544 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3545 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3546 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3547 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3550 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3552 * Return node id of node used for "local" allocations.
3553 * I.e., first node id of first zone in arg node's generic zonelist.
3554 * Used for initializing percpu 'numa_mem', which is used primarily
3555 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3557 int local_memory_node(int node
)
3561 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3562 gfp_zone(GFP_KERNEL
),
3569 #else /* CONFIG_NUMA */
3571 static void set_zonelist_order(void)
3573 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3576 static void build_zonelists(pg_data_t
*pgdat
)
3578 int node
, local_node
;
3580 struct zonelist
*zonelist
;
3582 local_node
= pgdat
->node_id
;
3584 zonelist
= &pgdat
->node_zonelists
[0];
3585 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3588 * Now we build the zonelist so that it contains the zones
3589 * of all the other nodes.
3590 * We don't want to pressure a particular node, so when
3591 * building the zones for node N, we make sure that the
3592 * zones coming right after the local ones are those from
3593 * node N+1 (modulo N)
3595 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3596 if (!node_online(node
))
3598 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3601 for (node
= 0; node
< local_node
; node
++) {
3602 if (!node_online(node
))
3604 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3608 zonelist
->_zonerefs
[j
].zone
= NULL
;
3609 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3612 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3613 static void build_zonelist_cache(pg_data_t
*pgdat
)
3615 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3618 #endif /* CONFIG_NUMA */
3621 * Boot pageset table. One per cpu which is going to be used for all
3622 * zones and all nodes. The parameters will be set in such a way
3623 * that an item put on a list will immediately be handed over to
3624 * the buddy list. This is safe since pageset manipulation is done
3625 * with interrupts disabled.
3627 * The boot_pagesets must be kept even after bootup is complete for
3628 * unused processors and/or zones. They do play a role for bootstrapping
3629 * hotplugged processors.
3631 * zoneinfo_show() and maybe other functions do
3632 * not check if the processor is online before following the pageset pointer.
3633 * Other parts of the kernel may not check if the zone is available.
3635 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3636 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3637 static void setup_zone_pageset(struct zone
*zone
);
3640 * Global mutex to protect against size modification of zonelists
3641 * as well as to serialize pageset setup for the new populated zone.
3643 DEFINE_MUTEX(zonelists_mutex
);
3645 /* return values int ....just for stop_machine() */
3646 static int __build_all_zonelists(void *data
)
3650 pg_data_t
*self
= data
;
3653 memset(node_load
, 0, sizeof(node_load
));
3656 if (self
&& !node_online(self
->node_id
)) {
3657 build_zonelists(self
);
3658 build_zonelist_cache(self
);
3661 for_each_online_node(nid
) {
3662 pg_data_t
*pgdat
= NODE_DATA(nid
);
3664 build_zonelists(pgdat
);
3665 build_zonelist_cache(pgdat
);
3669 * Initialize the boot_pagesets that are going to be used
3670 * for bootstrapping processors. The real pagesets for
3671 * each zone will be allocated later when the per cpu
3672 * allocator is available.
3674 * boot_pagesets are used also for bootstrapping offline
3675 * cpus if the system is already booted because the pagesets
3676 * are needed to initialize allocators on a specific cpu too.
3677 * F.e. the percpu allocator needs the page allocator which
3678 * needs the percpu allocator in order to allocate its pagesets
3679 * (a chicken-egg dilemma).
3681 for_each_possible_cpu(cpu
) {
3682 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3684 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3686 * We now know the "local memory node" for each node--
3687 * i.e., the node of the first zone in the generic zonelist.
3688 * Set up numa_mem percpu variable for on-line cpus. During
3689 * boot, only the boot cpu should be on-line; we'll init the
3690 * secondary cpus' numa_mem as they come on-line. During
3691 * node/memory hotplug, we'll fixup all on-line cpus.
3693 if (cpu_online(cpu
))
3694 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3702 * Called with zonelists_mutex held always
3703 * unless system_state == SYSTEM_BOOTING.
3705 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3707 set_zonelist_order();
3709 if (system_state
== SYSTEM_BOOTING
) {
3710 __build_all_zonelists(NULL
);
3711 mminit_verify_zonelist();
3712 cpuset_init_current_mems_allowed();
3714 /* we have to stop all cpus to guarantee there is no user
3716 #ifdef CONFIG_MEMORY_HOTPLUG
3718 setup_zone_pageset(zone
);
3720 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3721 /* cpuset refresh routine should be here */
3723 vm_total_pages
= nr_free_pagecache_pages();
3725 * Disable grouping by mobility if the number of pages in the
3726 * system is too low to allow the mechanism to work. It would be
3727 * more accurate, but expensive to check per-zone. This check is
3728 * made on memory-hotadd so a system can start with mobility
3729 * disabled and enable it later
3731 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3732 page_group_by_mobility_disabled
= 1;
3734 page_group_by_mobility_disabled
= 0;
3736 printk("Built %i zonelists in %s order, mobility grouping %s. "
3737 "Total pages: %ld\n",
3739 zonelist_order_name
[current_zonelist_order
],
3740 page_group_by_mobility_disabled
? "off" : "on",
3743 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3748 * Helper functions to size the waitqueue hash table.
3749 * Essentially these want to choose hash table sizes sufficiently
3750 * large so that collisions trying to wait on pages are rare.
3751 * But in fact, the number of active page waitqueues on typical
3752 * systems is ridiculously low, less than 200. So this is even
3753 * conservative, even though it seems large.
3755 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3756 * waitqueues, i.e. the size of the waitq table given the number of pages.
3758 #define PAGES_PER_WAITQUEUE 256
3760 #ifndef CONFIG_MEMORY_HOTPLUG
3761 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3763 unsigned long size
= 1;
3765 pages
/= PAGES_PER_WAITQUEUE
;
3767 while (size
< pages
)
3771 * Once we have dozens or even hundreds of threads sleeping
3772 * on IO we've got bigger problems than wait queue collision.
3773 * Limit the size of the wait table to a reasonable size.
3775 size
= min(size
, 4096UL);
3777 return max(size
, 4UL);
3781 * A zone's size might be changed by hot-add, so it is not possible to determine
3782 * a suitable size for its wait_table. So we use the maximum size now.
3784 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3786 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3787 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3788 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3790 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3791 * or more by the traditional way. (See above). It equals:
3793 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3794 * ia64(16K page size) : = ( 8G + 4M)byte.
3795 * powerpc (64K page size) : = (32G +16M)byte.
3797 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3804 * This is an integer logarithm so that shifts can be used later
3805 * to extract the more random high bits from the multiplicative
3806 * hash function before the remainder is taken.
3808 static inline unsigned long wait_table_bits(unsigned long size
)
3813 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3816 * Check if a pageblock contains reserved pages
3818 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3822 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3823 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3830 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3831 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3832 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3833 * higher will lead to a bigger reserve which will get freed as contiguous
3834 * blocks as reclaim kicks in
3836 static void setup_zone_migrate_reserve(struct zone
*zone
)
3838 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3840 unsigned long block_migratetype
;
3844 * Get the start pfn, end pfn and the number of blocks to reserve
3845 * We have to be careful to be aligned to pageblock_nr_pages to
3846 * make sure that we always check pfn_valid for the first page in
3849 start_pfn
= zone
->zone_start_pfn
;
3850 end_pfn
= zone_end_pfn(zone
);
3851 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
3852 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3856 * Reserve blocks are generally in place to help high-order atomic
3857 * allocations that are short-lived. A min_free_kbytes value that
3858 * would result in more than 2 reserve blocks for atomic allocations
3859 * is assumed to be in place to help anti-fragmentation for the
3860 * future allocation of hugepages at runtime.
