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
;
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
)
238 if (unlikely(page_group_by_mobility_disabled
))
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
)
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
));
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
;
373 set_page_count(p
, 0);
374 p
->first_page
= page
;
378 /* update __split_huge_page_refcount if you change this function */
379 static int destroy_compound_page(struct page
*page
, unsigned long order
)
382 int nr_pages
= 1 << order
;
385 if (unlikely(compound_order(page
) != order
)) {
390 __ClearPageHead(page
);
392 for (i
= 1; i
< nr_pages
; i
++) {
393 struct page
*p
= page
+ i
;
395 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
405 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
410 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
411 * and __GFP_HIGHMEM from hard or soft interrupt context.
413 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
414 for (i
= 0; i
< (1 << order
); i
++)
415 clear_highpage(page
+ i
);
418 #ifdef CONFIG_DEBUG_PAGEALLOC
419 unsigned int _debug_guardpage_minorder
;
421 static int __init
debug_guardpage_minorder_setup(char *buf
)
425 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
426 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
429 _debug_guardpage_minorder
= res
;
430 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
433 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
435 static inline void set_page_guard_flag(struct page
*page
)
437 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
440 static inline void clear_page_guard_flag(struct page
*page
)
442 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
445 static inline void set_page_guard_flag(struct page
*page
) { }
446 static inline void clear_page_guard_flag(struct page
*page
) { }
449 static inline void set_page_order(struct page
*page
, int order
)
451 set_page_private(page
, order
);
452 __SetPageBuddy(page
);
455 static inline void rmv_page_order(struct page
*page
)
457 __ClearPageBuddy(page
);
458 set_page_private(page
, 0);
462 * Locate the struct page for both the matching buddy in our
463 * pair (buddy1) and the combined O(n+1) page they form (page).
465 * 1) Any buddy B1 will have an order O twin B2 which satisfies
466 * the following equation:
468 * For example, if the starting buddy (buddy2) is #8 its order
470 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
472 * 2) Any buddy B will have an order O+1 parent P which
473 * satisfies the following equation:
476 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
478 static inline unsigned long
479 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
481 return page_idx
^ (1 << order
);
485 * This function checks whether a page is free && is the buddy
486 * we can do coalesce a page and its buddy if
487 * (a) the buddy is not in a hole &&
488 * (b) the buddy is in the buddy system &&
489 * (c) a page and its buddy have the same order &&
490 * (d) a page and its buddy are in the same zone.
492 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
493 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
495 * For recording page's order, we use page_private(page).
497 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
500 if (!pfn_valid_within(page_to_pfn(buddy
)))
503 if (page_zone_id(page
) != page_zone_id(buddy
))
506 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
507 VM_BUG_ON(page_count(buddy
) != 0);
511 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
512 VM_BUG_ON(page_count(buddy
) != 0);
519 * Freeing function for a buddy system allocator.
521 * The concept of a buddy system is to maintain direct-mapped table
522 * (containing bit values) for memory blocks of various "orders".
523 * The bottom level table contains the map for the smallest allocatable
524 * units of memory (here, pages), and each level above it describes
525 * pairs of units from the levels below, hence, "buddies".
526 * At a high level, all that happens here is marking the table entry
527 * at the bottom level available, and propagating the changes upward
528 * as necessary, plus some accounting needed to play nicely with other
529 * parts of the VM system.
530 * At each level, we keep a list of pages, which are heads of continuous
531 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
532 * order is recorded in page_private(page) field.
533 * So when we are allocating or freeing one, we can derive the state of the
534 * other. That is, if we allocate a small block, and both were
535 * free, the remainder of the region must be split into blocks.
536 * If a block is freed, and its buddy is also free, then this
537 * triggers coalescing into a block of larger size.
542 static inline void __free_one_page(struct page
*page
,
543 struct zone
*zone
, unsigned int order
,
546 unsigned long page_idx
;
547 unsigned long combined_idx
;
548 unsigned long uninitialized_var(buddy_idx
);
551 VM_BUG_ON(!zone_is_initialized(zone
));
553 if (unlikely(PageCompound(page
)))
554 if (unlikely(destroy_compound_page(page
, order
)))
557 VM_BUG_ON(migratetype
== -1);
559 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
561 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
562 VM_BUG_ON(bad_range(zone
, page
));
564 while (order
< MAX_ORDER
-1) {
565 buddy_idx
= __find_buddy_index(page_idx
, order
);
566 buddy
= page
+ (buddy_idx
- page_idx
);
567 if (!page_is_buddy(page
, buddy
, order
))
570 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
571 * merge with it and move up one order.
573 if (page_is_guard(buddy
)) {
574 clear_page_guard_flag(buddy
);
575 set_page_private(page
, 0);
576 __mod_zone_freepage_state(zone
, 1 << order
,
579 list_del(&buddy
->lru
);
580 zone
->free_area
[order
].nr_free
--;
581 rmv_page_order(buddy
);
583 combined_idx
= buddy_idx
& page_idx
;
584 page
= page
+ (combined_idx
- page_idx
);
585 page_idx
= combined_idx
;
588 set_page_order(page
, order
);
591 * If this is not the largest possible page, check if the buddy
592 * of the next-highest order is free. If it is, it's possible
593 * that pages are being freed that will coalesce soon. In case,
594 * that is happening, add the free page to the tail of the list
595 * so it's less likely to be used soon and more likely to be merged
596 * as a higher order page
598 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
599 struct page
*higher_page
, *higher_buddy
;
600 combined_idx
= buddy_idx
& page_idx
;
601 higher_page
= page
+ (combined_idx
- page_idx
);
602 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
603 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
604 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
605 list_add_tail(&page
->lru
,
606 &zone
->free_area
[order
].free_list
[migratetype
]);
611 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
613 zone
->free_area
[order
].nr_free
++;
616 static inline int free_pages_check(struct page
*page
)
618 if (unlikely(page_mapcount(page
) |
619 (page
->mapping
!= NULL
) |
620 (atomic_read(&page
->_count
) != 0) |
621 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
) |
622 (mem_cgroup_bad_page_check(page
)))) {
626 page_nid_reset_last(page
);
627 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
628 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
633 * Frees a number of pages from the PCP lists
634 * Assumes all pages on list are in same zone, and of same order.
635 * count is the number of pages to free.
637 * If the zone was previously in an "all pages pinned" state then look to
638 * see if this freeing clears that state.
640 * And clear the zone's pages_scanned counter, to hold off the "all pages are
641 * pinned" detection logic.
643 static void free_pcppages_bulk(struct zone
*zone
, int count
,
644 struct per_cpu_pages
*pcp
)
650 spin_lock(&zone
->lock
);
651 zone
->pages_scanned
= 0;
655 struct list_head
*list
;
658 * Remove pages from lists in a round-robin fashion. A
659 * batch_free count is maintained that is incremented when an
660 * empty list is encountered. This is so more pages are freed
661 * off fuller lists instead of spinning excessively around empty
666 if (++migratetype
== MIGRATE_PCPTYPES
)
668 list
= &pcp
->lists
[migratetype
];
669 } while (list_empty(list
));
671 /* This is the only non-empty list. Free them all. */
672 if (batch_free
== MIGRATE_PCPTYPES
)
673 batch_free
= to_free
;
676 int mt
; /* migratetype of the to-be-freed page */
678 page
= list_entry(list
->prev
, struct page
, lru
);
679 /* must delete as __free_one_page list manipulates */
680 list_del(&page
->lru
);
681 mt
= get_freepage_migratetype(page
);
682 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
683 __free_one_page(page
, zone
, 0, mt
);
684 trace_mm_page_pcpu_drain(page
, 0, mt
);
685 if (likely(!is_migrate_isolate_page(page
))) {
686 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1);
687 if (is_migrate_cma(mt
))
688 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
, 1);
690 } while (--to_free
&& --batch_free
&& !list_empty(list
));
692 spin_unlock(&zone
->lock
);
695 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
698 spin_lock(&zone
->lock
);
699 zone
->pages_scanned
= 0;
701 __free_one_page(page
, zone
, order
, migratetype
);
702 if (unlikely(!is_migrate_isolate(migratetype
)))
703 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
704 spin_unlock(&zone
->lock
);
707 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
712 trace_mm_page_free(page
, order
);
713 kmemcheck_free_shadow(page
, order
);
716 page
->mapping
= NULL
;
717 for (i
= 0; i
< (1 << order
); i
++)
718 bad
+= free_pages_check(page
+ i
);
722 if (!PageHighMem(page
)) {
723 debug_check_no_locks_freed(page_address(page
),
725 debug_check_no_obj_freed(page_address(page
),
728 arch_free_page(page
, order
);
729 kernel_map_pages(page
, 1 << order
, 0);
734 static void __free_pages_ok(struct page
*page
, unsigned int order
)
739 if (!free_pages_prepare(page
, order
))
742 local_irq_save(flags
);
743 __count_vm_events(PGFREE
, 1 << order
);
744 migratetype
= get_pageblock_migratetype(page
);
745 set_freepage_migratetype(page
, migratetype
);
746 free_one_page(page_zone(page
), page
, order
, migratetype
);
747 local_irq_restore(flags
);
750 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
752 unsigned int nr_pages
= 1 << order
;
753 struct page
*p
= page
;
757 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
759 __ClearPageReserved(p
);
760 set_page_count(p
, 0);
762 __ClearPageReserved(p
);
763 set_page_count(p
, 0);
765 page_zone(page
)->managed_pages
+= nr_pages
;
766 set_page_refcounted(page
);
767 __free_pages(page
, order
);
771 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
772 void __init
init_cma_reserved_pageblock(struct page
*page
)
774 unsigned i
= pageblock_nr_pages
;
775 struct page
*p
= page
;
778 __ClearPageReserved(p
);
779 set_page_count(p
, 0);
782 set_page_refcounted(page
);
783 set_pageblock_migratetype(page
, MIGRATE_CMA
);
784 __free_pages(page
, pageblock_order
);
785 adjust_managed_page_count(page
, pageblock_nr_pages
);
790 * The order of subdivision here is critical for the IO subsystem.
791 * Please do not alter this order without good reasons and regression
792 * testing. Specifically, as large blocks of memory are subdivided,
793 * the order in which smaller blocks are delivered depends on the order
794 * they're subdivided in this function. This is the primary factor
795 * influencing the order in which pages are delivered to the IO
796 * subsystem according to empirical testing, and this is also justified
797 * by considering the behavior of a buddy system containing a single
798 * large block of memory acted on by a series of small allocations.
799 * This behavior is a critical factor in sglist merging's success.
803 static inline void expand(struct zone
*zone
, struct page
*page
,
804 int low
, int high
, struct free_area
*area
,
807 unsigned long size
= 1 << high
;
813 VM_BUG_ON(bad_range(zone
, &page
[size
]));
815 #ifdef CONFIG_DEBUG_PAGEALLOC
816 if (high
< debug_guardpage_minorder()) {
818 * Mark as guard pages (or page), that will allow to
819 * merge back to allocator when buddy will be freed.
820 * Corresponding page table entries will not be touched,
821 * pages will stay not present in virtual address space
823 INIT_LIST_HEAD(&page
[size
].lru
);
824 set_page_guard_flag(&page
[size
]);
825 set_page_private(&page
[size
], high
);
826 /* Guard pages are not available for any usage */
827 __mod_zone_freepage_state(zone
, -(1 << high
),
832 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
834 set_page_order(&page
[size
], high
);
839 * This page is about to be returned from the page allocator
841 static inline int check_new_page(struct page
*page
)
843 if (unlikely(page_mapcount(page
) |
844 (page
->mapping
!= NULL
) |
845 (atomic_read(&page
->_count
) != 0) |
846 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
) |
847 (mem_cgroup_bad_page_check(page
)))) {
854 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
858 for (i
= 0; i
< (1 << order
); i
++) {
859 struct page
*p
= page
+ i
;
860 if (unlikely(check_new_page(p
)))
864 set_page_private(page
, 0);
865 set_page_refcounted(page
);
867 arch_alloc_page(page
, order
);
868 kernel_map_pages(page
, 1 << order
, 1);
870 if (gfp_flags
& __GFP_ZERO
)
871 prep_zero_page(page
, order
, gfp_flags
);
873 if (order
&& (gfp_flags
& __GFP_COMP
))
874 prep_compound_page(page
, order
);
880 * Go through the free lists for the given migratetype and remove
881 * the smallest available page from the freelists
884 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
887 unsigned int current_order
;
888 struct free_area
*area
;
891 /* Find a page of the appropriate size in the preferred list */
892 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
893 area
= &(zone
->free_area
[current_order
]);
894 if (list_empty(&area
->free_list
[migratetype
]))
897 page
= list_entry(area
->free_list
[migratetype
].next
,
899 list_del(&page
->lru
);
900 rmv_page_order(page
);
902 expand(zone
, page
, order
, current_order
, area
, migratetype
);
911 * This array describes the order lists are fallen back to when
912 * the free lists for the desirable migrate type are depleted
914 static int fallbacks
[MIGRATE_TYPES
][4] = {
915 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
916 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
918 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
919 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
921 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
923 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
924 #ifdef CONFIG_MEMORY_ISOLATION
925 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
930 * Move the free pages in a range to the free lists of the requested type.
931 * Note that start_page and end_pages are not aligned on a pageblock
932 * boundary. If alignment is required, use move_freepages_block()
934 int move_freepages(struct zone
*zone
,
935 struct page
*start_page
, struct page
*end_page
,
942 #ifndef CONFIG_HOLES_IN_ZONE
944 * page_zone is not safe to call in this context when
945 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
946 * anyway as we check zone boundaries in move_freepages_block().
947 * Remove at a later date when no bug reports exist related to
948 * grouping pages by mobility
950 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
953 for (page
= start_page
; page
<= end_page
;) {
954 /* Make sure we are not inadvertently changing nodes */
955 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
957 if (!pfn_valid_within(page_to_pfn(page
))) {
962 if (!PageBuddy(page
)) {
967 order
= page_order(page
);
968 list_move(&page
->lru
,
969 &zone
->free_area
[order
].free_list
[migratetype
]);
970 set_freepage_migratetype(page
, migratetype
);
972 pages_moved
+= 1 << order
;
978 int move_freepages_block(struct zone
*zone
, struct page
*page
,
981 unsigned long start_pfn
, end_pfn
;
982 struct page
*start_page
, *end_page
;
984 start_pfn
= page_to_pfn(page
);
985 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
986 start_page
= pfn_to_page(start_pfn
);
987 end_page
= start_page
+ pageblock_nr_pages
- 1;
988 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
990 /* Do not cross zone boundaries */
991 if (!zone_spans_pfn(zone
, start_pfn
))
993 if (!zone_spans_pfn(zone
, end_pfn
))
996 return move_freepages(zone
, start_page
, end_page
, migratetype
);
999 static void change_pageblock_range(struct page
*pageblock_page
,
1000 int start_order
, int migratetype
)
1002 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1004 while (nr_pageblocks
--) {
1005 set_pageblock_migratetype(pageblock_page
, migratetype
);
1006 pageblock_page
+= pageblock_nr_pages
;
1011 * If breaking a large block of pages, move all free pages to the preferred
1012 * allocation list. If falling back for a reclaimable kernel allocation, be
1013 * more aggressive about taking ownership of free pages.
1015 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1016 * nor move CMA pages to different free lists. We don't want unmovable pages
1017 * to be allocated from MIGRATE_CMA areas.
1019 * Returns the new migratetype of the pageblock (or the same old migratetype
1020 * if it was unchanged).
1022 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1023 int start_type
, int fallback_type
)
1025 int current_order
= page_order(page
);
1027 if (is_migrate_cma(fallback_type
))
1028 return fallback_type
;
1030 /* Take ownership for orders >= pageblock_order */
1031 if (current_order
>= pageblock_order
) {
1032 change_pageblock_range(page
, current_order
, start_type
);
1036 if (current_order
>= pageblock_order
/ 2 ||
1037 start_type
== MIGRATE_RECLAIMABLE
||
1038 page_group_by_mobility_disabled
) {
1041 pages
= move_freepages_block(zone
, page
, start_type
);
1043 /* Claim the whole block if over half of it is free */
1044 if (pages
>= (1 << (pageblock_order
-1)) ||
1045 page_group_by_mobility_disabled
) {
1047 set_pageblock_migratetype(page
, start_type
);
1053 return fallback_type
;
1056 /* Remove an element from the buddy allocator from the fallback list */
1057 static inline struct page
*
1058 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
1060 struct free_area
*area
;
1063 int migratetype
, new_type
, i
;
1065 /* Find the largest possible block of pages in the other list */
1066 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
1069 migratetype
= fallbacks
[start_migratetype
][i
];
1071 /* MIGRATE_RESERVE handled later if necessary */
1072 if (migratetype
== MIGRATE_RESERVE
)
1075 area
= &(zone
->free_area
[current_order
]);
1076 if (list_empty(&area
->free_list
[migratetype
]))
1079 page
= list_entry(area
->free_list
[migratetype
].next
,
1083 new_type
= try_to_steal_freepages(zone
, page
,
1087 /* Remove the page from the freelists */
1088 list_del(&page
->lru
);
1089 rmv_page_order(page
);
1092 * Borrow the excess buddy pages as well, irrespective
1093 * of whether we stole freepages, or took ownership of
1094 * the pageblock or not.
1096 * Exception: When borrowing from MIGRATE_CMA, release
1097 * the excess buddy pages to CMA itself.
1099 expand(zone
, page
, order
, current_order
, area
,
1100 is_migrate_cma(migratetype
)
1101 ? migratetype
: start_migratetype
);
1103 trace_mm_page_alloc_extfrag(page
, order
,
1104 current_order
, start_migratetype
, migratetype
,
1105 new_type
== start_migratetype
);
1115 * Do the hard work of removing an element from the buddy allocator.
