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/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock
);
73 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74 DEFINE_PER_CPU(int, numa_node
);
75 EXPORT_PER_CPU_SYMBOL(numa_node
);
78 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
85 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
86 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
90 * Array of node states.
92 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
93 [N_POSSIBLE
] = NODE_MASK_ALL
,
94 [N_ONLINE
] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY
] = { { [0] = 1UL } },
103 [N_CPU
] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states
);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock
);
111 unsigned long totalram_pages __read_mostly
;
112 unsigned long totalreserve_pages __read_mostly
;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly
;
121 int percpu_pagelist_fraction
;
122 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask
;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex
));
139 if (saved_gfp_mask
) {
140 gfp_allowed_mask
= saved_gfp_mask
;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex
));
148 WARN_ON(saved_gfp_mask
);
149 saved_gfp_mask
= gfp_allowed_mask
;
150 gfp_allowed_mask
&= ~GFP_IOFS
;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly
;
165 static void __free_pages_ok(struct page
*page
, unsigned int order
);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages
);
193 static char * const zone_names
[MAX_NR_ZONES
] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes
= 1024;
208 int user_min_free_kbytes
= -1;
210 static unsigned long __meminitdata nr_kernel_pages
;
211 static unsigned long __meminitdata nr_all_pages
;
212 static unsigned long __meminitdata dma_reserve
;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
217 static unsigned long __initdata required_kernelcore
;
218 static unsigned long __initdata required_movablecore
;
219 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone
);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
228 int nr_online_nodes __read_mostly
= 1;
229 EXPORT_SYMBOL(nr_node_ids
);
230 EXPORT_SYMBOL(nr_online_nodes
);
233 int page_group_by_mobility_disabled __read_mostly
;
235 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
237 if (unlikely(page_group_by_mobility_disabled
&&
238 migratetype
< MIGRATE_PCPTYPES
))
239 migratetype
= MIGRATE_UNMOVABLE
;
241 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
242 PB_migrate
, PB_migrate_end
);
245 bool oom_killer_disabled __read_mostly
;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
252 unsigned long pfn
= page_to_pfn(page
);
253 unsigned long sp
, start_pfn
;
256 seq
= zone_span_seqbegin(zone
);
257 start_pfn
= zone
->zone_start_pfn
;
258 sp
= zone
->spanned_pages
;
259 if (!zone_spans_pfn(zone
, pfn
))
261 } while (zone_span_seqretry(zone
, seq
));
264 pr_err("page %lu outside zone [ %lu - %lu ]\n",
265 pfn
, start_pfn
, start_pfn
+ sp
);
270 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
272 if (!pfn_valid_within(page_to_pfn(page
)))
274 if (zone
!= page_zone(page
))
280 * Temporary debugging check for pages not lying within a given zone.
282 static int bad_range(struct zone
*zone
, struct page
*page
)
284 if (page_outside_zone_boundaries(zone
, page
))
286 if (!page_is_consistent(zone
, page
))
292 static inline int bad_range(struct zone
*zone
, struct page
*page
)
298 static void bad_page(struct page
*page
, char *reason
, unsigned long bad_flags
)
300 static unsigned long resume
;
301 static unsigned long nr_shown
;
302 static unsigned long nr_unshown
;
304 /* Don't complain about poisoned pages */
305 if (PageHWPoison(page
)) {
306 page_mapcount_reset(page
); /* remove PageBuddy */
311 * Allow a burst of 60 reports, then keep quiet for that minute;
312 * or allow a steady drip of one report per second.
314 if (nr_shown
== 60) {
315 if (time_before(jiffies
, resume
)) {
321 "BUG: Bad page state: %lu messages suppressed\n",
328 resume
= jiffies
+ 60 * HZ
;
330 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
331 current
->comm
, page_to_pfn(page
));
332 dump_page_badflags(page
, reason
, bad_flags
);
337 /* Leave bad fields for debug, except PageBuddy could make trouble */
338 page_mapcount_reset(page
); /* remove PageBuddy */
339 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
343 * Higher-order pages are called "compound pages". They are structured thusly:
345 * The first PAGE_SIZE page is called the "head page".
347 * The remaining PAGE_SIZE pages are called "tail pages".
349 * All pages have PG_compound set. All tail pages have their ->first_page
350 * pointing at the head page.
352 * The first tail page's ->lru.next holds the address of the compound page's
353 * put_page() function. Its ->lru.prev holds the order of allocation.
354 * This usage means that zero-order pages may not be compound.
357 static void free_compound_page(struct page
*page
)
359 __free_pages_ok(page
, compound_order(page
));
362 void prep_compound_page(struct page
*page
, unsigned long order
)
365 int nr_pages
= 1 << order
;
367 set_compound_page_dtor(page
, free_compound_page
);
368 set_compound_order(page
, order
);
370 for (i
= 1; i
< nr_pages
; i
++) {
371 struct page
*p
= page
+ i
;
372 set_page_count(p
, 0);
373 p
->first_page
= page
;
374 /* Make sure p->first_page is always valid for PageTail() */
380 /* update __split_huge_page_refcount if you change this function */
381 static int destroy_compound_page(struct page
*page
, unsigned long order
)
384 int nr_pages
= 1 << order
;
387 if (unlikely(compound_order(page
) != order
)) {
388 bad_page(page
, "wrong compound order", 0);
392 __ClearPageHead(page
);
394 for (i
= 1; i
< nr_pages
; i
++) {
395 struct page
*p
= page
+ i
;
397 if (unlikely(!PageTail(p
))) {
398 bad_page(page
, "PageTail not set", 0);
400 } else if (unlikely(p
->first_page
!= page
)) {
401 bad_page(page
, "first_page not consistent", 0);
410 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
415 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
416 * and __GFP_HIGHMEM from hard or soft interrupt context.
418 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
419 for (i
= 0; i
< (1 << order
); i
++)
420 clear_highpage(page
+ i
);
423 #ifdef CONFIG_DEBUG_PAGEALLOC
424 unsigned int _debug_guardpage_minorder
;
426 static int __init
debug_guardpage_minorder_setup(char *buf
)
430 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
431 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
434 _debug_guardpage_minorder
= res
;
435 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
438 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
440 static inline void set_page_guard_flag(struct page
*page
)
442 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
445 static inline void clear_page_guard_flag(struct page
*page
)
447 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
450 static inline void set_page_guard_flag(struct page
*page
) { }
451 static inline void clear_page_guard_flag(struct page
*page
) { }
454 static inline void set_page_order(struct page
*page
, int order
)
456 set_page_private(page
, order
);
457 __SetPageBuddy(page
);
460 static inline void rmv_page_order(struct page
*page
)
462 __ClearPageBuddy(page
);
463 set_page_private(page
, 0);
467 * Locate the struct page for both the matching buddy in our
468 * pair (buddy1) and the combined O(n+1) page they form (page).
470 * 1) Any buddy B1 will have an order O twin B2 which satisfies
471 * the following equation:
473 * For example, if the starting buddy (buddy2) is #8 its order
475 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
477 * 2) Any buddy B will have an order O+1 parent P which
478 * satisfies the following equation:
481 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
483 static inline unsigned long
484 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
486 return page_idx
^ (1 << order
);
490 * This function checks whether a page is free && is the buddy
491 * we can do coalesce a page and its buddy if
492 * (a) the buddy is not in a hole &&
493 * (b) the buddy is in the buddy system &&
494 * (c) a page and its buddy have the same order &&
495 * (d) a page and its buddy are in the same zone.
497 * For recording whether a page is in the buddy system, we set ->_mapcount
498 * PAGE_BUDDY_MAPCOUNT_VALUE.
499 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
500 * serialized by zone->lock.
502 * For recording page's order, we use page_private(page).
504 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
507 if (!pfn_valid_within(page_to_pfn(buddy
)))
510 if (page_zone_id(page
) != page_zone_id(buddy
))
513 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
514 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
518 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
519 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
526 * Freeing function for a buddy system allocator.
528 * The concept of a buddy system is to maintain direct-mapped table
529 * (containing bit values) for memory blocks of various "orders".
530 * The bottom level table contains the map for the smallest allocatable
531 * units of memory (here, pages), and each level above it describes
532 * pairs of units from the levels below, hence, "buddies".
533 * At a high level, all that happens here is marking the table entry
534 * at the bottom level available, and propagating the changes upward
535 * as necessary, plus some accounting needed to play nicely with other
536 * parts of the VM system.
537 * At each level, we keep a list of pages, which are heads of continuous
538 * free pages of length of (1 << order) and marked with _mapcount
539 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
541 * So when we are allocating or freeing one, we can derive the state of the
542 * other. That is, if we allocate a small block, and both were
543 * free, the remainder of the region must be split into blocks.
544 * If a block is freed, and its buddy is also free, then this
545 * triggers coalescing into a block of larger size.
550 static inline void __free_one_page(struct page
*page
,
551 struct zone
*zone
, unsigned int order
,
554 unsigned long page_idx
;
555 unsigned long combined_idx
;
556 unsigned long uninitialized_var(buddy_idx
);
559 VM_BUG_ON(!zone_is_initialized(zone
));
561 if (unlikely(PageCompound(page
)))
562 if (unlikely(destroy_compound_page(page
, order
)))
565 VM_BUG_ON(migratetype
== -1);
567 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
569 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
570 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
572 while (order
< MAX_ORDER
-1) {
573 buddy_idx
= __find_buddy_index(page_idx
, order
);
574 buddy
= page
+ (buddy_idx
- page_idx
);
575 if (!page_is_buddy(page
, buddy
, order
))
578 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
579 * merge with it and move up one order.
581 if (page_is_guard(buddy
)) {
582 clear_page_guard_flag(buddy
);
583 set_page_private(page
, 0);
584 __mod_zone_freepage_state(zone
, 1 << order
,
587 list_del(&buddy
->lru
);
588 zone
->free_area
[order
].nr_free
--;
589 rmv_page_order(buddy
);
591 combined_idx
= buddy_idx
& page_idx
;
592 page
= page
+ (combined_idx
- page_idx
);
593 page_idx
= combined_idx
;
596 set_page_order(page
, order
);
599 * If this is not the largest possible page, check if the buddy
600 * of the next-highest order is free. If it is, it's possible
601 * that pages are being freed that will coalesce soon. In case,
602 * that is happening, add the free page to the tail of the list
603 * so it's less likely to be used soon and more likely to be merged
604 * as a higher order page
606 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
607 struct page
*higher_page
, *higher_buddy
;
608 combined_idx
= buddy_idx
& page_idx
;
609 higher_page
= page
+ (combined_idx
- page_idx
);
610 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
611 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
612 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
613 list_add_tail(&page
->lru
,
614 &zone
->free_area
[order
].free_list
[migratetype
]);
619 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
621 zone
->free_area
[order
].nr_free
++;
624 static inline int free_pages_check(struct page
*page
)
626 char *bad_reason
= NULL
;
627 unsigned long bad_flags
= 0;
629 if (unlikely(page_mapcount(page
)))
630 bad_reason
= "nonzero mapcount";
631 if (unlikely(page
->mapping
!= NULL
))
632 bad_reason
= "non-NULL mapping";
633 if (unlikely(atomic_read(&page
->_count
) != 0))
634 bad_reason
= "nonzero _count";
635 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
636 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
637 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
639 if (unlikely(mem_cgroup_bad_page_check(page
)))
640 bad_reason
= "cgroup check failed";
641 if (unlikely(bad_reason
)) {
642 bad_page(page
, bad_reason
, bad_flags
);
645 page_cpupid_reset_last(page
);
646 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
647 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
652 * Frees a number of pages from the PCP lists
653 * Assumes all pages on list are in same zone, and of same order.
654 * count is the number of pages to free.
656 * If the zone was previously in an "all pages pinned" state then look to
657 * see if this freeing clears that state.
659 * And clear the zone's pages_scanned counter, to hold off the "all pages are
660 * pinned" detection logic.
662 static void free_pcppages_bulk(struct zone
*zone
, int count
,
663 struct per_cpu_pages
*pcp
)
669 spin_lock(&zone
->lock
);
670 zone
->pages_scanned
= 0;
674 struct list_head
*list
;
677 * Remove pages from lists in a round-robin fashion. A
678 * batch_free count is maintained that is incremented when an
679 * empty list is encountered. This is so more pages are freed
680 * off fuller lists instead of spinning excessively around empty
685 if (++migratetype
== MIGRATE_PCPTYPES
)
687 list
= &pcp
->lists
[migratetype
];
688 } while (list_empty(list
));
690 /* This is the only non-empty list. Free them all. */
691 if (batch_free
== MIGRATE_PCPTYPES
)
692 batch_free
= to_free
;
695 int mt
; /* migratetype of the to-be-freed page */
697 page
= list_entry(list
->prev
, struct page
, lru
);
698 /* must delete as __free_one_page list manipulates */
699 list_del(&page
->lru
);
700 mt
= get_freepage_migratetype(page
);
701 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
702 __free_one_page(page
, zone
, 0, mt
);
703 trace_mm_page_pcpu_drain(page
, 0, mt
);
704 if (likely(!is_migrate_isolate_page(page
))) {
705 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1);
706 if (is_migrate_cma(mt
))
707 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
, 1);
709 } while (--to_free
&& --batch_free
&& !list_empty(list
));
711 spin_unlock(&zone
->lock
);
714 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
717 spin_lock(&zone
->lock
);
718 zone
->pages_scanned
= 0;
720 __free_one_page(page
, zone
, order
, migratetype
);
721 if (unlikely(!is_migrate_isolate(migratetype
)))
722 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
723 spin_unlock(&zone
->lock
);
726 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
731 trace_mm_page_free(page
, order
);
732 kmemcheck_free_shadow(page
, order
);
735 page
->mapping
= NULL
;
736 for (i
= 0; i
< (1 << order
); i
++)
737 bad
+= free_pages_check(page
+ i
);
741 if (!PageHighMem(page
)) {
742 debug_check_no_locks_freed(page_address(page
),
744 debug_check_no_obj_freed(page_address(page
),
747 arch_free_page(page
, order
);
748 kernel_map_pages(page
, 1 << order
, 0);
753 static void __free_pages_ok(struct page
*page
, unsigned int order
)
758 if (!free_pages_prepare(page
, order
))
761 local_irq_save(flags
);
762 __count_vm_events(PGFREE
, 1 << order
);
763 migratetype
= get_pageblock_migratetype(page
);
764 set_freepage_migratetype(page
, migratetype
);
765 free_one_page(page_zone(page
), page
, order
, migratetype
);
766 local_irq_restore(flags
);
769 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
771 unsigned int nr_pages
= 1 << order
;
772 struct page
*p
= page
;
776 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
778 __ClearPageReserved(p
);
779 set_page_count(p
, 0);
781 __ClearPageReserved(p
);
782 set_page_count(p
, 0);
784 page_zone(page
)->managed_pages
+= nr_pages
;
785 set_page_refcounted(page
);
786 __free_pages(page
, order
);
790 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
791 void __init
init_cma_reserved_pageblock(struct page
*page
)
793 unsigned i
= pageblock_nr_pages
;
794 struct page
*p
= page
;
797 __ClearPageReserved(p
);
798 set_page_count(p
, 0);
801 set_page_refcounted(page
);
802 set_pageblock_migratetype(page
, MIGRATE_CMA
);
803 __free_pages(page
, pageblock_order
);
804 adjust_managed_page_count(page
, pageblock_nr_pages
);
809 * The order of subdivision here is critical for the IO subsystem.
810 * Please do not alter this order without good reasons and regression
811 * testing. Specifically, as large blocks of memory are subdivided,
812 * the order in which smaller blocks are delivered depends on the order
813 * they're subdivided in this function. This is the primary factor
814 * influencing the order in which pages are delivered to the IO
815 * subsystem according to empirical testing, and this is also justified
816 * by considering the behavior of a buddy system containing a single
817 * large block of memory acted on by a series of small allocations.
818 * This behavior is a critical factor in sglist merging's success.
822 static inline void expand(struct zone
*zone
, struct page
*page
,
823 int low
, int high
, struct free_area
*area
,
826 unsigned long size
= 1 << high
;
832 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
834 #ifdef CONFIG_DEBUG_PAGEALLOC
835 if (high
< debug_guardpage_minorder()) {
837 * Mark as guard pages (or page), that will allow to
838 * merge back to allocator when buddy will be freed.
839 * Corresponding page table entries will not be touched,
840 * pages will stay not present in virtual address space
842 INIT_LIST_HEAD(&page
[size
].lru
);
843 set_page_guard_flag(&page
[size
]);
844 set_page_private(&page
[size
], high
);
845 /* Guard pages are not available for any usage */
846 __mod_zone_freepage_state(zone
, -(1 << high
),
851 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
853 set_page_order(&page
[size
], high
);
858 * This page is about to be returned from the page allocator
860 static inline int check_new_page(struct page
*page
)
862 char *bad_reason
= NULL
;
863 unsigned long bad_flags
= 0;
865 if (unlikely(page_mapcount(page
)))
866 bad_reason
= "nonzero mapcount";
867 if (unlikely(page
->mapping
!= NULL
))
868 bad_reason
= "non-NULL mapping";
869 if (unlikely(atomic_read(&page
->_count
) != 0))
870 bad_reason
= "nonzero _count";
871 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
872 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
873 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
875 if (unlikely(mem_cgroup_bad_page_check(page
)))
876 bad_reason
= "cgroup check failed";
877 if (unlikely(bad_reason
)) {
878 bad_page(page
, bad_reason
, bad_flags
);
884 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
888 for (i
= 0; i
< (1 << order
); i
++) {
889 struct page
*p
= page
+ i
;
890 if (unlikely(check_new_page(p
)))
894 set_page_private(page
, 0);
895 set_page_refcounted(page
);
897 arch_alloc_page(page
, order
);
898 kernel_map_pages(page
, 1 << order
, 1);
900 if (gfp_flags
& __GFP_ZERO
)
901 prep_zero_page(page
, order
, gfp_flags
);
903 if (order
&& (gfp_flags
& __GFP_COMP
))
904 prep_compound_page(page
, order
);
910 * Go through the free lists for the given migratetype and remove
911 * the smallest available page from the freelists
914 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
917 unsigned int current_order
;
918 struct free_area
*area
;
921 /* Find a page of the appropriate size in the preferred list */
922 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
923 area
= &(zone
->free_area
[current_order
]);
924 if (list_empty(&area
->free_list
[migratetype
]))
927 page
= list_entry(area
->free_list
[migratetype
].next
,
929 list_del(&page
->lru
);
930 rmv_page_order(page
);
932 expand(zone
, page
, order
, current_order
, area
, migratetype
);
941 * This array describes the order lists are fallen back to when
942 * the free lists for the desirable migrate type are depleted
944 static int fallbacks
[MIGRATE_TYPES
][4] = {
945 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
946 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
948 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
949 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
951 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
953 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
954 #ifdef CONFIG_MEMORY_ISOLATION
955 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
960 * Move the free pages in a range to the free lists of the requested type.