3862 reserve
= min(2, reserve
);
3864 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3865 if (!pfn_valid(pfn
))
3867 page
= pfn_to_page(pfn
);
3869 /* Watch out for overlapping nodes */
3870 if (page_to_nid(page
) != zone_to_nid(zone
))
3873 block_migratetype
= get_pageblock_migratetype(page
);
3875 /* Only test what is necessary when the reserves are not met */
3878 * Blocks with reserved pages will never free, skip
3881 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
3882 if (pageblock_is_reserved(pfn
, block_end_pfn
))
3885 /* If this block is reserved, account for it */
3886 if (block_migratetype
== MIGRATE_RESERVE
) {
3891 /* Suitable for reserving if this block is movable */
3892 if (block_migratetype
== MIGRATE_MOVABLE
) {
3893 set_pageblock_migratetype(page
,
3895 move_freepages_block(zone
, page
,
3903 * If the reserve is met and this is a previous reserved block,
3906 if (block_migratetype
== MIGRATE_RESERVE
) {
3907 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3908 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3914 * Initially all pages are reserved - free ones are freed
3915 * up by free_all_bootmem() once the early boot process is
3916 * done. Non-atomic initialization, single-pass.
3918 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3919 unsigned long start_pfn
, enum memmap_context context
)
3922 unsigned long end_pfn
= start_pfn
+ size
;
3926 if (highest_memmap_pfn
< end_pfn
- 1)
3927 highest_memmap_pfn
= end_pfn
- 1;
3929 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3930 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3932 * There can be holes in boot-time mem_map[]s
3933 * handed to this function. They do not
3934 * exist on hotplugged memory.
3936 if (context
== MEMMAP_EARLY
) {
3937 if (!early_pfn_valid(pfn
))
3939 if (!early_pfn_in_nid(pfn
, nid
))
3942 page
= pfn_to_page(pfn
);
3943 set_page_links(page
, zone
, nid
, pfn
);
3944 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3945 init_page_count(page
);
3946 page_mapcount_reset(page
);
3947 page_nid_reset_last(page
);
3948 SetPageReserved(page
);
3950 * Mark the block movable so that blocks are reserved for
3951 * movable at startup. This will force kernel allocations
3952 * to reserve their blocks rather than leaking throughout
3953 * the address space during boot when many long-lived
3954 * kernel allocations are made. Later some blocks near
3955 * the start are marked MIGRATE_RESERVE by
3956 * setup_zone_migrate_reserve()
3958 * bitmap is created for zone's valid pfn range. but memmap
3959 * can be created for invalid pages (for alignment)
3960 * check here not to call set_pageblock_migratetype() against
3963 if ((z
->zone_start_pfn
<= pfn
)
3964 && (pfn
< zone_end_pfn(z
))
3965 && !(pfn
& (pageblock_nr_pages
- 1)))
3966 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3968 INIT_LIST_HEAD(&page
->lru
);
3969 #ifdef WANT_PAGE_VIRTUAL
3970 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3971 if (!is_highmem_idx(zone
))
3972 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3977 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3980 for_each_migratetype_order(order
, t
) {
3981 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3982 zone
->free_area
[order
].nr_free
= 0;
3986 #ifndef __HAVE_ARCH_MEMMAP_INIT
3987 #define memmap_init(size, nid, zone, start_pfn) \
3988 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3991 static int __meminit
zone_batchsize(struct zone
*zone
)
3997 * The per-cpu-pages pools are set to around 1000th of the
3998 * size of the zone. But no more than 1/2 of a meg.
4000 * OK, so we don't know how big the cache is. So guess.
4002 batch
= zone
->managed_pages
/ 1024;
4003 if (batch
* PAGE_SIZE
> 512 * 1024)
4004 batch
= (512 * 1024) / PAGE_SIZE
;
4005 batch
/= 4; /* We effectively *= 4 below */
4010 * Clamp the batch to a 2^n - 1 value. Having a power
4011 * of 2 value was found to be more likely to have
4012 * suboptimal cache aliasing properties in some cases.
4014 * For example if 2 tasks are alternately allocating
4015 * batches of pages, one task can end up with a lot
4016 * of pages of one half of the possible page colors
4017 * and the other with pages of the other colors.
4019 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4024 /* The deferral and batching of frees should be suppressed under NOMMU
4027 * The problem is that NOMMU needs to be able to allocate large chunks
4028 * of contiguous memory as there's no hardware page translation to
4029 * assemble apparent contiguous memory from discontiguous pages.
4031 * Queueing large contiguous runs of pages for batching, however,
4032 * causes the pages to actually be freed in smaller chunks. As there
4033 * can be a significant delay between the individual batches being
4034 * recycled, this leads to the once large chunks of space being
4035 * fragmented and becoming unavailable for high-order allocations.
4042 * pcp->high and pcp->batch values are related and dependent on one another:
4043 * ->batch must never be higher then ->high.
4044 * The following function updates them in a safe manner without read side
4047 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4048 * those fields changing asynchronously (acording the the above rule).
4050 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4051 * outside of boot time (or some other assurance that no concurrent updaters
4054 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4055 unsigned long batch
)
4057 /* start with a fail safe value for batch */
4061 /* Update high, then batch, in order */
4068 /* a companion to setup_pagelist_highmark() */
4069 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4071 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4074 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4076 struct per_cpu_pages
*pcp
;
4079 memset(p
, 0, sizeof(*p
));
4083 pageset_set_batch(p
, batch
);
4084 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4085 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4089 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4090 * to the value high for the pageset p.
4092 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
4095 unsigned long batch
= max(1UL, high
/ 4);
4096 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4097 batch
= PAGE_SHIFT
* 8;
4099 pageset_update(&p
->pcp
, high
, batch
);
4102 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4106 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4108 for_each_possible_cpu(cpu
) {
4109 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4111 setup_pageset(pcp
, zone_batchsize(zone
));
4113 if (percpu_pagelist_fraction
)
4114 setup_pagelist_highmark(pcp
,
4115 (zone
->managed_pages
/
4116 percpu_pagelist_fraction
));
4121 * Allocate per cpu pagesets and initialize them.
4122 * Before this call only boot pagesets were available.
4124 void __init
setup_per_cpu_pageset(void)
4128 for_each_populated_zone(zone
)
4129 setup_zone_pageset(zone
);
4132 static noinline __init_refok
4133 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4136 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4140 * The per-page waitqueue mechanism uses hashed waitqueues
4143 zone
->wait_table_hash_nr_entries
=
4144 wait_table_hash_nr_entries(zone_size_pages
);
4145 zone
->wait_table_bits
=
4146 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4147 alloc_size
= zone
->wait_table_hash_nr_entries
4148 * sizeof(wait_queue_head_t
);
4150 if (!slab_is_available()) {
4151 zone
->wait_table
= (wait_queue_head_t
*)
4152 alloc_bootmem_node_nopanic(pgdat
, alloc_size
);
4155 * This case means that a zone whose size was 0 gets new memory
4156 * via memory hot-add.
4157 * But it may be the case that a new node was hot-added. In
4158 * this case vmalloc() will not be able to use this new node's
4159 * memory - this wait_table must be initialized to use this new
4160 * node itself as well.
4161 * To use this new node's memory, further consideration will be
4164 zone
->wait_table
= vmalloc(alloc_size
);
4166 if (!zone
->wait_table
)
4169 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4170 init_waitqueue_head(zone
->wait_table
+ i
);
4175 static __meminit
void zone_pcp_init(struct zone
*zone
)
4178 * per cpu subsystem is not up at this point. The following code
4179 * relies on the ability of the linker to provide the
4180 * offset of a (static) per cpu variable into the per cpu area.
4182 zone
->pageset
= &boot_pageset
;
4184 if (zone
->present_pages
)
4185 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4186 zone
->name
, zone
->present_pages
,
4187 zone_batchsize(zone
));
4190 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4191 unsigned long zone_start_pfn
,
4193 enum memmap_context context
)
4195 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4197 ret
= zone_wait_table_init(zone
, size
);
4200 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4202 zone
->zone_start_pfn
= zone_start_pfn
;
4204 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4205 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4207 (unsigned long)zone_idx(zone
),
4208 zone_start_pfn
, (zone_start_pfn
+ size
));
4210 zone_init_free_lists(zone
);
4215 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4216 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4218 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4219 * Architectures may implement their own version but if add_active_range()
4220 * was used and there are no special requirements, this is a convenient
4223 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4225 unsigned long start_pfn
, end_pfn
;
4228 * NOTE: The following SMP-unsafe globals are only used early in boot
4229 * when the kernel is running single-threaded.