1116 * Call me with the zone->lock already held.
1118 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1124 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1126 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1127 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1130 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1131 * is used because __rmqueue_smallest is an inline function
1132 * and we want just one call site
1135 migratetype
= MIGRATE_RESERVE
;
1140 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1145 * Obtain a specified number of elements from the buddy allocator, all under
1146 * a single hold of the lock, for efficiency. Add them to the supplied list.
1147 * Returns the number of new pages which were placed at *list.
1149 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1150 unsigned long count
, struct list_head
*list
,
1151 int migratetype
, int cold
)
1153 int mt
= migratetype
, i
;
1155 spin_lock(&zone
->lock
);
1156 for (i
= 0; i
< count
; ++i
) {
1157 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1158 if (unlikely(page
== NULL
))
1162 * Split buddy pages returned by expand() are received here
1163 * in physical page order. The page is added to the callers and
1164 * list and the list head then moves forward. From the callers
1165 * perspective, the linked list is ordered by page number in
1166 * some conditions. This is useful for IO devices that can
1167 * merge IO requests if the physical pages are ordered
1170 if (likely(cold
== 0))
1171 list_add(&page
->lru
, list
);
1173 list_add_tail(&page
->lru
, list
);
1174 if (IS_ENABLED(CONFIG_CMA
)) {
1175 mt
= get_pageblock_migratetype(page
);
1176 if (!is_migrate_cma(mt
) && !is_migrate_isolate(mt
))
1179 set_freepage_migratetype(page
, mt
);
1181 if (is_migrate_cma(mt
))
1182 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1185 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1186 spin_unlock(&zone
->lock
);
1192 * Called from the vmstat counter updater to drain pagesets of this
1193 * currently executing processor on remote nodes after they have
1196 * Note that this function must be called with the thread pinned to
1197 * a single processor.
1199 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1201 unsigned long flags
;
1203 unsigned long batch
;
1205 local_irq_save(flags
);
1206 batch
= ACCESS_ONCE(pcp
->batch
);
1207 if (pcp
->count
>= batch
)
1210 to_drain
= pcp
->count
;
1212 free_pcppages_bulk(zone
, to_drain
, pcp
);
1213 pcp
->count
-= to_drain
;
1215 local_irq_restore(flags
);
1220 * Drain pages of the indicated processor.
1222 * The processor must either be the current processor and the
1223 * thread pinned to the current processor or a processor that
1226 static void drain_pages(unsigned int cpu
)
1228 unsigned long flags
;
1231 for_each_populated_zone(zone
) {
1232 struct per_cpu_pageset
*pset
;
1233 struct per_cpu_pages
*pcp
;
1235 local_irq_save(flags
);
1236 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1240 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1243 local_irq_restore(flags
);
1248 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1250 void drain_local_pages(void *arg
)
1252 drain_pages(smp_processor_id());
1256 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1258 * Note that this code is protected against sending an IPI to an offline
1259 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1260 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1261 * nothing keeps CPUs from showing up after we populated the cpumask and
1262 * before the call to on_each_cpu_mask().
1264 void drain_all_pages(void)
1267 struct per_cpu_pageset
*pcp
;
1271 * Allocate in the BSS so we wont require allocation in
1272 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1274 static cpumask_t cpus_with_pcps
;
1277 * We don't care about racing with CPU hotplug event
1278 * as offline notification will cause the notified
1279 * cpu to drain that CPU pcps and on_each_cpu_mask
1280 * disables preemption as part of its processing
1282 for_each_online_cpu(cpu
) {
1283 bool has_pcps
= false;
1284 for_each_populated_zone(zone
) {
1285 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1286 if (pcp
->pcp
.count
) {
1292 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1294 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1296 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1299 #ifdef CONFIG_HIBERNATION
1301 void mark_free_pages(struct zone
*zone
)
1303 unsigned long pfn
, max_zone_pfn
;
1304 unsigned long flags
;
1306 struct list_head
*curr
;
1308 if (zone_is_empty(zone
))
1311 spin_lock_irqsave(&zone
->lock
, flags
);
1313 max_zone_pfn
= zone_end_pfn(zone
);
1314 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1315 if (pfn_valid(pfn
)) {
1316 struct page
*page
= pfn_to_page(pfn
);
1318 if (!swsusp_page_is_forbidden(page
))
1319 swsusp_unset_page_free(page
);
1322 for_each_migratetype_order(order
, t
) {
1323 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1326 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1327 for (i
= 0; i
< (1UL << order
); i
++)
1328 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1331 spin_unlock_irqrestore(&zone
->lock
, flags
);
1333 #endif /* CONFIG_PM */
1336 * Free a 0-order page
1337 * cold == 1 ? free a cold page : free a hot page
1339 void free_hot_cold_page(struct page
*page
, int cold
)
1341 struct zone
*zone
= page_zone(page
);
1342 struct per_cpu_pages
*pcp
;
1343 unsigned long flags
;
1346 if (!free_pages_prepare(page
, 0))
1349 migratetype
= get_pageblock_migratetype(page
);
1350 set_freepage_migratetype(page
, migratetype
);
1351 local_irq_save(flags
);
1352 __count_vm_event(PGFREE
);
1355 * We only track unmovable, reclaimable and movable on pcp lists.
1356 * Free ISOLATE pages back to the allocator because they are being
1357 * offlined but treat RESERVE as movable pages so we can get those
1358 * areas back if necessary. Otherwise, we may have to free
1359 * excessively into the page allocator
1361 if (migratetype
>= MIGRATE_PCPTYPES
) {
1362 if (unlikely(is_migrate_isolate(migratetype
))) {
1363 free_one_page(zone
, page
, 0, migratetype
);
1366 migratetype
= MIGRATE_MOVABLE
;
1369 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1371 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1373 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1375 if (pcp
->count
>= pcp
->high
) {
1376 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1377 free_pcppages_bulk(zone
, batch
, pcp
);
1378 pcp
->count
-= batch
;
1382 local_irq_restore(flags
);
1386 * Free a list of 0-order pages
1388 void free_hot_cold_page_list(struct list_head
*list
, int cold
)
1390 struct page
*page
, *next
;
1392 list_for_each_entry_safe(page
, next
, list
, lru
) {
1393 trace_mm_page_free_batched(page
, cold
);
1394 free_hot_cold_page(page
, cold
);
1399 * split_page takes a non-compound higher-order page, and splits it into
1400 * n (1<<order) sub-pages: page[0..n]
1401 * Each sub-page must be freed individually.
1403 * Note: this is probably too low level an operation for use in drivers.
1404 * Please consult with lkml before using this in your driver.
1406 void split_page(struct page
*page
, unsigned int order
)
1410 VM_BUG_ON(PageCompound(page
));
1411 VM_BUG_ON(!page_count(page
));
1413 #ifdef CONFIG_KMEMCHECK
1415 * Split shadow pages too, because free(page[0]) would
1416 * otherwise free the whole shadow.
1418 if (kmemcheck_page_is_tracked(page
))
1419 split_page(virt_to_page(page
[0].shadow
), order
);
1422 for (i
= 1; i
< (1 << order
); i
++)
1423 set_page_refcounted(page
+ i
);
1425 EXPORT_SYMBOL_GPL(split_page
);
1427 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1429 unsigned long watermark
;
1433 BUG_ON(!PageBuddy(page
));
1435 zone
= page_zone(page
);
1436 mt
= get_pageblock_migratetype(page
);
1438 if (!is_migrate_isolate(mt
)) {
1439 /* Obey watermarks as if the page was being allocated */
1440 watermark
= low_wmark_pages(zone
) + (1 << order
);
1441 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1444 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1447 /* Remove page from free list */
1448 list_del(&page
->lru
);
1449 zone
->free_area
[order
].nr_free
--;
1450 rmv_page_order(page
);
1452 /* Set the pageblock if the isolated page is at least a pageblock */
1453 if (order
>= pageblock_order
- 1) {
1454 struct page
*endpage
= page
+ (1 << order
) - 1;
1455 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1456 int mt
= get_pageblock_migratetype(page
);
1457 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1458 set_pageblock_migratetype(page
,
1463 return 1UL << order
;
1467 * Similar to split_page except the page is already free. As this is only
1468 * being used for migration, the migratetype of the block also changes.
1469 * As this is called with interrupts disabled, the caller is responsible
1470 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1473 * Note: this is probably too low level an operation for use in drivers.
1474 * Please consult with lkml before using this in your driver.
1476 int split_free_page(struct page
*page
)
1481 order
= page_order(page
);
1483 nr_pages
= __isolate_free_page(page
, order
);
1487 /* Split into individual pages */
1488 set_page_refcounted(page
);
1489 split_page(page
, order
);
1494 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1495 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1499 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1500 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1503 unsigned long flags
;
1505 int cold
= !!(gfp_flags
& __GFP_COLD
);
1508 if (likely(order
== 0)) {
1509 struct per_cpu_pages
*pcp
;
1510 struct list_head
*list
;
1512 local_irq_save(flags
);
1513 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1514 list
= &pcp
->lists
[migratetype
];
1515 if (list_empty(list
)) {
1516 pcp
->count
+= rmqueue_bulk(zone
, 0,
1519 if (unlikely(list_empty(list
)))
1524 page
= list_entry(list
->prev
, struct page
, lru
);
1526 page
= list_entry(list
->next
, struct page
, lru
);
1528 list_del(&page
->lru
);
1531 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1533 * __GFP_NOFAIL is not to be used in new code.
1535 * All __GFP_NOFAIL callers should be fixed so that they
1536 * properly detect and handle allocation failures.
1538 * We most definitely don't want callers attempting to
1539 * allocate greater than order-1 page units with
1542 WARN_ON_ONCE(order
> 1);
1544 spin_lock_irqsave(&zone
->lock
, flags
);
1545 page
= __rmqueue(zone
, order
, migratetype
);
1546 spin_unlock(&zone
->lock
);
1549 __mod_zone_freepage_state(zone
, -(1 << order
),
1550 get_pageblock_migratetype(page
));
1553 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1554 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1555 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1556 local_irq_restore(flags
);
1558 VM_BUG_ON(bad_range(zone
, page
));
1559 if (prep_new_page(page
, order
, gfp_flags
))
1564 local_irq_restore(flags
);
1568 #ifdef CONFIG_FAIL_PAGE_ALLOC
1571 struct fault_attr attr
;
1573 u32 ignore_gfp_highmem
;
1574 u32 ignore_gfp_wait
;
1576 } fail_page_alloc
= {
1577 .attr
= FAULT_ATTR_INITIALIZER
,
1578 .ignore_gfp_wait
= 1,
1579 .ignore_gfp_highmem
= 1,
1583 static int __init
setup_fail_page_alloc(char *str
)
1585 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1587 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1589 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1591 if (order
< fail_page_alloc
.min_order
)
1593 if (gfp_mask
& __GFP_NOFAIL
)
1595 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1597 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1600 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1603 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1605 static int __init
fail_page_alloc_debugfs(void)
1607 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1610 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1611 &fail_page_alloc
.attr
);
1613 return PTR_ERR(dir
);
1615 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1616 &fail_page_alloc
.ignore_gfp_wait
))
1618 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1619 &fail_page_alloc
.ignore_gfp_highmem
))
1621 if (!debugfs_create_u32("min-order", mode
, dir
,
1622 &fail_page_alloc
.min_order
))
1627 debugfs_remove_recursive(dir
);
1632 late_initcall(fail_page_alloc_debugfs
);
1634 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1636 #else /* CONFIG_FAIL_PAGE_ALLOC */
1638 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1643 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1646 * Return true if free pages are above 'mark'. This takes into account the order
1647 * of the allocation.
1649 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1650 int classzone_idx
, int alloc_flags
, long free_pages
)
1652 /* free_pages my go negative - that's OK */
1654 long lowmem_reserve
= z
->lowmem_reserve
[classzone_idx
];
1658 free_pages
-= (1 << order
) - 1;
1659 if (alloc_flags
& ALLOC_HIGH
)
1661 if (alloc_flags
& ALLOC_HARDER
)
1664 /* If allocation can't use CMA areas don't use free CMA pages */
1665 if (!(alloc_flags
& ALLOC_CMA
))
1666 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1669 if (free_pages
- free_cma
<= min
+ lowmem_reserve
)
1671 for (o
= 0; o
< order
; o
++) {
1672 /* At the next order, this order's pages become unavailable */
1673 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1675 /* Require fewer higher order pages to be free */
1678 if (free_pages
<= min
)
1684 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1685 int classzone_idx
, int alloc_flags
)
1687 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1688 zone_page_state(z
, NR_FREE_PAGES
));
1691 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1692 int classzone_idx
, int alloc_flags
)
1694 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1696 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1697 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1699 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1705 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1706 * skip over zones that are not allowed by the cpuset, or that have
1707 * been recently (in last second) found to be nearly full. See further
1708 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1709 * that have to skip over a lot of full or unallowed zones.
1711 * If the zonelist cache is present in the passed in zonelist, then
1712 * returns a pointer to the allowed node mask (either the current
1713 * tasks mems_allowed, or node_states[N_MEMORY].)
1715 * If the zonelist cache is not available for this zonelist, does
1716 * nothing and returns NULL.
1718 * If the fullzones BITMAP in the zonelist cache is stale (more than
1719 * a second since last zap'd) then we zap it out (clear its bits.)
1721 * We hold off even calling zlc_setup, until after we've checked the
1722 * first zone in the zonelist, on the theory that most allocations will
1723 * be satisfied from that first zone, so best to examine that zone as
1724 * quickly as we can.
1726 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1728 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1729 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1731 zlc
= zonelist
->zlcache_ptr
;
1735 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1736 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1737 zlc
->last_full_zap
= jiffies
;
1740 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1741 &cpuset_current_mems_allowed
:
1742 &node_states
[N_MEMORY
];
1743 return allowednodes
;
1747 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1748 * if it is worth looking at further for free memory:
1749 * 1) Check that the zone isn't thought to be full (doesn't have its
1750 * bit set in the zonelist_cache fullzones BITMAP).
1751 * 2) Check that the zones node (obtained from the zonelist_cache
1752 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1753 * Return true (non-zero) if zone is worth looking at further, or
1754 * else return false (zero) if it is not.
1756 * This check -ignores- the distinction between various watermarks,
1757 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1758 * found to be full for any variation of these watermarks, it will
1759 * be considered full for up to one second by all requests, unless
1760 * we are so low on memory on all allowed nodes that we are forced
1761 * into the second scan of the zonelist.
1763 * In the second scan we ignore this zonelist cache and exactly
1764 * apply the watermarks to all zones, even it is slower to do so.
1765 * We are low on memory in the second scan, and should leave no stone
1766 * unturned looking for a free page.
1768 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1769 nodemask_t
*allowednodes
)
1771 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1772 int i
; /* index of *z in zonelist zones */
1773 int n
; /* node that zone *z is on */
1775 zlc
= zonelist
->zlcache_ptr
;
1779 i
= z
- zonelist
->_zonerefs
;
1782 /* This zone is worth trying if it is allowed but not full */
1783 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1787 * Given 'z' scanning a zonelist, set the corresponding bit in
1788 * zlc->fullzones, so that subsequent attempts to allocate a page
1789 * from that zone don't waste time re-examining it.
1791 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1793 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1794 int i
; /* index of *z in zonelist zones */
1796 zlc
= zonelist
->zlcache_ptr
;
1800 i
= z
- zonelist
->_zonerefs
;
1802 set_bit(i
, zlc
->fullzones
);
1806 * clear all zones full, called after direct reclaim makes progress so that
1807 * a zone that was recently full is not skipped over for up to a second
1809 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1811 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1813 zlc
= zonelist
->zlcache_ptr
;
1817 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1820 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1822 return node_distance(local_zone
->node
, zone
->node
) == LOCAL_DISTANCE
;
1825 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1827 return node_isset(local_zone
->node
, zone
->zone_pgdat
->reclaim_nodes
);
1830 static void __paginginit
init_zone_allows_reclaim(int nid
)
1834 for_each_online_node(i
)
1835 if (node_distance(nid
, i
) <= RECLAIM_DISTANCE
)
1836 node_set(i
, NODE_DATA(nid
)->reclaim_nodes
);
1838 zone_reclaim_mode
= 1;
1841 #else /* CONFIG_NUMA */
1843 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1848 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1849 nodemask_t
*allowednodes
)
1854 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1858 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1862 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1867 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1872 static inline void init_zone_allows_reclaim(int nid
)
1875 #endif /* CONFIG_NUMA */
1878 * get_page_from_freelist goes through the zonelist trying to allocate
1881 static struct page
*
1882 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1883 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1884 struct zone
*preferred_zone
, int migratetype
)
1887 struct page
*page
= NULL
;
1890 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1891 int zlc_active
= 0; /* set if using zonelist_cache */
1892 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1894 classzone_idx
= zone_idx(preferred_zone
);
1897 * Scan zonelist, looking for a zone with enough free.
1898 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1900 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1901 high_zoneidx
, nodemask
) {
1904 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1905 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1907 if ((alloc_flags
& ALLOC_CPUSET
) &&
1908 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1910 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1911 if (unlikely(alloc_flags
& ALLOC_NO_WATERMARKS
))
1914 * Distribute pages in proportion to the individual
1915 * zone size to ensure fair page aging. The zone a
1916 * page was allocated in should have no effect on the
1917 * time the page has in memory before being reclaimed.
1919 * When zone_reclaim_mode is enabled, try to stay in
1920 * local zones in the fastpath. If that fails, the
1921 * slowpath is entered, which will do another pass
1922 * starting with the local zones, but ultimately fall
1923 * back to remote zones that do not partake in the
1924 * fairness round-robin cycle of this zonelist.