961 * Note that start_page and end_pages are not aligned on a pageblock
962 * boundary. If alignment is required, use move_freepages_block()
964 int move_freepages(struct zone
*zone
,
965 struct page
*start_page
, struct page
*end_page
,
972 #ifndef CONFIG_HOLES_IN_ZONE
974 * page_zone is not safe to call in this context when
975 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
976 * anyway as we check zone boundaries in move_freepages_block().
977 * Remove at a later date when no bug reports exist related to
978 * grouping pages by mobility
980 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
983 for (page
= start_page
; page
<= end_page
;) {
984 /* Make sure we are not inadvertently changing nodes */
985 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
987 if (!pfn_valid_within(page_to_pfn(page
))) {
992 if (!PageBuddy(page
)) {
997 order
= page_order(page
);
998 list_move(&page
->lru
,
999 &zone
->free_area
[order
].free_list
[migratetype
]);
1000 set_freepage_migratetype(page
, migratetype
);
1002 pages_moved
+= 1 << order
;
1008 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1011 unsigned long start_pfn
, end_pfn
;
1012 struct page
*start_page
, *end_page
;
1014 start_pfn
= page_to_pfn(page
);
1015 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1016 start_page
= pfn_to_page(start_pfn
);
1017 end_page
= start_page
+ pageblock_nr_pages
- 1;
1018 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1020 /* Do not cross zone boundaries */
1021 if (!zone_spans_pfn(zone
, start_pfn
))
1023 if (!zone_spans_pfn(zone
, end_pfn
))
1026 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1029 static void change_pageblock_range(struct page
*pageblock_page
,
1030 int start_order
, int migratetype
)
1032 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1034 while (nr_pageblocks
--) {
1035 set_pageblock_migratetype(pageblock_page
, migratetype
);
1036 pageblock_page
+= pageblock_nr_pages
;
1041 * If breaking a large block of pages, move all free pages to the preferred
1042 * allocation list. If falling back for a reclaimable kernel allocation, be
1043 * more aggressive about taking ownership of free pages.
1045 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1046 * nor move CMA pages to different free lists. We don't want unmovable pages
1047 * to be allocated from MIGRATE_CMA areas.
1049 * Returns the new migratetype of the pageblock (or the same old migratetype
1050 * if it was unchanged).
1052 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1053 int start_type
, int fallback_type
)
1055 int current_order
= page_order(page
);
1058 * When borrowing from MIGRATE_CMA, we need to release the excess
1059 * buddy pages to CMA itself.
1061 if (is_migrate_cma(fallback_type
))
1062 return fallback_type
;
1064 /* Take ownership for orders >= pageblock_order */
1065 if (current_order
>= pageblock_order
) {
1066 change_pageblock_range(page
, current_order
, start_type
);
1070 if (current_order
>= pageblock_order
/ 2 ||
1071 start_type
== MIGRATE_RECLAIMABLE
||
1072 page_group_by_mobility_disabled
) {
1075 pages
= move_freepages_block(zone
, page
, start_type
);
1077 /* Claim the whole block if over half of it is free */
1078 if (pages
>= (1 << (pageblock_order
-1)) ||
1079 page_group_by_mobility_disabled
) {
1081 set_pageblock_migratetype(page
, start_type
);
1087 return fallback_type
;
1090 /* Remove an element from the buddy allocator from the fallback list */
1091 static inline struct page
*
1092 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
1094 struct free_area
*area
;
1097 int migratetype
, new_type
, i
;
1099 /* Find the largest possible block of pages in the other list */
1100 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
1103 migratetype
= fallbacks
[start_migratetype
][i
];
1105 /* MIGRATE_RESERVE handled later if necessary */
1106 if (migratetype
== MIGRATE_RESERVE
)
1109 area
= &(zone
->free_area
[current_order
]);
1110 if (list_empty(&area
->free_list
[migratetype
]))
1113 page
= list_entry(area
->free_list
[migratetype
].next
,
1117 new_type
= try_to_steal_freepages(zone
, page
,
1121 /* Remove the page from the freelists */
1122 list_del(&page
->lru
);
1123 rmv_page_order(page
);
1125 expand(zone
, page
, order
, current_order
, area
,
1128 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1129 start_migratetype
, migratetype
, new_type
);
1139 * Do the hard work of removing an element from the buddy allocator.
1140 * Call me with the zone->lock already held.
1142 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1148 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1150 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1151 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1154 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1155 * is used because __rmqueue_smallest is an inline function
1156 * and we want just one call site
1159 migratetype
= MIGRATE_RESERVE
;
1164 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1169 * Obtain a specified number of elements from the buddy allocator, all under
1170 * a single hold of the lock, for efficiency. Add them to the supplied list.
1171 * Returns the number of new pages which were placed at *list.
1173 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1174 unsigned long count
, struct list_head
*list
,
1175 int migratetype
, int cold
)
1177 int mt
= migratetype
, i
;
1179 spin_lock(&zone
->lock
);
1180 for (i
= 0; i
< count
; ++i
) {
1181 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1182 if (unlikely(page
== NULL
))
1186 * Split buddy pages returned by expand() are received here
1187 * in physical page order. The page is added to the callers and
1188 * list and the list head then moves forward. From the callers
1189 * perspective, the linked list is ordered by page number in
1190 * some conditions. This is useful for IO devices that can
1191 * merge IO requests if the physical pages are ordered
1194 if (likely(cold
== 0))
1195 list_add(&page
->lru
, list
);
1197 list_add_tail(&page
->lru
, list
);
1198 if (IS_ENABLED(CONFIG_CMA
)) {
1199 mt
= get_pageblock_migratetype(page
);
1200 if (!is_migrate_cma(mt
) && !is_migrate_isolate(mt
))
1203 set_freepage_migratetype(page
, mt
);
1205 if (is_migrate_cma(mt
))
1206 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1209 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1210 spin_unlock(&zone
->lock
);
1216 * Called from the vmstat counter updater to drain pagesets of this
1217 * currently executing processor on remote nodes after they have
1220 * Note that this function must be called with the thread pinned to
1221 * a single processor.
1223 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1225 unsigned long flags
;
1227 unsigned long batch
;
1229 local_irq_save(flags
);
1230 batch
= ACCESS_ONCE(pcp
->batch
);
1231 if (pcp
->count
>= batch
)
1234 to_drain
= pcp
->count
;
1236 free_pcppages_bulk(zone
, to_drain
, pcp
);
1237 pcp
->count
-= to_drain
;
1239 local_irq_restore(flags
);
1244 * Drain pages of the indicated processor.
1246 * The processor must either be the current processor and the
1247 * thread pinned to the current processor or a processor that
1250 static void drain_pages(unsigned int cpu
)
1252 unsigned long flags
;
1255 for_each_populated_zone(zone
) {
1256 struct per_cpu_pageset
*pset
;
1257 struct per_cpu_pages
*pcp
;
1259 local_irq_save(flags
);
1260 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1264 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1267 local_irq_restore(flags
);
1272 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1274 void drain_local_pages(void *arg
)
1276 drain_pages(smp_processor_id());
1280 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1282 * Note that this code is protected against sending an IPI to an offline
1283 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1284 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1285 * nothing keeps CPUs from showing up after we populated the cpumask and
1286 * before the call to on_each_cpu_mask().
1288 void drain_all_pages(void)
1291 struct per_cpu_pageset
*pcp
;
1295 * Allocate in the BSS so we wont require allocation in
1296 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1298 static cpumask_t cpus_with_pcps
;
1301 * We don't care about racing with CPU hotplug event
1302 * as offline notification will cause the notified
1303 * cpu to drain that CPU pcps and on_each_cpu_mask
1304 * disables preemption as part of its processing
1306 for_each_online_cpu(cpu
) {
1307 bool has_pcps
= false;
1308 for_each_populated_zone(zone
) {
1309 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1310 if (pcp
->pcp
.count
) {
1316 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1318 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1320 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1323 #ifdef CONFIG_HIBERNATION
1325 void mark_free_pages(struct zone
*zone
)
1327 unsigned long pfn
, max_zone_pfn
;
1328 unsigned long flags
;
1330 struct list_head
*curr
;
1332 if (zone_is_empty(zone
))
1335 spin_lock_irqsave(&zone
->lock
, flags
);
1337 max_zone_pfn
= zone_end_pfn(zone
);
1338 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1339 if (pfn_valid(pfn
)) {
1340 struct page
*page
= pfn_to_page(pfn
);
1342 if (!swsusp_page_is_forbidden(page
))
1343 swsusp_unset_page_free(page
);
1346 for_each_migratetype_order(order
, t
) {
1347 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1350 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1351 for (i
= 0; i
< (1UL << order
); i
++)
1352 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1355 spin_unlock_irqrestore(&zone
->lock
, flags
);
1357 #endif /* CONFIG_PM */
1360 * Free a 0-order page
1361 * cold == 1 ? free a cold page : free a hot page
1363 void free_hot_cold_page(struct page
*page
, int cold
)
1365 struct zone
*zone
= page_zone(page
);
1366 struct per_cpu_pages
*pcp
;
1367 unsigned long flags
;
1370 if (!free_pages_prepare(page
, 0))
1373 migratetype
= get_pageblock_migratetype(page
);
1374 set_freepage_migratetype(page
, migratetype
);
1375 local_irq_save(flags
);
1376 __count_vm_event(PGFREE
);
1379 * We only track unmovable, reclaimable and movable on pcp lists.
1380 * Free ISOLATE pages back to the allocator because they are being
1381 * offlined but treat RESERVE as movable pages so we can get those
1382 * areas back if necessary. Otherwise, we may have to free
1383 * excessively into the page allocator
1385 if (migratetype
>= MIGRATE_PCPTYPES
) {
1386 if (unlikely(is_migrate_isolate(migratetype
))) {
1387 free_one_page(zone
, page
, 0, migratetype
);
1390 migratetype
= MIGRATE_MOVABLE
;
1393 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1395 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1397 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1399 if (pcp
->count
>= pcp
->high
) {
1400 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1401 free_pcppages_bulk(zone
, batch
, pcp
);
1402 pcp
->count
-= batch
;
1406 local_irq_restore(flags
);
1410 * Free a list of 0-order pages
1412 void free_hot_cold_page_list(struct list_head
*list
, int cold
)
1414 struct page
*page
, *next
;
1416 list_for_each_entry_safe(page
, next
, list
, lru
) {
1417 trace_mm_page_free_batched(page
, cold
);
1418 free_hot_cold_page(page
, cold
);
1423 * split_page takes a non-compound higher-order page, and splits it into
1424 * n (1<<order) sub-pages: page[0..n]
1425 * Each sub-page must be freed individually.
1427 * Note: this is probably too low level an operation for use in drivers.
1428 * Please consult with lkml before using this in your driver.
1430 void split_page(struct page
*page
, unsigned int order
)
1434 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1435 VM_BUG_ON_PAGE(!page_count(page
), page
);
1437 #ifdef CONFIG_KMEMCHECK
1439 * Split shadow pages too, because free(page[0]) would
1440 * otherwise free the whole shadow.
1442 if (kmemcheck_page_is_tracked(page
))
1443 split_page(virt_to_page(page
[0].shadow
), order
);
1446 for (i
= 1; i
< (1 << order
); i
++)
1447 set_page_refcounted(page
+ i
);
1449 EXPORT_SYMBOL_GPL(split_page
);
1451 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1453 unsigned long watermark
;
1457 BUG_ON(!PageBuddy(page
));
1459 zone
= page_zone(page
);
1460 mt
= get_pageblock_migratetype(page
);
1462 if (!is_migrate_isolate(mt
)) {
1463 /* Obey watermarks as if the page was being allocated */
1464 watermark
= low_wmark_pages(zone
) + (1 << order
);
1465 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1468 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1471 /* Remove page from free list */
1472 list_del(&page
->lru
);
1473 zone
->free_area
[order
].nr_free
--;
1474 rmv_page_order(page
);
1476 /* Set the pageblock if the isolated page is at least a pageblock */
1477 if (order
>= pageblock_order
- 1) {
1478 struct page
*endpage
= page
+ (1 << order
) - 1;
1479 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1480 int mt
= get_pageblock_migratetype(page
);
1481 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1482 set_pageblock_migratetype(page
,
1487 return 1UL << order
;
1491 * Similar to split_page except the page is already free. As this is only
1492 * being used for migration, the migratetype of the block also changes.
1493 * As this is called with interrupts disabled, the caller is responsible
1494 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1497 * Note: this is probably too low level an operation for use in drivers.
1498 * Please consult with lkml before using this in your driver.
1500 int split_free_page(struct page
*page
)
1505 order
= page_order(page
);
1507 nr_pages
= __isolate_free_page(page
, order
);
1511 /* Split into individual pages */
1512 set_page_refcounted(page
);
1513 split_page(page
, order
);
1518 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1519 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1523 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1524 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1527 unsigned long flags
;
1529 int cold
= !!(gfp_flags
& __GFP_COLD
);
1532 if (likely(order
== 0)) {
1533 struct per_cpu_pages
*pcp
;
1534 struct list_head
*list
;
1536 local_irq_save(flags
);
1537 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1538 list
= &pcp
->lists
[migratetype
];
1539 if (list_empty(list
)) {
1540 pcp
->count
+= rmqueue_bulk(zone
, 0,
1543 if (unlikely(list_empty(list
)))
1548 page
= list_entry(list
->prev
, struct page
, lru
);
1550 page
= list_entry(list
->next
, struct page
, lru
);
1552 list_del(&page
->lru
);
1555 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1557 * __GFP_NOFAIL is not to be used in new code.
1559 * All __GFP_NOFAIL callers should be fixed so that they
1560 * properly detect and handle allocation failures.
1562 * We most definitely don't want callers attempting to
1563 * allocate greater than order-1 page units with
1566 WARN_ON_ONCE(order
> 1);
1568 spin_lock_irqsave(&zone
->lock
, flags
);
1569 page
= __rmqueue(zone
, order
, migratetype
);
1570 spin_unlock(&zone
->lock
);
1573 __mod_zone_freepage_state(zone
, -(1 << order
),
1574 get_pageblock_migratetype(page
));
1577 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1578 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1579 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1580 local_irq_restore(flags
);
1582 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1583 if (prep_new_page(page
, order
, gfp_flags
))
1588 local_irq_restore(flags
);
1592 #ifdef CONFIG_FAIL_PAGE_ALLOC
1595 struct fault_attr attr
;
1597 u32 ignore_gfp_highmem
;
1598 u32 ignore_gfp_wait
;
1600 } fail_page_alloc
= {
1601 .attr
= FAULT_ATTR_INITIALIZER
,
1602 .ignore_gfp_wait
= 1,
1603 .ignore_gfp_highmem
= 1,
1607 static int __init
setup_fail_page_alloc(char *str
)
1609 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1611 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1613 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1615 if (order
< fail_page_alloc
.min_order
)
1617 if (gfp_mask
& __GFP_NOFAIL
)
1619 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1621 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1624 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1627 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1629 static int __init
fail_page_alloc_debugfs(void)
1631 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1634 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1635 &fail_page_alloc
.attr
);
1637 return PTR_ERR(dir
);
1639 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1640 &fail_page_alloc
.ignore_gfp_wait
))
1642 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1643 &fail_page_alloc
.ignore_gfp_highmem
))
1645 if (!debugfs_create_u32("min-order", mode
, dir
,
1646 &fail_page_alloc
.min_order
))
1651 debugfs_remove_recursive(dir
);
1656 late_initcall(fail_page_alloc_debugfs
);
1658 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1660 #else /* CONFIG_FAIL_PAGE_ALLOC */
1662 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1667 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1670 * Return true if free pages are above 'mark'. This takes into account the order
1671 * of the allocation.
1673 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1674 int classzone_idx
, int alloc_flags
, long free_pages
)
1676 /* free_pages my go negative - that's OK */
1678 long lowmem_reserve
= z
->lowmem_reserve
[classzone_idx
];
1682 free_pages
-= (1 << order
) - 1;
1683 if (alloc_flags
& ALLOC_HIGH
)
1685 if (alloc_flags
& ALLOC_HARDER
)
1688 /* If allocation can't use CMA areas don't use free CMA pages */
1689 if (!(alloc_flags
& ALLOC_CMA
))
1690 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1693 if (free_pages
- free_cma
<= min
+ lowmem_reserve
)
1695 for (o
= 0; o
< order
; o
++) {
1696 /* At the next order, this order's pages become unavailable */
1697 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1699 /* Require fewer higher order pages to be free */
1702 if (free_pages
<= min
)
1708 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1709 int classzone_idx
, int alloc_flags
)
1711 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1712 zone_page_state(z
, NR_FREE_PAGES
));
1715 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1716 int classzone_idx
, int alloc_flags
)
1718 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1720 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1721 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1723 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1729 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1730 * skip over zones that are not allowed by the cpuset, or that have
1731 * been recently (in last second) found to be nearly full. See further
1732 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1733 * that have to skip over a lot of full or unallowed zones.
1735 * If the zonelist cache is present in the passed zonelist, then
1736 * returns a pointer to the allowed node mask (either the current
1737 * tasks mems_allowed, or node_states[N_MEMORY].)
1739 * If the zonelist cache is not available for this zonelist, does
1740 * nothing and returns NULL.
1742 * If the fullzones BITMAP in the zonelist cache is stale (more than
1743 * a second since last zap'd) then we zap it out (clear its bits.)
1745 * We hold off even calling zlc_setup, until after we've checked the
1746 * first zone in the zonelist, on the theory that most allocations will
1747 * be satisfied from that first zone, so best to examine that zone as
1748 * quickly as we can.
1750 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1752 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1753 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1755 zlc
= zonelist
->zlcache_ptr
;
1759 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1760 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1761 zlc
->last_full_zap
= jiffies
;
1764 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1765 &cpuset_current_mems_allowed
:
1766 &node_states
[N_MEMORY
];
1767 return allowednodes
;
1771 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1772 * if it is worth looking at further for free memory:
1773 * 1) Check that the zone isn't thought to be full (doesn't have its
1774 * bit set in the zonelist_cache fullzones BITMAP).