4231 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4232 static int __meminitdata last_nid
;
4234 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4237 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
4238 if (start_pfn
<= pfn
&& pfn
< end_pfn
) {
4239 last_start_pfn
= start_pfn
;
4240 last_end_pfn
= end_pfn
;
4244 /* This is a memory hole */
4247 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4249 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4253 nid
= __early_pfn_to_nid(pfn
);
4256 /* just returns 0 */
4260 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4261 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4265 nid
= __early_pfn_to_nid(pfn
);
4266 if (nid
>= 0 && nid
!= node
)
4273 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4274 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4275 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4277 * If an architecture guarantees that all ranges registered with
4278 * add_active_ranges() contain no holes and may be freed, this
4279 * this function may be used instead of calling free_bootmem() manually.
4281 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4283 unsigned long start_pfn
, end_pfn
;
4286 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4287 start_pfn
= min(start_pfn
, max_low_pfn
);
4288 end_pfn
= min(end_pfn
, max_low_pfn
);
4290 if (start_pfn
< end_pfn
)
4291 free_bootmem_node(NODE_DATA(this_nid
),
4292 PFN_PHYS(start_pfn
),
4293 (end_pfn
- start_pfn
) << PAGE_SHIFT
);
4298 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4299 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4301 * If an architecture guarantees that all ranges registered with
4302 * add_active_ranges() contain no holes and may be freed, this
4303 * function may be used instead of calling memory_present() manually.
4305 void __init
sparse_memory_present_with_active_regions(int nid
)
4307 unsigned long start_pfn
, end_pfn
;
4310 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4311 memory_present(this_nid
, start_pfn
, end_pfn
);
4315 * get_pfn_range_for_nid - Return the start and end page frames for a node
4316 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4317 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4318 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4320 * It returns the start and end page frame of a node based on information
4321 * provided by an arch calling add_active_range(). If called for a node
4322 * with no available memory, a warning is printed and the start and end
4325 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4326 unsigned long *start_pfn
, unsigned long *end_pfn
)
4328 unsigned long this_start_pfn
, this_end_pfn
;
4334 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4335 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4336 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4339 if (*start_pfn
== -1UL)
4344 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4345 * assumption is made that zones within a node are ordered in monotonic
4346 * increasing memory addresses so that the "highest" populated zone is used
4348 static void __init
find_usable_zone_for_movable(void)
4351 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4352 if (zone_index
== ZONE_MOVABLE
)
4355 if (arch_zone_highest_possible_pfn
[zone_index
] >
4356 arch_zone_lowest_possible_pfn
[zone_index
])
4360 VM_BUG_ON(zone_index
== -1);
4361 movable_zone
= zone_index
;
4365 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4366 * because it is sized independent of architecture. Unlike the other zones,
4367 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4368 * in each node depending on the size of each node and how evenly kernelcore
4369 * is distributed. This helper function adjusts the zone ranges
4370 * provided by the architecture for a given node by using the end of the
4371 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4372 * zones within a node are in order of monotonic increases memory addresses
4374 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4375 unsigned long zone_type
,
4376 unsigned long node_start_pfn
,
4377 unsigned long node_end_pfn
,
4378 unsigned long *zone_start_pfn
,
4379 unsigned long *zone_end_pfn
)
4381 /* Only adjust if ZONE_MOVABLE is on this node */
4382 if (zone_movable_pfn
[nid
]) {
4383 /* Size ZONE_MOVABLE */
4384 if (zone_type
== ZONE_MOVABLE
) {
4385 *zone_start_pfn
= zone_movable_pfn
[nid
];
4386 *zone_end_pfn
= min(node_end_pfn
,
4387 arch_zone_highest_possible_pfn
[movable_zone
]);
4389 /* Adjust for ZONE_MOVABLE starting within this range */
4390 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4391 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4392 *zone_end_pfn
= zone_movable_pfn
[nid
];
4394 /* Check if this whole range is within ZONE_MOVABLE */
4395 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4396 *zone_start_pfn
= *zone_end_pfn
;
4401 * Return the number of pages a zone spans in a node, including holes
4402 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4404 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4405 unsigned long zone_type
,
4406 unsigned long *ignored
)
4408 unsigned long node_start_pfn
, node_end_pfn
;
4409 unsigned long zone_start_pfn
, zone_end_pfn
;
4411 /* Get the start and end of the node and zone */
4412 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4413 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4414 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4415 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4416 node_start_pfn
, node_end_pfn
,
4417 &zone_start_pfn
, &zone_end_pfn
);
4419 /* Check that this node has pages within the zone's required range */
4420 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4423 /* Move the zone boundaries inside the node if necessary */
4424 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4425 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4427 /* Return the spanned pages */
4428 return zone_end_pfn
- zone_start_pfn
;
4432 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4433 * then all holes in the requested range will be accounted for.
4435 unsigned long __meminit
__absent_pages_in_range(int nid
,
4436 unsigned long range_start_pfn
,
4437 unsigned long range_end_pfn
)
4439 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4440 unsigned long start_pfn
, end_pfn
;
4443 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4444 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4445 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4446 nr_absent
-= end_pfn
- start_pfn
;
4452 * absent_pages_in_range - Return number of page frames in holes within a range
4453 * @start_pfn: The start PFN to start searching for holes
4454 * @end_pfn: The end PFN to stop searching for holes
4456 * It returns the number of pages frames in memory holes within a range.
4458 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4459 unsigned long end_pfn
)
4461 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4464 /* Return the number of page frames in holes in a zone on a node */
4465 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4466 unsigned long zone_type
,
4467 unsigned long *ignored
)
4469 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4470 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4471 unsigned long node_start_pfn
, node_end_pfn
;
4472 unsigned long zone_start_pfn
, zone_end_pfn
;
4474 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4475 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4476 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4478 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4479 node_start_pfn
, node_end_pfn
,
4480 &zone_start_pfn
, &zone_end_pfn
);
4481 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4484 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4485 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4486 unsigned long zone_type
,
4487 unsigned long *zones_size
)
4489 return zones_size
[zone_type
];
4492 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4493 unsigned long zone_type
,
4494 unsigned long *zholes_size
)
4499 return zholes_size
[zone_type
];
4502 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4504 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4505 unsigned long *zones_size
, unsigned long *zholes_size
)
4507 unsigned long realtotalpages
, totalpages
= 0;
4510 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4511 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4513 pgdat
->node_spanned_pages
= totalpages
;
4515 realtotalpages
= totalpages
;
4516 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4518 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4520 pgdat
->node_present_pages
= realtotalpages
;
4521 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4525 #ifndef CONFIG_SPARSEMEM
4527 * Calculate the size of the zone->blockflags rounded to an unsigned long
4528 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4529 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4530 * round what is now in bits to nearest long in bits, then return it in
4533 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4535 unsigned long usemapsize
;
4537 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4538 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4539 usemapsize
= usemapsize
>> pageblock_order
;
4540 usemapsize
*= NR_PAGEBLOCK_BITS
;
4541 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4543 return usemapsize
/ 8;
4546 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4548 unsigned long zone_start_pfn
,
4549 unsigned long zonesize
)
4551 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4552 zone
->pageblock_flags
= NULL
;
4554 zone
->pageblock_flags
= alloc_bootmem_node_nopanic(pgdat
,
4558 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4559 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4560 #endif /* CONFIG_SPARSEMEM */
4562 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4564 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4565 void __init
set_pageblock_order(void)
4569 /* Check that pageblock_nr_pages has not already been setup */
4570 if (pageblock_order
)
4573 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4574 order
= HUGETLB_PAGE_ORDER
;
4576 order
= MAX_ORDER
- 1;
4579 * Assume the largest contiguous order of interest is a huge page.
4580 * This value may be variable depending on boot parameters on IA64 and
4583 pageblock_order
= order
;
4585 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4588 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4589 * is unused as pageblock_order is set at compile-time. See
4590 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4593 void __init
set_pageblock_order(void)
4597 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4599 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4600 unsigned long present_pages
)
4602 unsigned long pages
= spanned_pages
;
4605 * Provide a more accurate estimation if there are holes within
4606 * the zone and SPARSEMEM is in use. If there are holes within the
4607 * zone, each populated memory region may cost us one or two extra
4608 * memmap pages due to alignment because memmap pages for each
4609 * populated regions may not naturally algined on page boundary.