1926 if (alloc_flags
& ALLOC_WMARK_LOW
) {
1927 if (zone_page_state(zone
, NR_ALLOC_BATCH
) <= 0)
1929 if (zone_reclaim_mode
&&
1930 !zone_local(preferred_zone
, zone
))
1934 * When allocating a page cache page for writing, we
1935 * want to get it from a zone that is within its dirty
1936 * limit, such that no single zone holds more than its
1937 * proportional share of globally allowed dirty pages.
1938 * The dirty limits take into account the zone's
1939 * lowmem reserves and high watermark so that kswapd
1940 * should be able to balance it without having to
1941 * write pages from its LRU list.
1943 * This may look like it could increase pressure on
1944 * lower zones by failing allocations in higher zones
1945 * before they are full. But the pages that do spill
1946 * over are limited as the lower zones are protected
1947 * by this very same mechanism. It should not become
1948 * a practical burden to them.
1950 * XXX: For now, allow allocations to potentially
1951 * exceed the per-zone dirty limit in the slowpath
1952 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1953 * which is important when on a NUMA setup the allowed
1954 * zones are together not big enough to reach the
1955 * global limit. The proper fix for these situations
1956 * will require awareness of zones in the
1957 * dirty-throttling and the flusher threads.
1959 if ((alloc_flags
& ALLOC_WMARK_LOW
) &&
1960 (gfp_mask
& __GFP_WRITE
) && !zone_dirty_ok(zone
))
1961 goto this_zone_full
;
1963 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1964 if (!zone_watermark_ok(zone
, order
, mark
,
1965 classzone_idx
, alloc_flags
)) {
1968 if (IS_ENABLED(CONFIG_NUMA
) &&
1969 !did_zlc_setup
&& nr_online_nodes
> 1) {
1971 * we do zlc_setup if there are multiple nodes
1972 * and before considering the first zone allowed
1975 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1980 if (zone_reclaim_mode
== 0 ||
1981 !zone_allows_reclaim(preferred_zone
, zone
))
1982 goto this_zone_full
;
1985 * As we may have just activated ZLC, check if the first
1986 * eligible zone has failed zone_reclaim recently.
1988 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1989 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1992 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1994 case ZONE_RECLAIM_NOSCAN
:
1997 case ZONE_RECLAIM_FULL
:
1998 /* scanned but unreclaimable */
2001 /* did we reclaim enough */
2002 if (zone_watermark_ok(zone
, order
, mark
,
2003 classzone_idx
, alloc_flags
))
2007 * Failed to reclaim enough to meet watermark.
2008 * Only mark the zone full if checking the min
2009 * watermark or if we failed to reclaim just
2010 * 1<<order pages or else the page allocator
2011 * fastpath will prematurely mark zones full
2012 * when the watermark is between the low and
2015 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2016 ret
== ZONE_RECLAIM_SOME
)
2017 goto this_zone_full
;
2024 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2025 gfp_mask
, migratetype
);
2029 if (IS_ENABLED(CONFIG_NUMA
))
2030 zlc_mark_zone_full(zonelist
, z
);
2033 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && page
== NULL
&& zlc_active
)) {
2034 /* Disable zlc cache for second zonelist scan */
2041 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2042 * necessary to allocate the page. The expectation is
2043 * that the caller is taking steps that will free more
2044 * memory. The caller should avoid the page being used
2045 * for !PFMEMALLOC purposes.
2047 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2053 * Large machines with many possible nodes should not always dump per-node
2054 * meminfo in irq context.
2056 static inline bool should_suppress_show_mem(void)
2061 ret
= in_interrupt();
2066 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2067 DEFAULT_RATELIMIT_INTERVAL
,
2068 DEFAULT_RATELIMIT_BURST
);
2070 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2072 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2074 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2075 debug_guardpage_minorder() > 0)
2079 * Walking all memory to count page types is very expensive and should
2080 * be inhibited in non-blockable contexts.
2082 if (!(gfp_mask
& __GFP_WAIT
))
2083 filter
|= SHOW_MEM_FILTER_PAGE_COUNT
;
2086 * This documents exceptions given to allocations in certain
2087 * contexts that are allowed to allocate outside current's set
2090 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2091 if (test_thread_flag(TIF_MEMDIE
) ||
2092 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2093 filter
&= ~SHOW_MEM_FILTER_NODES
;
2094 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2095 filter
&= ~SHOW_MEM_FILTER_NODES
;
2098 struct va_format vaf
;
2101 va_start(args
, fmt
);
2106 pr_warn("%pV", &vaf
);
2111 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2112 current
->comm
, order
, gfp_mask
);
2115 if (!should_suppress_show_mem())
2120 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2121 unsigned long did_some_progress
,
2122 unsigned long pages_reclaimed
)
2124 /* Do not loop if specifically requested */
2125 if (gfp_mask
& __GFP_NORETRY
)
2128 /* Always retry if specifically requested */
2129 if (gfp_mask
& __GFP_NOFAIL
)
2133 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2134 * making forward progress without invoking OOM. Suspend also disables
2135 * storage devices so kswapd will not help. Bail if we are suspending.
2137 if (!did_some_progress
&& pm_suspended_storage())
2141 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2142 * means __GFP_NOFAIL, but that may not be true in other
2145 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2149 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2150 * specified, then we retry until we no longer reclaim any pages
2151 * (above), or we've reclaimed an order of pages at least as
2152 * large as the allocation's order. In both cases, if the
2153 * allocation still fails, we stop retrying.
2155 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2161 static inline struct page
*
2162 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2163 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2164 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2169 /* Acquire the OOM killer lock for the zones in zonelist */
2170 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
2171 schedule_timeout_uninterruptible(1);
2176 * Go through the zonelist yet one more time, keep very high watermark
2177 * here, this is only to catch a parallel oom killing, we must fail if
2178 * we're still under heavy pressure.
2180 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2181 order
, zonelist
, high_zoneidx
,
2182 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2183 preferred_zone
, migratetype
);
2187 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2188 /* The OOM killer will not help higher order allocs */
2189 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2191 /* The OOM killer does not needlessly kill tasks for lowmem */
2192 if (high_zoneidx
< ZONE_NORMAL
)
2195 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2196 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2197 * The caller should handle page allocation failure by itself if
2198 * it specifies __GFP_THISNODE.
2199 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2201 if (gfp_mask
& __GFP_THISNODE
)
2204 /* Exhausted what can be done so it's blamo time */
2205 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2208 clear_zonelist_oom(zonelist
, gfp_mask
);
2212 #ifdef CONFIG_COMPACTION
2213 /* Try memory compaction for high-order allocations before reclaim */
2214 static struct page
*
2215 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2216 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2217 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2218 int migratetype
, bool sync_migration
,
2219 bool *contended_compaction
, bool *deferred_compaction
,
2220 unsigned long *did_some_progress
)
2225 if (compaction_deferred(preferred_zone
, order
)) {
2226 *deferred_compaction
= true;
2230 current
->flags
|= PF_MEMALLOC
;
2231 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2232 nodemask
, sync_migration
,
2233 contended_compaction
);
2234 current
->flags
&= ~PF_MEMALLOC
;
2236 if (*did_some_progress
!= COMPACT_SKIPPED
) {
2239 /* Page migration frees to the PCP lists but we want merging */
2240 drain_pages(get_cpu());
2243 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2244 order
, zonelist
, high_zoneidx
,
2245 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2246 preferred_zone
, migratetype
);
2248 preferred_zone
->compact_blockskip_flush
= false;
2249 preferred_zone
->compact_considered
= 0;
2250 preferred_zone
->compact_defer_shift
= 0;
2251 if (order
>= preferred_zone
->compact_order_failed
)
2252 preferred_zone
->compact_order_failed
= order
+ 1;
2253 count_vm_event(COMPACTSUCCESS
);
2258 * It's bad if compaction run occurs and fails.
2259 * The most likely reason is that pages exist,
2260 * but not enough to satisfy watermarks.
2262 count_vm_event(COMPACTFAIL
);
2265 * As async compaction considers a subset of pageblocks, only
2266 * defer if the failure was a sync compaction failure.
2269 defer_compaction(preferred_zone
, order
);
2277 static inline struct page
*
2278 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2279 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2280 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2281 int migratetype
, bool sync_migration
,
2282 bool *contended_compaction
, bool *deferred_compaction
,
2283 unsigned long *did_some_progress
)
2287 #endif /* CONFIG_COMPACTION */
2289 /* Perform direct synchronous page reclaim */
2291 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2292 nodemask_t
*nodemask
)
2294 struct reclaim_state reclaim_state
;
2299 /* We now go into synchronous reclaim */
2300 cpuset_memory_pressure_bump();
2301 current
->flags
|= PF_MEMALLOC
;
2302 lockdep_set_current_reclaim_state(gfp_mask
);
2303 reclaim_state
.reclaimed_slab
= 0;
2304 current
->reclaim_state
= &reclaim_state
;
2306 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2308 current
->reclaim_state
= NULL
;
2309 lockdep_clear_current_reclaim_state();
2310 current
->flags
&= ~PF_MEMALLOC
;
2317 /* The really slow allocator path where we enter direct reclaim */
2318 static inline struct page
*
2319 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2320 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2321 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2322 int migratetype
, unsigned long *did_some_progress
)
2324 struct page
*page
= NULL
;
2325 bool drained
= false;
2327 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2329 if (unlikely(!(*did_some_progress
)))
2332 /* After successful reclaim, reconsider all zones for allocation */
2333 if (IS_ENABLED(CONFIG_NUMA
))
2334 zlc_clear_zones_full(zonelist
);
2337 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2338 zonelist
, high_zoneidx
,
2339 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2340 preferred_zone
, migratetype
);
2343 * If an allocation failed after direct reclaim, it could be because
2344 * pages are pinned on the per-cpu lists. Drain them and try again
2346 if (!page
&& !drained
) {
2356 * This is called in the allocator slow-path if the allocation request is of
2357 * sufficient urgency to ignore watermarks and take other desperate measures
2359 static inline struct page
*
2360 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2361 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2362 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2368 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2369 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2370 preferred_zone
, migratetype
);
2372 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2373 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2374 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2379 static void prepare_slowpath(gfp_t gfp_mask
, unsigned int order
,
2380 struct zonelist
*zonelist
,
2381 enum zone_type high_zoneidx
,
2382 struct zone
*preferred_zone
)
2387 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
2388 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2389 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2391 * Only reset the batches of zones that were actually
2392 * considered in the fast path, we don't want to
2393 * thrash fairness information for zones that are not
2394 * actually part of this zonelist's round-robin cycle.
2396 if (zone_reclaim_mode
&& !zone_local(preferred_zone
, zone
))
2398 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2399 high_wmark_pages(zone
) -
2400 low_wmark_pages(zone
) -
2401 zone_page_state(zone
, NR_ALLOC_BATCH
));
2406 gfp_to_alloc_flags(gfp_t gfp_mask
)
2408 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2409 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2411 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2412 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2415 * The caller may dip into page reserves a bit more if the caller
2416 * cannot run direct reclaim, or if the caller has realtime scheduling
2417 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2418 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2420 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2424 * Not worth trying to allocate harder for
2425 * __GFP_NOMEMALLOC even if it can't schedule.
2427 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2428 alloc_flags
|= ALLOC_HARDER
;
2430 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2431 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2433 alloc_flags
&= ~ALLOC_CPUSET
;
2434 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2435 alloc_flags
|= ALLOC_HARDER
;
2437 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2438 if (gfp_mask
& __GFP_MEMALLOC
)
2439 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2440 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2441 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2442 else if (!in_interrupt() &&
2443 ((current
->flags
& PF_MEMALLOC
) ||
2444 unlikely(test_thread_flag(TIF_MEMDIE
))))
2445 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2448 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2449 alloc_flags
|= ALLOC_CMA
;
2454 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2456 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2459 static inline struct page
*
2460 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2461 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2462 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2465 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2466 struct page
*page
= NULL
;
2468 unsigned long pages_reclaimed
= 0;
2469 unsigned long did_some_progress
;
2470 bool sync_migration
= false;
2471 bool deferred_compaction
= false;
2472 bool contended_compaction
= false;
2475 * In the slowpath, we sanity check order to avoid ever trying to
2476 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2477 * be using allocators in order of preference for an area that is
2480 if (order
>= MAX_ORDER
) {
2481 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2486 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2487 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2488 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2489 * using a larger set of nodes after it has established that the
2490 * allowed per node queues are empty and that nodes are
2493 if (IS_ENABLED(CONFIG_NUMA
) &&
2494 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2498 prepare_slowpath(gfp_mask
, order
, zonelist
,
2499 high_zoneidx
, preferred_zone
);
2502 * OK, we're below the kswapd watermark and have kicked background
2503 * reclaim. Now things get more complex, so set up alloc_flags according
2504 * to how we want to proceed.
2506 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2509 * Find the true preferred zone if the allocation is unconstrained by
2512 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2513 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2517 /* This is the last chance, in general, before the goto nopage. */
2518 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2519 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2520 preferred_zone
, migratetype
);
2524 /* Allocate without watermarks if the context allows */
2525 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2527 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2528 * the allocation is high priority and these type of
2529 * allocations are system rather than user orientated
2531 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2533 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2534 zonelist
, high_zoneidx
, nodemask
,
2535 preferred_zone
, migratetype
);
2541 /* Atomic allocations - we can't balance anything */
2545 /* Avoid recursion of direct reclaim */
2546 if (current
->flags
& PF_MEMALLOC
)
2549 /* Avoid allocations with no watermarks from looping endlessly */
2550 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2554 * Try direct compaction. The first pass is asynchronous. Subsequent
2555 * attempts after direct reclaim are synchronous
2557 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2558 zonelist
, high_zoneidx
,
2560 alloc_flags
, preferred_zone
,
2561 migratetype
, sync_migration
,
2562 &contended_compaction
,
2563 &deferred_compaction
,
2564 &did_some_progress
);
2567 sync_migration
= true;
2570 * If compaction is deferred for high-order allocations, it is because
2571 * sync compaction recently failed. In this is the case and the caller
2572 * requested a movable allocation that does not heavily disrupt the
2573 * system then fail the allocation instead of entering direct reclaim.
2575 if ((deferred_compaction
|| contended_compaction
) &&
2576 (gfp_mask
& __GFP_NO_KSWAPD
))
2579 /* Try direct reclaim and then allocating */
2580 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2581 zonelist
, high_zoneidx
,
2583 alloc_flags
, preferred_zone
,
2584 migratetype
, &did_some_progress
);
2589 * If we failed to make any progress reclaiming, then we are
2590 * running out of options and have to consider going OOM
2592 if (!did_some_progress
) {
2593 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2594 if (oom_killer_disabled
)
2596 /* Coredumps can quickly deplete all memory reserves */
2597 if ((current
->flags
& PF_DUMPCORE
) &&
2598 !(gfp_mask
& __GFP_NOFAIL
))
2600 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2601 zonelist
, high_zoneidx
,
2602 nodemask
, preferred_zone
,
2607 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2609 * The oom killer is not called for high-order
2610 * allocations that may fail, so if no progress
2611 * is being made, there are no other options and
2612 * retrying is unlikely to help.
2614 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2617 * The oom killer is not called for lowmem
2618 * allocations to prevent needlessly killing
2621 if (high_zoneidx
< ZONE_NORMAL
)
2629 /* Check if we should retry the allocation */
2630 pages_reclaimed
+= did_some_progress
;
2631 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2633 /* Wait for some write requests to complete then retry */
2634 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2638 * High-order allocations do not necessarily loop after
2639 * direct reclaim and reclaim/compaction depends on compaction
2640 * being called after reclaim so call directly if necessary
2642 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2643 zonelist
, high_zoneidx
,
2645 alloc_flags
, preferred_zone
,
2646 migratetype
, sync_migration
,
2647 &contended_compaction
,
2648 &deferred_compaction
,
2649 &did_some_progress
);
2655 warn_alloc_failed(gfp_mask
, order
, NULL
);
2658 if (kmemcheck_enabled
)
2659 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2665 * This is the 'heart' of the zoned buddy allocator.
2668 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2669 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2671 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2672 struct zone
*preferred_zone
;
2673 struct page
*page
= NULL
;
2674 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2675 unsigned int cpuset_mems_cookie
;
2676 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
;
2677 struct mem_cgroup
*memcg
= NULL
;
2679 gfp_mask
&= gfp_allowed_mask
;
2681 lockdep_trace_alloc(gfp_mask
);
2683 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2685 if (should_fail_alloc_page(gfp_mask
, order
))
2689 * Check the zones suitable for the gfp_mask contain at least one
2690 * valid zone. It's possible to have an empty zonelist as a result
2691 * of GFP_THISNODE and a memoryless node
2693 if (unlikely(!zonelist
->_zonerefs
->zone
))
2697 * Will only have any effect when __GFP_KMEMCG is set. This is
2698 * verified in the (always inline) callee
2700 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2704 cpuset_mems_cookie
= get_mems_allowed();
2706 /* The preferred zone is used for statistics later */
2707 first_zones_zonelist(zonelist
, high_zoneidx
,
2708 nodemask
? : &cpuset_current_mems_allowed
,
2710 if (!preferred_zone
)
2714 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2715 alloc_flags
|= ALLOC_CMA
;
2717 /* First allocation attempt */
2718 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2719 zonelist
, high_zoneidx
, alloc_flags
,
2720 preferred_zone
, migratetype
);
2721 if (unlikely(!page
)) {
2723 * Runtime PM, block IO and its error handling path
2724 * can deadlock because I/O on the device might not
2727 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2728 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2729 zonelist
, high_zoneidx
, nodemask
,
2730 preferred_zone
, migratetype
);
2733 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2737 * When updating a task's mems_allowed, it is possible to race with
2738 * parallel threads in such a way that an allocation can fail while
2739 * the mask is being updated. If a page allocation is about to fail,
2740 * check if the cpuset changed during allocation and if so, retry.