1775 * 2) Check that the zones node (obtained from the zonelist_cache
1776 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1777 * Return true (non-zero) if zone is worth looking at further, or
1778 * else return false (zero) if it is not.
1780 * This check -ignores- the distinction between various watermarks,
1781 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1782 * found to be full for any variation of these watermarks, it will
1783 * be considered full for up to one second by all requests, unless
1784 * we are so low on memory on all allowed nodes that we are forced
1785 * into the second scan of the zonelist.
1787 * In the second scan we ignore this zonelist cache and exactly
1788 * apply the watermarks to all zones, even it is slower to do so.
1789 * We are low on memory in the second scan, and should leave no stone
1790 * unturned looking for a free page.
1792 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1793 nodemask_t
*allowednodes
)
1795 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1796 int i
; /* index of *z in zonelist zones */
1797 int n
; /* node that zone *z is on */
1799 zlc
= zonelist
->zlcache_ptr
;
1803 i
= z
- zonelist
->_zonerefs
;
1806 /* This zone is worth trying if it is allowed but not full */
1807 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1811 * Given 'z' scanning a zonelist, set the corresponding bit in
1812 * zlc->fullzones, so that subsequent attempts to allocate a page
1813 * from that zone don't waste time re-examining it.
1815 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1817 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1818 int i
; /* index of *z in zonelist zones */
1820 zlc
= zonelist
->zlcache_ptr
;
1824 i
= z
- zonelist
->_zonerefs
;
1826 set_bit(i
, zlc
->fullzones
);
1830 * clear all zones full, called after direct reclaim makes progress so that
1831 * a zone that was recently full is not skipped over for up to a second
1833 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1835 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1837 zlc
= zonelist
->zlcache_ptr
;
1841 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1844 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1846 return local_zone
->node
== zone
->node
;
1849 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1851 return node_isset(local_zone
->node
, zone
->zone_pgdat
->reclaim_nodes
);
1854 static void __paginginit
init_zone_allows_reclaim(int nid
)
1858 for_each_online_node(i
)
1859 if (node_distance(nid
, i
) <= RECLAIM_DISTANCE
)
1860 node_set(i
, NODE_DATA(nid
)->reclaim_nodes
);
1862 zone_reclaim_mode
= 1;
1865 #else /* CONFIG_NUMA */
1867 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1872 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1873 nodemask_t
*allowednodes
)
1878 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1882 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1886 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1891 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1896 static inline void init_zone_allows_reclaim(int nid
)
1899 #endif /* CONFIG_NUMA */
1902 * get_page_from_freelist goes through the zonelist trying to allocate
1905 static struct page
*
1906 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1907 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1908 struct zone
*preferred_zone
, int migratetype
)
1911 struct page
*page
= NULL
;
1914 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1915 int zlc_active
= 0; /* set if using zonelist_cache */
1916 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1918 classzone_idx
= zone_idx(preferred_zone
);
1921 * Scan zonelist, looking for a zone with enough free.
1922 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1924 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1925 high_zoneidx
, nodemask
) {
1928 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1929 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1931 if ((alloc_flags
& ALLOC_CPUSET
) &&
1932 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1934 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1935 if (unlikely(alloc_flags
& ALLOC_NO_WATERMARKS
))
1938 * Distribute pages in proportion to the individual
1939 * zone size to ensure fair page aging. The zone a
1940 * page was allocated in should have no effect on the
1941 * time the page has in memory before being reclaimed.
1943 * Try to stay in local zones in the fastpath. If
1944 * that fails, the slowpath is entered, which will do
1945 * another pass starting with the local zones, but
1946 * ultimately fall back to remote zones that do not
1947 * partake in the fairness round-robin cycle of this
1950 if (alloc_flags
& ALLOC_WMARK_LOW
) {
1951 if (zone_page_state(zone
, NR_ALLOC_BATCH
) <= 0)
1953 if (!zone_local(preferred_zone
, zone
))
1957 * When allocating a page cache page for writing, we
1958 * want to get it from a zone that is within its dirty
1959 * limit, such that no single zone holds more than its
1960 * proportional share of globally allowed dirty pages.
1961 * The dirty limits take into account the zone's
1962 * lowmem reserves and high watermark so that kswapd
1963 * should be able to balance it without having to
1964 * write pages from its LRU list.
1966 * This may look like it could increase pressure on
1967 * lower zones by failing allocations in higher zones
1968 * before they are full. But the pages that do spill
1969 * over are limited as the lower zones are protected
1970 * by this very same mechanism. It should not become
1971 * a practical burden to them.
1973 * XXX: For now, allow allocations to potentially
1974 * exceed the per-zone dirty limit in the slowpath
1975 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1976 * which is important when on a NUMA setup the allowed
1977 * zones are together not big enough to reach the
1978 * global limit. The proper fix for these situations
1979 * will require awareness of zones in the
1980 * dirty-throttling and the flusher threads.
1982 if ((alloc_flags
& ALLOC_WMARK_LOW
) &&
1983 (gfp_mask
& __GFP_WRITE
) && !zone_dirty_ok(zone
))
1984 goto this_zone_full
;
1986 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1987 if (!zone_watermark_ok(zone
, order
, mark
,
1988 classzone_idx
, alloc_flags
)) {
1991 if (IS_ENABLED(CONFIG_NUMA
) &&
1992 !did_zlc_setup
&& nr_online_nodes
> 1) {
1994 * we do zlc_setup if there are multiple nodes
1995 * and before considering the first zone allowed
1998 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
2003 if (zone_reclaim_mode
== 0 ||
2004 !zone_allows_reclaim(preferred_zone
, zone
))
2005 goto this_zone_full
;
2008 * As we may have just activated ZLC, check if the first
2009 * eligible zone has failed zone_reclaim recently.
2011 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
2012 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2015 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2017 case ZONE_RECLAIM_NOSCAN
:
2020 case ZONE_RECLAIM_FULL
:
2021 /* scanned but unreclaimable */
2024 /* did we reclaim enough */
2025 if (zone_watermark_ok(zone
, order
, mark
,
2026 classzone_idx
, alloc_flags
))
2030 * Failed to reclaim enough to meet watermark.
2031 * Only mark the zone full if checking the min
2032 * watermark or if we failed to reclaim just
2033 * 1<<order pages or else the page allocator
2034 * fastpath will prematurely mark zones full
2035 * when the watermark is between the low and
2038 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2039 ret
== ZONE_RECLAIM_SOME
)
2040 goto this_zone_full
;
2047 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2048 gfp_mask
, migratetype
);
2052 if (IS_ENABLED(CONFIG_NUMA
))
2053 zlc_mark_zone_full(zonelist
, z
);
2056 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && page
== NULL
&& zlc_active
)) {
2057 /* Disable zlc cache for second zonelist scan */
2064 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2065 * necessary to allocate the page. The expectation is
2066 * that the caller is taking steps that will free more
2067 * memory. The caller should avoid the page being used
2068 * for !PFMEMALLOC purposes.
2070 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2076 * Large machines with many possible nodes should not always dump per-node
2077 * meminfo in irq context.
2079 static inline bool should_suppress_show_mem(void)
2084 ret
= in_interrupt();
2089 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2090 DEFAULT_RATELIMIT_INTERVAL
,
2091 DEFAULT_RATELIMIT_BURST
);
2093 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2095 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2097 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2098 debug_guardpage_minorder() > 0)
2102 * This documents exceptions given to allocations in certain
2103 * contexts that are allowed to allocate outside current's set
2106 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2107 if (test_thread_flag(TIF_MEMDIE
) ||
2108 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2109 filter
&= ~SHOW_MEM_FILTER_NODES
;
2110 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2111 filter
&= ~SHOW_MEM_FILTER_NODES
;
2114 struct va_format vaf
;
2117 va_start(args
, fmt
);
2122 pr_warn("%pV", &vaf
);
2127 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2128 current
->comm
, order
, gfp_mask
);
2131 if (!should_suppress_show_mem())
2136 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2137 unsigned long did_some_progress
,
2138 unsigned long pages_reclaimed
)
2140 /* Do not loop if specifically requested */
2141 if (gfp_mask
& __GFP_NORETRY
)
2144 /* Always retry if specifically requested */
2145 if (gfp_mask
& __GFP_NOFAIL
)
2149 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2150 * making forward progress without invoking OOM. Suspend also disables
2151 * storage devices so kswapd will not help. Bail if we are suspending.
2153 if (!did_some_progress
&& pm_suspended_storage())
2157 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2158 * means __GFP_NOFAIL, but that may not be true in other
2161 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2165 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2166 * specified, then we retry until we no longer reclaim any pages
2167 * (above), or we've reclaimed an order of pages at least as
2168 * large as the allocation's order. In both cases, if the
2169 * allocation still fails, we stop retrying.
2171 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2177 static inline struct page
*
2178 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2179 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2180 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2185 /* Acquire the OOM killer lock for the zones in zonelist */
2186 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
2187 schedule_timeout_uninterruptible(1);
2192 * Go through the zonelist yet one more time, keep very high watermark
2193 * here, this is only to catch a parallel oom killing, we must fail if
2194 * we're still under heavy pressure.
2196 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2197 order
, zonelist
, high_zoneidx
,
2198 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2199 preferred_zone
, migratetype
);
2203 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2204 /* The OOM killer will not help higher order allocs */
2205 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2207 /* The OOM killer does not needlessly kill tasks for lowmem */
2208 if (high_zoneidx
< ZONE_NORMAL
)
2211 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2212 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2213 * The caller should handle page allocation failure by itself if
2214 * it specifies __GFP_THISNODE.
2215 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2217 if (gfp_mask
& __GFP_THISNODE
)
2220 /* Exhausted what can be done so it's blamo time */
2221 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2224 clear_zonelist_oom(zonelist
, gfp_mask
);
2228 #ifdef CONFIG_COMPACTION
2229 /* Try memory compaction for high-order allocations before reclaim */
2230 static struct page
*
2231 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2232 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2233 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2234 int migratetype
, bool sync_migration
,
2235 bool *contended_compaction
, bool *deferred_compaction
,
2236 unsigned long *did_some_progress
)
2241 if (compaction_deferred(preferred_zone
, order
)) {
2242 *deferred_compaction
= true;
2246 current
->flags
|= PF_MEMALLOC
;
2247 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2248 nodemask
, sync_migration
,
2249 contended_compaction
);
2250 current
->flags
&= ~PF_MEMALLOC
;
2252 if (*did_some_progress
!= COMPACT_SKIPPED
) {
2255 /* Page migration frees to the PCP lists but we want merging */
2256 drain_pages(get_cpu());
2259 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2260 order
, zonelist
, high_zoneidx
,
2261 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2262 preferred_zone
, migratetype
);
2264 preferred_zone
->compact_blockskip_flush
= false;
2265 compaction_defer_reset(preferred_zone
, order
, true);
2266 count_vm_event(COMPACTSUCCESS
);
2271 * It's bad if compaction run occurs and fails.
2272 * The most likely reason is that pages exist,
2273 * but not enough to satisfy watermarks.
2275 count_vm_event(COMPACTFAIL
);
2278 * As async compaction considers a subset of pageblocks, only
2279 * defer if the failure was a sync compaction failure.
2282 defer_compaction(preferred_zone
, order
);
2290 static inline struct page
*
2291 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2292 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2293 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2294 int migratetype
, bool sync_migration
,
2295 bool *contended_compaction
, bool *deferred_compaction
,
2296 unsigned long *did_some_progress
)
2300 #endif /* CONFIG_COMPACTION */
2302 /* Perform direct synchronous page reclaim */
2304 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2305 nodemask_t
*nodemask
)
2307 struct reclaim_state reclaim_state
;
2312 /* We now go into synchronous reclaim */
2313 cpuset_memory_pressure_bump();
2314 current
->flags
|= PF_MEMALLOC
;
2315 lockdep_set_current_reclaim_state(gfp_mask
);
2316 reclaim_state
.reclaimed_slab
= 0;
2317 current
->reclaim_state
= &reclaim_state
;
2319 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2321 current
->reclaim_state
= NULL
;
2322 lockdep_clear_current_reclaim_state();
2323 current
->flags
&= ~PF_MEMALLOC
;
2330 /* The really slow allocator path where we enter direct reclaim */
2331 static inline struct page
*
2332 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2333 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2334 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2335 int migratetype
, unsigned long *did_some_progress
)
2337 struct page
*page
= NULL
;
2338 bool drained
= false;
2340 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2342 if (unlikely(!(*did_some_progress
)))
2345 /* After successful reclaim, reconsider all zones for allocation */
2346 if (IS_ENABLED(CONFIG_NUMA
))
2347 zlc_clear_zones_full(zonelist
);
2350 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2351 zonelist
, high_zoneidx
,
2352 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2353 preferred_zone
, migratetype
);
2356 * If an allocation failed after direct reclaim, it could be because
2357 * pages are pinned on the per-cpu lists. Drain them and try again
2359 if (!page
&& !drained
) {
2369 * This is called in the allocator slow-path if the allocation request is of
2370 * sufficient urgency to ignore watermarks and take other desperate measures
2372 static inline struct page
*
2373 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2374 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2375 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2381 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2382 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2383 preferred_zone
, migratetype
);
2385 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2386 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2387 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2392 static void prepare_slowpath(gfp_t gfp_mask
, unsigned int order
,
2393 struct zonelist
*zonelist
,
2394 enum zone_type high_zoneidx
,
2395 struct zone
*preferred_zone
)
2400 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
2401 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2402 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2404 * Only reset the batches of zones that were actually
2405 * considered in the fast path, we don't want to
2406 * thrash fairness information for zones that are not
2407 * actually part of this zonelist's round-robin cycle.
2409 if (!zone_local(preferred_zone
, zone
))
2411 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2412 high_wmark_pages(zone
) -
2413 low_wmark_pages(zone
) -
2414 zone_page_state(zone
, NR_ALLOC_BATCH
));
2419 gfp_to_alloc_flags(gfp_t gfp_mask
)
2421 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2422 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2424 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2425 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2428 * The caller may dip into page reserves a bit more if the caller
2429 * cannot run direct reclaim, or if the caller has realtime scheduling
2430 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2431 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2433 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2437 * Not worth trying to allocate harder for
2438 * __GFP_NOMEMALLOC even if it can't schedule.
2440 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2441 alloc_flags
|= ALLOC_HARDER
;
2443 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2444 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2446 alloc_flags
&= ~ALLOC_CPUSET
;
2447 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2448 alloc_flags
|= ALLOC_HARDER
;
2450 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2451 if (gfp_mask
& __GFP_MEMALLOC
)
2452 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2453 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2454 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2455 else if (!in_interrupt() &&
2456 ((current
->flags
& PF_MEMALLOC
) ||
2457 unlikely(test_thread_flag(TIF_MEMDIE
))))
2458 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2461 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2462 alloc_flags
|= ALLOC_CMA
;
2467 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2469 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2472 static inline struct page
*
2473 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2474 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2475 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2478 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2479 struct page
*page
= NULL
;
2481 unsigned long pages_reclaimed
= 0;
2482 unsigned long did_some_progress
;
2483 bool sync_migration
= false;
2484 bool deferred_compaction
= false;
2485 bool contended_compaction
= false;
2488 * In the slowpath, we sanity check order to avoid ever trying to
2489 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2490 * be using allocators in order of preference for an area that is
2493 if (order
>= MAX_ORDER
) {
2494 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2499 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2500 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2501 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2502 * using a larger set of nodes after it has established that the
2503 * allowed per node queues are empty and that nodes are
2506 if (IS_ENABLED(CONFIG_NUMA
) &&
2507 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2511 prepare_slowpath(gfp_mask
, order
, zonelist
,
2512 high_zoneidx
, preferred_zone
);
2515 * OK, we're below the kswapd watermark and have kicked background
2516 * reclaim. Now things get more complex, so set up alloc_flags according
2517 * to how we want to proceed.
2519 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2522 * Find the true preferred zone if the allocation is unconstrained by
2525 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2526 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2530 /* This is the last chance, in general, before the goto nopage. */
2531 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2532 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2533 preferred_zone
, migratetype
);
2537 /* Allocate without watermarks if the context allows */
2538 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2540 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2541 * the allocation is high priority and these type of
2542 * allocations are system rather than user orientated
2544 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2546 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2547 zonelist
, high_zoneidx
, nodemask
,
2548 preferred_zone
, migratetype
);
2554 /* Atomic allocations - we can't balance anything */
2557 * All existing users of the deprecated __GFP_NOFAIL are
2558 * blockable, so warn of any new users that actually allow this
2559 * type of allocation to fail.
2561 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2565 /* Avoid recursion of direct reclaim */
2566 if (current
->flags
& PF_MEMALLOC
)
2569 /* Avoid allocations with no watermarks from looping endlessly */
2570 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2574 * Try direct compaction. The first pass is asynchronous. Subsequent
2575 * attempts after direct reclaim are synchronous
2577 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2578 zonelist
, high_zoneidx
,
2580 alloc_flags
, preferred_zone
,
2581 migratetype
, sync_migration
,
2582 &contended_compaction
,
2583 &deferred_compaction
,
2584 &did_some_progress
);
2587 sync_migration
= true;
2590 * If compaction is deferred for high-order allocations, it is because
2591 * sync compaction recently failed. In this is the case and the caller
2592 * requested a movable allocation that does not heavily disrupt the
2593 * system then fail the allocation instead of entering direct reclaim.
2595 if ((deferred_compaction
|| contended_compaction
) &&
2596 (gfp_mask
& __GFP_NO_KSWAPD
))
2599 /* Try direct reclaim and then allocating */
2600 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2601 zonelist
, high_zoneidx
,
2603 alloc_flags
, preferred_zone
,
2604 migratetype
, &did_some_progress
);
2609 * If we failed to make any progress reclaiming, then we are
2610 * running out of options and have to consider going OOM
2612 if (!did_some_progress
) {
2613 if (oom_gfp_allowed(gfp_mask
)) {
2614 if (oom_killer_disabled
)
2616 /* Coredumps can quickly deplete all memory reserves */
2617 if ((current
->flags
& PF_DUMPCORE
) &&
2618 !(gfp_mask
& __GFP_NOFAIL
))
2620 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2621 zonelist
, high_zoneidx
,
2622 nodemask
, preferred_zone
,
2627 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2629 * The oom killer is not called for high-order
2630 * allocations that may fail, so if no progress
2631 * is being made, there are no other options and
2632 * retrying is unlikely to help.