4610 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4612 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4613 IS_ENABLED(CONFIG_SPARSEMEM
))
4614 pages
= present_pages
;
4616 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4620 * Set up the zone data structures:
4621 * - mark all pages reserved
4622 * - mark all memory queues empty
4623 * - clear the memory bitmaps
4625 * NOTE: pgdat should get zeroed by caller.
4627 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4628 unsigned long *zones_size
, unsigned long *zholes_size
)
4631 int nid
= pgdat
->node_id
;
4632 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4635 pgdat_resize_init(pgdat
);
4636 #ifdef CONFIG_NUMA_BALANCING
4637 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4638 pgdat
->numabalancing_migrate_nr_pages
= 0;
4639 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4641 init_waitqueue_head(&pgdat
->kswapd_wait
);
4642 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4643 pgdat_page_cgroup_init(pgdat
);
4645 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4646 struct zone
*zone
= pgdat
->node_zones
+ j
;
4647 unsigned long size
, realsize
, freesize
, memmap_pages
;
4649 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4650 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4654 * Adjust freesize so that it accounts for how much memory
4655 * is used by this zone for memmap. This affects the watermark
4656 * and per-cpu initialisations
4658 memmap_pages
= calc_memmap_size(size
, realsize
);
4659 if (freesize
>= memmap_pages
) {
4660 freesize
-= memmap_pages
;
4663 " %s zone: %lu pages used for memmap\n",
4664 zone_names
[j
], memmap_pages
);
4667 " %s zone: %lu pages exceeds freesize %lu\n",
4668 zone_names
[j
], memmap_pages
, freesize
);
4670 /* Account for reserved pages */
4671 if (j
== 0 && freesize
> dma_reserve
) {
4672 freesize
-= dma_reserve
;
4673 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4674 zone_names
[0], dma_reserve
);
4677 if (!is_highmem_idx(j
))
4678 nr_kernel_pages
+= freesize
;
4679 /* Charge for highmem memmap if there are enough kernel pages */
4680 else if (nr_kernel_pages
> memmap_pages
* 2)
4681 nr_kernel_pages
-= memmap_pages
;
4682 nr_all_pages
+= freesize
;
4684 zone
->spanned_pages
= size
;
4685 zone
->present_pages
= realsize
;
4687 * Set an approximate value for lowmem here, it will be adjusted
4688 * when the bootmem allocator frees pages into the buddy system.
4689 * And all highmem pages will be managed by the buddy system.
4691 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4694 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4696 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4698 zone
->name
= zone_names
[j
];
4699 spin_lock_init(&zone
->lock
);
4700 spin_lock_init(&zone
->lru_lock
);
4701 zone_seqlock_init(zone
);
4702 zone
->zone_pgdat
= pgdat
;
4704 zone_pcp_init(zone
);
4705 lruvec_init(&zone
->lruvec
);
4709 set_pageblock_order();
4710 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4711 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4712 size
, MEMMAP_EARLY
);
4714 memmap_init(size
, nid
, j
, zone_start_pfn
);
4715 zone_start_pfn
+= size
;
4719 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4721 /* Skip empty nodes */
4722 if (!pgdat
->node_spanned_pages
)
4725 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4726 /* ia64 gets its own node_mem_map, before this, without bootmem */
4727 if (!pgdat
->node_mem_map
) {
4728 unsigned long size
, start
, end
;
4732 * The zone's endpoints aren't required to be MAX_ORDER
4733 * aligned but the node_mem_map endpoints must be in order
4734 * for the buddy allocator to function correctly.
4736 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4737 end
= pgdat_end_pfn(pgdat
);
4738 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4739 size
= (end
- start
) * sizeof(struct page
);
4740 map
= alloc_remap(pgdat
->node_id
, size
);
4742 map
= alloc_bootmem_node_nopanic(pgdat
, size
);
4743 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4745 #ifndef CONFIG_NEED_MULTIPLE_NODES
4747 * With no DISCONTIG, the global mem_map is just set as node 0's
4749 if (pgdat
== NODE_DATA(0)) {
4750 mem_map
= NODE_DATA(0)->node_mem_map
;
4751 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4752 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4753 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4754 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4757 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4760 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4761 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4763 pg_data_t
*pgdat
= NODE_DATA(nid
);
4765 /* pg_data_t should be reset to zero when it's allocated */
4766 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4768 pgdat
->node_id
= nid
;
4769 pgdat
->node_start_pfn
= node_start_pfn
;
4770 init_zone_allows_reclaim(nid
);
4771 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4773 alloc_node_mem_map(pgdat
);
4774 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4775 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4776 nid
, (unsigned long)pgdat
,
4777 (unsigned long)pgdat
->node_mem_map
);
4780 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4783 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4785 #if MAX_NUMNODES > 1
4787 * Figure out the number of possible node ids.
4789 void __init
setup_nr_node_ids(void)
4792 unsigned int highest
= 0;
4794 for_each_node_mask(node
, node_possible_map
)
4796 nr_node_ids
= highest
+ 1;
4801 * node_map_pfn_alignment - determine the maximum internode alignment
4803 * This function should be called after node map is populated and sorted.
4804 * It calculates the maximum power of two alignment which can distinguish
4807 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4808 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4809 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4810 * shifted, 1GiB is enough and this function will indicate so.
4812 * This is used to test whether pfn -> nid mapping of the chosen memory
4813 * model has fine enough granularity to avoid incorrect mapping for the
4814 * populated node map.
4816 * Returns the determined alignment in pfn's. 0 if there is no alignment
4817 * requirement (single node).
4819 unsigned long __init
node_map_pfn_alignment(void)
4821 unsigned long accl_mask
= 0, last_end
= 0;
4822 unsigned long start
, end
, mask
;
4826 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
4827 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
4834 * Start with a mask granular enough to pin-point to the
4835 * start pfn and tick off bits one-by-one until it becomes
4836 * too coarse to separate the current node from the last.
4838 mask
= ~((1 << __ffs(start
)) - 1);
4839 while (mask
&& last_end
<= (start
& (mask
<< 1)))
4842 /* accumulate all internode masks */
4846 /* convert mask to number of pages */
4847 return ~accl_mask
+ 1;
4850 /* Find the lowest pfn for a node */
4851 static unsigned long __init
find_min_pfn_for_node(int nid
)
4853 unsigned long min_pfn
= ULONG_MAX
;
4854 unsigned long start_pfn
;
4857 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
4858 min_pfn
= min(min_pfn
, start_pfn
);
4860 if (min_pfn
== ULONG_MAX
) {
4862 "Could not find start_pfn for node %d\n", nid
);
4870 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4872 * It returns the minimum PFN based on information provided via
4873 * add_active_range().
4875 unsigned long __init
find_min_pfn_with_active_regions(void)
4877 return find_min_pfn_for_node(MAX_NUMNODES
);
4881 * early_calculate_totalpages()
4882 * Sum pages in active regions for movable zone.
4883 * Populate N_MEMORY for calculating usable_nodes.
4885 static unsigned long __init
early_calculate_totalpages(void)
4887 unsigned long totalpages
= 0;
4888 unsigned long start_pfn
, end_pfn
;
4891 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4892 unsigned long pages
= end_pfn
- start_pfn
;
4894 totalpages
+= pages
;
4896 node_set_state(nid
, N_MEMORY
);
4902 * Find the PFN the Movable zone begins in each node. Kernel memory
4903 * is spread evenly between nodes as long as the nodes have enough
4904 * memory. When they don't, some nodes will have more kernelcore than
4907 static void __init
find_zone_movable_pfns_for_nodes(void)
4910 unsigned long usable_startpfn
;
4911 unsigned long kernelcore_node
, kernelcore_remaining
;
4912 /* save the state before borrow the nodemask */
4913 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
4914 unsigned long totalpages
= early_calculate_totalpages();
4915 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
4918 * If movablecore was specified, calculate what size of
4919 * kernelcore that corresponds so that memory usable for
4920 * any allocation type is evenly spread. If both kernelcore
4921 * and movablecore are specified, then the value of kernelcore
4922 * will be used for required_kernelcore if it's greater than
4923 * what movablecore would have allowed.
