2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
61 #include <linux/sched/rt.h>
63 #include <asm/tlbflush.h>
64 #include <asm/div64.h>
67 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
68 DEFINE_PER_CPU(int, numa_node
);
69 EXPORT_PER_CPU_SYMBOL(numa_node
);
72 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
74 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
75 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
76 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
77 * defined in <linux/topology.h>.
79 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
80 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
84 * Array of node states.
86 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
87 [N_POSSIBLE
] = NODE_MASK_ALL
,
88 [N_ONLINE
] = { { [0] = 1UL } },
90 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
92 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
94 #ifdef CONFIG_MOVABLE_NODE
95 [N_MEMORY
] = { { [0] = 1UL } },
97 [N_CPU
] = { { [0] = 1UL } },
100 EXPORT_SYMBOL(node_states
);
102 unsigned long totalram_pages __read_mostly
;
103 unsigned long totalreserve_pages __read_mostly
;
105 * When calculating the number of globally allowed dirty pages, there
106 * is a certain number of per-zone reserves that should not be
107 * considered dirtyable memory. This is the sum of those reserves
108 * over all existing zones that contribute dirtyable memory.
110 unsigned long dirty_balance_reserve __read_mostly
;
112 int percpu_pagelist_fraction
;
113 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
115 #ifdef CONFIG_PM_SLEEP
117 * The following functions are used by the suspend/hibernate code to temporarily
118 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
119 * while devices are suspended. To avoid races with the suspend/hibernate code,
120 * they should always be called with pm_mutex held (gfp_allowed_mask also should
121 * only be modified with pm_mutex held, unless the suspend/hibernate code is
122 * guaranteed not to run in parallel with that modification).
125 static gfp_t saved_gfp_mask
;
127 void pm_restore_gfp_mask(void)
129 WARN_ON(!mutex_is_locked(&pm_mutex
));
130 if (saved_gfp_mask
) {
131 gfp_allowed_mask
= saved_gfp_mask
;
136 void pm_restrict_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex
));
139 WARN_ON(saved_gfp_mask
);
140 saved_gfp_mask
= gfp_allowed_mask
;
141 gfp_allowed_mask
&= ~GFP_IOFS
;
144 bool pm_suspended_storage(void)
146 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
150 #endif /* CONFIG_PM_SLEEP */
152 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
153 int pageblock_order __read_mostly
;
156 static void __free_pages_ok(struct page
*page
, unsigned int order
);
159 * results with 256, 32 in the lowmem_reserve sysctl:
160 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
161 * 1G machine -> (16M dma, 784M normal, 224M high)
162 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
163 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
164 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
166 * TBD: should special case ZONE_DMA32 machines here - in those we normally
167 * don't need any ZONE_NORMAL reservation
169 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
170 #ifdef CONFIG_ZONE_DMA
173 #ifdef CONFIG_ZONE_DMA32
176 #ifdef CONFIG_HIGHMEM
182 EXPORT_SYMBOL(totalram_pages
);
184 static char * const zone_names
[MAX_NR_ZONES
] = {
185 #ifdef CONFIG_ZONE_DMA
188 #ifdef CONFIG_ZONE_DMA32
192 #ifdef CONFIG_HIGHMEM
198 int min_free_kbytes
= 1024;
200 static unsigned long __meminitdata nr_kernel_pages
;
201 static unsigned long __meminitdata nr_all_pages
;
202 static unsigned long __meminitdata dma_reserve
;
204 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
205 /* Movable memory ranges, will also be used by memblock subsystem. */
206 struct movablemem_map movablemem_map
= {
211 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
212 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
213 static unsigned long __initdata required_kernelcore
;
214 static unsigned long __initdata required_movablecore
;
215 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
216 static unsigned long __meminitdata zone_movable_limit
[MAX_NUMNODES
];
218 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
220 EXPORT_SYMBOL(movable_zone
);
221 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
224 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
225 int nr_online_nodes __read_mostly
= 1;
226 EXPORT_SYMBOL(nr_node_ids
);
227 EXPORT_SYMBOL(nr_online_nodes
);
230 int page_group_by_mobility_disabled __read_mostly
;
232 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
235 if (unlikely(page_group_by_mobility_disabled
))
236 migratetype
= MIGRATE_UNMOVABLE
;
238 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
239 PB_migrate
, PB_migrate_end
);
242 bool oom_killer_disabled __read_mostly
;
244 #ifdef CONFIG_DEBUG_VM
245 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
249 unsigned long pfn
= page_to_pfn(page
);
252 seq
= zone_span_seqbegin(zone
);
253 if (pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
255 else if (pfn
< zone
->zone_start_pfn
)
257 } while (zone_span_seqretry(zone
, seq
));
262 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
264 if (!pfn_valid_within(page_to_pfn(page
)))
266 if (zone
!= page_zone(page
))
272 * Temporary debugging check for pages not lying within a given zone.