2634 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2637 * The oom killer is not called for lowmem
2638 * allocations to prevent needlessly killing
2641 if (high_zoneidx
< ZONE_NORMAL
)
2649 /* Check if we should retry the allocation */
2650 pages_reclaimed
+= did_some_progress
;
2651 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2653 /* Wait for some write requests to complete then retry */
2654 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2658 * High-order allocations do not necessarily loop after
2659 * direct reclaim and reclaim/compaction depends on compaction
2660 * being called after reclaim so call directly if necessary
2662 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2663 zonelist
, high_zoneidx
,
2665 alloc_flags
, preferred_zone
,
2666 migratetype
, sync_migration
,
2667 &contended_compaction
,
2668 &deferred_compaction
,
2669 &did_some_progress
);
2675 warn_alloc_failed(gfp_mask
, order
, NULL
);
2678 if (kmemcheck_enabled
)
2679 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2685 * This is the 'heart' of the zoned buddy allocator.
2688 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2689 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2691 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2692 struct zone
*preferred_zone
;
2693 struct page
*page
= NULL
;
2694 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2695 unsigned int cpuset_mems_cookie
;
2696 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
;
2697 struct mem_cgroup
*memcg
= NULL
;
2699 gfp_mask
&= gfp_allowed_mask
;
2701 lockdep_trace_alloc(gfp_mask
);
2703 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2705 if (should_fail_alloc_page(gfp_mask
, order
))
2709 * Check the zones suitable for the gfp_mask contain at least one
2710 * valid zone. It's possible to have an empty zonelist as a result
2711 * of GFP_THISNODE and a memoryless node
2713 if (unlikely(!zonelist
->_zonerefs
->zone
))
2717 * Will only have any effect when __GFP_KMEMCG is set. This is
2718 * verified in the (always inline) callee
2720 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2724 cpuset_mems_cookie
= get_mems_allowed();
2726 /* The preferred zone is used for statistics later */
2727 first_zones_zonelist(zonelist
, high_zoneidx
,
2728 nodemask
? : &cpuset_current_mems_allowed
,
2730 if (!preferred_zone
)
2734 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2735 alloc_flags
|= ALLOC_CMA
;
2737 /* First allocation attempt */
2738 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2739 zonelist
, high_zoneidx
, alloc_flags
,
2740 preferred_zone
, migratetype
);
2741 if (unlikely(!page
)) {
2743 * Runtime PM, block IO and its error handling path
2744 * can deadlock because I/O on the device might not
2747 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2748 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2749 zonelist
, high_zoneidx
, nodemask
,
2750 preferred_zone
, migratetype
);
2753 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2757 * When updating a task's mems_allowed, it is possible to race with
2758 * parallel threads in such a way that an allocation can fail while
2759 * the mask is being updated. If a page allocation is about to fail,
2760 * check if the cpuset changed during allocation and if so, retry.
2762 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2765 memcg_kmem_commit_charge(page
, memcg
, order
);
2769 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2772 * Common helper functions.
2774 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2779 * __get_free_pages() returns a 32-bit address, which cannot represent
2782 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2784 page
= alloc_pages(gfp_mask
, order
);
2787 return (unsigned long) page_address(page
);
2789 EXPORT_SYMBOL(__get_free_pages
);
2791 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2793 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2795 EXPORT_SYMBOL(get_zeroed_page
);
2797 void __free_pages(struct page
*page
, unsigned int order
)
2799 if (put_page_testzero(page
)) {
2801 free_hot_cold_page(page
, 0);
2803 __free_pages_ok(page
, order
);
2807 EXPORT_SYMBOL(__free_pages
);
2809 void free_pages(unsigned long addr
, unsigned int order
)
2812 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2813 __free_pages(virt_to_page((void *)addr
), order
);
2817 EXPORT_SYMBOL(free_pages
);
2820 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2821 * pages allocated with __GFP_KMEMCG.
2823 * Those pages are accounted to a particular memcg, embedded in the
2824 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2825 * for that information only to find out that it is NULL for users who have no
2826 * interest in that whatsoever, we provide these functions.
2828 * The caller knows better which flags it relies on.
2830 void __free_memcg_kmem_pages(struct page
*page
, unsigned int order
)
2832 memcg_kmem_uncharge_pages(page
, order
);
2833 __free_pages(page
, order
);
2836 void free_memcg_kmem_pages(unsigned long addr
, unsigned int order
)
2839 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2840 __free_memcg_kmem_pages(virt_to_page((void *)addr
), order
);
2844 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2847 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2848 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2850 split_page(virt_to_page((void *)addr
), order
);
2851 while (used
< alloc_end
) {
2856 return (void *)addr
;
2860 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2861 * @size: the number of bytes to allocate
2862 * @gfp_mask: GFP flags for the allocation
2864 * This function is similar to alloc_pages(), except that it allocates the
2865 * minimum number of pages to satisfy the request. alloc_pages() can only
2866 * allocate memory in power-of-two pages.
2868 * This function is also limited by MAX_ORDER.
2870 * Memory allocated by this function must be released by free_pages_exact().
2872 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2874 unsigned int order
= get_order(size
);
2877 addr
= __get_free_pages(gfp_mask
, order
);
2878 return make_alloc_exact(addr
, order
, size
);
2880 EXPORT_SYMBOL(alloc_pages_exact
);
2883 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2885 * @nid: the preferred node ID where memory should be allocated
2886 * @size: the number of bytes to allocate
2887 * @gfp_mask: GFP flags for the allocation
2889 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2891 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2894 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2896 unsigned order
= get_order(size
);
2897 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2900 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2902 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2905 * free_pages_exact - release memory allocated via alloc_pages_exact()
2906 * @virt: the value returned by alloc_pages_exact.
2907 * @size: size of allocation, same value as passed to alloc_pages_exact().
2909 * Release the memory allocated by a previous call to alloc_pages_exact.
2911 void free_pages_exact(void *virt
, size_t size
)
2913 unsigned long addr
= (unsigned long)virt
;
2914 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2916 while (addr
< end
) {
2921 EXPORT_SYMBOL(free_pages_exact
);
2924 * nr_free_zone_pages - count number of pages beyond high watermark
2925 * @offset: The zone index of the highest zone
2927 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2928 * high watermark within all zones at or below a given zone index. For each
2929 * zone, the number of pages is calculated as:
2930 * managed_pages - high_pages
2932 static unsigned long nr_free_zone_pages(int offset
)
2937 /* Just pick one node, since fallback list is circular */
2938 unsigned long sum
= 0;
2940 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2942 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2943 unsigned long size
= zone
->managed_pages
;
2944 unsigned long high
= high_wmark_pages(zone
);
2953 * nr_free_buffer_pages - count number of pages beyond high watermark
2955 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2956 * watermark within ZONE_DMA and ZONE_NORMAL.
2958 unsigned long nr_free_buffer_pages(void)
2960 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2962 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2965 * nr_free_pagecache_pages - count number of pages beyond high watermark
2967 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2968 * high watermark within all zones.
2970 unsigned long nr_free_pagecache_pages(void)
2972 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2975 static inline void show_node(struct zone
*zone
)
2977 if (IS_ENABLED(CONFIG_NUMA
))
2978 printk("Node %d ", zone_to_nid(zone
));
2981 void si_meminfo(struct sysinfo
*val
)
2983 val
->totalram
= totalram_pages
;
2985 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2986 val
->bufferram
= nr_blockdev_pages();
2987 val
->totalhigh
= totalhigh_pages
;
2988 val
->freehigh
= nr_free_highpages();
2989 val
->mem_unit
= PAGE_SIZE
;
2992 EXPORT_SYMBOL(si_meminfo
);
2995 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2997 int zone_type
; /* needs to be signed */
2998 unsigned long managed_pages
= 0;
2999 pg_data_t
*pgdat
= NODE_DATA(nid
);
3001 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3002 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3003 val
->totalram
= managed_pages
;
3004 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3005 #ifdef CONFIG_HIGHMEM
3006 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3007 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3013 val
->mem_unit
= PAGE_SIZE
;
3018 * Determine whether the node should be displayed or not, depending on whether
3019 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3021 bool skip_free_areas_node(unsigned int flags
, int nid
)
3024 unsigned int cpuset_mems_cookie
;
3026 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3030 cpuset_mems_cookie
= get_mems_allowed();
3031 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3032 } while (!put_mems_allowed(cpuset_mems_cookie
));
3037 #define K(x) ((x) << (PAGE_SHIFT-10))
3039 static void show_migration_types(unsigned char type
)
3041 static const char types
[MIGRATE_TYPES
] = {
3042 [MIGRATE_UNMOVABLE
] = 'U',
3043 [MIGRATE_RECLAIMABLE
] = 'E',
3044 [MIGRATE_MOVABLE
] = 'M',
3045 [MIGRATE_RESERVE
] = 'R',
3047 [MIGRATE_CMA
] = 'C',
3049 #ifdef CONFIG_MEMORY_ISOLATION
3050 [MIGRATE_ISOLATE
] = 'I',
3053 char tmp
[MIGRATE_TYPES
+ 1];
3057 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3058 if (type
& (1 << i
))
3063 printk("(%s) ", tmp
);
3067 * Show free area list (used inside shift_scroll-lock stuff)
3068 * We also calculate the percentage fragmentation. We do this by counting the
3069 * memory on each free list with the exception of the first item on the list.
3070 * Suppresses nodes that are not allowed by current's cpuset if
3071 * SHOW_MEM_FILTER_NODES is passed.
3073 void show_free_areas(unsigned int filter
)
3078 for_each_populated_zone(zone
) {
3079 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3082 printk("%s per-cpu:\n", zone
->name
);
3084 for_each_online_cpu(cpu
) {
3085 struct per_cpu_pageset
*pageset
;
3087 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3089 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3090 cpu
, pageset
->pcp
.high
,
3091 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3095 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3096 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3098 " dirty:%lu writeback:%lu unstable:%lu\n"
3099 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3100 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3102 global_page_state(NR_ACTIVE_ANON
),
3103 global_page_state(NR_INACTIVE_ANON
),
3104 global_page_state(NR_ISOLATED_ANON
),
3105 global_page_state(NR_ACTIVE_FILE
),
3106 global_page_state(NR_INACTIVE_FILE
),
3107 global_page_state(NR_ISOLATED_FILE
),
3108 global_page_state(NR_UNEVICTABLE
),
3109 global_page_state(NR_FILE_DIRTY
),
3110 global_page_state(NR_WRITEBACK
),
3111 global_page_state(NR_UNSTABLE_NFS
),
3112 global_page_state(NR_FREE_PAGES
),
3113 global_page_state(NR_SLAB_RECLAIMABLE
),
3114 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3115 global_page_state(NR_FILE_MAPPED
),
3116 global_page_state(NR_SHMEM
),
3117 global_page_state(NR_PAGETABLE
),
3118 global_page_state(NR_BOUNCE
),
3119 global_page_state(NR_FREE_CMA_PAGES
));
3121 for_each_populated_zone(zone
) {
3124 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3132 " active_anon:%lukB"
3133 " inactive_anon:%lukB"
3134 " active_file:%lukB"
3135 " inactive_file:%lukB"
3136 " unevictable:%lukB"
3137 " isolated(anon):%lukB"
3138 " isolated(file):%lukB"
3146 " slab_reclaimable:%lukB"
3147 " slab_unreclaimable:%lukB"
3148 " kernel_stack:%lukB"
3153 " writeback_tmp:%lukB"
3154 " pages_scanned:%lu"
3155 " all_unreclaimable? %s"
3158 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3159 K(min_wmark_pages(zone
)),
3160 K(low_wmark_pages(zone
)),
3161 K(high_wmark_pages(zone
)),
3162 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3163 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3164 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3165 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3166 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3167 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3168 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3169 K(zone
->present_pages
),
3170 K(zone
->managed_pages
),
3171 K(zone_page_state(zone
, NR_MLOCK
)),
3172 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3173 K(zone_page_state(zone
, NR_WRITEBACK
)),
3174 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3175 K(zone_page_state(zone
, NR_SHMEM
)),
3176 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3177 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3178 zone_page_state(zone
, NR_KERNEL_STACK
) *
3180 K(zone_page_state(zone
, NR_PAGETABLE
)),
3181 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3182 K(zone_page_state(zone
, NR_BOUNCE
)),
3183 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3184 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3185 zone
->pages_scanned
,
3186 (!zone_reclaimable(zone
) ? "yes" : "no")
3188 printk("lowmem_reserve[]:");
3189 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3190 printk(" %lu", zone
->lowmem_reserve
[i
]);
3194 for_each_populated_zone(zone
) {
3195 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3196 unsigned char types
[MAX_ORDER
];
3198 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3201 printk("%s: ", zone
->name
);
3203 spin_lock_irqsave(&zone
->lock
, flags
);
3204 for (order
= 0; order
< MAX_ORDER
; order
++) {
3205 struct free_area
*area
= &zone
->free_area
[order
];
3208 nr
[order
] = area
->nr_free
;
3209 total
+= nr
[order
] << order
;
3212 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3213 if (!list_empty(&area
->free_list
[type
]))
3214 types
[order
] |= 1 << type
;
3217 spin_unlock_irqrestore(&zone
->lock
, flags
);
3218 for (order
= 0; order
< MAX_ORDER
; order
++) {
3219 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3221 show_migration_types(types
[order
]);
3223 printk("= %lukB\n", K(total
));
3226 hugetlb_show_meminfo();
3228 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3230 show_swap_cache_info();
3233 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3235 zoneref
->zone
= zone
;
3236 zoneref
->zone_idx
= zone_idx(zone
);
3240 * Builds allocation fallback zone lists.
3242 * Add all populated zones of a node to the zonelist.
3244 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3248 enum zone_type zone_type
= MAX_NR_ZONES
;
3252 zone
= pgdat
->node_zones
+ zone_type
;
3253 if (populated_zone(zone
)) {
3254 zoneref_set_zone(zone
,
3255 &zonelist
->_zonerefs
[nr_zones
++]);
3256 check_highest_zone(zone_type
);
3258 } while (zone_type
);
3266 * 0 = automatic detection of better ordering.
3267 * 1 = order by ([node] distance, -zonetype)
3268 * 2 = order by (-zonetype, [node] distance)
3270 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3271 * the same zonelist. So only NUMA can configure this param.
3273 #define ZONELIST_ORDER_DEFAULT 0
3274 #define ZONELIST_ORDER_NODE 1
3275 #define ZONELIST_ORDER_ZONE 2
3277 /* zonelist order in the kernel.
3278 * set_zonelist_order() will set this to NODE or ZONE.
3280 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3281 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3285 /* The value user specified ....changed by config */
3286 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3287 /* string for sysctl */
3288 #define NUMA_ZONELIST_ORDER_LEN 16
3289 char numa_zonelist_order
[16] = "default";
3292 * interface for configure zonelist ordering.
3293 * command line option "numa_zonelist_order"
3294 * = "[dD]efault - default, automatic configuration.
3295 * = "[nN]ode - order by node locality, then by zone within node
3296 * = "[zZ]one - order by zone, then by locality within zone
3299 static int __parse_numa_zonelist_order(char *s
)
3301 if (*s
== 'd' || *s
== 'D') {
3302 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3303 } else if (*s
== 'n' || *s
== 'N') {
3304 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3305 } else if (*s
== 'z' || *s
== 'Z') {
3306 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3309 "Ignoring invalid numa_zonelist_order value: "
3316 static __init
int setup_numa_zonelist_order(char *s
)
3323 ret
= __parse_numa_zonelist_order(s
);
3325 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3329 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3332 * sysctl handler for numa_zonelist_order
3334 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
3335 void __user
*buffer
, size_t *length
,
3338 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3340 static DEFINE_MUTEX(zl_order_mutex
);
3342 mutex_lock(&zl_order_mutex
);
3344 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3348 strcpy(saved_string
, (char *)table
->data
);
3350 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3354 int oldval
= user_zonelist_order
;
3356 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3359 * bogus value. restore saved string
3361 strncpy((char *)table
->data
, saved_string
,
3362 NUMA_ZONELIST_ORDER_LEN
);
3363 user_zonelist_order
= oldval
;
3364 } else if (oldval
!= user_zonelist_order
) {
3365 mutex_lock(&zonelists_mutex
);
3366 build_all_zonelists(NULL
, NULL
);
3367 mutex_unlock(&zonelists_mutex
);
3371 mutex_unlock(&zl_order_mutex
);
3376 #define MAX_NODE_LOAD (nr_online_nodes)
3377 static int node_load
[MAX_NUMNODES
];
3380 * find_next_best_node - find the next node that should appear in a given node's fallback list
3381 * @node: node whose fallback list we're appending
3382 * @used_node_mask: nodemask_t of already used nodes
3384 * We use a number of factors to determine which is the next node that should
3385 * appear on a given node's fallback list. The node should not have appeared
3386 * already in @node's fallback list, and it should be the next closest node
3387 * according to the distance array (which contains arbitrary distance values
3388 * from each node to each node in the system), and should also prefer nodes
3389 * with no CPUs, since presumably they'll have very little allocation pressure
3390 * on them otherwise.
3391 * It returns -1 if no node is found.
3393 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3396 int min_val
= INT_MAX
;
3397 int best_node
= NUMA_NO_NODE
;
3398 const struct cpumask
*tmp
= cpumask_of_node(0);
3400 /* Use the local node if we haven't already */
3401 if (!node_isset(node
, *used_node_mask
)) {
3402 node_set(node
, *used_node_mask
);
3406 for_each_node_state(n
, N_MEMORY
) {
3408 /* Don't want a node to appear more than once */
3409 if (node_isset(n
, *used_node_mask
))
3412 /* Use the distance array to find the distance */
3413 val
= node_distance(node
, n
);
3415 /* Penalize nodes under us ("prefer the next node") */
3418 /* Give preference to headless and unused nodes */
3419 tmp
= cpumask_of_node(n
);
3420 if (!cpumask_empty(tmp
))
3421 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3423 /* Slight preference for less loaded node */
3424 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3425 val
+= node_load
[n
];
3427 if (val
< min_val
) {
3434 node_set(best_node
, *used_node_mask
);
3441 * Build zonelists ordered by node and zones within node.
3442 * This results in maximum locality--normal zone overflows into local
3443 * DMA zone, if any--but risks exhausting DMA zone.
3445 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3448 struct zonelist
*zonelist
;
3450 zonelist
= &pgdat
->node_zonelists
[0];
3451 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3453 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3454 zonelist
->_zonerefs
[j
].zone
= NULL
;
3455 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3459 * Build gfp_thisnode zonelists
3461 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3464 struct zonelist
*zonelist
;
3466 zonelist
= &pgdat
->node_zonelists
[1];
3467 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3468 zonelist
->_zonerefs
[j
].zone
= NULL
;
3469 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3473 * Build zonelists ordered by zone and nodes within zones.