2742 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2745 memcg_kmem_commit_charge(page
, memcg
, order
);
2749 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2752 * Common helper functions.
2754 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2759 * __get_free_pages() returns a 32-bit address, which cannot represent
2762 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2764 page
= alloc_pages(gfp_mask
, order
);
2767 return (unsigned long) page_address(page
);
2769 EXPORT_SYMBOL(__get_free_pages
);
2771 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2773 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2775 EXPORT_SYMBOL(get_zeroed_page
);
2777 void __free_pages(struct page
*page
, unsigned int order
)
2779 if (put_page_testzero(page
)) {
2781 free_hot_cold_page(page
, 0);
2783 __free_pages_ok(page
, order
);
2787 EXPORT_SYMBOL(__free_pages
);
2789 void free_pages(unsigned long addr
, unsigned int order
)
2792 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2793 __free_pages(virt_to_page((void *)addr
), order
);
2797 EXPORT_SYMBOL(free_pages
);
2800 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2801 * pages allocated with __GFP_KMEMCG.
2803 * Those pages are accounted to a particular memcg, embedded in the
2804 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2805 * for that information only to find out that it is NULL for users who have no
2806 * interest in that whatsoever, we provide these functions.
2808 * The caller knows better which flags it relies on.
2810 void __free_memcg_kmem_pages(struct page
*page
, unsigned int order
)
2812 memcg_kmem_uncharge_pages(page
, order
);
2813 __free_pages(page
, order
);
2816 void free_memcg_kmem_pages(unsigned long addr
, unsigned int order
)
2819 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2820 __free_memcg_kmem_pages(virt_to_page((void *)addr
), order
);
2824 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2827 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2828 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2830 split_page(virt_to_page((void *)addr
), order
);
2831 while (used
< alloc_end
) {
2836 return (void *)addr
;
2840 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2841 * @size: the number of bytes to allocate
2842 * @gfp_mask: GFP flags for the allocation
2844 * This function is similar to alloc_pages(), except that it allocates the
2845 * minimum number of pages to satisfy the request. alloc_pages() can only
2846 * allocate memory in power-of-two pages.
2848 * This function is also limited by MAX_ORDER.
2850 * Memory allocated by this function must be released by free_pages_exact().
2852 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2854 unsigned int order
= get_order(size
);
2857 addr
= __get_free_pages(gfp_mask
, order
);
2858 return make_alloc_exact(addr
, order
, size
);
2860 EXPORT_SYMBOL(alloc_pages_exact
);
2863 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2865 * @nid: the preferred node ID where memory should be allocated
2866 * @size: the number of bytes to allocate
2867 * @gfp_mask: GFP flags for the allocation
2869 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2871 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2874 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2876 unsigned order
= get_order(size
);
2877 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2880 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2882 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2885 * free_pages_exact - release memory allocated via alloc_pages_exact()
2886 * @virt: the value returned by alloc_pages_exact.
2887 * @size: size of allocation, same value as passed to alloc_pages_exact().
2889 * Release the memory allocated by a previous call to alloc_pages_exact.
2891 void free_pages_exact(void *virt
, size_t size
)
2893 unsigned long addr
= (unsigned long)virt
;
2894 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2896 while (addr
< end
) {
2901 EXPORT_SYMBOL(free_pages_exact
);
2904 * nr_free_zone_pages - count number of pages beyond high watermark
2905 * @offset: The zone index of the highest zone
2907 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2908 * high watermark within all zones at or below a given zone index. For each
2909 * zone, the number of pages is calculated as:
2910 * managed_pages - high_pages
2912 static unsigned long nr_free_zone_pages(int offset
)
2917 /* Just pick one node, since fallback list is circular */
2918 unsigned long sum
= 0;
2920 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2922 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2923 unsigned long size
= zone
->managed_pages
;
2924 unsigned long high
= high_wmark_pages(zone
);
2933 * nr_free_buffer_pages - count number of pages beyond high watermark
2935 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2936 * watermark within ZONE_DMA and ZONE_NORMAL.
2938 unsigned long nr_free_buffer_pages(void)
2940 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2942 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2945 * nr_free_pagecache_pages - count number of pages beyond high watermark
2947 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2948 * high watermark within all zones.
2950 unsigned long nr_free_pagecache_pages(void)
2952 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2955 static inline void show_node(struct zone
*zone
)
2957 if (IS_ENABLED(CONFIG_NUMA
))
2958 printk("Node %d ", zone_to_nid(zone
));
2961 void si_meminfo(struct sysinfo
*val
)
2963 val
->totalram
= totalram_pages
;
2965 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2966 val
->bufferram
= nr_blockdev_pages();
2967 val
->totalhigh
= totalhigh_pages
;
2968 val
->freehigh
= nr_free_highpages();
2969 val
->mem_unit
= PAGE_SIZE
;
2972 EXPORT_SYMBOL(si_meminfo
);
2975 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2977 int zone_type
; /* needs to be signed */
2978 unsigned long managed_pages
= 0;
2979 pg_data_t
*pgdat
= NODE_DATA(nid
);
2981 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
2982 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
2983 val
->totalram
= managed_pages
;
2984 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2985 #ifdef CONFIG_HIGHMEM
2986 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
2987 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2993 val
->mem_unit
= PAGE_SIZE
;
2998 * Determine whether the node should be displayed or not, depending on whether
2999 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3001 bool skip_free_areas_node(unsigned int flags
, int nid
)
3004 unsigned int cpuset_mems_cookie
;
3006 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3010 cpuset_mems_cookie
= get_mems_allowed();
3011 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3012 } while (!put_mems_allowed(cpuset_mems_cookie
));
3017 #define K(x) ((x) << (PAGE_SHIFT-10))
3019 static void show_migration_types(unsigned char type
)
3021 static const char types
[MIGRATE_TYPES
] = {
3022 [MIGRATE_UNMOVABLE
] = 'U',
3023 [MIGRATE_RECLAIMABLE
] = 'E',
3024 [MIGRATE_MOVABLE
] = 'M',
3025 [MIGRATE_RESERVE
] = 'R',
3027 [MIGRATE_CMA
] = 'C',
3029 #ifdef CONFIG_MEMORY_ISOLATION
3030 [MIGRATE_ISOLATE
] = 'I',
3033 char tmp
[MIGRATE_TYPES
+ 1];
3037 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3038 if (type
& (1 << i
))
3043 printk("(%s) ", tmp
);
3047 * Show free area list (used inside shift_scroll-lock stuff)
3048 * We also calculate the percentage fragmentation. We do this by counting the
3049 * memory on each free list with the exception of the first item on the list.
3050 * Suppresses nodes that are not allowed by current's cpuset if
3051 * SHOW_MEM_FILTER_NODES is passed.
3053 void show_free_areas(unsigned int filter
)
3058 for_each_populated_zone(zone
) {
3059 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3062 printk("%s per-cpu:\n", zone
->name
);
3064 for_each_online_cpu(cpu
) {
3065 struct per_cpu_pageset
*pageset
;
3067 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3069 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3070 cpu
, pageset
->pcp
.high
,
3071 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3075 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3076 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3078 " dirty:%lu writeback:%lu unstable:%lu\n"
3079 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3080 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3082 global_page_state(NR_ACTIVE_ANON
),
3083 global_page_state(NR_INACTIVE_ANON
),
3084 global_page_state(NR_ISOLATED_ANON
),
3085 global_page_state(NR_ACTIVE_FILE
),
3086 global_page_state(NR_INACTIVE_FILE
),
3087 global_page_state(NR_ISOLATED_FILE
),
3088 global_page_state(NR_UNEVICTABLE
),
3089 global_page_state(NR_FILE_DIRTY
),
3090 global_page_state(NR_WRITEBACK
),
3091 global_page_state(NR_UNSTABLE_NFS
),
3092 global_page_state(NR_FREE_PAGES
),
3093 global_page_state(NR_SLAB_RECLAIMABLE
),
3094 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3095 global_page_state(NR_FILE_MAPPED
),
3096 global_page_state(NR_SHMEM
),
3097 global_page_state(NR_PAGETABLE
),
3098 global_page_state(NR_BOUNCE
),
3099 global_page_state(NR_FREE_CMA_PAGES
));
3101 for_each_populated_zone(zone
) {
3104 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3112 " active_anon:%lukB"
3113 " inactive_anon:%lukB"
3114 " active_file:%lukB"
3115 " inactive_file:%lukB"
3116 " unevictable:%lukB"
3117 " isolated(anon):%lukB"
3118 " isolated(file):%lukB"
3126 " slab_reclaimable:%lukB"
3127 " slab_unreclaimable:%lukB"
3128 " kernel_stack:%lukB"
3133 " writeback_tmp:%lukB"
3134 " pages_scanned:%lu"
3135 " all_unreclaimable? %s"
3138 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3139 K(min_wmark_pages(zone
)),
3140 K(low_wmark_pages(zone
)),
3141 K(high_wmark_pages(zone
)),
3142 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3143 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3144 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3145 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3146 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3147 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3148 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3149 K(zone
->present_pages
),
3150 K(zone
->managed_pages
),
3151 K(zone_page_state(zone
, NR_MLOCK
)),
3152 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3153 K(zone_page_state(zone
, NR_WRITEBACK
)),
3154 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3155 K(zone_page_state(zone
, NR_SHMEM
)),
3156 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3157 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3158 zone_page_state(zone
, NR_KERNEL_STACK
) *
3160 K(zone_page_state(zone
, NR_PAGETABLE
)),
3161 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3162 K(zone_page_state(zone
, NR_BOUNCE
)),
3163 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3164 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3165 zone
->pages_scanned
,
3166 (!zone_reclaimable(zone
) ? "yes" : "no")
3168 printk("lowmem_reserve[]:");
3169 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3170 printk(" %lu", zone
->lowmem_reserve
[i
]);
3174 for_each_populated_zone(zone
) {
3175 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3176 unsigned char types
[MAX_ORDER
];
3178 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3181 printk("%s: ", zone
->name
);
3183 spin_lock_irqsave(&zone
->lock
, flags
);
3184 for (order
= 0; order
< MAX_ORDER
; order
++) {
3185 struct free_area
*area
= &zone
->free_area
[order
];
3188 nr
[order
] = area
->nr_free
;
3189 total
+= nr
[order
] << order
;
3192 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3193 if (!list_empty(&area
->free_list
[type
]))
3194 types
[order
] |= 1 << type
;
3197 spin_unlock_irqrestore(&zone
->lock
, flags
);
3198 for (order
= 0; order
< MAX_ORDER
; order
++) {
3199 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3201 show_migration_types(types
[order
]);
3203 printk("= %lukB\n", K(total
));
3206 hugetlb_show_meminfo();
3208 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3210 show_swap_cache_info();
3213 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3215 zoneref
->zone
= zone
;
3216 zoneref
->zone_idx
= zone_idx(zone
);
3220 * Builds allocation fallback zone lists.
3222 * Add all populated zones of a node to the zonelist.
3224 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3228 enum zone_type zone_type
= MAX_NR_ZONES
;
3232 zone
= pgdat
->node_zones
+ zone_type
;
3233 if (populated_zone(zone
)) {
3234 zoneref_set_zone(zone
,
3235 &zonelist
->_zonerefs
[nr_zones
++]);
3236 check_highest_zone(zone_type
);
3238 } while (zone_type
);
3246 * 0 = automatic detection of better ordering.
3247 * 1 = order by ([node] distance, -zonetype)
3248 * 2 = order by (-zonetype, [node] distance)
3250 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3251 * the same zonelist. So only NUMA can configure this param.
3253 #define ZONELIST_ORDER_DEFAULT 0
3254 #define ZONELIST_ORDER_NODE 1
3255 #define ZONELIST_ORDER_ZONE 2
3257 /* zonelist order in the kernel.
3258 * set_zonelist_order() will set this to NODE or ZONE.
3260 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3261 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3265 /* The value user specified ....changed by config */
3266 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3267 /* string for sysctl */
3268 #define NUMA_ZONELIST_ORDER_LEN 16
3269 char numa_zonelist_order
[16] = "default";
3272 * interface for configure zonelist ordering.
3273 * command line option "numa_zonelist_order"
3274 * = "[dD]efault - default, automatic configuration.
3275 * = "[nN]ode - order by node locality, then by zone within node
3276 * = "[zZ]one - order by zone, then by locality within zone
3279 static int __parse_numa_zonelist_order(char *s
)
3281 if (*s
== 'd' || *s
== 'D') {
3282 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3283 } else if (*s
== 'n' || *s
== 'N') {
3284 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3285 } else if (*s
== 'z' || *s
== 'Z') {
3286 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3289 "Ignoring invalid numa_zonelist_order value: "
3296 static __init
int setup_numa_zonelist_order(char *s
)
3303 ret
= __parse_numa_zonelist_order(s
);
3305 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3309 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3312 * sysctl handler for numa_zonelist_order
3314 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
3315 void __user
*buffer
, size_t *length
,
3318 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3320 static DEFINE_MUTEX(zl_order_mutex
);
3322 mutex_lock(&zl_order_mutex
);
3324 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3328 strcpy(saved_string
, (char *)table
->data
);
3330 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3334 int oldval
= user_zonelist_order
;
3336 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3339 * bogus value. restore saved string
3341 strncpy((char *)table
->data
, saved_string
,
3342 NUMA_ZONELIST_ORDER_LEN
);
3343 user_zonelist_order
= oldval
;
3344 } else if (oldval
!= user_zonelist_order
) {
3345 mutex_lock(&zonelists_mutex
);
3346 build_all_zonelists(NULL
, NULL
);
3347 mutex_unlock(&zonelists_mutex
);
3351 mutex_unlock(&zl_order_mutex
);
3356 #define MAX_NODE_LOAD (nr_online_nodes)
3357 static int node_load
[MAX_NUMNODES
];
3360 * find_next_best_node - find the next node that should appear in a given node's fallback list
3361 * @node: node whose fallback list we're appending
3362 * @used_node_mask: nodemask_t of already used nodes
3364 * We use a number of factors to determine which is the next node that should
3365 * appear on a given node's fallback list. The node should not have appeared
3366 * already in @node's fallback list, and it should be the next closest node
3367 * according to the distance array (which contains arbitrary distance values
3368 * from each node to each node in the system), and should also prefer nodes
3369 * with no CPUs, since presumably they'll have very little allocation pressure
3370 * on them otherwise.
3371 * It returns -1 if no node is found.
3373 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3376 int min_val
= INT_MAX
;
3377 int best_node
= NUMA_NO_NODE
;
3378 const struct cpumask
*tmp
= cpumask_of_node(0);
3380 /* Use the local node if we haven't already */
3381 if (!node_isset(node
, *used_node_mask
)) {
3382 node_set(node
, *used_node_mask
);
3386 for_each_node_state(n
, N_MEMORY
) {
3388 /* Don't want a node to appear more than once */
3389 if (node_isset(n
, *used_node_mask
))
3392 /* Use the distance array to find the distance */
3393 val
= node_distance(node
, n
);
3395 /* Penalize nodes under us ("prefer the next node") */
3398 /* Give preference to headless and unused nodes */
3399 tmp
= cpumask_of_node(n
);
3400 if (!cpumask_empty(tmp
))
3401 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3403 /* Slight preference for less loaded node */
3404 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3405 val
+= node_load
[n
];
3407 if (val
< min_val
) {
3414 node_set(best_node
, *used_node_mask
);
3421 * Build zonelists ordered by node and zones within node.
3422 * This results in maximum locality--normal zone overflows into local
3423 * DMA zone, if any--but risks exhausting DMA zone.
3425 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3428 struct zonelist
*zonelist
;
3430 zonelist
= &pgdat
->node_zonelists
[0];
3431 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3433 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3434 zonelist
->_zonerefs
[j
].zone
= NULL
;
3435 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3439 * Build gfp_thisnode zonelists
3441 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3444 struct zonelist
*zonelist
;
3446 zonelist
= &pgdat
->node_zonelists
[1];
3447 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3448 zonelist
->_zonerefs
[j
].zone
= NULL
;
3449 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3453 * Build zonelists ordered by zone and nodes within zones.
3454 * This results in conserving DMA zone[s] until all Normal memory is
3455 * exhausted, but results in overflowing to remote node while memory
3456 * may still exist in local DMA zone.
3458 static int node_order
[MAX_NUMNODES
];
3460 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3463 int zone_type
; /* needs to be signed */
3465 struct zonelist
*zonelist
;
3467 zonelist
= &pgdat
->node_zonelists
[0];
3469 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3470 for (j
= 0; j
< nr_nodes
; j
++) {
3471 node
= node_order
[j
];
3472 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3473 if (populated_zone(z
)) {
3475 &zonelist
->_zonerefs
[pos
++]);
3476 check_highest_zone(zone_type
);
3480 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3481 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3484 static int default_zonelist_order(void)
3487 unsigned long low_kmem_size
, total_size
;
3491 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3492 * If they are really small and used heavily, the system can fall
3493 * into OOM very easily.
3494 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3496 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3499 for_each_online_node(nid
) {
3500 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3501 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3502 if (populated_zone(z
)) {
3503 if (zone_type
< ZONE_NORMAL
)
3504 low_kmem_size
+= z
->managed_pages
;
3505 total_size
+= z
->managed_pages
;
3506 } else if (zone_type
== ZONE_NORMAL
) {
3508 * If any node has only lowmem, then node order
3509 * is preferred to allow kernel allocations
3510 * locally; otherwise, they can easily infringe
3511 * on other nodes when there is an abundance of
3512 * lowmem available to allocate from.
3514 return ZONELIST_ORDER_NODE
;
3518 if (!low_kmem_size
|| /* there are no DMA area. */
3519 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
3520 return ZONELIST_ORDER_NODE
;
3522 * look into each node's config.