4925 if (required_movablecore
) {
4926 unsigned long corepages
;
4929 * Round-up so that ZONE_MOVABLE is at least as large as what
4930 * was requested by the user
4932 required_movablecore
=
4933 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4934 corepages
= totalpages
- required_movablecore
;
4936 required_kernelcore
= max(required_kernelcore
, corepages
);
4939 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4940 if (!required_kernelcore
)
4943 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4944 find_usable_zone_for_movable();
4945 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4948 /* Spread kernelcore memory as evenly as possible throughout nodes */
4949 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4950 for_each_node_state(nid
, N_MEMORY
) {
4951 unsigned long start_pfn
, end_pfn
;
4954 * Recalculate kernelcore_node if the division per node
4955 * now exceeds what is necessary to satisfy the requested
4956 * amount of memory for the kernel
4958 if (required_kernelcore
< kernelcore_node
)
4959 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4962 * As the map is walked, we track how much memory is usable
4963 * by the kernel using kernelcore_remaining. When it is
4964 * 0, the rest of the node is usable by ZONE_MOVABLE
4966 kernelcore_remaining
= kernelcore_node
;
4968 /* Go through each range of PFNs within this node */
4969 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4970 unsigned long size_pages
;
4972 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
4973 if (start_pfn
>= end_pfn
)
4976 /* Account for what is only usable for kernelcore */
4977 if (start_pfn
< usable_startpfn
) {
4978 unsigned long kernel_pages
;
4979 kernel_pages
= min(end_pfn
, usable_startpfn
)
4982 kernelcore_remaining
-= min(kernel_pages
,
4983 kernelcore_remaining
);
4984 required_kernelcore
-= min(kernel_pages
,
4985 required_kernelcore
);
4987 /* Continue if range is now fully accounted */
4988 if (end_pfn
<= usable_startpfn
) {
4991 * Push zone_movable_pfn to the end so
4992 * that if we have to rebalance
4993 * kernelcore across nodes, we will
4994 * not double account here
4996 zone_movable_pfn
[nid
] = end_pfn
;
4999 start_pfn
= usable_startpfn
;
5003 * The usable PFN range for ZONE_MOVABLE is from
5004 * start_pfn->end_pfn. Calculate size_pages as the
5005 * number of pages used as kernelcore
5007 size_pages
= end_pfn
- start_pfn
;
5008 if (size_pages
> kernelcore_remaining
)
5009 size_pages
= kernelcore_remaining
;
5010 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5013 * Some kernelcore has been met, update counts and
5014 * break if the kernelcore for this node has been
5017 required_kernelcore
-= min(required_kernelcore
,
5019 kernelcore_remaining
-= size_pages
;
5020 if (!kernelcore_remaining
)
5026 * If there is still required_kernelcore, we do another pass with one
5027 * less node in the count. This will push zone_movable_pfn[nid] further
5028 * along on the nodes that still have memory until kernelcore is
5032 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5035 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5036 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5037 zone_movable_pfn
[nid
] =
5038 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5041 /* restore the node_state */
5042 node_states
[N_MEMORY
] = saved_node_state
;
5045 /* Any regular or high memory on that node ? */
5046 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5048 enum zone_type zone_type
;
5050 if (N_MEMORY
== N_NORMAL_MEMORY
)
5053 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5054 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5055 if (zone
->present_pages
) {
5056 node_set_state(nid
, N_HIGH_MEMORY
);
5057 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5058 zone_type
<= ZONE_NORMAL
)
5059 node_set_state(nid
, N_NORMAL_MEMORY
);
5066 * free_area_init_nodes - Initialise all pg_data_t and zone data
5067 * @max_zone_pfn: an array of max PFNs for each zone
5069 * This will call free_area_init_node() for each active node in the system.
5070 * Using the page ranges provided by add_active_range(), the size of each
5071 * zone in each node and their holes is calculated. If the maximum PFN
5072 * between two adjacent zones match, it is assumed that the zone is empty.
5073 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5074 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5075 * starts where the previous one ended. For example, ZONE_DMA32 starts
5076 * at arch_max_dma_pfn.
5078 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5080 unsigned long start_pfn
, end_pfn
;
5083 /* Record where the zone boundaries are */
5084 memset(arch_zone_lowest_possible_pfn
, 0,
5085 sizeof(arch_zone_lowest_possible_pfn
));
5086 memset(arch_zone_highest_possible_pfn
, 0,
5087 sizeof(arch_zone_highest_possible_pfn
));
5088 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5089 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5090 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5091 if (i
== ZONE_MOVABLE
)
5093 arch_zone_lowest_possible_pfn
[i
] =
5094 arch_zone_highest_possible_pfn
[i
-1];
5095 arch_zone_highest_possible_pfn
[i
] =
5096 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5098 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5099 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5101 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5102 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5103 find_zone_movable_pfns_for_nodes();
5105 /* Print out the zone ranges */
5106 printk("Zone ranges:\n");
5107 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5108 if (i
== ZONE_MOVABLE
)
5110 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5111 if (arch_zone_lowest_possible_pfn
[i
] ==
5112 arch_zone_highest_possible_pfn
[i
])
5113 printk(KERN_CONT
"empty\n");
5115 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5116 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5117 (arch_zone_highest_possible_pfn
[i
]
5118 << PAGE_SHIFT
) - 1);
5121 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5122 printk("Movable zone start for each node\n");
5123 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5124 if (zone_movable_pfn
[i
])
5125 printk(" Node %d: %#010lx\n", i
,
5126 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5129 /* Print out the early node map */
5130 printk("Early memory node ranges\n");
5131 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5132 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5133 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5135 /* Initialise every node */
5136 mminit_verify_pageflags_layout();
5137 setup_nr_node_ids();
5138 for_each_online_node(nid
) {
5139 pg_data_t
*pgdat
= NODE_DATA(nid
);
5140 free_area_init_node(nid
, NULL
,
5141 find_min_pfn_for_node(nid
), NULL
);
5143 /* Any memory on that node */
5144 if (pgdat
->node_present_pages
)
5145 node_set_state(nid
, N_MEMORY
);
5146 check_for_memory(pgdat
, nid
);
5150 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5152 unsigned long long coremem
;
5156 coremem
= memparse(p
, &p
);
5157 *core
= coremem
>> PAGE_SHIFT
;
5159 /* Paranoid check that UL is enough for the coremem value */
5160 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5166 * kernelcore=size sets the amount of memory for use for allocations that
5167 * cannot be reclaimed or migrated.
5169 static int __init
cmdline_parse_kernelcore(char *p
)
5171 return cmdline_parse_core(p
, &required_kernelcore
);
5175 * movablecore=size sets the amount of memory for use for allocations that
5176 * can be reclaimed or migrated.
5178 static int __init
cmdline_parse_movablecore(char *p
)
5180 return cmdline_parse_core(p
, &required_movablecore
);
5183 early_param("kernelcore", cmdline_parse_kernelcore
);
5184 early_param("movablecore", cmdline_parse_movablecore
);
5186 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5188 unsigned long free_reserved_area(unsigned long start
, unsigned long end
,
5189 int poison
, char *s
)
5191 unsigned long pages
, pos
;
5193 pos
= start
= PAGE_ALIGN(start
);
5195 for (pages
= 0; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5197 memset((void *)pos
, poison
, PAGE_SIZE
);
5198 free_reserved_page(virt_to_page((void *)pos
));
5202 pr_info("Freeing %s memory: %ldK (%lx - %lx)\n",
5203 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5208 #ifdef CONFIG_HIGHMEM
5209 void free_highmem_page(struct page
*page
)
5211 __free_reserved_page(page
);
5218 * set_dma_reserve - set the specified number of pages reserved in the first zone
5219 * @new_dma_reserve: The number of pages to mark reserved
5221 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5222 * In the DMA zone, a significant percentage may be consumed by kernel image
5223 * and other unfreeable allocations which can skew the watermarks badly. This
5224 * function may optionally be used to account for unfreeable pages in the
5225 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5226 * smaller per-cpu batchsize.
5228 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5230 dma_reserve
= new_dma_reserve
;
5233 void __init
free_area_init(unsigned long *zones_size
)
5235 free_area_init_node(0, zones_size
,
5236 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5239 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5240 unsigned long action
, void *hcpu
)
5242 int cpu
= (unsigned long)hcpu
;
5244 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5245 lru_add_drain_cpu(cpu
);
5249 * Spill the event counters of the dead processor
5250 * into the current processors event counters.