274 static int bad_range(struct zone
*zone
, struct page
*page
)
276 if (page_outside_zone_boundaries(zone
, page
))
278 if (!page_is_consistent(zone
, page
))
284 static inline int bad_range(struct zone
*zone
, struct page
*page
)
290 static void bad_page(struct page
*page
)
292 static unsigned long resume
;
293 static unsigned long nr_shown
;
294 static unsigned long nr_unshown
;
296 /* Don't complain about poisoned pages */
297 if (PageHWPoison(page
)) {
298 reset_page_mapcount(page
); /* remove PageBuddy */
303 * Allow a burst of 60 reports, then keep quiet for that minute;
304 * or allow a steady drip of one report per second.
306 if (nr_shown
== 60) {
307 if (time_before(jiffies
, resume
)) {
313 "BUG: Bad page state: %lu messages suppressed\n",
320 resume
= jiffies
+ 60 * HZ
;
322 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
323 current
->comm
, page_to_pfn(page
));
329 /* Leave bad fields for debug, except PageBuddy could make trouble */
330 reset_page_mapcount(page
); /* remove PageBuddy */
331 add_taint(TAINT_BAD_PAGE
);
335 * Higher-order pages are called "compound pages". They are structured thusly:
337 * The first PAGE_SIZE page is called the "head page".
339 * The remaining PAGE_SIZE pages are called "tail pages".
341 * All pages have PG_compound set. All tail pages have their ->first_page
342 * pointing at the head page.
344 * The first tail page's ->lru.next holds the address of the compound page's
345 * put_page() function. Its ->lru.prev holds the order of allocation.
346 * This usage means that zero-order pages may not be compound.
349 static void free_compound_page(struct page
*page
)
351 __free_pages_ok(page
, compound_order(page
));
354 void prep_compound_page(struct page
*page
, unsigned long order
)
357 int nr_pages
= 1 << order
;
359 set_compound_page_dtor(page
, free_compound_page
);
360 set_compound_order(page
, order
);
362 for (i
= 1; i
< nr_pages
; i
++) {
363 struct page
*p
= page
+ i
;
365 set_page_count(p
, 0);
366 p
->first_page
= page
;
370 /* update __split_huge_page_refcount if you change this function */
371 static int destroy_compound_page(struct page
*page
, unsigned long order
)
374 int nr_pages
= 1 << order
;
377 if (unlikely(compound_order(page
) != order
)) {
382 __ClearPageHead(page
);
384 for (i
= 1; i
< nr_pages
; i
++) {
385 struct page
*p
= page
+ i
;
387 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
397 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
402 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
403 * and __GFP_HIGHMEM from hard or soft interrupt context.
405 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
406 for (i
= 0; i
< (1 << order
); i
++)
407 clear_highpage(page
+ i
);
410 #ifdef CONFIG_DEBUG_PAGEALLOC
411 unsigned int _debug_guardpage_minorder
;
413 static int __init
debug_guardpage_minorder_setup(char *buf
)
417 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
418 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
421 _debug_guardpage_minorder
= res
;
422 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
425 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
427 static inline void set_page_guard_flag(struct page
*page
)
429 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
432 static inline void clear_page_guard_flag(struct page
*page
)
434 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
437 static inline void set_page_guard_flag(struct page
*page
) { }
438 static inline void clear_page_guard_flag(struct page
*page
) { }
441 static inline void set_page_order(struct page
*page
, int order
)
443 set_page_private(page
, order
);
444 __SetPageBuddy(page
);
447 static inline void rmv_page_order(struct page
*page
)
449 __ClearPageBuddy(page
);
450 set_page_private(page
, 0);
454 * Locate the struct page for both the matching buddy in our
455 * pair (buddy1) and the combined O(n+1) page they form (page).
457 * 1) Any buddy B1 will have an order O twin B2 which satisfies
458 * the following equation:
460 * For example, if the starting buddy (buddy2) is #8 its order
462 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
464 * 2) Any buddy B will have an order O+1 parent P which
465 * satisfies the following equation:
468 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
470 static inline unsigned long
471 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
473 return page_idx
^ (1 << order
);
477 * This function checks whether a page is free && is the buddy
478 * we can do coalesce a page and its buddy if
479 * (a) the buddy is not in a hole &&
480 * (b) the buddy is in the buddy system &&
481 * (c) a page and its buddy have the same order &&
482 * (d) a page and its buddy are in the same zone.
484 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
485 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
487 * For recording page's order, we use page_private(page).
489 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
492 if (!pfn_valid_within(page_to_pfn(buddy
)))
495 if (page_zone_id(page
) != page_zone_id(buddy
))
498 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
499 VM_BUG_ON(page_count(buddy
) != 0);
503 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
504 VM_BUG_ON(page_count(buddy
) != 0);
511 * Freeing function for a buddy system allocator.
513 * The concept of a buddy system is to maintain direct-mapped table
514 * (containing bit values) for memory blocks of various "orders".
515 * The bottom level table contains the map for the smallest allocatable
516 * units of memory (here, pages), and each level above it describes
517 * pairs of units from the levels below, hence, "buddies".