3523 * If there is a node whose DMA/DMA32 memory is very big area on
3524 * local memory, NODE_ORDER may be suitable.
3526 average_size
= total_size
/
3527 (nodes_weight(node_states
[N_MEMORY
]) + 1);
3528 for_each_online_node(nid
) {
3531 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3532 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3533 if (populated_zone(z
)) {
3534 if (zone_type
< ZONE_NORMAL
)
3535 low_kmem_size
+= z
->present_pages
;
3536 total_size
+= z
->present_pages
;
3539 if (low_kmem_size
&&
3540 total_size
> average_size
&& /* ignore small node */
3541 low_kmem_size
> total_size
* 70/100)
3542 return ZONELIST_ORDER_NODE
;
3544 return ZONELIST_ORDER_ZONE
;
3547 static void set_zonelist_order(void)
3549 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3550 current_zonelist_order
= default_zonelist_order();
3552 current_zonelist_order
= user_zonelist_order
;
3555 static void build_zonelists(pg_data_t
*pgdat
)
3559 nodemask_t used_mask
;
3560 int local_node
, prev_node
;
3561 struct zonelist
*zonelist
;
3562 int order
= current_zonelist_order
;
3564 /* initialize zonelists */
3565 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3566 zonelist
= pgdat
->node_zonelists
+ i
;
3567 zonelist
->_zonerefs
[0].zone
= NULL
;
3568 zonelist
->_zonerefs
[0].zone_idx
= 0;
3571 /* NUMA-aware ordering of nodes */
3572 local_node
= pgdat
->node_id
;
3573 load
= nr_online_nodes
;
3574 prev_node
= local_node
;
3575 nodes_clear(used_mask
);
3577 memset(node_order
, 0, sizeof(node_order
));
3580 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3582 * We don't want to pressure a particular node.
3583 * So adding penalty to the first node in same
3584 * distance group to make it round-robin.
3586 if (node_distance(local_node
, node
) !=
3587 node_distance(local_node
, prev_node
))
3588 node_load
[node
] = load
;
3592 if (order
== ZONELIST_ORDER_NODE
)
3593 build_zonelists_in_node_order(pgdat
, node
);
3595 node_order
[j
++] = node
; /* remember order */
3598 if (order
== ZONELIST_ORDER_ZONE
) {
3599 /* calculate node order -- i.e., DMA last! */
3600 build_zonelists_in_zone_order(pgdat
, j
);
3603 build_thisnode_zonelists(pgdat
);
3606 /* Construct the zonelist performance cache - see further mmzone.h */
3607 static void build_zonelist_cache(pg_data_t
*pgdat
)
3609 struct zonelist
*zonelist
;
3610 struct zonelist_cache
*zlc
;
3613 zonelist
= &pgdat
->node_zonelists
[0];
3614 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3615 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3616 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3617 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3620 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3622 * Return node id of node used for "local" allocations.
3623 * I.e., first node id of first zone in arg node's generic zonelist.
3624 * Used for initializing percpu 'numa_mem', which is used primarily
3625 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3627 int local_memory_node(int node
)
3631 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3632 gfp_zone(GFP_KERNEL
),
3639 #else /* CONFIG_NUMA */
3641 static void set_zonelist_order(void)
3643 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3646 static void build_zonelists(pg_data_t
*pgdat
)
3648 int node
, local_node
;
3650 struct zonelist
*zonelist
;
3652 local_node
= pgdat
->node_id
;
3654 zonelist
= &pgdat
->node_zonelists
[0];
3655 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3658 * Now we build the zonelist so that it contains the zones
3659 * of all the other nodes.
3660 * We don't want to pressure a particular node, so when
3661 * building the zones for node N, we make sure that the
3662 * zones coming right after the local ones are those from
3663 * node N+1 (modulo N)
3665 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3666 if (!node_online(node
))
3668 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3670 for (node
= 0; node
< local_node
; node
++) {
3671 if (!node_online(node
))
3673 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3676 zonelist
->_zonerefs
[j
].zone
= NULL
;
3677 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3680 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3681 static void build_zonelist_cache(pg_data_t
*pgdat
)
3683 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3686 #endif /* CONFIG_NUMA */
3689 * Boot pageset table. One per cpu which is going to be used for all
3690 * zones and all nodes. The parameters will be set in such a way
3691 * that an item put on a list will immediately be handed over to
3692 * the buddy list. This is safe since pageset manipulation is done
3693 * with interrupts disabled.
3695 * The boot_pagesets must be kept even after bootup is complete for
3696 * unused processors and/or zones. They do play a role for bootstrapping
3697 * hotplugged processors.
3699 * zoneinfo_show() and maybe other functions do
3700 * not check if the processor is online before following the pageset pointer.
3701 * Other parts of the kernel may not check if the zone is available.
3703 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3704 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3705 static void setup_zone_pageset(struct zone
*zone
);
3708 * Global mutex to protect against size modification of zonelists
3709 * as well as to serialize pageset setup for the new populated zone.
3711 DEFINE_MUTEX(zonelists_mutex
);
3713 /* return values int ....just for stop_machine() */
3714 static int __build_all_zonelists(void *data
)
3718 pg_data_t
*self
= data
;
3721 memset(node_load
, 0, sizeof(node_load
));
3724 if (self
&& !node_online(self
->node_id
)) {
3725 build_zonelists(self
);
3726 build_zonelist_cache(self
);
3729 for_each_online_node(nid
) {
3730 pg_data_t
*pgdat
= NODE_DATA(nid
);
3732 build_zonelists(pgdat
);
3733 build_zonelist_cache(pgdat
);
3737 * Initialize the boot_pagesets that are going to be used
3738 * for bootstrapping processors. The real pagesets for
3739 * each zone will be allocated later when the per cpu
3740 * allocator is available.
3742 * boot_pagesets are used also for bootstrapping offline
3743 * cpus if the system is already booted because the pagesets
3744 * are needed to initialize allocators on a specific cpu too.
3745 * F.e. the percpu allocator needs the page allocator which
3746 * needs the percpu allocator in order to allocate its pagesets
3747 * (a chicken-egg dilemma).
3749 for_each_possible_cpu(cpu
) {
3750 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3752 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3754 * We now know the "local memory node" for each node--
3755 * i.e., the node of the first zone in the generic zonelist.
3756 * Set up numa_mem percpu variable for on-line cpus. During
3757 * boot, only the boot cpu should be on-line; we'll init the
3758 * secondary cpus' numa_mem as they come on-line. During
3759 * node/memory hotplug, we'll fixup all on-line cpus.
3761 if (cpu_online(cpu
))
3762 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3770 * Called with zonelists_mutex held always
3771 * unless system_state == SYSTEM_BOOTING.
3773 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3775 set_zonelist_order();
3777 if (system_state
== SYSTEM_BOOTING
) {
3778 __build_all_zonelists(NULL
);
3779 mminit_verify_zonelist();
3780 cpuset_init_current_mems_allowed();
3782 #ifdef CONFIG_MEMORY_HOTPLUG
3784 setup_zone_pageset(zone
);
3786 /* we have to stop all cpus to guarantee there is no user
3788 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3789 /* cpuset refresh routine should be here */
3791 vm_total_pages
= nr_free_pagecache_pages();
3793 * Disable grouping by mobility if the number of pages in the
3794 * system is too low to allow the mechanism to work. It would be
3795 * more accurate, but expensive to check per-zone. This check is
3796 * made on memory-hotadd so a system can start with mobility
3797 * disabled and enable it later
3799 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3800 page_group_by_mobility_disabled
= 1;
3802 page_group_by_mobility_disabled
= 0;
3804 printk("Built %i zonelists in %s order, mobility grouping %s. "
3805 "Total pages: %ld\n",
3807 zonelist_order_name
[current_zonelist_order
],
3808 page_group_by_mobility_disabled
? "off" : "on",
3811 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3816 * Helper functions to size the waitqueue hash table.
3817 * Essentially these want to choose hash table sizes sufficiently
3818 * large so that collisions trying to wait on pages are rare.
3819 * But in fact, the number of active page waitqueues on typical
3820 * systems is ridiculously low, less than 200. So this is even
3821 * conservative, even though it seems large.
3823 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3824 * waitqueues, i.e. the size of the waitq table given the number of pages.
3826 #define PAGES_PER_WAITQUEUE 256
3828 #ifndef CONFIG_MEMORY_HOTPLUG
3829 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3831 unsigned long size
= 1;
3833 pages
/= PAGES_PER_WAITQUEUE
;
3835 while (size
< pages
)
3839 * Once we have dozens or even hundreds of threads sleeping
3840 * on IO we've got bigger problems than wait queue collision.
3841 * Limit the size of the wait table to a reasonable size.
3843 size
= min(size
, 4096UL);
3845 return max(size
, 4UL);
3849 * A zone's size might be changed by hot-add, so it is not possible to determine
3850 * a suitable size for its wait_table. So we use the maximum size now.
3852 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3854 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3855 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3856 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3858 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3859 * or more by the traditional way. (See above). It equals:
3861 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3862 * ia64(16K page size) : = ( 8G + 4M)byte.
3863 * powerpc (64K page size) : = (32G +16M)byte.
3865 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3872 * This is an integer logarithm so that shifts can be used later
3873 * to extract the more random high bits from the multiplicative
3874 * hash function before the remainder is taken.
3876 static inline unsigned long wait_table_bits(unsigned long size
)
3881 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3884 * Check if a pageblock contains reserved pages
3886 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3890 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3891 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3898 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3899 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3900 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3901 * higher will lead to a bigger reserve which will get freed as contiguous
3902 * blocks as reclaim kicks in
3904 static void setup_zone_migrate_reserve(struct zone
*zone
)
3906 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3908 unsigned long block_migratetype
;
3912 * Get the start pfn, end pfn and the number of blocks to reserve
3913 * We have to be careful to be aligned to pageblock_nr_pages to
3914 * make sure that we always check pfn_valid for the first page in
3917 start_pfn
= zone
->zone_start_pfn
;
3918 end_pfn
= zone_end_pfn(zone
);
3919 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
3920 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3924 * Reserve blocks are generally in place to help high-order atomic
3925 * allocations that are short-lived. A min_free_kbytes value that
3926 * would result in more than 2 reserve blocks for atomic allocations
3927 * is assumed to be in place to help anti-fragmentation for the
3928 * future allocation of hugepages at runtime.
3930 reserve
= min(2, reserve
);
3932 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3933 if (!pfn_valid(pfn
))
3935 page
= pfn_to_page(pfn
);
3937 /* Watch out for overlapping nodes */
3938 if (page_to_nid(page
) != zone_to_nid(zone
))
3941 block_migratetype
= get_pageblock_migratetype(page
);
3943 /* Only test what is necessary when the reserves are not met */
3946 * Blocks with reserved pages will never free, skip
3949 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
3950 if (pageblock_is_reserved(pfn
, block_end_pfn
))
3953 /* If this block is reserved, account for it */
3954 if (block_migratetype
== MIGRATE_RESERVE
) {
3959 /* Suitable for reserving if this block is movable */
3960 if (block_migratetype
== MIGRATE_MOVABLE
) {
3961 set_pageblock_migratetype(page
,
3963 move_freepages_block(zone
, page
,
3971 * If the reserve is met and this is a previous reserved block,
3974 if (block_migratetype
== MIGRATE_RESERVE
) {
3975 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3976 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3982 * Initially all pages are reserved - free ones are freed
3983 * up by free_all_bootmem() once the early boot process is
3984 * done. Non-atomic initialization, single-pass.
3986 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3987 unsigned long start_pfn
, enum memmap_context context
)
3990 unsigned long end_pfn
= start_pfn
+ size
;
3994 if (highest_memmap_pfn
< end_pfn
- 1)
3995 highest_memmap_pfn
= end_pfn
- 1;
3997 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3998 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4000 * There can be holes in boot-time mem_map[]s
4001 * handed to this function. They do not
4002 * exist on hotplugged memory.
4004 if (context
== MEMMAP_EARLY
) {
4005 if (!early_pfn_valid(pfn
))
4007 if (!early_pfn_in_nid(pfn
, nid
))
4010 page
= pfn_to_page(pfn
);
4011 set_page_links(page
, zone
, nid
, pfn
);
4012 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4013 init_page_count(page
);
4014 page_mapcount_reset(page
);
4015 page_nid_reset_last(page
);
4016 SetPageReserved(page
);
4018 * Mark the block movable so that blocks are reserved for
4019 * movable at startup. This will force kernel allocations
4020 * to reserve their blocks rather than leaking throughout
4021 * the address space during boot when many long-lived
4022 * kernel allocations are made. Later some blocks near
4023 * the start are marked MIGRATE_RESERVE by
4024 * setup_zone_migrate_reserve()
4026 * bitmap is created for zone's valid pfn range. but memmap
4027 * can be created for invalid pages (for alignment)
4028 * check here not to call set_pageblock_migratetype() against
4031 if ((z
->zone_start_pfn
<= pfn
)
4032 && (pfn
< zone_end_pfn(z
))
4033 && !(pfn
& (pageblock_nr_pages
- 1)))
4034 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4036 INIT_LIST_HEAD(&page
->lru
);
4037 #ifdef WANT_PAGE_VIRTUAL
4038 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4039 if (!is_highmem_idx(zone
))
4040 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4045 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4048 for_each_migratetype_order(order
, t
) {
4049 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4050 zone
->free_area
[order
].nr_free
= 0;
4054 #ifndef __HAVE_ARCH_MEMMAP_INIT
4055 #define memmap_init(size, nid, zone, start_pfn) \
4056 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4059 static int __meminit
zone_batchsize(struct zone
*zone
)
4065 * The per-cpu-pages pools are set to around 1000th of the
4066 * size of the zone. But no more than 1/2 of a meg.
4068 * OK, so we don't know how big the cache is. So guess.
4070 batch
= zone
->managed_pages
/ 1024;
4071 if (batch
* PAGE_SIZE
> 512 * 1024)
4072 batch
= (512 * 1024) / PAGE_SIZE
;
4073 batch
/= 4; /* We effectively *= 4 below */
4078 * Clamp the batch to a 2^n - 1 value. Having a power
4079 * of 2 value was found to be more likely to have
4080 * suboptimal cache aliasing properties in some cases.
4082 * For example if 2 tasks are alternately allocating
4083 * batches of pages, one task can end up with a lot
4084 * of pages of one half of the possible page colors
4085 * and the other with pages of the other colors.
4087 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4092 /* The deferral and batching of frees should be suppressed under NOMMU
4095 * The problem is that NOMMU needs to be able to allocate large chunks
4096 * of contiguous memory as there's no hardware page translation to
4097 * assemble apparent contiguous memory from discontiguous pages.
4099 * Queueing large contiguous runs of pages for batching, however,
4100 * causes the pages to actually be freed in smaller chunks. As there
4101 * can be a significant delay between the individual batches being
4102 * recycled, this leads to the once large chunks of space being
4103 * fragmented and becoming unavailable for high-order allocations.
4110 * pcp->high and pcp->batch values are related and dependent on one another:
4111 * ->batch must never be higher then ->high.
4112 * The following function updates them in a safe manner without read side
4115 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4116 * those fields changing asynchronously (acording the the above rule).
4118 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4119 * outside of boot time (or some other assurance that no concurrent updaters
4122 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4123 unsigned long batch
)
4125 /* start with a fail safe value for batch */
4129 /* Update high, then batch, in order */
4136 /* a companion to pageset_set_high() */
4137 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4139 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4142 static void pageset_init(struct per_cpu_pageset
*p
)
4144 struct per_cpu_pages
*pcp
;
4147 memset(p
, 0, sizeof(*p
));
4151 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4152 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4155 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4158 pageset_set_batch(p
, batch
);
4162 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4163 * to the value high for the pageset p.
4165 static void pageset_set_high(struct per_cpu_pageset
*p
,
4168 unsigned long batch
= max(1UL, high
/ 4);
4169 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4170 batch
= PAGE_SHIFT
* 8;
4172 pageset_update(&p
->pcp
, high
, batch
);
4175 static void __meminit
pageset_set_high_and_batch(struct zone
*zone
,
4176 struct per_cpu_pageset
*pcp
)
4178 if (percpu_pagelist_fraction
)
4179 pageset_set_high(pcp
,
4180 (zone
->managed_pages
/
4181 percpu_pagelist_fraction
));
4183 pageset_set_batch(pcp
, zone_batchsize(zone
));
4186 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4188 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4191 pageset_set_high_and_batch(zone
, pcp
);
4194 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4197 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4198 for_each_possible_cpu(cpu
)
4199 zone_pageset_init(zone
, cpu
);
4203 * Allocate per cpu pagesets and initialize them.
4204 * Before this call only boot pagesets were available.
4206 void __init
setup_per_cpu_pageset(void)
4210 for_each_populated_zone(zone
)
4211 setup_zone_pageset(zone
);
4214 static noinline __init_refok
4215 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4218 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4222 * The per-page waitqueue mechanism uses hashed waitqueues
4225 zone
->wait_table_hash_nr_entries
=
4226 wait_table_hash_nr_entries(zone_size_pages
);
4227 zone
->wait_table_bits
=
4228 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4229 alloc_size
= zone
->wait_table_hash_nr_entries
4230 * sizeof(wait_queue_head_t
);
4232 if (!slab_is_available()) {
4233 zone
->wait_table
= (wait_queue_head_t
*)
4234 alloc_bootmem_node_nopanic(pgdat
, alloc_size
);
4237 * This case means that a zone whose size was 0 gets new memory
4238 * via memory hot-add.
4239 * But it may be the case that a new node was hot-added. In
4240 * this case vmalloc() will not be able to use this new node's
4241 * memory - this wait_table must be initialized to use this new
4242 * node itself as well.