5251 * This artificially elevates the count of the current
5254 vm_events_fold_cpu(cpu
);
5257 * Zero the differential counters of the dead processor
5258 * so that the vm statistics are consistent.
5260 * This is only okay since the processor is dead and cannot
5261 * race with what we are doing.
5263 refresh_cpu_vm_stats(cpu
);
5268 void __init
page_alloc_init(void)
5270 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5274 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5275 * or min_free_kbytes changes.
5277 static void calculate_totalreserve_pages(void)
5279 struct pglist_data
*pgdat
;
5280 unsigned long reserve_pages
= 0;
5281 enum zone_type i
, j
;
5283 for_each_online_pgdat(pgdat
) {
5284 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5285 struct zone
*zone
= pgdat
->node_zones
+ i
;
5286 unsigned long max
= 0;
5288 /* Find valid and maximum lowmem_reserve in the zone */
5289 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5290 if (zone
->lowmem_reserve
[j
] > max
)
5291 max
= zone
->lowmem_reserve
[j
];
5294 /* we treat the high watermark as reserved pages. */
5295 max
+= high_wmark_pages(zone
);
5297 if (max
> zone
->managed_pages
)
5298 max
= zone
->managed_pages
;
5299 reserve_pages
+= max
;
5301 * Lowmem reserves are not available to
5302 * GFP_HIGHUSER page cache allocations and
5303 * kswapd tries to balance zones to their high
5304 * watermark. As a result, neither should be
5305 * regarded as dirtyable memory, to prevent a
5306 * situation where reclaim has to clean pages
5307 * in order to balance the zones.
5309 zone
->dirty_balance_reserve
= max
;
5312 dirty_balance_reserve
= reserve_pages
;
5313 totalreserve_pages
= reserve_pages
;
5317 * setup_per_zone_lowmem_reserve - called whenever
5318 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5319 * has a correct pages reserved value, so an adequate number of
5320 * pages are left in the zone after a successful __alloc_pages().
5322 static void setup_per_zone_lowmem_reserve(void)
5324 struct pglist_data
*pgdat
;
5325 enum zone_type j
, idx
;
5327 for_each_online_pgdat(pgdat
) {
5328 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5329 struct zone
*zone
= pgdat
->node_zones
+ j
;
5330 unsigned long managed_pages
= zone
->managed_pages
;
5332 zone
->lowmem_reserve
[j
] = 0;
5336 struct zone
*lower_zone
;
5340 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5341 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5343 lower_zone
= pgdat
->node_zones
+ idx
;
5344 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5345 sysctl_lowmem_reserve_ratio
[idx
];
5346 managed_pages
+= lower_zone
->managed_pages
;
5351 /* update totalreserve_pages */
5352 calculate_totalreserve_pages();
5355 static void __setup_per_zone_wmarks(void)
5357 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5358 unsigned long lowmem_pages
= 0;
5360 unsigned long flags
;
5362 /* Calculate total number of !ZONE_HIGHMEM pages */
5363 for_each_zone(zone
) {
5364 if (!is_highmem(zone
))
5365 lowmem_pages
+= zone
->managed_pages
;
5368 for_each_zone(zone
) {
5371 spin_lock_irqsave(&zone
->lock
, flags
);
5372 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5373 do_div(tmp
, lowmem_pages
);
5374 if (is_highmem(zone
)) {
5376 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5377 * need highmem pages, so cap pages_min to a small
5380 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5381 * deltas controls asynch page reclaim, and so should
5382 * not be capped for highmem.
5384 unsigned long min_pages
;
5386 min_pages
= zone
->managed_pages
/ 1024;
5387 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5388 zone
->watermark
[WMARK_MIN
] = min_pages
;
5391 * If it's a lowmem zone, reserve a number of pages
5392 * proportionate to the zone's size.
5394 zone
->watermark
[WMARK_MIN
] = tmp
;
5397 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5398 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5400 setup_zone_migrate_reserve(zone
);
5401 spin_unlock_irqrestore(&zone
->lock
, flags
);
5404 /* update totalreserve_pages */
5405 calculate_totalreserve_pages();
5409 * setup_per_zone_wmarks - called when min_free_kbytes changes
5410 * or when memory is hot-{added|removed}
5412 * Ensures that the watermark[min,low,high] values for each zone are set
5413 * correctly with respect to min_free_kbytes.
5415 void setup_per_zone_wmarks(void)
5417 mutex_lock(&zonelists_mutex
);
5418 __setup_per_zone_wmarks();
5419 mutex_unlock(&zonelists_mutex
);
5423 * The inactive anon list should be small enough that the VM never has to
5424 * do too much work, but large enough that each inactive page has a chance
5425 * to be referenced again before it is swapped out.
5427 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5428 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5429 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5430 * the anonymous pages are kept on the inactive list.
5433 * memory ratio inactive anon
5434 * -------------------------------------
5443 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5445 unsigned int gb
, ratio
;
5447 /* Zone size in gigabytes */
5448 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5450 ratio
= int_sqrt(10 * gb
);
5454 zone
->inactive_ratio
= ratio
;
5457 static void __meminit
setup_per_zone_inactive_ratio(void)
5462 calculate_zone_inactive_ratio(zone
);
5466 * Initialise min_free_kbytes.
5468 * For small machines we want it small (128k min). For large machines
5469 * we want it large (64MB max). But it is not linear, because network
5470 * bandwidth does not increase linearly with machine size. We use
5472 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5473 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5489 int __meminit
init_per_zone_wmark_min(void)
5491 unsigned long lowmem_kbytes
;
5493 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5495 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5496 if (min_free_kbytes
< 128)
5497 min_free_kbytes
= 128;
5498 if (min_free_kbytes
> 65536)
5499 min_free_kbytes
= 65536;
5500 setup_per_zone_wmarks();
5501 refresh_zone_stat_thresholds();
5502 setup_per_zone_lowmem_reserve();
5503 setup_per_zone_inactive_ratio();
5506 module_init(init_per_zone_wmark_min
)
5509 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5510 * that we can call two helper functions whenever min_free_kbytes
5513 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5514 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5516 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5518 setup_per_zone_wmarks();
5523 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5524 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5529 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5534 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5535 sysctl_min_unmapped_ratio
) / 100;
5539 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5540 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5545 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5550 zone
->min_slab_pages
= (zone
->managed_pages
*
5551 sysctl_min_slab_ratio
) / 100;
5557 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5558 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5559 * whenever sysctl_lowmem_reserve_ratio changes.
5561 * The reserve ratio obviously has absolutely no relation with the
5562 * minimum watermarks. The lowmem reserve ratio can only make sense
5563 * if in function of the boot time zone sizes.
5565 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5566 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5568 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5569 setup_per_zone_lowmem_reserve();
5574 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5575 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5576 * can have before it gets flushed back to buddy allocator.