3474 * This results in conserving DMA zone[s] until all Normal memory is
3475 * exhausted, but results in overflowing to remote node while memory
3476 * may still exist in local DMA zone.
3478 static int node_order
[MAX_NUMNODES
];
3480 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3483 int zone_type
; /* needs to be signed */
3485 struct zonelist
*zonelist
;
3487 zonelist
= &pgdat
->node_zonelists
[0];
3489 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3490 for (j
= 0; j
< nr_nodes
; j
++) {
3491 node
= node_order
[j
];
3492 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3493 if (populated_zone(z
)) {
3495 &zonelist
->_zonerefs
[pos
++]);
3496 check_highest_zone(zone_type
);
3500 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3501 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3504 static int default_zonelist_order(void)
3507 unsigned long low_kmem_size
, total_size
;
3511 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3512 * If they are really small and used heavily, the system can fall
3513 * into OOM very easily.
3514 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3516 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3519 for_each_online_node(nid
) {
3520 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3521 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3522 if (populated_zone(z
)) {
3523 if (zone_type
< ZONE_NORMAL
)
3524 low_kmem_size
+= z
->managed_pages
;
3525 total_size
+= z
->managed_pages
;
3526 } else if (zone_type
== ZONE_NORMAL
) {
3528 * If any node has only lowmem, then node order
3529 * is preferred to allow kernel allocations
3530 * locally; otherwise, they can easily infringe
3531 * on other nodes when there is an abundance of
3532 * lowmem available to allocate from.
3534 return ZONELIST_ORDER_NODE
;
3538 if (!low_kmem_size
|| /* there are no DMA area. */
3539 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
3540 return ZONELIST_ORDER_NODE
;
3542 * look into each node's config.
3543 * If there is a node whose DMA/DMA32 memory is very big area on
3544 * local memory, NODE_ORDER may be suitable.
3546 average_size
= total_size
/
3547 (nodes_weight(node_states
[N_MEMORY
]) + 1);
3548 for_each_online_node(nid
) {
3551 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3552 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3553 if (populated_zone(z
)) {
3554 if (zone_type
< ZONE_NORMAL
)
3555 low_kmem_size
+= z
->present_pages
;
3556 total_size
+= z
->present_pages
;
3559 if (low_kmem_size
&&
3560 total_size
> average_size
&& /* ignore small node */
3561 low_kmem_size
> total_size
* 70/100)
3562 return ZONELIST_ORDER_NODE
;
3564 return ZONELIST_ORDER_ZONE
;
3567 static void set_zonelist_order(void)
3569 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3570 current_zonelist_order
= default_zonelist_order();
3572 current_zonelist_order
= user_zonelist_order
;
3575 static void build_zonelists(pg_data_t
*pgdat
)
3579 nodemask_t used_mask
;
3580 int local_node
, prev_node
;
3581 struct zonelist
*zonelist
;
3582 int order
= current_zonelist_order
;
3584 /* initialize zonelists */
3585 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3586 zonelist
= pgdat
->node_zonelists
+ i
;
3587 zonelist
->_zonerefs
[0].zone
= NULL
;
3588 zonelist
->_zonerefs
[0].zone_idx
= 0;
3591 /* NUMA-aware ordering of nodes */
3592 local_node
= pgdat
->node_id
;
3593 load
= nr_online_nodes
;
3594 prev_node
= local_node
;
3595 nodes_clear(used_mask
);
3597 memset(node_order
, 0, sizeof(node_order
));
3600 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3602 * We don't want to pressure a particular node.
3603 * So adding penalty to the first node in same
3604 * distance group to make it round-robin.
3606 if (node_distance(local_node
, node
) !=
3607 node_distance(local_node
, prev_node
))
3608 node_load
[node
] = load
;
3612 if (order
== ZONELIST_ORDER_NODE
)
3613 build_zonelists_in_node_order(pgdat
, node
);
3615 node_order
[j
++] = node
; /* remember order */
3618 if (order
== ZONELIST_ORDER_ZONE
) {
3619 /* calculate node order -- i.e., DMA last! */
3620 build_zonelists_in_zone_order(pgdat
, j
);
3623 build_thisnode_zonelists(pgdat
);
3626 /* Construct the zonelist performance cache - see further mmzone.h */
3627 static void build_zonelist_cache(pg_data_t
*pgdat
)
3629 struct zonelist
*zonelist
;
3630 struct zonelist_cache
*zlc
;
3633 zonelist
= &pgdat
->node_zonelists
[0];
3634 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3635 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3636 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3637 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3640 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3642 * Return node id of node used for "local" allocations.
3643 * I.e., first node id of first zone in arg node's generic zonelist.
3644 * Used for initializing percpu 'numa_mem', which is used primarily
3645 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3647 int local_memory_node(int node
)
3651 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3652 gfp_zone(GFP_KERNEL
),
3659 #else /* CONFIG_NUMA */
3661 static void set_zonelist_order(void)
3663 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3666 static void build_zonelists(pg_data_t
*pgdat
)
3668 int node
, local_node
;
3670 struct zonelist
*zonelist
;
3672 local_node
= pgdat
->node_id
;
3674 zonelist
= &pgdat
->node_zonelists
[0];
3675 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3678 * Now we build the zonelist so that it contains the zones
3679 * of all the other nodes.
3680 * We don't want to pressure a particular node, so when
3681 * building the zones for node N, we make sure that the
3682 * zones coming right after the local ones are those from
3683 * node N+1 (modulo N)
3685 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3686 if (!node_online(node
))
3688 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3690 for (node
= 0; node
< local_node
; node
++) {
3691 if (!node_online(node
))
3693 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3696 zonelist
->_zonerefs
[j
].zone
= NULL
;
3697 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3700 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3701 static void build_zonelist_cache(pg_data_t
*pgdat
)
3703 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3706 #endif /* CONFIG_NUMA */
3709 * Boot pageset table. One per cpu which is going to be used for all
3710 * zones and all nodes. The parameters will be set in such a way
3711 * that an item put on a list will immediately be handed over to
3712 * the buddy list. This is safe since pageset manipulation is done
3713 * with interrupts disabled.
3715 * The boot_pagesets must be kept even after bootup is complete for
3716 * unused processors and/or zones. They do play a role for bootstrapping
3717 * hotplugged processors.
3719 * zoneinfo_show() and maybe other functions do
3720 * not check if the processor is online before following the pageset pointer.
3721 * Other parts of the kernel may not check if the zone is available.
3723 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3724 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3725 static void setup_zone_pageset(struct zone
*zone
);
3728 * Global mutex to protect against size modification of zonelists
3729 * as well as to serialize pageset setup for the new populated zone.
3731 DEFINE_MUTEX(zonelists_mutex
);
3733 /* return values int ....just for stop_machine() */
3734 static int __build_all_zonelists(void *data
)
3738 pg_data_t
*self
= data
;
3741 memset(node_load
, 0, sizeof(node_load
));
3744 if (self
&& !node_online(self
->node_id
)) {
3745 build_zonelists(self
);
3746 build_zonelist_cache(self
);
3749 for_each_online_node(nid
) {
3750 pg_data_t
*pgdat
= NODE_DATA(nid
);
3752 build_zonelists(pgdat
);
3753 build_zonelist_cache(pgdat
);
3757 * Initialize the boot_pagesets that are going to be used
3758 * for bootstrapping processors. The real pagesets for
3759 * each zone will be allocated later when the per cpu
3760 * allocator is available.
3762 * boot_pagesets are used also for bootstrapping offline
3763 * cpus if the system is already booted because the pagesets
3764 * are needed to initialize allocators on a specific cpu too.
3765 * F.e. the percpu allocator needs the page allocator which
3766 * needs the percpu allocator in order to allocate its pagesets
3767 * (a chicken-egg dilemma).
3769 for_each_possible_cpu(cpu
) {
3770 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3772 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3774 * We now know the "local memory node" for each node--
3775 * i.e., the node of the first zone in the generic zonelist.
3776 * Set up numa_mem percpu variable for on-line cpus. During
3777 * boot, only the boot cpu should be on-line; we'll init the
3778 * secondary cpus' numa_mem as they come on-line. During
3779 * node/memory hotplug, we'll fixup all on-line cpus.
3781 if (cpu_online(cpu
))
3782 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3790 * Called with zonelists_mutex held always
3791 * unless system_state == SYSTEM_BOOTING.
3793 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3795 set_zonelist_order();
3797 if (system_state
== SYSTEM_BOOTING
) {
3798 __build_all_zonelists(NULL
);
3799 mminit_verify_zonelist();
3800 cpuset_init_current_mems_allowed();
3802 #ifdef CONFIG_MEMORY_HOTPLUG
3804 setup_zone_pageset(zone
);
3806 /* we have to stop all cpus to guarantee there is no user
3808 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3809 /* cpuset refresh routine should be here */
3811 vm_total_pages
= nr_free_pagecache_pages();
3813 * Disable grouping by mobility if the number of pages in the
3814 * system is too low to allow the mechanism to work. It would be
3815 * more accurate, but expensive to check per-zone. This check is
3816 * made on memory-hotadd so a system can start with mobility
3817 * disabled and enable it later
3819 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3820 page_group_by_mobility_disabled
= 1;
3822 page_group_by_mobility_disabled
= 0;
3824 printk("Built %i zonelists in %s order, mobility grouping %s. "
3825 "Total pages: %ld\n",
3827 zonelist_order_name
[current_zonelist_order
],
3828 page_group_by_mobility_disabled
? "off" : "on",
3831 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3836 * Helper functions to size the waitqueue hash table.
3837 * Essentially these want to choose hash table sizes sufficiently
3838 * large so that collisions trying to wait on pages are rare.
3839 * But in fact, the number of active page waitqueues on typical
3840 * systems is ridiculously low, less than 200. So this is even
3841 * conservative, even though it seems large.
3843 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3844 * waitqueues, i.e. the size of the waitq table given the number of pages.
3846 #define PAGES_PER_WAITQUEUE 256
3848 #ifndef CONFIG_MEMORY_HOTPLUG
3849 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3851 unsigned long size
= 1;
3853 pages
/= PAGES_PER_WAITQUEUE
;
3855 while (size
< pages
)
3859 * Once we have dozens or even hundreds of threads sleeping
3860 * on IO we've got bigger problems than wait queue collision.
3861 * Limit the size of the wait table to a reasonable size.
3863 size
= min(size
, 4096UL);
3865 return max(size
, 4UL);
3869 * A zone's size might be changed by hot-add, so it is not possible to determine
3870 * a suitable size for its wait_table. So we use the maximum size now.
3872 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3874 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3875 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3876 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3878 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3879 * or more by the traditional way. (See above). It equals:
3881 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3882 * ia64(16K page size) : = ( 8G + 4M)byte.
3883 * powerpc (64K page size) : = (32G +16M)byte.
3885 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3892 * This is an integer logarithm so that shifts can be used later
3893 * to extract the more random high bits from the multiplicative
3894 * hash function before the remainder is taken.
3896 static inline unsigned long wait_table_bits(unsigned long size
)
3902 * Check if a pageblock contains reserved pages
3904 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3908 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3909 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3916 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3917 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3918 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3919 * higher will lead to a bigger reserve which will get freed as contiguous
3920 * blocks as reclaim kicks in
3922 static void setup_zone_migrate_reserve(struct zone
*zone
)
3924 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3926 unsigned long block_migratetype
;
3931 * Get the start pfn, end pfn and the number of blocks to reserve
3932 * We have to be careful to be aligned to pageblock_nr_pages to
3933 * make sure that we always check pfn_valid for the first page in
3936 start_pfn
= zone
->zone_start_pfn
;
3937 end_pfn
= zone_end_pfn(zone
);
3938 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
3939 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3943 * Reserve blocks are generally in place to help high-order atomic
3944 * allocations that are short-lived. A min_free_kbytes value that
3945 * would result in more than 2 reserve blocks for atomic allocations
3946 * is assumed to be in place to help anti-fragmentation for the
3947 * future allocation of hugepages at runtime.
3949 reserve
= min(2, reserve
);
3950 old_reserve
= zone
->nr_migrate_reserve_block
;
3952 /* When memory hot-add, we almost always need to do nothing */
3953 if (reserve
== old_reserve
)
3955 zone
->nr_migrate_reserve_block
= reserve
;
3957 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3958 if (!pfn_valid(pfn
))
3960 page
= pfn_to_page(pfn
);
3962 /* Watch out for overlapping nodes */
3963 if (page_to_nid(page
) != zone_to_nid(zone
))
3966 block_migratetype
= get_pageblock_migratetype(page
);
3968 /* Only test what is necessary when the reserves are not met */
3971 * Blocks with reserved pages will never free, skip
3974 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
3975 if (pageblock_is_reserved(pfn
, block_end_pfn
))
3978 /* If this block is reserved, account for it */
3979 if (block_migratetype
== MIGRATE_RESERVE
) {
3984 /* Suitable for reserving if this block is movable */
3985 if (block_migratetype
== MIGRATE_MOVABLE
) {
3986 set_pageblock_migratetype(page
,
3988 move_freepages_block(zone
, page
,
3993 } else if (!old_reserve
) {
3995 * At boot time we don't need to scan the whole zone
3996 * for turning off MIGRATE_RESERVE.
4002 * If the reserve is met and this is a previous reserved block,
4005 if (block_migratetype
== MIGRATE_RESERVE
) {
4006 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4007 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4013 * Initially all pages are reserved - free ones are freed
4014 * up by free_all_bootmem() once the early boot process is
4015 * done. Non-atomic initialization, single-pass.
4017 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4018 unsigned long start_pfn
, enum memmap_context context
)
4021 unsigned long end_pfn
= start_pfn
+ size
;
4025 if (highest_memmap_pfn
< end_pfn
- 1)
4026 highest_memmap_pfn
= end_pfn
- 1;
4028 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4029 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4031 * There can be holes in boot-time mem_map[]s
4032 * handed to this function. They do not
4033 * exist on hotplugged memory.
4035 if (context
== MEMMAP_EARLY
) {
4036 if (!early_pfn_valid(pfn
))
4038 if (!early_pfn_in_nid(pfn
, nid
))
4041 page
= pfn_to_page(pfn
);
4042 set_page_links(page
, zone
, nid
, pfn
);
4043 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4044 init_page_count(page
);
4045 page_mapcount_reset(page
);
4046 page_cpupid_reset_last(page
);
4047 SetPageReserved(page
);
4049 * Mark the block movable so that blocks are reserved for
4050 * movable at startup. This will force kernel allocations
4051 * to reserve their blocks rather than leaking throughout
4052 * the address space during boot when many long-lived
4053 * kernel allocations are made. Later some blocks near
4054 * the start are marked MIGRATE_RESERVE by
4055 * setup_zone_migrate_reserve()
4057 * bitmap is created for zone's valid pfn range. but memmap
4058 * can be created for invalid pages (for alignment)
4059 * check here not to call set_pageblock_migratetype() against
4062 if ((z
->zone_start_pfn
<= pfn
)
4063 && (pfn
< zone_end_pfn(z
))
4064 && !(pfn
& (pageblock_nr_pages
- 1)))
4065 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4067 INIT_LIST_HEAD(&page
->lru
);
4068 #ifdef WANT_PAGE_VIRTUAL
4069 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4070 if (!is_highmem_idx(zone
))
4071 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4076 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4079 for_each_migratetype_order(order
, t
) {
4080 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4081 zone
->free_area
[order
].nr_free
= 0;
4085 #ifndef __HAVE_ARCH_MEMMAP_INIT
4086 #define memmap_init(size, nid, zone, start_pfn) \
4087 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4090 static int __meminit
zone_batchsize(struct zone
*zone
)
4096 * The per-cpu-pages pools are set to around 1000th of the
4097 * size of the zone. But no more than 1/2 of a meg.
4099 * OK, so we don't know how big the cache is. So guess.
4101 batch
= zone
->managed_pages
/ 1024;
4102 if (batch
* PAGE_SIZE
> 512 * 1024)
4103 batch
= (512 * 1024) / PAGE_SIZE
;
4104 batch
/= 4; /* We effectively *= 4 below */
4109 * Clamp the batch to a 2^n - 1 value. Having a power
4110 * of 2 value was found to be more likely to have
4111 * suboptimal cache aliasing properties in some cases.
4113 * For example if 2 tasks are alternately allocating
4114 * batches of pages, one task can end up with a lot
4115 * of pages of one half of the possible page colors
4116 * and the other with pages of the other colors.
4118 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4123 /* The deferral and batching of frees should be suppressed under NOMMU
4126 * The problem is that NOMMU needs to be able to allocate large chunks
4127 * of contiguous memory as there's no hardware page translation to
4128 * assemble apparent contiguous memory from discontiguous pages.
4130 * Queueing large contiguous runs of pages for batching, however,
4131 * causes the pages to actually be freed in smaller chunks. As there
4132 * can be a significant delay between the individual batches being
4133 * recycled, this leads to the once large chunks of space being
4134 * fragmented and becoming unavailable for high-order allocations.
4141 * pcp->high and pcp->batch values are related and dependent on one another:
4142 * ->batch must never be higher then ->high.
4143 * The following function updates them in a safe manner without read side
4146 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4147 * those fields changing asynchronously (acording the the above rule).
4149 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4150 * outside of boot time (or some other assurance that no concurrent updaters
4153 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4154 unsigned long batch
)
4156 /* start with a fail safe value for batch */
4160 /* Update high, then batch, in order */
4167 /* a companion to pageset_set_high() */
4168 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4170 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4173 static void pageset_init(struct per_cpu_pageset
*p
)
4175 struct per_cpu_pages
*pcp
;
4178 memset(p
, 0, sizeof(*p
));
4182 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4183 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4186 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4189 pageset_set_batch(p
, batch
);
4193 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4194 * to the value high for the pageset p.
4196 static void pageset_set_high(struct per_cpu_pageset
*p
,
4199 unsigned long batch
= max(1UL, high
/ 4);
4200 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4201 batch
= PAGE_SHIFT
* 8;
4203 pageset_update(&p
->pcp
, high
, batch
);
4206 static void __meminit
pageset_set_high_and_batch(struct zone
*zone
,
4207 struct per_cpu_pageset
*pcp
)
4209 if (percpu_pagelist_fraction
)
4210 pageset_set_high(pcp
,
4211 (zone
->managed_pages
/
4212 percpu_pagelist_fraction
));
4214 pageset_set_batch(pcp
, zone_batchsize(zone
));
4217 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4219 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4222 pageset_set_high_and_batch(zone
, pcp
);
4225 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4228 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4229 for_each_possible_cpu(cpu
)
4230 zone_pageset_init(zone
, cpu
);
4234 * Allocate per cpu pagesets and initialize them.