4243 * To use this new node's memory, further consideration will be
4246 zone
->wait_table
= vmalloc(alloc_size
);
4248 if (!zone
->wait_table
)
4251 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4252 init_waitqueue_head(zone
->wait_table
+ i
);
4257 static __meminit
void zone_pcp_init(struct zone
*zone
)
4260 * per cpu subsystem is not up at this point. The following code
4261 * relies on the ability of the linker to provide the
4262 * offset of a (static) per cpu variable into the per cpu area.
4264 zone
->pageset
= &boot_pageset
;
4266 if (zone
->present_pages
)
4267 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4268 zone
->name
, zone
->present_pages
,
4269 zone_batchsize(zone
));
4272 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4273 unsigned long zone_start_pfn
,
4275 enum memmap_context context
)
4277 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4279 ret
= zone_wait_table_init(zone
, size
);
4282 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4284 zone
->zone_start_pfn
= zone_start_pfn
;
4286 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4287 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4289 (unsigned long)zone_idx(zone
),
4290 zone_start_pfn
, (zone_start_pfn
+ size
));
4292 zone_init_free_lists(zone
);
4297 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4298 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4300 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4301 * Architectures may implement their own version but if add_active_range()
4302 * was used and there are no special requirements, this is a convenient
4305 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4307 unsigned long start_pfn
, end_pfn
;
4310 * NOTE: The following SMP-unsafe globals are only used early in boot
4311 * when the kernel is running single-threaded.
4313 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4314 static int __meminitdata last_nid
;
4316 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4319 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4321 last_start_pfn
= start_pfn
;
4322 last_end_pfn
= end_pfn
;
4328 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4330 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4334 nid
= __early_pfn_to_nid(pfn
);
4337 /* just returns 0 */
4341 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4342 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4346 nid
= __early_pfn_to_nid(pfn
);
4347 if (nid
>= 0 && nid
!= node
)
4354 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4355 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4356 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4358 * If an architecture guarantees that all ranges registered with
4359 * add_active_ranges() contain no holes and may be freed, this
4360 * this function may be used instead of calling free_bootmem() manually.
4362 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4364 unsigned long start_pfn
, end_pfn
;
4367 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4368 start_pfn
= min(start_pfn
, max_low_pfn
);
4369 end_pfn
= min(end_pfn
, max_low_pfn
);
4371 if (start_pfn
< end_pfn
)
4372 free_bootmem_node(NODE_DATA(this_nid
),
4373 PFN_PHYS(start_pfn
),
4374 (end_pfn
- start_pfn
) << PAGE_SHIFT
);
4379 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4380 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4382 * If an architecture guarantees that all ranges registered with
4383 * add_active_ranges() contain no holes and may be freed, this
4384 * function may be used instead of calling memory_present() manually.
4386 void __init
sparse_memory_present_with_active_regions(int nid
)
4388 unsigned long start_pfn
, end_pfn
;
4391 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4392 memory_present(this_nid
, start_pfn
, end_pfn
);
4396 * get_pfn_range_for_nid - Return the start and end page frames for a node
4397 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4398 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4399 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4401 * It returns the start and end page frame of a node based on information
4402 * provided by an arch calling add_active_range(). If called for a node
4403 * with no available memory, a warning is printed and the start and end
4406 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4407 unsigned long *start_pfn
, unsigned long *end_pfn
)
4409 unsigned long this_start_pfn
, this_end_pfn
;
4415 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4416 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4417 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4420 if (*start_pfn
== -1UL)
4425 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4426 * assumption is made that zones within a node are ordered in monotonic
4427 * increasing memory addresses so that the "highest" populated zone is used
4429 static void __init
find_usable_zone_for_movable(void)
4432 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4433 if (zone_index
== ZONE_MOVABLE
)
4436 if (arch_zone_highest_possible_pfn
[zone_index
] >
4437 arch_zone_lowest_possible_pfn
[zone_index
])
4441 VM_BUG_ON(zone_index
== -1);
4442 movable_zone
= zone_index
;
4446 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4447 * because it is sized independent of architecture. Unlike the other zones,
4448 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4449 * in each node depending on the size of each node and how evenly kernelcore
4450 * is distributed. This helper function adjusts the zone ranges
4451 * provided by the architecture for a given node by using the end of the
4452 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4453 * zones within a node are in order of monotonic increases memory addresses
4455 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4456 unsigned long zone_type
,
4457 unsigned long node_start_pfn
,
4458 unsigned long node_end_pfn
,
4459 unsigned long *zone_start_pfn
,
4460 unsigned long *zone_end_pfn
)
4462 /* Only adjust if ZONE_MOVABLE is on this node */
4463 if (zone_movable_pfn
[nid
]) {
4464 /* Size ZONE_MOVABLE */
4465 if (zone_type
== ZONE_MOVABLE
) {
4466 *zone_start_pfn
= zone_movable_pfn
[nid
];
4467 *zone_end_pfn
= min(node_end_pfn
,
4468 arch_zone_highest_possible_pfn
[movable_zone
]);
4470 /* Adjust for ZONE_MOVABLE starting within this range */
4471 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4472 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4473 *zone_end_pfn
= zone_movable_pfn
[nid
];
4475 /* Check if this whole range is within ZONE_MOVABLE */
4476 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4477 *zone_start_pfn
= *zone_end_pfn
;
4482 * Return the number of pages a zone spans in a node, including holes
4483 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4485 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4486 unsigned long zone_type
,
4487 unsigned long node_start_pfn
,
4488 unsigned long node_end_pfn
,
4489 unsigned long *ignored
)
4491 unsigned long zone_start_pfn
, zone_end_pfn
;
4493 /* Get the start and end of the zone */
4494 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4495 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4496 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4497 node_start_pfn
, node_end_pfn
,
4498 &zone_start_pfn
, &zone_end_pfn
);
4500 /* Check that this node has pages within the zone's required range */
4501 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4504 /* Move the zone boundaries inside the node if necessary */
4505 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4506 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4508 /* Return the spanned pages */
4509 return zone_end_pfn
- zone_start_pfn
;
4513 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4514 * then all holes in the requested range will be accounted for.
4516 unsigned long __meminit
__absent_pages_in_range(int nid
,
4517 unsigned long range_start_pfn
,
4518 unsigned long range_end_pfn
)
4520 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4521 unsigned long start_pfn
, end_pfn
;
4524 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4525 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4526 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4527 nr_absent
-= end_pfn
- start_pfn
;
4533 * absent_pages_in_range - Return number of page frames in holes within a range
4534 * @start_pfn: The start PFN to start searching for holes
4535 * @end_pfn: The end PFN to stop searching for holes
4537 * It returns the number of pages frames in memory holes within a range.
4539 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4540 unsigned long end_pfn
)
4542 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4545 /* Return the number of page frames in holes in a zone on a node */
4546 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4547 unsigned long zone_type
,
4548 unsigned long node_start_pfn
,
4549 unsigned long node_end_pfn
,
4550 unsigned long *ignored
)
4552 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4553 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4554 unsigned long zone_start_pfn
, zone_end_pfn
;
4556 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4557 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4559 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4560 node_start_pfn
, node_end_pfn
,
4561 &zone_start_pfn
, &zone_end_pfn
);
4562 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4565 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4566 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4567 unsigned long zone_type
,
4568 unsigned long node_start_pfn
,
4569 unsigned long node_end_pfn
,
4570 unsigned long *zones_size
)
4572 return zones_size
[zone_type
];
4575 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4576 unsigned long zone_type
,
4577 unsigned long node_start_pfn
,
4578 unsigned long node_end_pfn
,
4579 unsigned long *zholes_size
)
4584 return zholes_size
[zone_type
];
4587 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4589 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4590 unsigned long node_start_pfn
,
4591 unsigned long node_end_pfn
,
4592 unsigned long *zones_size
,
4593 unsigned long *zholes_size
)
4595 unsigned long realtotalpages
, totalpages
= 0;
4598 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4599 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4603 pgdat
->node_spanned_pages
= totalpages
;
4605 realtotalpages
= totalpages
;
4606 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4608 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4609 node_start_pfn
, node_end_pfn
,
4611 pgdat
->node_present_pages
= realtotalpages
;
4612 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4616 #ifndef CONFIG_SPARSEMEM
4618 * Calculate the size of the zone->blockflags rounded to an unsigned long
4619 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4620 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4621 * round what is now in bits to nearest long in bits, then return it in
4624 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4626 unsigned long usemapsize
;
4628 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4629 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4630 usemapsize
= usemapsize
>> pageblock_order
;
4631 usemapsize
*= NR_PAGEBLOCK_BITS
;
4632 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4634 return usemapsize
/ 8;
4637 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4639 unsigned long zone_start_pfn
,
4640 unsigned long zonesize
)
4642 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4643 zone
->pageblock_flags
= NULL
;
4645 zone
->pageblock_flags
= alloc_bootmem_node_nopanic(pgdat
,
4649 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4650 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4651 #endif /* CONFIG_SPARSEMEM */
4653 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4655 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4656 void __paginginit
set_pageblock_order(void)
4660 /* Check that pageblock_nr_pages has not already been setup */
4661 if (pageblock_order
)
4664 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4665 order
= HUGETLB_PAGE_ORDER
;
4667 order
= MAX_ORDER
- 1;
4670 * Assume the largest contiguous order of interest is a huge page.
4671 * This value may be variable depending on boot parameters on IA64 and
4674 pageblock_order
= order
;
4676 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4679 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4680 * is unused as pageblock_order is set at compile-time. See
4681 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4684 void __paginginit
set_pageblock_order(void)
4688 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4690 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4691 unsigned long present_pages
)
4693 unsigned long pages
= spanned_pages
;
4696 * Provide a more accurate estimation if there are holes within
4697 * the zone and SPARSEMEM is in use. If there are holes within the
4698 * zone, each populated memory region may cost us one or two extra
4699 * memmap pages due to alignment because memmap pages for each
4700 * populated regions may not naturally algined on page boundary.
4701 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4703 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4704 IS_ENABLED(CONFIG_SPARSEMEM
))
4705 pages
= present_pages
;
4707 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4711 * Set up the zone data structures:
4712 * - mark all pages reserved
4713 * - mark all memory queues empty
4714 * - clear the memory bitmaps
4716 * NOTE: pgdat should get zeroed by caller.
4718 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4719 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4720 unsigned long *zones_size
, unsigned long *zholes_size
)
4723 int nid
= pgdat
->node_id
;
4724 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4727 pgdat_resize_init(pgdat
);
4728 #ifdef CONFIG_NUMA_BALANCING
4729 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4730 pgdat
->numabalancing_migrate_nr_pages
= 0;
4731 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4733 init_waitqueue_head(&pgdat
->kswapd_wait
);
4734 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4735 pgdat_page_cgroup_init(pgdat
);
4737 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4738 struct zone
*zone
= pgdat
->node_zones
+ j
;
4739 unsigned long size
, realsize
, freesize
, memmap_pages
;
4741 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4742 node_end_pfn
, zones_size
);
4743 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4749 * Adjust freesize so that it accounts for how much memory
4750 * is used by this zone for memmap. This affects the watermark
4751 * and per-cpu initialisations
4753 memmap_pages
= calc_memmap_size(size
, realsize
);
4754 if (freesize
>= memmap_pages
) {
4755 freesize
-= memmap_pages
;
4758 " %s zone: %lu pages used for memmap\n",
4759 zone_names
[j
], memmap_pages
);
4762 " %s zone: %lu pages exceeds freesize %lu\n",
4763 zone_names
[j
], memmap_pages
, freesize
);
4765 /* Account for reserved pages */
4766 if (j
== 0 && freesize
> dma_reserve
) {
4767 freesize
-= dma_reserve
;
4768 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4769 zone_names
[0], dma_reserve
);
4772 if (!is_highmem_idx(j
))
4773 nr_kernel_pages
+= freesize
;
4774 /* Charge for highmem memmap if there are enough kernel pages */
4775 else if (nr_kernel_pages
> memmap_pages
* 2)
4776 nr_kernel_pages
-= memmap_pages
;
4777 nr_all_pages
+= freesize
;
4779 zone
->spanned_pages
= size
;
4780 zone
->present_pages
= realsize
;
4782 * Set an approximate value for lowmem here, it will be adjusted
4783 * when the bootmem allocator frees pages into the buddy system.
4784 * And all highmem pages will be managed by the buddy system.
4786 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4789 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4791 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4793 zone
->name
= zone_names
[j
];
4794 spin_lock_init(&zone
->lock
);
4795 spin_lock_init(&zone
->lru_lock
);
4796 zone_seqlock_init(zone
);
4797 zone
->zone_pgdat
= pgdat
;
4798 zone_pcp_init(zone
);
4800 /* For bootup, initialized properly in watermark setup */
4801 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4803 lruvec_init(&zone
->lruvec
);
4807 set_pageblock_order();
4808 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4809 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4810 size
, MEMMAP_EARLY
);
4812 memmap_init(size
, nid
, j
, zone_start_pfn
);
4813 zone_start_pfn
+= size
;
4817 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4819 /* Skip empty nodes */
4820 if (!pgdat
->node_spanned_pages
)
4823 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4824 /* ia64 gets its own node_mem_map, before this, without bootmem */
4825 if (!pgdat
->node_mem_map
) {
4826 unsigned long size
, start
, end
;
4830 * The zone's endpoints aren't required to be MAX_ORDER
4831 * aligned but the node_mem_map endpoints must be in order
4832 * for the buddy allocator to function correctly.
4834 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4835 end
= pgdat_end_pfn(pgdat
);
4836 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4837 size
= (end
- start
) * sizeof(struct page
);
4838 map
= alloc_remap(pgdat
->node_id
, size
);
4840 map
= alloc_bootmem_node_nopanic(pgdat
, size
);
4841 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4843 #ifndef CONFIG_NEED_MULTIPLE_NODES
4845 * With no DISCONTIG, the global mem_map is just set as node 0's
4847 if (pgdat
== NODE_DATA(0)) {
4848 mem_map
= NODE_DATA(0)->node_mem_map
;
4849 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4850 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4851 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4852 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4855 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4858 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4859 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4861 pg_data_t
*pgdat
= NODE_DATA(nid
);
4862 unsigned long start_pfn
= 0;
4863 unsigned long end_pfn
= 0;
4865 /* pg_data_t should be reset to zero when it's allocated */
4866 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4868 pgdat
->node_id
= nid
;
4869 pgdat
->node_start_pfn
= node_start_pfn
;
4870 init_zone_allows_reclaim(nid
);
4871 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4872 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4874 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
4875 zones_size
, zholes_size
);
4877 alloc_node_mem_map(pgdat
);
4878 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4879 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4880 nid
, (unsigned long)pgdat
,
4881 (unsigned long)pgdat
->node_mem_map
);
4884 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
4885 zones_size
, zholes_size
);
4888 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4890 #if MAX_NUMNODES > 1
4892 * Figure out the number of possible node ids.
4894 void __init
setup_nr_node_ids(void)
4897 unsigned int highest
= 0;
4899 for_each_node_mask(node
, node_possible_map
)
4901 nr_node_ids
= highest
+ 1;
4906 * node_map_pfn_alignment - determine the maximum internode alignment
4908 * This function should be called after node map is populated and sorted.
4909 * It calculates the maximum power of two alignment which can distinguish
4912 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4913 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4914 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4915 * shifted, 1GiB is enough and this function will indicate so.
4917 * This is used to test whether pfn -> nid mapping of the chosen memory
4918 * model has fine enough granularity to avoid incorrect mapping for the
4919 * populated node map.
4921 * Returns the determined alignment in pfn's. 0 if there is no alignment
4922 * requirement (single node).
4924 unsigned long __init
node_map_pfn_alignment(void)
4926 unsigned long accl_mask
= 0, last_end
= 0;
4927 unsigned long start
, end
, mask
;
4931 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
4932 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
4939 * Start with a mask granular enough to pin-point to the
4940 * start pfn and tick off bits one-by-one until it becomes
4941 * too coarse to separate the current node from the last.
4943 mask
= ~((1 << __ffs(start
)) - 1);
4944 while (mask
&& last_end
<= (start
& (mask
<< 1)))
4947 /* accumulate all internode masks */
4951 /* convert mask to number of pages */
4952 return ~accl_mask
+ 1;
4955 /* Find the lowest pfn for a node */
4956 static unsigned long __init
find_min_pfn_for_node(int nid
)
4958 unsigned long min_pfn
= ULONG_MAX
;
4959 unsigned long start_pfn
;
4962 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
4963 min_pfn
= min(min_pfn
, start_pfn
);
4965 if (min_pfn
== ULONG_MAX
) {
4967 "Could not find start_pfn for node %d\n", nid
);
4975 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4977 * It returns the minimum PFN based on information provided via
4978 * add_active_range().
4980 unsigned long __init
find_min_pfn_with_active_regions(void)
4982 return find_min_pfn_for_node(MAX_NUMNODES
);
4986 * early_calculate_totalpages()
4987 * Sum pages in active regions for movable zone.
4988 * Populate N_MEMORY for calculating usable_nodes.
4990 static unsigned long __init
early_calculate_totalpages(void)
4992 unsigned long totalpages
= 0;
4993 unsigned long start_pfn
, end_pfn
;
4996 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4997 unsigned long pages
= end_pfn
- start_pfn
;
4999 totalpages
+= pages
;
5001 node_set_state(nid
, N_MEMORY
);
5007 * Find the PFN the Movable zone begins in each node. Kernel memory
5008 * is spread evenly between nodes as long as the nodes have enough
5009 * memory. When they don't, some nodes will have more kernelcore than
5012 static void __init
find_zone_movable_pfns_for_nodes(void)
5015 unsigned long usable_startpfn
;
5016 unsigned long kernelcore_node
, kernelcore_remaining
;
5017 /* save the state before borrow the nodemask */
5018 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5019 unsigned long totalpages
= early_calculate_totalpages();
5020 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5023 * If movablecore was specified, calculate what size of
5024 * kernelcore that corresponds so that memory usable for
5025 * any allocation type is evenly spread. If both kernelcore
5026 * and movablecore are specified, then the value of kernelcore
5027 * will be used for required_kernelcore if it's greater than
5028 * what movablecore would have allowed.