5579 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5580 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5586 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5587 if (!write
|| (ret
< 0))
5590 mutex_lock(&pcp_batch_high_lock
);
5591 for_each_populated_zone(zone
) {
5592 for_each_possible_cpu(cpu
) {
5594 high
= zone
->managed_pages
/ percpu_pagelist_fraction
;
5595 setup_pagelist_highmark(
5596 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5599 mutex_unlock(&pcp_batch_high_lock
);
5603 int hashdist
= HASHDIST_DEFAULT
;
5606 static int __init
set_hashdist(char *str
)
5610 hashdist
= simple_strtoul(str
, &str
, 0);
5613 __setup("hashdist=", set_hashdist
);
5617 * allocate a large system hash table from bootmem
5618 * - it is assumed that the hash table must contain an exact power-of-2
5619 * quantity of entries
5620 * - limit is the number of hash buckets, not the total allocation size
5622 void *__init
alloc_large_system_hash(const char *tablename
,
5623 unsigned long bucketsize
,
5624 unsigned long numentries
,
5627 unsigned int *_hash_shift
,
5628 unsigned int *_hash_mask
,
5629 unsigned long low_limit
,
5630 unsigned long high_limit
)
5632 unsigned long long max
= high_limit
;
5633 unsigned long log2qty
, size
;
5636 /* allow the kernel cmdline to have a say */
5638 /* round applicable memory size up to nearest megabyte */
5639 numentries
= nr_kernel_pages
;
5640 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5641 numentries
>>= 20 - PAGE_SHIFT
;
5642 numentries
<<= 20 - PAGE_SHIFT
;
5644 /* limit to 1 bucket per 2^scale bytes of low memory */
5645 if (scale
> PAGE_SHIFT
)
5646 numentries
>>= (scale
- PAGE_SHIFT
);
5648 numentries
<<= (PAGE_SHIFT
- scale
);
5650 /* Make sure we've got at least a 0-order allocation.. */
5651 if (unlikely(flags
& HASH_SMALL
)) {
5652 /* Makes no sense without HASH_EARLY */
5653 WARN_ON(!(flags
& HASH_EARLY
));
5654 if (!(numentries
>> *_hash_shift
)) {
5655 numentries
= 1UL << *_hash_shift
;
5656 BUG_ON(!numentries
);
5658 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5659 numentries
= PAGE_SIZE
/ bucketsize
;
5661 numentries
= roundup_pow_of_two(numentries
);
5663 /* limit allocation size to 1/16 total memory by default */
5665 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5666 do_div(max
, bucketsize
);
5668 max
= min(max
, 0x80000000ULL
);
5670 if (numentries
< low_limit
)
5671 numentries
= low_limit
;
5672 if (numentries
> max
)
5675 log2qty
= ilog2(numentries
);
5678 size
= bucketsize
<< log2qty
;
5679 if (flags
& HASH_EARLY
)
5680 table
= alloc_bootmem_nopanic(size
);
5682 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5685 * If bucketsize is not a power-of-two, we may free
5686 * some pages at the end of hash table which
5687 * alloc_pages_exact() automatically does
5689 if (get_order(size
) < MAX_ORDER
) {
5690 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5691 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5694 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5697 panic("Failed to allocate %s hash table\n", tablename
);
5699 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5702 ilog2(size
) - PAGE_SHIFT
,
5706 *_hash_shift
= log2qty
;
5708 *_hash_mask
= (1 << log2qty
) - 1;
5713 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5714 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5717 #ifdef CONFIG_SPARSEMEM
5718 return __pfn_to_section(pfn
)->pageblock_flags
;
5720 return zone
->pageblock_flags
;
5721 #endif /* CONFIG_SPARSEMEM */
5724 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5726 #ifdef CONFIG_SPARSEMEM
5727 pfn
&= (PAGES_PER_SECTION
-1);
5728 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5730 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
5731 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5732 #endif /* CONFIG_SPARSEMEM */
5736 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5737 * @page: The page within the block of interest
5738 * @start_bitidx: The first bit of interest to retrieve
5739 * @end_bitidx: The last bit of interest
5740 * returns pageblock_bits flags
5742 unsigned long get_pageblock_flags_group(struct page
*page
,
5743 int start_bitidx
, int end_bitidx
)
5746 unsigned long *bitmap
;
5747 unsigned long pfn
, bitidx
;
5748 unsigned long flags
= 0;
5749 unsigned long value
= 1;
5751 zone
= page_zone(page
);
5752 pfn
= page_to_pfn(page
);
5753 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5754 bitidx
= pfn_to_bitidx(zone
, pfn
);
5756 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5757 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5764 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5765 * @page: The page within the block of interest
5766 * @start_bitidx: The first bit of interest
5767 * @end_bitidx: The last bit of interest
5768 * @flags: The flags to set
5770 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5771 int start_bitidx
, int end_bitidx
)
5774 unsigned long *bitmap
;
5775 unsigned long pfn
, bitidx
;
5776 unsigned long value
= 1;
5778 zone
= page_zone(page
);
5779 pfn
= page_to_pfn(page
);
5780 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5781 bitidx
= pfn_to_bitidx(zone
, pfn
);
5782 VM_BUG_ON(!zone_spans_pfn(zone
, pfn
));
5784 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5786 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5788 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5792 * This function checks whether pageblock includes unmovable pages or not.
5793 * If @count is not zero, it is okay to include less @count unmovable pages
5795 * PageLRU check wihtout isolation or lru_lock could race so that
5796 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5797 * expect this function should be exact.
5799 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
5800 bool skip_hwpoisoned_pages
)
5802 unsigned long pfn
, iter
, found
;
5806 * For avoiding noise data, lru_add_drain_all() should be called
5807 * If ZONE_MOVABLE, the zone never contains unmovable pages
5809 if (zone_idx(zone
) == ZONE_MOVABLE
)
5811 mt
= get_pageblock_migratetype(page
);
5812 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
5815 pfn
= page_to_pfn(page
);
5816 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5817 unsigned long check
= pfn
+ iter
;
5819 if (!pfn_valid_within(check
))
5822 page
= pfn_to_page(check
);
5824 * We can't use page_count without pin a page
5825 * because another CPU can free compound page.
5826 * This check already skips compound tails of THP
5827 * because their page->_count is zero at all time.
5829 if (!atomic_read(&page
->_count
)) {
5830 if (PageBuddy(page
))
5831 iter
+= (1 << page_order(page
)) - 1;
5836 * The HWPoisoned page may be not in buddy system, and
5837 * page_count() is not 0.
5839 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
5845 * If there are RECLAIMABLE pages, we need to check it.
5846 * But now, memory offline itself doesn't call shrink_slab()
5847 * and it still to be fixed.
5850 * If the page is not RAM, page_count()should be 0.
5851 * we don't need more check. This is an _used_ not-movable page.
5853 * The problematic thing here is PG_reserved pages. PG_reserved
5854 * is set to both of a memory hole page and a _used_ kernel
5863 bool is_pageblock_removable_nolock(struct page
*page
)
5869 * We have to be careful here because we are iterating over memory
5870 * sections which are not zone aware so we might end up outside of
5871 * the zone but still within the section.
5872 * We have to take care about the node as well. If the node is offline
5873 * its NODE_DATA will be NULL - see page_zone.
5875 if (!node_online(page_to_nid(page
)))
5878 zone
= page_zone(page
);
5879 pfn
= page_to_pfn(page
);
5880 if (!zone_spans_pfn(zone
, pfn
))
5883 return !has_unmovable_pages(zone
, page
, 0, true);
5888 static unsigned long pfn_max_align_down(unsigned long pfn
)
5890 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
5891 pageblock_nr_pages
) - 1);
5894 static unsigned long pfn_max_align_up(unsigned long pfn
)
5896 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
5897 pageblock_nr_pages
));
5900 /* [start, end) must belong to a single zone. */
5901 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
5902 unsigned long start
, unsigned long end
)
5904 /* This function is based on compact_zone() from compaction.c. */
5905 unsigned long nr_reclaimed
;
5906 unsigned long pfn
= start
;
5907 unsigned int tries
= 0;
5912 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
5913 if (fatal_signal_pending(current
)) {
5918 if (list_empty(&cc
->migratepages
)) {
5919 cc
->nr_migratepages
= 0;
5920 pfn
= isolate_migratepages_range(cc
->zone
, cc
,
5927 } else if (++tries
== 5) {
5928 ret
= ret
< 0 ? ret
: -EBUSY
;
5932 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
5934 cc
->nr_migratepages
-= nr_reclaimed
;
5936 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
5937 0, MIGRATE_SYNC
, MR_CMA
);
5940 putback_movable_pages(&cc
->migratepages
);
5947 * alloc_contig_range() -- tries to allocate given range of pages
5948 * @start: start PFN to allocate
5949 * @end: one-past-the-last PFN to allocate
5950 * @migratetype: migratetype of the underlaying pageblocks (either
5951 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5952 * in range must have the same migratetype and it must
5953 * be either of the two.
5955 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5956 * aligned, however it's the caller's responsibility to guarantee that
5957 * we are the only thread that changes migrate type of pageblocks the
5960 * The PFN range must belong to a single zone.
5962 * Returns zero on success or negative error code. On success all
5963 * pages which PFN is in [start, end) are allocated for the caller and
5964 * need to be freed with free_contig_range().
5966 int alloc_contig_range(unsigned long start
, unsigned long end
,
5967 unsigned migratetype
)
5969 unsigned long outer_start
, outer_end
;
5972 struct compact_control cc
= {
5973 .nr_migratepages
= 0,
5975 .zone
= page_zone(pfn_to_page(start
)),
5977 .ignore_skip_hint
= true,
5979 INIT_LIST_HEAD(&cc
.migratepages
);
5982 * What we do here is we mark all pageblocks in range as
5983 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5984 * have different sizes, and due to the way page allocator
5985 * work, we align the range to biggest of the two pages so
5986 * that page allocator won't try to merge buddies from
5987 * different pageblocks and change MIGRATE_ISOLATE to some
5988 * other migration type.