4235 * Before this call only boot pagesets were available.
4237 void __init
setup_per_cpu_pageset(void)
4241 for_each_populated_zone(zone
)
4242 setup_zone_pageset(zone
);
4245 static noinline __init_refok
4246 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4252 * The per-page waitqueue mechanism uses hashed waitqueues
4255 zone
->wait_table_hash_nr_entries
=
4256 wait_table_hash_nr_entries(zone_size_pages
);
4257 zone
->wait_table_bits
=
4258 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4259 alloc_size
= zone
->wait_table_hash_nr_entries
4260 * sizeof(wait_queue_head_t
);
4262 if (!slab_is_available()) {
4263 zone
->wait_table
= (wait_queue_head_t
*)
4264 memblock_virt_alloc_node_nopanic(
4265 alloc_size
, zone
->zone_pgdat
->node_id
);
4268 * This case means that a zone whose size was 0 gets new memory
4269 * via memory hot-add.
4270 * But it may be the case that a new node was hot-added. In
4271 * this case vmalloc() will not be able to use this new node's
4272 * memory - this wait_table must be initialized to use this new
4273 * node itself as well.
4274 * To use this new node's memory, further consideration will be
4277 zone
->wait_table
= vmalloc(alloc_size
);
4279 if (!zone
->wait_table
)
4282 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4283 init_waitqueue_head(zone
->wait_table
+ i
);
4288 static __meminit
void zone_pcp_init(struct zone
*zone
)
4291 * per cpu subsystem is not up at this point. The following code
4292 * relies on the ability of the linker to provide the
4293 * offset of a (static) per cpu variable into the per cpu area.
4295 zone
->pageset
= &boot_pageset
;
4297 if (populated_zone(zone
))
4298 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4299 zone
->name
, zone
->present_pages
,
4300 zone_batchsize(zone
));
4303 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4304 unsigned long zone_start_pfn
,
4306 enum memmap_context context
)
4308 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4310 ret
= zone_wait_table_init(zone
, size
);
4313 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4315 zone
->zone_start_pfn
= zone_start_pfn
;
4317 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4318 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4320 (unsigned long)zone_idx(zone
),
4321 zone_start_pfn
, (zone_start_pfn
+ size
));
4323 zone_init_free_lists(zone
);
4328 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4329 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4331 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4332 * Architectures may implement their own version but if add_active_range()
4333 * was used and there are no special requirements, this is a convenient
4336 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4338 unsigned long start_pfn
, end_pfn
;
4341 * NOTE: The following SMP-unsafe globals are only used early in boot
4342 * when the kernel is running single-threaded.
4344 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4345 static int __meminitdata last_nid
;
4347 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4350 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4352 last_start_pfn
= start_pfn
;
4353 last_end_pfn
= end_pfn
;
4359 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4361 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4365 nid
= __early_pfn_to_nid(pfn
);
4368 /* just returns 0 */
4372 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4373 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4377 nid
= __early_pfn_to_nid(pfn
);
4378 if (nid
>= 0 && nid
!= node
)
4385 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4386 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4387 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4389 * If an architecture guarantees that all ranges registered with
4390 * add_active_ranges() contain no holes and may be freed, this
4391 * this function may be used instead of calling memblock_free_early_nid()
4394 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4396 unsigned long start_pfn
, end_pfn
;
4399 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4400 start_pfn
= min(start_pfn
, max_low_pfn
);
4401 end_pfn
= min(end_pfn
, max_low_pfn
);
4403 if (start_pfn
< end_pfn
)
4404 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4405 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4411 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4412 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4414 * If an architecture guarantees that all ranges registered with
4415 * add_active_ranges() contain no holes and may be freed, this
4416 * function may be used instead of calling memory_present() manually.
4418 void __init
sparse_memory_present_with_active_regions(int nid
)
4420 unsigned long start_pfn
, end_pfn
;
4423 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4424 memory_present(this_nid
, start_pfn
, end_pfn
);
4428 * get_pfn_range_for_nid - Return the start and end page frames for a node
4429 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4430 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4431 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4433 * It returns the start and end page frame of a node based on information
4434 * provided by an arch calling add_active_range(). If called for a node
4435 * with no available memory, a warning is printed and the start and end
4438 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4439 unsigned long *start_pfn
, unsigned long *end_pfn
)
4441 unsigned long this_start_pfn
, this_end_pfn
;
4447 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4448 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4449 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4452 if (*start_pfn
== -1UL)
4457 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4458 * assumption is made that zones within a node are ordered in monotonic
4459 * increasing memory addresses so that the "highest" populated zone is used
4461 static void __init
find_usable_zone_for_movable(void)
4464 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4465 if (zone_index
== ZONE_MOVABLE
)
4468 if (arch_zone_highest_possible_pfn
[zone_index
] >
4469 arch_zone_lowest_possible_pfn
[zone_index
])
4473 VM_BUG_ON(zone_index
== -1);
4474 movable_zone
= zone_index
;
4478 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4479 * because it is sized independent of architecture. Unlike the other zones,
4480 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4481 * in each node depending on the size of each node and how evenly kernelcore
4482 * is distributed. This helper function adjusts the zone ranges
4483 * provided by the architecture for a given node by using the end of the
4484 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4485 * zones within a node are in order of monotonic increases memory addresses
4487 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4488 unsigned long zone_type
,
4489 unsigned long node_start_pfn
,
4490 unsigned long node_end_pfn
,
4491 unsigned long *zone_start_pfn
,
4492 unsigned long *zone_end_pfn
)
4494 /* Only adjust if ZONE_MOVABLE is on this node */
4495 if (zone_movable_pfn
[nid
]) {
4496 /* Size ZONE_MOVABLE */
4497 if (zone_type
== ZONE_MOVABLE
) {
4498 *zone_start_pfn
= zone_movable_pfn
[nid
];
4499 *zone_end_pfn
= min(node_end_pfn
,
4500 arch_zone_highest_possible_pfn
[movable_zone
]);
4502 /* Adjust for ZONE_MOVABLE starting within this range */
4503 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4504 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4505 *zone_end_pfn
= zone_movable_pfn
[nid
];
4507 /* Check if this whole range is within ZONE_MOVABLE */
4508 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4509 *zone_start_pfn
= *zone_end_pfn
;
4514 * Return the number of pages a zone spans in a node, including holes
4515 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4517 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4518 unsigned long zone_type
,
4519 unsigned long node_start_pfn
,
4520 unsigned long node_end_pfn
,
4521 unsigned long *ignored
)
4523 unsigned long zone_start_pfn
, zone_end_pfn
;
4525 /* Get the start and end of the zone */
4526 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4527 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4528 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4529 node_start_pfn
, node_end_pfn
,
4530 &zone_start_pfn
, &zone_end_pfn
);
4532 /* Check that this node has pages within the zone's required range */
4533 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4536 /* Move the zone boundaries inside the node if necessary */
4537 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4538 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4540 /* Return the spanned pages */
4541 return zone_end_pfn
- zone_start_pfn
;
4545 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4546 * then all holes in the requested range will be accounted for.
4548 unsigned long __meminit
__absent_pages_in_range(int nid
,
4549 unsigned long range_start_pfn
,
4550 unsigned long range_end_pfn
)
4552 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4553 unsigned long start_pfn
, end_pfn
;
4556 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4557 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4558 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4559 nr_absent
-= end_pfn
- start_pfn
;
4565 * absent_pages_in_range - Return number of page frames in holes within a range
4566 * @start_pfn: The start PFN to start searching for holes
4567 * @end_pfn: The end PFN to stop searching for holes
4569 * It returns the number of pages frames in memory holes within a range.
4571 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4572 unsigned long end_pfn
)
4574 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4577 /* Return the number of page frames in holes in a zone on a node */
4578 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4579 unsigned long zone_type
,
4580 unsigned long node_start_pfn
,
4581 unsigned long node_end_pfn
,
4582 unsigned long *ignored
)
4584 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4585 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4586 unsigned long zone_start_pfn
, zone_end_pfn
;
4588 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4589 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4591 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4592 node_start_pfn
, node_end_pfn
,
4593 &zone_start_pfn
, &zone_end_pfn
);
4594 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4597 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4598 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4599 unsigned long zone_type
,
4600 unsigned long node_start_pfn
,
4601 unsigned long node_end_pfn
,
4602 unsigned long *zones_size
)
4604 return zones_size
[zone_type
];
4607 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4608 unsigned long zone_type
,
4609 unsigned long node_start_pfn
,
4610 unsigned long node_end_pfn
,
4611 unsigned long *zholes_size
)
4616 return zholes_size
[zone_type
];
4619 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4621 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4622 unsigned long node_start_pfn
,
4623 unsigned long node_end_pfn
,
4624 unsigned long *zones_size
,
4625 unsigned long *zholes_size
)
4627 unsigned long realtotalpages
, totalpages
= 0;
4630 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4631 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4635 pgdat
->node_spanned_pages
= totalpages
;
4637 realtotalpages
= totalpages
;
4638 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4640 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4641 node_start_pfn
, node_end_pfn
,
4643 pgdat
->node_present_pages
= realtotalpages
;
4644 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4648 #ifndef CONFIG_SPARSEMEM
4650 * Calculate the size of the zone->blockflags rounded to an unsigned long
4651 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4652 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4653 * round what is now in bits to nearest long in bits, then return it in
4656 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4658 unsigned long usemapsize
;
4660 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4661 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4662 usemapsize
= usemapsize
>> pageblock_order
;
4663 usemapsize
*= NR_PAGEBLOCK_BITS
;
4664 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4666 return usemapsize
/ 8;
4669 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4671 unsigned long zone_start_pfn
,
4672 unsigned long zonesize
)
4674 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4675 zone
->pageblock_flags
= NULL
;
4677 zone
->pageblock_flags
=
4678 memblock_virt_alloc_node_nopanic(usemapsize
,
4682 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4683 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4684 #endif /* CONFIG_SPARSEMEM */
4686 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4688 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4689 void __paginginit
set_pageblock_order(void)
4693 /* Check that pageblock_nr_pages has not already been setup */
4694 if (pageblock_order
)
4697 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4698 order
= HUGETLB_PAGE_ORDER
;
4700 order
= MAX_ORDER
- 1;
4703 * Assume the largest contiguous order of interest is a huge page.
4704 * This value may be variable depending on boot parameters on IA64 and
4707 pageblock_order
= order
;
4709 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4712 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4713 * is unused as pageblock_order is set at compile-time. See
4714 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4717 void __paginginit
set_pageblock_order(void)
4721 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4723 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4724 unsigned long present_pages
)
4726 unsigned long pages
= spanned_pages
;
4729 * Provide a more accurate estimation if there are holes within
4730 * the zone and SPARSEMEM is in use. If there are holes within the
4731 * zone, each populated memory region may cost us one or two extra
4732 * memmap pages due to alignment because memmap pages for each
4733 * populated regions may not naturally algined on page boundary.
4734 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4736 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4737 IS_ENABLED(CONFIG_SPARSEMEM
))
4738 pages
= present_pages
;
4740 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4744 * Set up the zone data structures:
4745 * - mark all pages reserved
4746 * - mark all memory queues empty
4747 * - clear the memory bitmaps
4749 * NOTE: pgdat should get zeroed by caller.
4751 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4752 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4753 unsigned long *zones_size
, unsigned long *zholes_size
)
4756 int nid
= pgdat
->node_id
;
4757 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4760 pgdat_resize_init(pgdat
);
4761 #ifdef CONFIG_NUMA_BALANCING
4762 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4763 pgdat
->numabalancing_migrate_nr_pages
= 0;
4764 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4766 init_waitqueue_head(&pgdat
->kswapd_wait
);
4767 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4768 pgdat_page_cgroup_init(pgdat
);
4770 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4771 struct zone
*zone
= pgdat
->node_zones
+ j
;
4772 unsigned long size
, realsize
, freesize
, memmap_pages
;
4774 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4775 node_end_pfn
, zones_size
);
4776 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4782 * Adjust freesize so that it accounts for how much memory
4783 * is used by this zone for memmap. This affects the watermark
4784 * and per-cpu initialisations
4786 memmap_pages
= calc_memmap_size(size
, realsize
);
4787 if (freesize
>= memmap_pages
) {
4788 freesize
-= memmap_pages
;
4791 " %s zone: %lu pages used for memmap\n",
4792 zone_names
[j
], memmap_pages
);
4795 " %s zone: %lu pages exceeds freesize %lu\n",
4796 zone_names
[j
], memmap_pages
, freesize
);
4798 /* Account for reserved pages */
4799 if (j
== 0 && freesize
> dma_reserve
) {
4800 freesize
-= dma_reserve
;
4801 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4802 zone_names
[0], dma_reserve
);
4805 if (!is_highmem_idx(j
))
4806 nr_kernel_pages
+= freesize
;
4807 /* Charge for highmem memmap if there are enough kernel pages */
4808 else if (nr_kernel_pages
> memmap_pages
* 2)
4809 nr_kernel_pages
-= memmap_pages
;
4810 nr_all_pages
+= freesize
;
4812 zone
->spanned_pages
= size
;
4813 zone
->present_pages
= realsize
;
4815 * Set an approximate value for lowmem here, it will be adjusted
4816 * when the bootmem allocator frees pages into the buddy system.
4817 * And all highmem pages will be managed by the buddy system.
4819 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4822 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4824 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4826 zone
->name
= zone_names
[j
];
4827 spin_lock_init(&zone
->lock
);
4828 spin_lock_init(&zone
->lru_lock
);
4829 zone_seqlock_init(zone
);
4830 zone
->zone_pgdat
= pgdat
;
4831 zone_pcp_init(zone
);
4833 /* For bootup, initialized properly in watermark setup */
4834 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4836 lruvec_init(&zone
->lruvec
);
4840 set_pageblock_order();
4841 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4842 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4843 size
, MEMMAP_EARLY
);
4845 memmap_init(size
, nid
, j
, zone_start_pfn
);
4846 zone_start_pfn
+= size
;
4850 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4852 /* Skip empty nodes */
4853 if (!pgdat
->node_spanned_pages
)
4856 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4857 /* ia64 gets its own node_mem_map, before this, without bootmem */
4858 if (!pgdat
->node_mem_map
) {
4859 unsigned long size
, start
, end
;
4863 * The zone's endpoints aren't required to be MAX_ORDER
4864 * aligned but the node_mem_map endpoints must be in order
4865 * for the buddy allocator to function correctly.
4867 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4868 end
= pgdat_end_pfn(pgdat
);
4869 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4870 size
= (end
- start
) * sizeof(struct page
);
4871 map
= alloc_remap(pgdat
->node_id
, size
);
4873 map
= memblock_virt_alloc_node_nopanic(size
,
4875 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4877 #ifndef CONFIG_NEED_MULTIPLE_NODES
4879 * With no DISCONTIG, the global mem_map is just set as node 0's
4881 if (pgdat
== NODE_DATA(0)) {
4882 mem_map
= NODE_DATA(0)->node_mem_map
;
4883 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4884 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4885 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4886 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4889 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4892 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4893 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4895 pg_data_t
*pgdat
= NODE_DATA(nid
);
4896 unsigned long start_pfn
= 0;
4897 unsigned long end_pfn
= 0;
4899 /* pg_data_t should be reset to zero when it's allocated */
4900 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4902 pgdat
->node_id
= nid
;
4903 pgdat
->node_start_pfn
= node_start_pfn
;
4904 init_zone_allows_reclaim(nid
);
4905 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4906 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4908 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
4909 zones_size
, zholes_size
);
4911 alloc_node_mem_map(pgdat
);
4912 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4913 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4914 nid
, (unsigned long)pgdat
,
4915 (unsigned long)pgdat
->node_mem_map
);
4918 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
4919 zones_size
, zholes_size
);
4922 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4924 #if MAX_NUMNODES > 1
4926 * Figure out the number of possible node ids.
4928 void __init
setup_nr_node_ids(void)
4931 unsigned int highest
= 0;
4933 for_each_node_mask(node
, node_possible_map
)
4935 nr_node_ids
= highest
+ 1;
4940 * node_map_pfn_alignment - determine the maximum internode alignment
4942 * This function should be called after node map is populated and sorted.
4943 * It calculates the maximum power of two alignment which can distinguish
4946 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4947 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4948 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4949 * shifted, 1GiB is enough and this function will indicate so.
4951 * This is used to test whether pfn -> nid mapping of the chosen memory
4952 * model has fine enough granularity to avoid incorrect mapping for the
4953 * populated node map.
4955 * Returns the determined alignment in pfn's. 0 if there is no alignment
4956 * requirement (single node).
4958 unsigned long __init
node_map_pfn_alignment(void)
4960 unsigned long accl_mask
= 0, last_end
= 0;
4961 unsigned long start
, end
, mask
;
4965 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
4966 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
4973 * Start with a mask granular enough to pin-point to the
4974 * start pfn and tick off bits one-by-one until it becomes
4975 * too coarse to separate the current node from the last.
4977 mask
= ~((1 << __ffs(start
)) - 1);
4978 while (mask
&& last_end
<= (start
& (mask
<< 1)))
4981 /* accumulate all internode masks */
4985 /* convert mask to number of pages */
4986 return ~accl_mask
+ 1;
4989 /* Find the lowest pfn for a node */
4990 static unsigned long __init
find_min_pfn_for_node(int nid
)
4992 unsigned long min_pfn
= ULONG_MAX
;
4993 unsigned long start_pfn
;
4996 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
4997 min_pfn
= min(min_pfn
, start_pfn
);
4999 if (min_pfn
== ULONG_MAX
) {
5001 "Could not find start_pfn for node %d\n", nid
);
5009 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5011 * It returns the minimum PFN based on information provided via
5012 * add_active_range().
5014 unsigned long __init
find_min_pfn_with_active_regions(void)
5016 return find_min_pfn_for_node(MAX_NUMNODES
);
5020 * early_calculate_totalpages()
5021 * Sum pages in active regions for movable zone.
5022 * Populate N_MEMORY for calculating usable_nodes.