5030 if (required_movablecore
) {
5031 unsigned long corepages
;
5034 * Round-up so that ZONE_MOVABLE is at least as large as what
5035 * was requested by the user
5037 required_movablecore
=
5038 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5039 corepages
= totalpages
- required_movablecore
;
5041 required_kernelcore
= max(required_kernelcore
, corepages
);
5044 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5045 if (!required_kernelcore
)
5048 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5049 find_usable_zone_for_movable();
5050 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5053 /* Spread kernelcore memory as evenly as possible throughout nodes */
5054 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5055 for_each_node_state(nid
, N_MEMORY
) {
5056 unsigned long start_pfn
, end_pfn
;
5059 * Recalculate kernelcore_node if the division per node
5060 * now exceeds what is necessary to satisfy the requested
5061 * amount of memory for the kernel
5063 if (required_kernelcore
< kernelcore_node
)
5064 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5067 * As the map is walked, we track how much memory is usable
5068 * by the kernel using kernelcore_remaining. When it is
5069 * 0, the rest of the node is usable by ZONE_MOVABLE
5071 kernelcore_remaining
= kernelcore_node
;
5073 /* Go through each range of PFNs within this node */
5074 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5075 unsigned long size_pages
;
5077 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5078 if (start_pfn
>= end_pfn
)
5081 /* Account for what is only usable for kernelcore */
5082 if (start_pfn
< usable_startpfn
) {
5083 unsigned long kernel_pages
;
5084 kernel_pages
= min(end_pfn
, usable_startpfn
)
5087 kernelcore_remaining
-= min(kernel_pages
,
5088 kernelcore_remaining
);
5089 required_kernelcore
-= min(kernel_pages
,
5090 required_kernelcore
);
5092 /* Continue if range is now fully accounted */
5093 if (end_pfn
<= usable_startpfn
) {
5096 * Push zone_movable_pfn to the end so
5097 * that if we have to rebalance
5098 * kernelcore across nodes, we will
5099 * not double account here
5101 zone_movable_pfn
[nid
] = end_pfn
;
5104 start_pfn
= usable_startpfn
;
5108 * The usable PFN range for ZONE_MOVABLE is from
5109 * start_pfn->end_pfn. Calculate size_pages as the
5110 * number of pages used as kernelcore
5112 size_pages
= end_pfn
- start_pfn
;
5113 if (size_pages
> kernelcore_remaining
)
5114 size_pages
= kernelcore_remaining
;
5115 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5118 * Some kernelcore has been met, update counts and
5119 * break if the kernelcore for this node has been
5122 required_kernelcore
-= min(required_kernelcore
,
5124 kernelcore_remaining
-= size_pages
;
5125 if (!kernelcore_remaining
)
5131 * If there is still required_kernelcore, we do another pass with one
5132 * less node in the count. This will push zone_movable_pfn[nid] further
5133 * along on the nodes that still have memory until kernelcore is
5137 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5140 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5141 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5142 zone_movable_pfn
[nid
] =
5143 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5146 /* restore the node_state */
5147 node_states
[N_MEMORY
] = saved_node_state
;
5150 /* Any regular or high memory on that node ? */
5151 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5153 enum zone_type zone_type
;
5155 if (N_MEMORY
== N_NORMAL_MEMORY
)
5158 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5159 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5160 if (zone
->present_pages
) {
5161 node_set_state(nid
, N_HIGH_MEMORY
);
5162 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5163 zone_type
<= ZONE_NORMAL
)
5164 node_set_state(nid
, N_NORMAL_MEMORY
);
5171 * free_area_init_nodes - Initialise all pg_data_t and zone data
5172 * @max_zone_pfn: an array of max PFNs for each zone
5174 * This will call free_area_init_node() for each active node in the system.
5175 * Using the page ranges provided by add_active_range(), the size of each
5176 * zone in each node and their holes is calculated. If the maximum PFN
5177 * between two adjacent zones match, it is assumed that the zone is empty.
5178 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5179 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5180 * starts where the previous one ended. For example, ZONE_DMA32 starts
5181 * at arch_max_dma_pfn.
5183 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5185 unsigned long start_pfn
, end_pfn
;
5188 /* Record where the zone boundaries are */
5189 memset(arch_zone_lowest_possible_pfn
, 0,
5190 sizeof(arch_zone_lowest_possible_pfn
));
5191 memset(arch_zone_highest_possible_pfn
, 0,
5192 sizeof(arch_zone_highest_possible_pfn
));
5193 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5194 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5195 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5196 if (i
== ZONE_MOVABLE
)
5198 arch_zone_lowest_possible_pfn
[i
] =
5199 arch_zone_highest_possible_pfn
[i
-1];
5200 arch_zone_highest_possible_pfn
[i
] =
5201 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5203 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5204 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5206 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5207 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5208 find_zone_movable_pfns_for_nodes();
5210 /* Print out the zone ranges */
5211 printk("Zone ranges:\n");
5212 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5213 if (i
== ZONE_MOVABLE
)
5215 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5216 if (arch_zone_lowest_possible_pfn
[i
] ==
5217 arch_zone_highest_possible_pfn
[i
])
5218 printk(KERN_CONT
"empty\n");
5220 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5221 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5222 (arch_zone_highest_possible_pfn
[i
]
5223 << PAGE_SHIFT
) - 1);
5226 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5227 printk("Movable zone start for each node\n");
5228 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5229 if (zone_movable_pfn
[i
])
5230 printk(" Node %d: %#010lx\n", i
,
5231 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5234 /* Print out the early node map */
5235 printk("Early memory node ranges\n");
5236 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5237 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5238 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5240 /* Initialise every node */
5241 mminit_verify_pageflags_layout();
5242 setup_nr_node_ids();
5243 for_each_online_node(nid
) {
5244 pg_data_t
*pgdat
= NODE_DATA(nid
);
5245 free_area_init_node(nid
, NULL
,
5246 find_min_pfn_for_node(nid
), NULL
);
5248 /* Any memory on that node */
5249 if (pgdat
->node_present_pages
)
5250 node_set_state(nid
, N_MEMORY
);
5251 check_for_memory(pgdat
, nid
);
5255 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5257 unsigned long long coremem
;
5261 coremem
= memparse(p
, &p
);
5262 *core
= coremem
>> PAGE_SHIFT
;
5264 /* Paranoid check that UL is enough for the coremem value */
5265 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5271 * kernelcore=size sets the amount of memory for use for allocations that
5272 * cannot be reclaimed or migrated.
5274 static int __init
cmdline_parse_kernelcore(char *p
)
5276 return cmdline_parse_core(p
, &required_kernelcore
);
5280 * movablecore=size sets the amount of memory for use for allocations that
5281 * can be reclaimed or migrated.
5283 static int __init
cmdline_parse_movablecore(char *p
)
5285 return cmdline_parse_core(p
, &required_movablecore
);
5288 early_param("kernelcore", cmdline_parse_kernelcore
);
5289 early_param("movablecore", cmdline_parse_movablecore
);
5291 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5293 void adjust_managed_page_count(struct page
*page
, long count
)
5295 spin_lock(&managed_page_count_lock
);
5296 page_zone(page
)->managed_pages
+= count
;
5297 totalram_pages
+= count
;
5298 #ifdef CONFIG_HIGHMEM
5299 if (PageHighMem(page
))
5300 totalhigh_pages
+= count
;
5302 spin_unlock(&managed_page_count_lock
);
5304 EXPORT_SYMBOL(adjust_managed_page_count
);
5306 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5309 unsigned long pages
= 0;
5311 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5312 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5313 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5314 if ((unsigned int)poison
<= 0xFF)
5315 memset(pos
, poison
, PAGE_SIZE
);
5316 free_reserved_page(virt_to_page(pos
));
5320 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5321 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5325 EXPORT_SYMBOL(free_reserved_area
);
5327 #ifdef CONFIG_HIGHMEM
5328 void free_highmem_page(struct page
*page
)
5330 __free_reserved_page(page
);
5332 page_zone(page
)->managed_pages
++;
5338 void __init
mem_init_print_info(const char *str
)
5340 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5341 unsigned long init_code_size
, init_data_size
;
5343 physpages
= get_num_physpages();
5344 codesize
= _etext
- _stext
;
5345 datasize
= _edata
- _sdata
;
5346 rosize
= __end_rodata
- __start_rodata
;
5347 bss_size
= __bss_stop
- __bss_start
;
5348 init_data_size
= __init_end
- __init_begin
;
5349 init_code_size
= _einittext
- _sinittext
;
5352 * Detect special cases and adjust section sizes accordingly:
5353 * 1) .init.* may be embedded into .data sections
5354 * 2) .init.text.* may be out of [__init_begin, __init_end],
5355 * please refer to arch/tile/kernel/vmlinux.lds.S.
5356 * 3) .rodata.* may be embedded into .text or .data sections.
5358 #define adj_init_size(start, end, size, pos, adj) \
5360 if (start <= pos && pos < end && size > adj) \
5364 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5365 _sinittext
, init_code_size
);
5366 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5367 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5368 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5369 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5371 #undef adj_init_size
5373 printk("Memory: %luK/%luK available "
5374 "(%luK kernel code, %luK rwdata, %luK rodata, "
5375 "%luK init, %luK bss, %luK reserved"
5376 #ifdef CONFIG_HIGHMEM
5380 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5381 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5382 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5383 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5384 #ifdef CONFIG_HIGHMEM
5385 totalhigh_pages
<< (PAGE_SHIFT
-10),
5387 str
? ", " : "", str
? str
: "");
5391 * set_dma_reserve - set the specified number of pages reserved in the first zone
5392 * @new_dma_reserve: The number of pages to mark reserved
5394 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5395 * In the DMA zone, a significant percentage may be consumed by kernel image
5396 * and other unfreeable allocations which can skew the watermarks badly. This
5397 * function may optionally be used to account for unfreeable pages in the
5398 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5399 * smaller per-cpu batchsize.
5401 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5403 dma_reserve
= new_dma_reserve
;
5406 void __init
free_area_init(unsigned long *zones_size
)
5408 free_area_init_node(0, zones_size
,
5409 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5412 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5413 unsigned long action
, void *hcpu
)
5415 int cpu
= (unsigned long)hcpu
;
5417 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5418 lru_add_drain_cpu(cpu
);
5422 * Spill the event counters of the dead processor
5423 * into the current processors event counters.
5424 * This artificially elevates the count of the current
5427 vm_events_fold_cpu(cpu
);
5430 * Zero the differential counters of the dead processor
5431 * so that the vm statistics are consistent.
5433 * This is only okay since the processor is dead and cannot
5434 * race with what we are doing.
5436 cpu_vm_stats_fold(cpu
);
5441 void __init
page_alloc_init(void)
5443 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5447 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5448 * or min_free_kbytes changes.
5450 static void calculate_totalreserve_pages(void)
5452 struct pglist_data
*pgdat
;
5453 unsigned long reserve_pages
= 0;
5454 enum zone_type i
, j
;
5456 for_each_online_pgdat(pgdat
) {
5457 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5458 struct zone
*zone
= pgdat
->node_zones
+ i
;
5459 unsigned long max
= 0;
5461 /* Find valid and maximum lowmem_reserve in the zone */
5462 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5463 if (zone
->lowmem_reserve
[j
] > max
)
5464 max
= zone
->lowmem_reserve
[j
];
5467 /* we treat the high watermark as reserved pages. */
5468 max
+= high_wmark_pages(zone
);
5470 if (max
> zone
->managed_pages
)
5471 max
= zone
->managed_pages
;
5472 reserve_pages
+= max
;
5474 * Lowmem reserves are not available to
5475 * GFP_HIGHUSER page cache allocations and
5476 * kswapd tries to balance zones to their high
5477 * watermark. As a result, neither should be
5478 * regarded as dirtyable memory, to prevent a
5479 * situation where reclaim has to clean pages
5480 * in order to balance the zones.
5482 zone
->dirty_balance_reserve
= max
;
5485 dirty_balance_reserve
= reserve_pages
;
5486 totalreserve_pages
= reserve_pages
;
5490 * setup_per_zone_lowmem_reserve - called whenever
5491 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5492 * has a correct pages reserved value, so an adequate number of
5493 * pages are left in the zone after a successful __alloc_pages().
5495 static void setup_per_zone_lowmem_reserve(void)
5497 struct pglist_data
*pgdat
;
5498 enum zone_type j
, idx
;
5500 for_each_online_pgdat(pgdat
) {
5501 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5502 struct zone
*zone
= pgdat
->node_zones
+ j
;
5503 unsigned long managed_pages
= zone
->managed_pages
;
5505 zone
->lowmem_reserve
[j
] = 0;
5509 struct zone
*lower_zone
;
5513 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5514 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5516 lower_zone
= pgdat
->node_zones
+ idx
;
5517 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5518 sysctl_lowmem_reserve_ratio
[idx
];
5519 managed_pages
+= lower_zone
->managed_pages
;
5524 /* update totalreserve_pages */
5525 calculate_totalreserve_pages();
5528 static void __setup_per_zone_wmarks(void)
5530 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5531 unsigned long lowmem_pages
= 0;
5533 unsigned long flags
;
5535 /* Calculate total number of !ZONE_HIGHMEM pages */
5536 for_each_zone(zone
) {
5537 if (!is_highmem(zone
))
5538 lowmem_pages
+= zone
->managed_pages
;
5541 for_each_zone(zone
) {
5544 spin_lock_irqsave(&zone
->lock
, flags
);
5545 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5546 do_div(tmp
, lowmem_pages
);
5547 if (is_highmem(zone
)) {
5549 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5550 * need highmem pages, so cap pages_min to a small
5553 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5554 * deltas controls asynch page reclaim, and so should
5555 * not be capped for highmem.
5557 unsigned long min_pages
;
5559 min_pages
= zone
->managed_pages
/ 1024;
5560 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5561 zone
->watermark
[WMARK_MIN
] = min_pages
;
5564 * If it's a lowmem zone, reserve a number of pages
5565 * proportionate to the zone's size.
5567 zone
->watermark
[WMARK_MIN
] = tmp
;
5570 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5571 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5573 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5574 high_wmark_pages(zone
) -
5575 low_wmark_pages(zone
) -
5576 zone_page_state(zone
, NR_ALLOC_BATCH
));
5578 setup_zone_migrate_reserve(zone
);
5579 spin_unlock_irqrestore(&zone
->lock
, flags
);
5582 /* update totalreserve_pages */
5583 calculate_totalreserve_pages();
5587 * setup_per_zone_wmarks - called when min_free_kbytes changes
5588 * or when memory is hot-{added|removed}
5590 * Ensures that the watermark[min,low,high] values for each zone are set
5591 * correctly with respect to min_free_kbytes.
5593 void setup_per_zone_wmarks(void)
5595 mutex_lock(&zonelists_mutex
);
5596 __setup_per_zone_wmarks();
5597 mutex_unlock(&zonelists_mutex
);
5601 * The inactive anon list should be small enough that the VM never has to
5602 * do too much work, but large enough that each inactive page has a chance
5603 * to be referenced again before it is swapped out.
5605 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5606 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5607 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5608 * the anonymous pages are kept on the inactive list.
5611 * memory ratio inactive anon
5612 * -------------------------------------
5621 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5623 unsigned int gb
, ratio
;
5625 /* Zone size in gigabytes */
5626 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5628 ratio
= int_sqrt(10 * gb
);
5632 zone
->inactive_ratio
= ratio
;
5635 static void __meminit
setup_per_zone_inactive_ratio(void)
5640 calculate_zone_inactive_ratio(zone
);
5644 * Initialise min_free_kbytes.
5646 * For small machines we want it small (128k min). For large machines
5647 * we want it large (64MB max). But it is not linear, because network
5648 * bandwidth does not increase linearly with machine size. We use
5650 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5651 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5667 int __meminit
init_per_zone_wmark_min(void)
5669 unsigned long lowmem_kbytes
;
5670 int new_min_free_kbytes
;
5672 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5673 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5675 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5676 min_free_kbytes
= new_min_free_kbytes
;
5677 if (min_free_kbytes
< 128)
5678 min_free_kbytes
= 128;
5679 if (min_free_kbytes
> 65536)
5680 min_free_kbytes
= 65536;
5682 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5683 new_min_free_kbytes
, user_min_free_kbytes
);
5685 setup_per_zone_wmarks();
5686 refresh_zone_stat_thresholds();
5687 setup_per_zone_lowmem_reserve();
5688 setup_per_zone_inactive_ratio();
5691 module_init(init_per_zone_wmark_min
)
5694 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5695 * that we can call two helper functions whenever min_free_kbytes
5698 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5699 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5701 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5703 user_min_free_kbytes
= min_free_kbytes
;
5704 setup_per_zone_wmarks();
5710 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5711 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5716 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5721 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5722 sysctl_min_unmapped_ratio
) / 100;
5726 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5727 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5732 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5737 zone
->min_slab_pages
= (zone
->managed_pages
*
5738 sysctl_min_slab_ratio
) / 100;
5744 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5745 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5746 * whenever sysctl_lowmem_reserve_ratio changes.
5748 * The reserve ratio obviously has absolutely no relation with the
5749 * minimum watermarks. The lowmem reserve ratio can only make sense
5750 * if in function of the boot time zone sizes.
5752 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5753 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5755 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5756 setup_per_zone_lowmem_reserve();
5761 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5762 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5763 * pagelist can have before it gets flushed back to buddy allocator.