5990 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5991 * migrate the pages from an unaligned range (ie. pages that
5992 * we are interested in). This will put all the pages in
5993 * range back to page allocator as MIGRATE_ISOLATE.
5995 * When this is done, we take the pages in range from page
5996 * allocator removing them from the buddy system. This way
5997 * page allocator will never consider using them.
5999 * This lets us mark the pageblocks back as
6000 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6001 * aligned range but not in the unaligned, original range are
6002 * put back to page allocator so that buddy can use them.
6005 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6006 pfn_max_align_up(end
), migratetype
,
6011 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6016 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6017 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6018 * more, all pages in [start, end) are free in page allocator.
6019 * What we are going to do is to allocate all pages from
6020 * [start, end) (that is remove them from page allocator).
6022 * The only problem is that pages at the beginning and at the
6023 * end of interesting range may be not aligned with pages that
6024 * page allocator holds, ie. they can be part of higher order
6025 * pages. Because of this, we reserve the bigger range and
6026 * once this is done free the pages we are not interested in.
6028 * We don't have to hold zone->lock here because the pages are
6029 * isolated thus they won't get removed from buddy.
6032 lru_add_drain_all();
6036 outer_start
= start
;
6037 while (!PageBuddy(pfn_to_page(outer_start
))) {
6038 if (++order
>= MAX_ORDER
) {
6042 outer_start
&= ~0UL << order
;
6045 /* Make sure the range is really isolated. */
6046 if (test_pages_isolated(outer_start
, end
, false)) {
6047 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6054 /* Grab isolated pages from freelists. */
6055 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6061 /* Free head and tail (if any) */
6062 if (start
!= outer_start
)
6063 free_contig_range(outer_start
, start
- outer_start
);
6064 if (end
!= outer_end
)
6065 free_contig_range(end
, outer_end
- end
);
6068 undo_isolate_page_range(pfn_max_align_down(start
),
6069 pfn_max_align_up(end
), migratetype
);
6073 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6075 unsigned int count
= 0;
6077 for (; nr_pages
--; pfn
++) {
6078 struct page
*page
= pfn_to_page(pfn
);
6080 count
+= page_count(page
) != 1;
6083 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6087 #ifdef CONFIG_MEMORY_HOTPLUG
6088 static int __meminit
__zone_pcp_update(void *data
)
6090 struct zone
*zone
= data
;
6092 unsigned long batch
= zone_batchsize(zone
), flags
;
6094 for_each_possible_cpu(cpu
) {
6095 struct per_cpu_pageset
*pset
;
6096 struct per_cpu_pages
*pcp
;
6098 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6101 local_irq_save(flags
);
6103 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
6104 drain_zonestat(zone
, pset
);
6105 setup_pageset(pset
, batch
);
6106 local_irq_restore(flags
);
6111 void __meminit
zone_pcp_update(struct zone
*zone
)
6113 mutex_lock(&pcp_batch_high_lock
);
6114 stop_machine(__zone_pcp_update
, zone
, NULL
);
6115 mutex_unlock(&pcp_batch_high_lock
);
6119 void zone_pcp_reset(struct zone
*zone
)
6121 unsigned long flags
;
6123 struct per_cpu_pageset
*pset
;
6125 /* avoid races with drain_pages() */
6126 local_irq_save(flags
);
6127 if (zone
->pageset
!= &boot_pageset
) {
6128 for_each_online_cpu(cpu
) {
6129 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6130 drain_zonestat(zone
, pset
);
6132 free_percpu(zone
->pageset
);
6133 zone
->pageset
= &boot_pageset
;
6135 local_irq_restore(flags
);
6138 #ifdef CONFIG_MEMORY_HOTREMOVE
6140 * All pages in the range must be isolated before calling this.
6143 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6149 unsigned long flags
;
6150 /* find the first valid pfn */
6151 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6156 zone
= page_zone(pfn_to_page(pfn
));
6157 spin_lock_irqsave(&zone
->lock
, flags
);
6159 while (pfn
< end_pfn
) {
6160 if (!pfn_valid(pfn
)) {
6164 page
= pfn_to_page(pfn
);
6166 * The HWPoisoned page may be not in buddy system, and
6167 * page_count() is not 0.
6169 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6171 SetPageReserved(page
);
6175 BUG_ON(page_count(page
));
6176 BUG_ON(!PageBuddy(page
));
6177 order
= page_order(page
);
6178 #ifdef CONFIG_DEBUG_VM
6179 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6180 pfn
, 1 << order
, end_pfn
);
6182 list_del(&page
->lru
);
6183 rmv_page_order(page
);
6184 zone
->free_area
[order
].nr_free
--;
6185 for (i
= 0; i
< (1 << order
); i
++)
6186 SetPageReserved((page
+i
));
6187 pfn
+= (1 << order
);
6189 spin_unlock_irqrestore(&zone
->lock
, flags
);
6193 #ifdef CONFIG_MEMORY_FAILURE
6194 bool is_free_buddy_page(struct page
*page
)
6196 struct zone
*zone
= page_zone(page
);
6197 unsigned long pfn
= page_to_pfn(page
);
6198 unsigned long flags
;
6201 spin_lock_irqsave(&zone
->lock
, flags
);
6202 for (order
= 0; order
< MAX_ORDER
; order
++) {
6203 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6205 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6208 spin_unlock_irqrestore(&zone
->lock
, flags
);
6210 return order
< MAX_ORDER
;
6214 static const struct trace_print_flags pageflag_names
[] = {
6215 {1UL << PG_locked
, "locked" },
6216 {1UL << PG_error
, "error" },
6217 {1UL << PG_referenced
, "referenced" },
6218 {1UL << PG_uptodate
, "uptodate" },
6219 {1UL << PG_dirty
, "dirty" },
6220 {1UL << PG_lru
, "lru" },
6221 {1UL << PG_active
, "active" },
6222 {1UL << PG_slab
, "slab" },
6223 {1UL << PG_owner_priv_1
, "owner_priv_1" },
6224 {1UL << PG_arch_1
, "arch_1" },
6225 {1UL << PG_reserved
, "reserved" },
6226 {1UL << PG_private
, "private" },
6227 {1UL << PG_private_2
, "private_2" },
6228 {1UL << PG_writeback
, "writeback" },
6229 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6230 {1UL << PG_head
, "head" },
6231 {1UL << PG_tail
, "tail" },
6233 {1UL << PG_compound
, "compound" },
6235 {1UL << PG_swapcache
, "swapcache" },
6236 {1UL << PG_mappedtodisk
, "mappedtodisk" },
6237 {1UL << PG_reclaim
, "reclaim" },
6238 {1UL << PG_swapbacked
, "swapbacked" },
6239 {1UL << PG_unevictable
, "unevictable" },
6241 {1UL << PG_mlocked
, "mlocked" },
6243 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6244 {1UL << PG_uncached
, "uncached" },
6246 #ifdef CONFIG_MEMORY_FAILURE
6247 {1UL << PG_hwpoison
, "hwpoison" },
6249 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6250 {1UL << PG_compound_lock
, "compound_lock" },
6254 static void dump_page_flags(unsigned long flags
)
6256 const char *delim
= "";
6260 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names
) != __NR_PAGEFLAGS
);
6262 printk(KERN_ALERT
"page flags: %#lx(", flags
);
6264 /* remove zone id */
6265 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
6267 for (i
= 0; i
< ARRAY_SIZE(pageflag_names
) && flags
; i
++) {
6269 mask
= pageflag_names
[i
].mask
;
6270 if ((flags
& mask
) != mask
)
6274 printk("%s%s", delim
, pageflag_names
[i
].name
);
6278 /* check for left over flags */
6280 printk("%s%#lx", delim
, flags
);
6285 void dump_page(struct page
*page
)
6288 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6289 page
, atomic_read(&page
->_count
), page_mapcount(page
),
6290 page
->mapping
, page
->index
);
6291 dump_page_flags(page
->flags
);
6292 mem_cgroup_print_bad_page(page
);