5024 static unsigned long __init
early_calculate_totalpages(void)
5026 unsigned long totalpages
= 0;
5027 unsigned long start_pfn
, end_pfn
;
5030 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5031 unsigned long pages
= end_pfn
- start_pfn
;
5033 totalpages
+= pages
;
5035 node_set_state(nid
, N_MEMORY
);
5041 * Find the PFN the Movable zone begins in each node. Kernel memory
5042 * is spread evenly between nodes as long as the nodes have enough
5043 * memory. When they don't, some nodes will have more kernelcore than
5046 static void __init
find_zone_movable_pfns_for_nodes(void)
5049 unsigned long usable_startpfn
;
5050 unsigned long kernelcore_node
, kernelcore_remaining
;
5051 /* save the state before borrow the nodemask */
5052 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5053 unsigned long totalpages
= early_calculate_totalpages();
5054 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5055 struct memblock_type
*type
= &memblock
.memory
;
5057 /* Need to find movable_zone earlier when movable_node is specified. */
5058 find_usable_zone_for_movable();
5061 * If movable_node is specified, ignore kernelcore and movablecore
5064 if (movable_node_is_enabled()) {
5065 for (i
= 0; i
< type
->cnt
; i
++) {
5066 if (!memblock_is_hotpluggable(&type
->regions
[i
]))
5069 nid
= type
->regions
[i
].nid
;
5071 usable_startpfn
= PFN_DOWN(type
->regions
[i
].base
);
5072 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5073 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5081 * If movablecore=nn[KMG] was specified, calculate what size of
5082 * kernelcore that corresponds so that memory usable for
5083 * any allocation type is evenly spread. If both kernelcore
5084 * and movablecore are specified, then the value of kernelcore
5085 * will be used for required_kernelcore if it's greater than
5086 * what movablecore would have allowed.
5088 if (required_movablecore
) {
5089 unsigned long corepages
;
5092 * Round-up so that ZONE_MOVABLE is at least as large as what
5093 * was requested by the user
5095 required_movablecore
=
5096 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5097 corepages
= totalpages
- required_movablecore
;
5099 required_kernelcore
= max(required_kernelcore
, corepages
);
5102 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5103 if (!required_kernelcore
)
5106 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5107 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5110 /* Spread kernelcore memory as evenly as possible throughout nodes */
5111 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5112 for_each_node_state(nid
, N_MEMORY
) {
5113 unsigned long start_pfn
, end_pfn
;
5116 * Recalculate kernelcore_node if the division per node
5117 * now exceeds what is necessary to satisfy the requested
5118 * amount of memory for the kernel
5120 if (required_kernelcore
< kernelcore_node
)
5121 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5124 * As the map is walked, we track how much memory is usable
5125 * by the kernel using kernelcore_remaining. When it is
5126 * 0, the rest of the node is usable by ZONE_MOVABLE
5128 kernelcore_remaining
= kernelcore_node
;
5130 /* Go through each range of PFNs within this node */
5131 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5132 unsigned long size_pages
;
5134 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5135 if (start_pfn
>= end_pfn
)
5138 /* Account for what is only usable for kernelcore */
5139 if (start_pfn
< usable_startpfn
) {
5140 unsigned long kernel_pages
;
5141 kernel_pages
= min(end_pfn
, usable_startpfn
)
5144 kernelcore_remaining
-= min(kernel_pages
,
5145 kernelcore_remaining
);
5146 required_kernelcore
-= min(kernel_pages
,
5147 required_kernelcore
);
5149 /* Continue if range is now fully accounted */
5150 if (end_pfn
<= usable_startpfn
) {
5153 * Push zone_movable_pfn to the end so
5154 * that if we have to rebalance
5155 * kernelcore across nodes, we will
5156 * not double account here
5158 zone_movable_pfn
[nid
] = end_pfn
;
5161 start_pfn
= usable_startpfn
;
5165 * The usable PFN range for ZONE_MOVABLE is from
5166 * start_pfn->end_pfn. Calculate size_pages as the
5167 * number of pages used as kernelcore
5169 size_pages
= end_pfn
- start_pfn
;
5170 if (size_pages
> kernelcore_remaining
)
5171 size_pages
= kernelcore_remaining
;
5172 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5175 * Some kernelcore has been met, update counts and
5176 * break if the kernelcore for this node has been
5179 required_kernelcore
-= min(required_kernelcore
,
5181 kernelcore_remaining
-= size_pages
;
5182 if (!kernelcore_remaining
)
5188 * If there is still required_kernelcore, we do another pass with one
5189 * less node in the count. This will push zone_movable_pfn[nid] further
5190 * along on the nodes that still have memory until kernelcore is
5194 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5198 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5199 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5200 zone_movable_pfn
[nid
] =
5201 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5204 /* restore the node_state */
5205 node_states
[N_MEMORY
] = saved_node_state
;
5208 /* Any regular or high memory on that node ? */
5209 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5211 enum zone_type zone_type
;
5213 if (N_MEMORY
== N_NORMAL_MEMORY
)
5216 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5217 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5218 if (populated_zone(zone
)) {
5219 node_set_state(nid
, N_HIGH_MEMORY
);
5220 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5221 zone_type
<= ZONE_NORMAL
)
5222 node_set_state(nid
, N_NORMAL_MEMORY
);
5229 * free_area_init_nodes - Initialise all pg_data_t and zone data
5230 * @max_zone_pfn: an array of max PFNs for each zone
5232 * This will call free_area_init_node() for each active node in the system.
5233 * Using the page ranges provided by add_active_range(), the size of each
5234 * zone in each node and their holes is calculated. If the maximum PFN
5235 * between two adjacent zones match, it is assumed that the zone is empty.
5236 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5237 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5238 * starts where the previous one ended. For example, ZONE_DMA32 starts
5239 * at arch_max_dma_pfn.
5241 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5243 unsigned long start_pfn
, end_pfn
;
5246 /* Record where the zone boundaries are */
5247 memset(arch_zone_lowest_possible_pfn
, 0,
5248 sizeof(arch_zone_lowest_possible_pfn
));
5249 memset(arch_zone_highest_possible_pfn
, 0,
5250 sizeof(arch_zone_highest_possible_pfn
));
5251 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5252 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5253 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5254 if (i
== ZONE_MOVABLE
)
5256 arch_zone_lowest_possible_pfn
[i
] =
5257 arch_zone_highest_possible_pfn
[i
-1];
5258 arch_zone_highest_possible_pfn
[i
] =
5259 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5261 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5262 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5264 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5265 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5266 find_zone_movable_pfns_for_nodes();
5268 /* Print out the zone ranges */
5269 printk("Zone ranges:\n");
5270 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5271 if (i
== ZONE_MOVABLE
)
5273 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5274 if (arch_zone_lowest_possible_pfn
[i
] ==
5275 arch_zone_highest_possible_pfn
[i
])
5276 printk(KERN_CONT
"empty\n");
5278 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5279 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5280 (arch_zone_highest_possible_pfn
[i
]
5281 << PAGE_SHIFT
) - 1);
5284 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5285 printk("Movable zone start for each node\n");
5286 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5287 if (zone_movable_pfn
[i
])
5288 printk(" Node %d: %#010lx\n", i
,
5289 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5292 /* Print out the early node map */
5293 printk("Early memory node ranges\n");
5294 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5295 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5296 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5298 /* Initialise every node */
5299 mminit_verify_pageflags_layout();
5300 setup_nr_node_ids();
5301 for_each_online_node(nid
) {
5302 pg_data_t
*pgdat
= NODE_DATA(nid
);
5303 free_area_init_node(nid
, NULL
,
5304 find_min_pfn_for_node(nid
), NULL
);
5306 /* Any memory on that node */
5307 if (pgdat
->node_present_pages
)
5308 node_set_state(nid
, N_MEMORY
);
5309 check_for_memory(pgdat
, nid
);
5313 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5315 unsigned long long coremem
;
5319 coremem
= memparse(p
, &p
);
5320 *core
= coremem
>> PAGE_SHIFT
;
5322 /* Paranoid check that UL is enough for the coremem value */
5323 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5329 * kernelcore=size sets the amount of memory for use for allocations that
5330 * cannot be reclaimed or migrated.
5332 static int __init
cmdline_parse_kernelcore(char *p
)
5334 return cmdline_parse_core(p
, &required_kernelcore
);
5338 * movablecore=size sets the amount of memory for use for allocations that
5339 * can be reclaimed or migrated.
5341 static int __init
cmdline_parse_movablecore(char *p
)
5343 return cmdline_parse_core(p
, &required_movablecore
);
5346 early_param("kernelcore", cmdline_parse_kernelcore
);
5347 early_param("movablecore", cmdline_parse_movablecore
);
5349 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5351 void adjust_managed_page_count(struct page
*page
, long count
)
5353 spin_lock(&managed_page_count_lock
);
5354 page_zone(page
)->managed_pages
+= count
;
5355 totalram_pages
+= count
;
5356 #ifdef CONFIG_HIGHMEM
5357 if (PageHighMem(page
))
5358 totalhigh_pages
+= count
;
5360 spin_unlock(&managed_page_count_lock
);
5362 EXPORT_SYMBOL(adjust_managed_page_count
);
5364 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5367 unsigned long pages
= 0;
5369 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5370 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5371 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5372 if ((unsigned int)poison
<= 0xFF)
5373 memset(pos
, poison
, PAGE_SIZE
);
5374 free_reserved_page(virt_to_page(pos
));
5378 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5379 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5383 EXPORT_SYMBOL(free_reserved_area
);
5385 #ifdef CONFIG_HIGHMEM
5386 void free_highmem_page(struct page
*page
)
5388 __free_reserved_page(page
);
5390 page_zone(page
)->managed_pages
++;
5396 void __init
mem_init_print_info(const char *str
)
5398 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5399 unsigned long init_code_size
, init_data_size
;
5401 physpages
= get_num_physpages();
5402 codesize
= _etext
- _stext
;
5403 datasize
= _edata
- _sdata
;
5404 rosize
= __end_rodata
- __start_rodata
;
5405 bss_size
= __bss_stop
- __bss_start
;
5406 init_data_size
= __init_end
- __init_begin
;
5407 init_code_size
= _einittext
- _sinittext
;
5410 * Detect special cases and adjust section sizes accordingly:
5411 * 1) .init.* may be embedded into .data sections
5412 * 2) .init.text.* may be out of [__init_begin, __init_end],
5413 * please refer to arch/tile/kernel/vmlinux.lds.S.
5414 * 3) .rodata.* may be embedded into .text or .data sections.
5416 #define adj_init_size(start, end, size, pos, adj) \
5418 if (start <= pos && pos < end && size > adj) \
5422 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5423 _sinittext
, init_code_size
);
5424 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5425 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5426 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5427 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5429 #undef adj_init_size
5431 printk("Memory: %luK/%luK available "
5432 "(%luK kernel code, %luK rwdata, %luK rodata, "
5433 "%luK init, %luK bss, %luK reserved"
5434 #ifdef CONFIG_HIGHMEM
5438 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5439 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5440 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5441 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5442 #ifdef CONFIG_HIGHMEM
5443 totalhigh_pages
<< (PAGE_SHIFT
-10),
5445 str
? ", " : "", str
? str
: "");
5449 * set_dma_reserve - set the specified number of pages reserved in the first zone
5450 * @new_dma_reserve: The number of pages to mark reserved
5452 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5453 * In the DMA zone, a significant percentage may be consumed by kernel image
5454 * and other unfreeable allocations which can skew the watermarks badly. This
5455 * function may optionally be used to account for unfreeable pages in the
5456 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5457 * smaller per-cpu batchsize.
5459 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5461 dma_reserve
= new_dma_reserve
;
5464 void __init
free_area_init(unsigned long *zones_size
)
5466 free_area_init_node(0, zones_size
,
5467 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5470 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5471 unsigned long action
, void *hcpu
)
5473 int cpu
= (unsigned long)hcpu
;
5475 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5476 lru_add_drain_cpu(cpu
);
5480 * Spill the event counters of the dead processor
5481 * into the current processors event counters.
5482 * This artificially elevates the count of the current
5485 vm_events_fold_cpu(cpu
);
5488 * Zero the differential counters of the dead processor
5489 * so that the vm statistics are consistent.
5491 * This is only okay since the processor is dead and cannot
5492 * race with what we are doing.
5494 cpu_vm_stats_fold(cpu
);
5499 void __init
page_alloc_init(void)
5501 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5505 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5506 * or min_free_kbytes changes.
5508 static void calculate_totalreserve_pages(void)
5510 struct pglist_data
*pgdat
;
5511 unsigned long reserve_pages
= 0;
5512 enum zone_type i
, j
;
5514 for_each_online_pgdat(pgdat
) {
5515 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5516 struct zone
*zone
= pgdat
->node_zones
+ i
;
5517 unsigned long max
= 0;
5519 /* Find valid and maximum lowmem_reserve in the zone */
5520 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5521 if (zone
->lowmem_reserve
[j
] > max
)
5522 max
= zone
->lowmem_reserve
[j
];
5525 /* we treat the high watermark as reserved pages. */
5526 max
+= high_wmark_pages(zone
);
5528 if (max
> zone
->managed_pages
)
5529 max
= zone
->managed_pages
;
5530 reserve_pages
+= max
;
5532 * Lowmem reserves are not available to
5533 * GFP_HIGHUSER page cache allocations and
5534 * kswapd tries to balance zones to their high
5535 * watermark. As a result, neither should be
5536 * regarded as dirtyable memory, to prevent a
5537 * situation where reclaim has to clean pages
5538 * in order to balance the zones.
5540 zone
->dirty_balance_reserve
= max
;
5543 dirty_balance_reserve
= reserve_pages
;
5544 totalreserve_pages
= reserve_pages
;
5548 * setup_per_zone_lowmem_reserve - called whenever
5549 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5550 * has a correct pages reserved value, so an adequate number of
5551 * pages are left in the zone after a successful __alloc_pages().
5553 static void setup_per_zone_lowmem_reserve(void)
5555 struct pglist_data
*pgdat
;
5556 enum zone_type j
, idx
;
5558 for_each_online_pgdat(pgdat
) {
5559 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5560 struct zone
*zone
= pgdat
->node_zones
+ j
;
5561 unsigned long managed_pages
= zone
->managed_pages
;
5563 zone
->lowmem_reserve
[j
] = 0;
5567 struct zone
*lower_zone
;
5571 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5572 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5574 lower_zone
= pgdat
->node_zones
+ idx
;
5575 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5576 sysctl_lowmem_reserve_ratio
[idx
];
5577 managed_pages
+= lower_zone
->managed_pages
;
5582 /* update totalreserve_pages */
5583 calculate_totalreserve_pages();
5586 static void __setup_per_zone_wmarks(void)
5588 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5589 unsigned long lowmem_pages
= 0;
5591 unsigned long flags
;
5593 /* Calculate total number of !ZONE_HIGHMEM pages */
5594 for_each_zone(zone
) {
5595 if (!is_highmem(zone
))
5596 lowmem_pages
+= zone
->managed_pages
;
5599 for_each_zone(zone
) {
5602 spin_lock_irqsave(&zone
->lock
, flags
);
5603 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5604 do_div(tmp
, lowmem_pages
);
5605 if (is_highmem(zone
)) {
5607 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5608 * need highmem pages, so cap pages_min to a small
5611 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5612 * deltas controls asynch page reclaim, and so should
5613 * not be capped for highmem.
5615 unsigned long min_pages
;
5617 min_pages
= zone
->managed_pages
/ 1024;
5618 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5619 zone
->watermark
[WMARK_MIN
] = min_pages
;
5622 * If it's a lowmem zone, reserve a number of pages
5623 * proportionate to the zone's size.
5625 zone
->watermark
[WMARK_MIN
] = tmp
;
5628 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5629 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5631 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5632 high_wmark_pages(zone
) -
5633 low_wmark_pages(zone
) -
5634 zone_page_state(zone
, NR_ALLOC_BATCH
));
5636 setup_zone_migrate_reserve(zone
);
5637 spin_unlock_irqrestore(&zone
->lock
, flags
);
5640 /* update totalreserve_pages */
5641 calculate_totalreserve_pages();
5645 * setup_per_zone_wmarks - called when min_free_kbytes changes
5646 * or when memory is hot-{added|removed}
5648 * Ensures that the watermark[min,low,high] values for each zone are set
5649 * correctly with respect to min_free_kbytes.
5651 void setup_per_zone_wmarks(void)
5653 mutex_lock(&zonelists_mutex
);
5654 __setup_per_zone_wmarks();
5655 mutex_unlock(&zonelists_mutex
);
5659 * The inactive anon list should be small enough that the VM never has to
5660 * do too much work, but large enough that each inactive page has a chance
5661 * to be referenced again before it is swapped out.
5663 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5664 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5665 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5666 * the anonymous pages are kept on the inactive list.
5669 * memory ratio inactive anon
5670 * -------------------------------------
5679 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5681 unsigned int gb
, ratio
;
5683 /* Zone size in gigabytes */
5684 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5686 ratio
= int_sqrt(10 * gb
);
5690 zone
->inactive_ratio
= ratio
;
5693 static void __meminit
setup_per_zone_inactive_ratio(void)
5698 calculate_zone_inactive_ratio(zone
);
5702 * Initialise min_free_kbytes.
5704 * For small machines we want it small (128k min). For large machines
5705 * we want it large (64MB max). But it is not linear, because network
5706 * bandwidth does not increase linearly with machine size. We use
5708 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5709 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5725 int __meminit
init_per_zone_wmark_min(void)
5727 unsigned long lowmem_kbytes
;
5728 int new_min_free_kbytes
;
5730 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5731 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5733 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5734 min_free_kbytes
= new_min_free_kbytes
;
5735 if (min_free_kbytes
< 128)
5736 min_free_kbytes
= 128;
5737 if (min_free_kbytes
> 65536)
5738 min_free_kbytes
= 65536;
5740 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5741 new_min_free_kbytes
, user_min_free_kbytes
);
5743 setup_per_zone_wmarks();
5744 refresh_zone_stat_thresholds();
5745 setup_per_zone_lowmem_reserve();
5746 setup_per_zone_inactive_ratio();
5749 module_init(init_per_zone_wmark_min
)
5752 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5753 * that we can call two helper functions whenever min_free_kbytes
5756 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5757 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5761 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5766 user_min_free_kbytes
= min_free_kbytes
;
5767 setup_per_zone_wmarks();
5773 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5774 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5779 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5784 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5785 sysctl_min_unmapped_ratio
) / 100;
5789 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5790 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5795 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5800 zone
->min_slab_pages
= (zone
->managed_pages
*
5801 sysctl_min_slab_ratio
) / 100;
5807 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5808 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5809 * whenever sysctl_lowmem_reserve_ratio changes.