5765 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5766 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5772 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5773 if (!write
|| (ret
< 0))
5776 mutex_lock(&pcp_batch_high_lock
);
5777 for_each_populated_zone(zone
) {
5779 high
= zone
->managed_pages
/ percpu_pagelist_fraction
;
5780 for_each_possible_cpu(cpu
)
5781 pageset_set_high(per_cpu_ptr(zone
->pageset
, cpu
),
5784 mutex_unlock(&pcp_batch_high_lock
);
5788 int hashdist
= HASHDIST_DEFAULT
;
5791 static int __init
set_hashdist(char *str
)
5795 hashdist
= simple_strtoul(str
, &str
, 0);
5798 __setup("hashdist=", set_hashdist
);
5802 * allocate a large system hash table from bootmem
5803 * - it is assumed that the hash table must contain an exact power-of-2
5804 * quantity of entries
5805 * - limit is the number of hash buckets, not the total allocation size
5807 void *__init
alloc_large_system_hash(const char *tablename
,
5808 unsigned long bucketsize
,
5809 unsigned long numentries
,
5812 unsigned int *_hash_shift
,
5813 unsigned int *_hash_mask
,
5814 unsigned long low_limit
,
5815 unsigned long high_limit
)
5817 unsigned long long max
= high_limit
;
5818 unsigned long log2qty
, size
;
5821 /* allow the kernel cmdline to have a say */
5823 /* round applicable memory size up to nearest megabyte */
5824 numentries
= nr_kernel_pages
;
5826 /* It isn't necessary when PAGE_SIZE >= 1MB */
5827 if (PAGE_SHIFT
< 20)
5828 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
5830 /* limit to 1 bucket per 2^scale bytes of low memory */
5831 if (scale
> PAGE_SHIFT
)
5832 numentries
>>= (scale
- PAGE_SHIFT
);
5834 numentries
<<= (PAGE_SHIFT
- scale
);
5836 /* Make sure we've got at least a 0-order allocation.. */
5837 if (unlikely(flags
& HASH_SMALL
)) {
5838 /* Makes no sense without HASH_EARLY */
5839 WARN_ON(!(flags
& HASH_EARLY
));
5840 if (!(numentries
>> *_hash_shift
)) {
5841 numentries
= 1UL << *_hash_shift
;
5842 BUG_ON(!numentries
);
5844 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5845 numentries
= PAGE_SIZE
/ bucketsize
;
5847 numentries
= roundup_pow_of_two(numentries
);
5849 /* limit allocation size to 1/16 total memory by default */
5851 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5852 do_div(max
, bucketsize
);
5854 max
= min(max
, 0x80000000ULL
);
5856 if (numentries
< low_limit
)
5857 numentries
= low_limit
;
5858 if (numentries
> max
)
5861 log2qty
= ilog2(numentries
);
5864 size
= bucketsize
<< log2qty
;
5865 if (flags
& HASH_EARLY
)
5866 table
= alloc_bootmem_nopanic(size
);
5868 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5871 * If bucketsize is not a power-of-two, we may free
5872 * some pages at the end of hash table which
5873 * alloc_pages_exact() automatically does
5875 if (get_order(size
) < MAX_ORDER
) {
5876 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5877 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5880 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5883 panic("Failed to allocate %s hash table\n", tablename
);
5885 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5888 ilog2(size
) - PAGE_SHIFT
,
5892 *_hash_shift
= log2qty
;
5894 *_hash_mask
= (1 << log2qty
) - 1;
5899 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5900 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5903 #ifdef CONFIG_SPARSEMEM
5904 return __pfn_to_section(pfn
)->pageblock_flags
;
5906 return zone
->pageblock_flags
;
5907 #endif /* CONFIG_SPARSEMEM */
5910 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5912 #ifdef CONFIG_SPARSEMEM
5913 pfn
&= (PAGES_PER_SECTION
-1);
5914 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5916 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
5917 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5918 #endif /* CONFIG_SPARSEMEM */
5922 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5923 * @page: The page within the block of interest
5924 * @start_bitidx: The first bit of interest to retrieve
5925 * @end_bitidx: The last bit of interest
5926 * returns pageblock_bits flags
5928 unsigned long get_pageblock_flags_group(struct page
*page
,
5929 int start_bitidx
, int end_bitidx
)
5932 unsigned long *bitmap
;
5933 unsigned long pfn
, bitidx
;
5934 unsigned long flags
= 0;
5935 unsigned long value
= 1;
5937 zone
= page_zone(page
);
5938 pfn
= page_to_pfn(page
);
5939 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5940 bitidx
= pfn_to_bitidx(zone
, pfn
);
5942 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5943 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5950 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5951 * @page: The page within the block of interest
5952 * @start_bitidx: The first bit of interest
5953 * @end_bitidx: The last bit of interest
5954 * @flags: The flags to set
5956 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5957 int start_bitidx
, int end_bitidx
)
5960 unsigned long *bitmap
;
5961 unsigned long pfn
, bitidx
;
5962 unsigned long value
= 1;
5964 zone
= page_zone(page
);
5965 pfn
= page_to_pfn(page
);
5966 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5967 bitidx
= pfn_to_bitidx(zone
, pfn
);
5968 VM_BUG_ON(!zone_spans_pfn(zone
, pfn
));
5970 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5972 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5974 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5978 * This function checks whether pageblock includes unmovable pages or not.
5979 * If @count is not zero, it is okay to include less @count unmovable pages
5981 * PageLRU check without isolation or lru_lock could race so that
5982 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5983 * expect this function should be exact.
5985 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
5986 bool skip_hwpoisoned_pages
)
5988 unsigned long pfn
, iter
, found
;
5992 * For avoiding noise data, lru_add_drain_all() should be called
5993 * If ZONE_MOVABLE, the zone never contains unmovable pages
5995 if (zone_idx(zone
) == ZONE_MOVABLE
)
5997 mt
= get_pageblock_migratetype(page
);
5998 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6001 pfn
= page_to_pfn(page
);
6002 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6003 unsigned long check
= pfn
+ iter
;
6005 if (!pfn_valid_within(check
))
6008 page
= pfn_to_page(check
);
6011 * Hugepages are not in LRU lists, but they're movable.
6012 * We need not scan over tail pages bacause we don't
6013 * handle each tail page individually in migration.
6015 if (PageHuge(page
)) {
6016 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6021 * We can't use page_count without pin a page
6022 * because another CPU can free compound page.
6023 * This check already skips compound tails of THP
6024 * because their page->_count is zero at all time.
6026 if (!atomic_read(&page
->_count
)) {
6027 if (PageBuddy(page
))
6028 iter
+= (1 << page_order(page
)) - 1;
6033 * The HWPoisoned page may be not in buddy system, and
6034 * page_count() is not 0.
6036 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6042 * If there are RECLAIMABLE pages, we need to check it.
6043 * But now, memory offline itself doesn't call shrink_slab()
6044 * and it still to be fixed.
6047 * If the page is not RAM, page_count()should be 0.
6048 * we don't need more check. This is an _used_ not-movable page.
6050 * The problematic thing here is PG_reserved pages. PG_reserved
6051 * is set to both of a memory hole page and a _used_ kernel
6060 bool is_pageblock_removable_nolock(struct page
*page
)
6066 * We have to be careful here because we are iterating over memory
6067 * sections which are not zone aware so we might end up outside of
6068 * the zone but still within the section.
6069 * We have to take care about the node as well. If the node is offline
6070 * its NODE_DATA will be NULL - see page_zone.
6072 if (!node_online(page_to_nid(page
)))
6075 zone
= page_zone(page
);
6076 pfn
= page_to_pfn(page
);
6077 if (!zone_spans_pfn(zone
, pfn
))
6080 return !has_unmovable_pages(zone
, page
, 0, true);
6085 static unsigned long pfn_max_align_down(unsigned long pfn
)
6087 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6088 pageblock_nr_pages
) - 1);
6091 static unsigned long pfn_max_align_up(unsigned long pfn
)
6093 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6094 pageblock_nr_pages
));
6097 /* [start, end) must belong to a single zone. */
6098 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6099 unsigned long start
, unsigned long end
)
6101 /* This function is based on compact_zone() from compaction.c. */
6102 unsigned long nr_reclaimed
;
6103 unsigned long pfn
= start
;
6104 unsigned int tries
= 0;
6109 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6110 if (fatal_signal_pending(current
)) {
6115 if (list_empty(&cc
->migratepages
)) {
6116 cc
->nr_migratepages
= 0;
6117 pfn
= isolate_migratepages_range(cc
->zone
, cc
,
6124 } else if (++tries
== 5) {
6125 ret
= ret
< 0 ? ret
: -EBUSY
;
6129 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6131 cc
->nr_migratepages
-= nr_reclaimed
;
6133 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6134 0, MIGRATE_SYNC
, MR_CMA
);
6137 putback_movable_pages(&cc
->migratepages
);
6144 * alloc_contig_range() -- tries to allocate given range of pages
6145 * @start: start PFN to allocate
6146 * @end: one-past-the-last PFN to allocate
6147 * @migratetype: migratetype of the underlaying pageblocks (either
6148 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6149 * in range must have the same migratetype and it must
6150 * be either of the two.
6152 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6153 * aligned, however it's the caller's responsibility to guarantee that
6154 * we are the only thread that changes migrate type of pageblocks the
6157 * The PFN range must belong to a single zone.
6159 * Returns zero on success or negative error code. On success all
6160 * pages which PFN is in [start, end) are allocated for the caller and
6161 * need to be freed with free_contig_range().
6163 int alloc_contig_range(unsigned long start
, unsigned long end
,
6164 unsigned migratetype
)
6166 unsigned long outer_start
, outer_end
;
6169 struct compact_control cc
= {
6170 .nr_migratepages
= 0,
6172 .zone
= page_zone(pfn_to_page(start
)),
6174 .ignore_skip_hint
= true,
6176 INIT_LIST_HEAD(&cc
.migratepages
);
6179 * What we do here is we mark all pageblocks in range as
6180 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6181 * have different sizes, and due to the way page allocator
6182 * work, we align the range to biggest of the two pages so
6183 * that page allocator won't try to merge buddies from
6184 * different pageblocks and change MIGRATE_ISOLATE to some
6185 * other migration type.
6187 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6188 * migrate the pages from an unaligned range (ie. pages that
6189 * we are interested in). This will put all the pages in
6190 * range back to page allocator as MIGRATE_ISOLATE.
6192 * When this is done, we take the pages in range from page
6193 * allocator removing them from the buddy system. This way
6194 * page allocator will never consider using them.
6196 * This lets us mark the pageblocks back as
6197 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6198 * aligned range but not in the unaligned, original range are
6199 * put back to page allocator so that buddy can use them.
6202 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6203 pfn_max_align_up(end
), migratetype
,
6208 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6213 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6214 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6215 * more, all pages in [start, end) are free in page allocator.
6216 * What we are going to do is to allocate all pages from
6217 * [start, end) (that is remove them from page allocator).
6219 * The only problem is that pages at the beginning and at the
6220 * end of interesting range may be not aligned with pages that
6221 * page allocator holds, ie. they can be part of higher order
6222 * pages. Because of this, we reserve the bigger range and
6223 * once this is done free the pages we are not interested in.
6225 * We don't have to hold zone->lock here because the pages are
6226 * isolated thus they won't get removed from buddy.
6229 lru_add_drain_all();
6233 outer_start
= start
;
6234 while (!PageBuddy(pfn_to_page(outer_start
))) {
6235 if (++order
>= MAX_ORDER
) {
6239 outer_start
&= ~0UL << order
;
6242 /* Make sure the range is really isolated. */
6243 if (test_pages_isolated(outer_start
, end
, false)) {
6244 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6251 /* Grab isolated pages from freelists. */
6252 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6258 /* Free head and tail (if any) */
6259 if (start
!= outer_start
)
6260 free_contig_range(outer_start
, start
- outer_start
);
6261 if (end
!= outer_end
)
6262 free_contig_range(end
, outer_end
- end
);
6265 undo_isolate_page_range(pfn_max_align_down(start
),
6266 pfn_max_align_up(end
), migratetype
);
6270 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6272 unsigned int count
= 0;
6274 for (; nr_pages
--; pfn
++) {
6275 struct page
*page
= pfn_to_page(pfn
);
6277 count
+= page_count(page
) != 1;
6280 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6284 #ifdef CONFIG_MEMORY_HOTPLUG
6286 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6287 * page high values need to be recalulated.
6289 void __meminit
zone_pcp_update(struct zone
*zone
)
6292 mutex_lock(&pcp_batch_high_lock
);
6293 for_each_possible_cpu(cpu
)
6294 pageset_set_high_and_batch(zone
,
6295 per_cpu_ptr(zone
->pageset
, cpu
));
6296 mutex_unlock(&pcp_batch_high_lock
);
6300 void zone_pcp_reset(struct zone
*zone
)
6302 unsigned long flags
;
6304 struct per_cpu_pageset
*pset
;
6306 /* avoid races with drain_pages() */
6307 local_irq_save(flags
);
6308 if (zone
->pageset
!= &boot_pageset
) {
6309 for_each_online_cpu(cpu
) {
6310 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6311 drain_zonestat(zone
, pset
);
6313 free_percpu(zone
->pageset
);
6314 zone
->pageset
= &boot_pageset
;
6316 local_irq_restore(flags
);
6319 #ifdef CONFIG_MEMORY_HOTREMOVE
6321 * All pages in the range must be isolated before calling this.
6324 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6330 unsigned long flags
;
6331 /* find the first valid pfn */
6332 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6337 zone
= page_zone(pfn_to_page(pfn
));
6338 spin_lock_irqsave(&zone
->lock
, flags
);
6340 while (pfn
< end_pfn
) {
6341 if (!pfn_valid(pfn
)) {
6345 page
= pfn_to_page(pfn
);
6347 * The HWPoisoned page may be not in buddy system, and
6348 * page_count() is not 0.
6350 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6352 SetPageReserved(page
);
6356 BUG_ON(page_count(page
));
6357 BUG_ON(!PageBuddy(page
));
6358 order
= page_order(page
);
6359 #ifdef CONFIG_DEBUG_VM
6360 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6361 pfn
, 1 << order
, end_pfn
);
6363 list_del(&page
->lru
);
6364 rmv_page_order(page
);
6365 zone
->free_area
[order
].nr_free
--;
6366 #ifdef CONFIG_HIGHMEM
6367 if (PageHighMem(page
))
6368 totalhigh_pages
-= 1 << order
;
6370 for (i
= 0; i
< (1 << order
); i
++)
6371 SetPageReserved((page
+i
));
6372 pfn
+= (1 << order
);
6374 spin_unlock_irqrestore(&zone
->lock
, flags
);
6378 #ifdef CONFIG_MEMORY_FAILURE
6379 bool is_free_buddy_page(struct page
*page
)
6381 struct zone
*zone
= page_zone(page
);
6382 unsigned long pfn
= page_to_pfn(page
);
6383 unsigned long flags
;
6386 spin_lock_irqsave(&zone
->lock
, flags
);
6387 for (order
= 0; order
< MAX_ORDER
; order
++) {
6388 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6390 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6393 spin_unlock_irqrestore(&zone
->lock
, flags
);
6395 return order
< MAX_ORDER
;
6399 static const struct trace_print_flags pageflag_names
[] = {
6400 {1UL << PG_locked
, "locked" },
6401 {1UL << PG_error
, "error" },
6402 {1UL << PG_referenced
, "referenced" },
6403 {1UL << PG_uptodate
, "uptodate" },
6404 {1UL << PG_dirty
, "dirty" },
6405 {1UL << PG_lru
, "lru" },
6406 {1UL << PG_active
, "active" },
6407 {1UL << PG_slab
, "slab" },
6408 {1UL << PG_owner_priv_1
, "owner_priv_1" },
6409 {1UL << PG_arch_1
, "arch_1" },
6410 {1UL << PG_reserved
, "reserved" },
6411 {1UL << PG_private
, "private" },
6412 {1UL << PG_private_2
, "private_2" },
6413 {1UL << PG_writeback
, "writeback" },
6414 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6415 {1UL << PG_head
, "head" },
6416 {1UL << PG_tail
, "tail" },
6418 {1UL << PG_compound
, "compound" },
6420 {1UL << PG_swapcache
, "swapcache" },
6421 {1UL << PG_mappedtodisk
, "mappedtodisk" },
6422 {1UL << PG_reclaim
, "reclaim" },
6423 {1UL << PG_swapbacked
, "swapbacked" },
6424 {1UL << PG_unevictable
, "unevictable" },
6426 {1UL << PG_mlocked
, "mlocked" },
6428 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6429 {1UL << PG_uncached
, "uncached" },
6431 #ifdef CONFIG_MEMORY_FAILURE
6432 {1UL << PG_hwpoison
, "hwpoison" },
6434 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6435 {1UL << PG_compound_lock
, "compound_lock" },
6439 static void dump_page_flags(unsigned long flags
)
6441 const char *delim
= "";
6445 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names
) != __NR_PAGEFLAGS
);
6447 printk(KERN_ALERT
"page flags: %#lx(", flags
);
6449 /* remove zone id */
6450 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
6452 for (i
= 0; i
< ARRAY_SIZE(pageflag_names
) && flags
; i
++) {
6454 mask
= pageflag_names
[i
].mask
;
6455 if ((flags
& mask
) != mask
)
6459 printk("%s%s", delim
, pageflag_names
[i
].name
);
6463 /* check for left over flags */
6465 printk("%s%#lx", delim
, flags
);
6470 void dump_page(struct page
*page
)
6473 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6474 page
, atomic_read(&page
->_count
), page_mapcount(page
),
6475 page
->mapping
, page
->index
);
6476 dump_page_flags(page
->flags
);
6477 mem_cgroup_print_bad_page(page
);