5811 * The reserve ratio obviously has absolutely no relation with the
5812 * minimum watermarks. The lowmem reserve ratio can only make sense
5813 * if in function of the boot time zone sizes.
5815 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5816 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5818 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5819 setup_per_zone_lowmem_reserve();
5824 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5825 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5826 * pagelist can have before it gets flushed back to buddy allocator.
5828 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5829 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5835 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5836 if (!write
|| (ret
< 0))
5839 mutex_lock(&pcp_batch_high_lock
);
5840 for_each_populated_zone(zone
) {
5842 high
= zone
->managed_pages
/ percpu_pagelist_fraction
;
5843 for_each_possible_cpu(cpu
)
5844 pageset_set_high(per_cpu_ptr(zone
->pageset
, cpu
),
5847 mutex_unlock(&pcp_batch_high_lock
);
5851 int hashdist
= HASHDIST_DEFAULT
;
5854 static int __init
set_hashdist(char *str
)
5858 hashdist
= simple_strtoul(str
, &str
, 0);
5861 __setup("hashdist=", set_hashdist
);
5865 * allocate a large system hash table from bootmem
5866 * - it is assumed that the hash table must contain an exact power-of-2
5867 * quantity of entries
5868 * - limit is the number of hash buckets, not the total allocation size
5870 void *__init
alloc_large_system_hash(const char *tablename
,
5871 unsigned long bucketsize
,
5872 unsigned long numentries
,
5875 unsigned int *_hash_shift
,
5876 unsigned int *_hash_mask
,
5877 unsigned long low_limit
,
5878 unsigned long high_limit
)
5880 unsigned long long max
= high_limit
;
5881 unsigned long log2qty
, size
;
5884 /* allow the kernel cmdline to have a say */
5886 /* round applicable memory size up to nearest megabyte */
5887 numentries
= nr_kernel_pages
;
5889 /* It isn't necessary when PAGE_SIZE >= 1MB */
5890 if (PAGE_SHIFT
< 20)
5891 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
5893 /* limit to 1 bucket per 2^scale bytes of low memory */
5894 if (scale
> PAGE_SHIFT
)
5895 numentries
>>= (scale
- PAGE_SHIFT
);
5897 numentries
<<= (PAGE_SHIFT
- scale
);
5899 /* Make sure we've got at least a 0-order allocation.. */
5900 if (unlikely(flags
& HASH_SMALL
)) {
5901 /* Makes no sense without HASH_EARLY */
5902 WARN_ON(!(flags
& HASH_EARLY
));
5903 if (!(numentries
>> *_hash_shift
)) {
5904 numentries
= 1UL << *_hash_shift
;
5905 BUG_ON(!numentries
);
5907 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5908 numentries
= PAGE_SIZE
/ bucketsize
;
5910 numentries
= roundup_pow_of_two(numentries
);
5912 /* limit allocation size to 1/16 total memory by default */
5914 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5915 do_div(max
, bucketsize
);
5917 max
= min(max
, 0x80000000ULL
);
5919 if (numentries
< low_limit
)
5920 numentries
= low_limit
;
5921 if (numentries
> max
)
5924 log2qty
= ilog2(numentries
);
5927 size
= bucketsize
<< log2qty
;
5928 if (flags
& HASH_EARLY
)
5929 table
= memblock_virt_alloc_nopanic(size
, 0);
5931 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5934 * If bucketsize is not a power-of-two, we may free
5935 * some pages at the end of hash table which
5936 * alloc_pages_exact() automatically does
5938 if (get_order(size
) < MAX_ORDER
) {
5939 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5940 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5943 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5946 panic("Failed to allocate %s hash table\n", tablename
);
5948 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5951 ilog2(size
) - PAGE_SHIFT
,
5955 *_hash_shift
= log2qty
;
5957 *_hash_mask
= (1 << log2qty
) - 1;
5962 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5963 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5966 #ifdef CONFIG_SPARSEMEM
5967 return __pfn_to_section(pfn
)->pageblock_flags
;
5969 return zone
->pageblock_flags
;
5970 #endif /* CONFIG_SPARSEMEM */
5973 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5975 #ifdef CONFIG_SPARSEMEM
5976 pfn
&= (PAGES_PER_SECTION
-1);
5977 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5979 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
5980 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5981 #endif /* CONFIG_SPARSEMEM */
5985 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5986 * @page: The page within the block of interest
5987 * @start_bitidx: The first bit of interest to retrieve
5988 * @end_bitidx: The last bit of interest
5989 * returns pageblock_bits flags
5991 unsigned long get_pageblock_flags_group(struct page
*page
,
5992 int start_bitidx
, int end_bitidx
)
5995 unsigned long *bitmap
;
5996 unsigned long pfn
, bitidx
;
5997 unsigned long flags
= 0;
5998 unsigned long value
= 1;
6000 zone
= page_zone(page
);
6001 pfn
= page_to_pfn(page
);
6002 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6003 bitidx
= pfn_to_bitidx(zone
, pfn
);
6005 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
6006 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
6013 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
6014 * @page: The page within the block of interest
6015 * @start_bitidx: The first bit of interest
6016 * @end_bitidx: The last bit of interest
6017 * @flags: The flags to set
6019 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
6020 int start_bitidx
, int end_bitidx
)
6023 unsigned long *bitmap
;
6024 unsigned long pfn
, bitidx
;
6025 unsigned long value
= 1;
6027 zone
= page_zone(page
);
6028 pfn
= page_to_pfn(page
);
6029 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6030 bitidx
= pfn_to_bitidx(zone
, pfn
);
6031 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6033 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
6035 __set_bit(bitidx
+ start_bitidx
, bitmap
);
6037 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
6041 * This function checks whether pageblock includes unmovable pages or not.
6042 * If @count is not zero, it is okay to include less @count unmovable pages
6044 * PageLRU check without isolation or lru_lock could race so that
6045 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6046 * expect this function should be exact.
6048 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6049 bool skip_hwpoisoned_pages
)
6051 unsigned long pfn
, iter
, found
;
6055 * For avoiding noise data, lru_add_drain_all() should be called
6056 * If ZONE_MOVABLE, the zone never contains unmovable pages
6058 if (zone_idx(zone
) == ZONE_MOVABLE
)
6060 mt
= get_pageblock_migratetype(page
);
6061 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6064 pfn
= page_to_pfn(page
);
6065 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6066 unsigned long check
= pfn
+ iter
;
6068 if (!pfn_valid_within(check
))
6071 page
= pfn_to_page(check
);
6074 * Hugepages are not in LRU lists, but they're movable.
6075 * We need not scan over tail pages bacause we don't
6076 * handle each tail page individually in migration.
6078 if (PageHuge(page
)) {
6079 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6084 * We can't use page_count without pin a page
6085 * because another CPU can free compound page.
6086 * This check already skips compound tails of THP
6087 * because their page->_count is zero at all time.
6089 if (!atomic_read(&page
->_count
)) {
6090 if (PageBuddy(page
))
6091 iter
+= (1 << page_order(page
)) - 1;
6096 * The HWPoisoned page may be not in buddy system, and
6097 * page_count() is not 0.
6099 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6105 * If there are RECLAIMABLE pages, we need to check it.
6106 * But now, memory offline itself doesn't call shrink_slab()
6107 * and it still to be fixed.
6110 * If the page is not RAM, page_count()should be 0.
6111 * we don't need more check. This is an _used_ not-movable page.
6113 * The problematic thing here is PG_reserved pages. PG_reserved
6114 * is set to both of a memory hole page and a _used_ kernel
6123 bool is_pageblock_removable_nolock(struct page
*page
)
6129 * We have to be careful here because we are iterating over memory
6130 * sections which are not zone aware so we might end up outside of
6131 * the zone but still within the section.
6132 * We have to take care about the node as well. If the node is offline
6133 * its NODE_DATA will be NULL - see page_zone.
6135 if (!node_online(page_to_nid(page
)))
6138 zone
= page_zone(page
);
6139 pfn
= page_to_pfn(page
);
6140 if (!zone_spans_pfn(zone
, pfn
))
6143 return !has_unmovable_pages(zone
, page
, 0, true);
6148 static unsigned long pfn_max_align_down(unsigned long pfn
)
6150 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6151 pageblock_nr_pages
) - 1);
6154 static unsigned long pfn_max_align_up(unsigned long pfn
)
6156 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6157 pageblock_nr_pages
));
6160 /* [start, end) must belong to a single zone. */
6161 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6162 unsigned long start
, unsigned long end
)
6164 /* This function is based on compact_zone() from compaction.c. */
6165 unsigned long nr_reclaimed
;
6166 unsigned long pfn
= start
;
6167 unsigned int tries
= 0;
6172 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6173 if (fatal_signal_pending(current
)) {
6178 if (list_empty(&cc
->migratepages
)) {
6179 cc
->nr_migratepages
= 0;
6180 pfn
= isolate_migratepages_range(cc
->zone
, cc
,
6187 } else if (++tries
== 5) {
6188 ret
= ret
< 0 ? ret
: -EBUSY
;
6192 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6194 cc
->nr_migratepages
-= nr_reclaimed
;
6196 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6197 0, MIGRATE_SYNC
, MR_CMA
);
6200 putback_movable_pages(&cc
->migratepages
);
6207 * alloc_contig_range() -- tries to allocate given range of pages
6208 * @start: start PFN to allocate
6209 * @end: one-past-the-last PFN to allocate
6210 * @migratetype: migratetype of the underlaying pageblocks (either
6211 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6212 * in range must have the same migratetype and it must
6213 * be either of the two.
6215 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6216 * aligned, however it's the caller's responsibility to guarantee that
6217 * we are the only thread that changes migrate type of pageblocks the
6220 * The PFN range must belong to a single zone.
6222 * Returns zero on success or negative error code. On success all
6223 * pages which PFN is in [start, end) are allocated for the caller and
6224 * need to be freed with free_contig_range().
6226 int alloc_contig_range(unsigned long start
, unsigned long end
,
6227 unsigned migratetype
)
6229 unsigned long outer_start
, outer_end
;
6232 struct compact_control cc
= {
6233 .nr_migratepages
= 0,
6235 .zone
= page_zone(pfn_to_page(start
)),
6237 .ignore_skip_hint
= true,
6239 INIT_LIST_HEAD(&cc
.migratepages
);
6242 * What we do here is we mark all pageblocks in range as
6243 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6244 * have different sizes, and due to the way page allocator
6245 * work, we align the range to biggest of the two pages so
6246 * that page allocator won't try to merge buddies from
6247 * different pageblocks and change MIGRATE_ISOLATE to some
6248 * other migration type.
6250 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6251 * migrate the pages from an unaligned range (ie. pages that
6252 * we are interested in). This will put all the pages in
6253 * range back to page allocator as MIGRATE_ISOLATE.
6255 * When this is done, we take the pages in range from page
6256 * allocator removing them from the buddy system. This way
6257 * page allocator will never consider using them.
6259 * This lets us mark the pageblocks back as
6260 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6261 * aligned range but not in the unaligned, original range are
6262 * put back to page allocator so that buddy can use them.
6265 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6266 pfn_max_align_up(end
), migratetype
,
6271 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6276 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6277 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6278 * more, all pages in [start, end) are free in page allocator.
6279 * What we are going to do is to allocate all pages from
6280 * [start, end) (that is remove them from page allocator).
6282 * The only problem is that pages at the beginning and at the
6283 * end of interesting range may be not aligned with pages that
6284 * page allocator holds, ie. they can be part of higher order
6285 * pages. Because of this, we reserve the bigger range and
6286 * once this is done free the pages we are not interested in.
6288 * We don't have to hold zone->lock here because the pages are
6289 * isolated thus they won't get removed from buddy.
6292 lru_add_drain_all();
6296 outer_start
= start
;
6297 while (!PageBuddy(pfn_to_page(outer_start
))) {
6298 if (++order
>= MAX_ORDER
) {
6302 outer_start
&= ~0UL << order
;
6305 /* Make sure the range is really isolated. */
6306 if (test_pages_isolated(outer_start
, end
, false)) {
6307 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6314 /* Grab isolated pages from freelists. */
6315 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6321 /* Free head and tail (if any) */
6322 if (start
!= outer_start
)
6323 free_contig_range(outer_start
, start
- outer_start
);
6324 if (end
!= outer_end
)
6325 free_contig_range(end
, outer_end
- end
);
6328 undo_isolate_page_range(pfn_max_align_down(start
),
6329 pfn_max_align_up(end
), migratetype
);
6333 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6335 unsigned int count
= 0;
6337 for (; nr_pages
--; pfn
++) {
6338 struct page
*page
= pfn_to_page(pfn
);
6340 count
+= page_count(page
) != 1;
6343 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6347 #ifdef CONFIG_MEMORY_HOTPLUG
6349 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6350 * page high values need to be recalulated.
6352 void __meminit
zone_pcp_update(struct zone
*zone
)
6355 mutex_lock(&pcp_batch_high_lock
);
6356 for_each_possible_cpu(cpu
)
6357 pageset_set_high_and_batch(zone
,
6358 per_cpu_ptr(zone
->pageset
, cpu
));
6359 mutex_unlock(&pcp_batch_high_lock
);
6363 void zone_pcp_reset(struct zone
*zone
)
6365 unsigned long flags
;
6367 struct per_cpu_pageset
*pset
;
6369 /* avoid races with drain_pages() */
6370 local_irq_save(flags
);
6371 if (zone
->pageset
!= &boot_pageset
) {
6372 for_each_online_cpu(cpu
) {
6373 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6374 drain_zonestat(zone
, pset
);
6376 free_percpu(zone
->pageset
);
6377 zone
->pageset
= &boot_pageset
;
6379 local_irq_restore(flags
);
6382 #ifdef CONFIG_MEMORY_HOTREMOVE
6384 * All pages in the range must be isolated before calling this.
6387 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6393 unsigned long flags
;
6394 /* find the first valid pfn */
6395 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6400 zone
= page_zone(pfn_to_page(pfn
));
6401 spin_lock_irqsave(&zone
->lock
, flags
);
6403 while (pfn
< end_pfn
) {
6404 if (!pfn_valid(pfn
)) {
6408 page
= pfn_to_page(pfn
);
6410 * The HWPoisoned page may be not in buddy system, and
6411 * page_count() is not 0.
6413 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6415 SetPageReserved(page
);
6419 BUG_ON(page_count(page
));
6420 BUG_ON(!PageBuddy(page
));
6421 order
= page_order(page
);
6422 #ifdef CONFIG_DEBUG_VM
6423 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6424 pfn
, 1 << order
, end_pfn
);
6426 list_del(&page
->lru
);
6427 rmv_page_order(page
);
6428 zone
->free_area
[order
].nr_free
--;
6429 for (i
= 0; i
< (1 << order
); i
++)
6430 SetPageReserved((page
+i
));
6431 pfn
+= (1 << order
);
6433 spin_unlock_irqrestore(&zone
->lock
, flags
);
6437 #ifdef CONFIG_MEMORY_FAILURE
6438 bool is_free_buddy_page(struct page
*page
)
6440 struct zone
*zone
= page_zone(page
);
6441 unsigned long pfn
= page_to_pfn(page
);
6442 unsigned long flags
;
6445 spin_lock_irqsave(&zone
->lock
, flags
);
6446 for (order
= 0; order
< MAX_ORDER
; order
++) {
6447 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6449 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6452 spin_unlock_irqrestore(&zone
->lock
, flags
);
6454 return order
< MAX_ORDER
;
6458 static const struct trace_print_flags pageflag_names
[] = {
6459 {1UL << PG_locked
, "locked" },
6460 {1UL << PG_error
, "error" },
6461 {1UL << PG_referenced
, "referenced" },
6462 {1UL << PG_uptodate
, "uptodate" },
6463 {1UL << PG_dirty
, "dirty" },
6464 {1UL << PG_lru
, "lru" },
6465 {1UL << PG_active
, "active" },
6466 {1UL << PG_slab
, "slab" },
6467 {1UL << PG_owner_priv_1
, "owner_priv_1" },
6468 {1UL << PG_arch_1
, "arch_1" },
6469 {1UL << PG_reserved
, "reserved" },
6470 {1UL << PG_private
, "private" },
6471 {1UL << PG_private_2
, "private_2" },
6472 {1UL << PG_writeback
, "writeback" },
6473 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6474 {1UL << PG_head
, "head" },
6475 {1UL << PG_tail
, "tail" },
6477 {1UL << PG_compound
, "compound" },
6479 {1UL << PG_swapcache
, "swapcache" },
6480 {1UL << PG_mappedtodisk
, "mappedtodisk" },
6481 {1UL << PG_reclaim
, "reclaim" },
6482 {1UL << PG_swapbacked
, "swapbacked" },
6483 {1UL << PG_unevictable
, "unevictable" },
6485 {1UL << PG_mlocked
, "mlocked" },
6487 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6488 {1UL << PG_uncached
, "uncached" },
6490 #ifdef CONFIG_MEMORY_FAILURE
6491 {1UL << PG_hwpoison
, "hwpoison" },
6493 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6494 {1UL << PG_compound_lock
, "compound_lock" },
6498 static void dump_page_flags(unsigned long flags
)
6500 const char *delim
= "";
6504 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names
) != __NR_PAGEFLAGS
);
6506 printk(KERN_ALERT
"page flags: %#lx(", flags
);
6508 /* remove zone id */
6509 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
6511 for (i
= 0; i
< ARRAY_SIZE(pageflag_names
) && flags
; i
++) {
6513 mask
= pageflag_names
[i
].mask
;
6514 if ((flags
& mask
) != mask
)
6518 printk("%s%s", delim
, pageflag_names
[i
].name
);
6522 /* check for left over flags */
6524 printk("%s%#lx", delim
, flags
);
6529 void dump_page_badflags(struct page
*page
, char *reason
, unsigned long badflags
)
6532 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6533 page
, atomic_read(&page
->_count
), page_mapcount(page
),
6534 page
->mapping
, page
->index
);
6535 dump_page_flags(page
->flags
);
6537 pr_alert("page dumped because: %s\n", reason
);
6538 if (page
->flags
& badflags
) {
6539 pr_alert("bad because of flags:\n");
6540 dump_page_flags(page
->flags
& badflags
);
6542 mem_cgroup_print_bad_page(page
);
6545 void dump_page(struct page
*page
, char *reason
)
6547 dump_page_badflags(page
, reason
, 0);
6549 EXPORT_SYMBOL_GPL(dump_page
);