518 * At a high level, all that happens here is marking the table entry
519 * at the bottom level available, and propagating the changes upward
520 * as necessary, plus some accounting needed to play nicely with other
521 * parts of the VM system.
522 * At each level, we keep a list of pages, which are heads of continuous
523 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
524 * order is recorded in page_private(page) field.
525 * So when we are allocating or freeing one, we can derive the state of the
526 * other. That is, if we allocate a small block, and both were
527 * free, the remainder of the region must be split into blocks.
528 * If a block is freed, and its buddy is also free, then this
529 * triggers coalescing into a block of larger size.
534 static inline void __free_one_page(struct page
*page
,
535 struct zone
*zone
, unsigned int order
,
538 unsigned long page_idx
;
539 unsigned long combined_idx
;
540 unsigned long uninitialized_var(buddy_idx
);
543 if (unlikely(PageCompound(page
)))
544 if (unlikely(destroy_compound_page(page
, order
)))
547 VM_BUG_ON(migratetype
== -1);
549 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
551 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
552 VM_BUG_ON(bad_range(zone
, page
));
554 while (order
< MAX_ORDER
-1) {
555 buddy_idx
= __find_buddy_index(page_idx
, order
);
556 buddy
= page
+ (buddy_idx
- page_idx
);
557 if (!page_is_buddy(page
, buddy
, order
))
560 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
561 * merge with it and move up one order.
563 if (page_is_guard(buddy
)) {
564 clear_page_guard_flag(buddy
);
565 set_page_private(page
, 0);
566 __mod_zone_freepage_state(zone
, 1 << order
,
569 list_del(&buddy
->lru
);
570 zone
->free_area
[order
].nr_free
--;
571 rmv_page_order(buddy
);
573 combined_idx
= buddy_idx
& page_idx
;
574 page
= page
+ (combined_idx
- page_idx
);
575 page_idx
= combined_idx
;
578 set_page_order(page
, order
);
581 * If this is not the largest possible page, check if the buddy
582 * of the next-highest order is free. If it is, it's possible
583 * that pages are being freed that will coalesce soon. In case,
584 * that is happening, add the free page to the tail of the list
585 * so it's less likely to be used soon and more likely to be merged
586 * as a higher order page
588 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
589 struct page
*higher_page
, *higher_buddy
;
590 combined_idx
= buddy_idx
& page_idx
;
591 higher_page
= page
+ (combined_idx
- page_idx
);
592 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
593 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
594 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
595 list_add_tail(&page
->lru
,
596 &zone
->free_area
[order
].free_list
[migratetype
]);
601 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
603 zone
->free_area
[order
].nr_free
++;
606 static inline int free_pages_check(struct page
*page
)
608 if (unlikely(page_mapcount(page
) |
609 (page
->mapping
!= NULL
) |
610 (atomic_read(&page
->_count
) != 0) |
611 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
) |
612 (mem_cgroup_bad_page_check(page
)))) {
616 reset_page_last_nid(page
);
617 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
618 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
623 * Frees a number of pages from the PCP lists
624 * Assumes all pages on list are in same zone, and of same order.
625 * count is the number of pages to free.
627 * If the zone was previously in an "all pages pinned" state then look to
628 * see if this freeing clears that state.
630 * And clear the zone's pages_scanned counter, to hold off the "all pages are
631 * pinned" detection logic.
633 static void free_pcppages_bulk(struct zone
*zone
, int count
,
634 struct per_cpu_pages
*pcp
)
640 spin_lock(&zone
->lock
);
641 zone
->all_unreclaimable
= 0;
642 zone
->pages_scanned
= 0;
646 struct list_head
*list
;
649 * Remove pages from lists in a round-robin fashion. A
650 * batch_free count is maintained that is incremented when an
651 * empty list is encountered. This is so more pages are freed
652 * off fuller lists instead of spinning excessively around empty
657 if (++migratetype
== MIGRATE_PCPTYPES
)
659 list
= &pcp
->lists
[migratetype
];
660 } while (list_empty(list
));
662 /* This is the only non-empty list. Free them all. */
663 if (batch_free
== MIGRATE_PCPTYPES
)
664 batch_free
= to_free
;
667 int mt
; /* migratetype of the to-be-freed page */
669 page
= list_entry(list
->prev
, struct page
, lru
);
670 /* must delete as __free_one_page list manipulates */
671 list_del(&page
->lru
);
672 mt
= get_freepage_migratetype(page
);
673 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
674 __free_one_page(page
, zone
, 0, mt
);
675 trace_mm_page_pcpu_drain(page
, 0, mt
);
676 if (likely(get_pageblock_migratetype(page
) != MIGRATE_ISOLATE
)) {
677 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1);
678 if (is_migrate_cma(mt
))
679 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
, 1);
681 } while (--to_free
&& --batch_free
&& !list_empty(list
));
683 spin_unlock(&zone
->lock
);
686 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
689 spin_lock(&zone
->lock
);
690 zone
->all_unreclaimable
= 0;
691 zone
->pages_scanned
= 0;
693 __free_one_page(page
, zone
, order
, migratetype
);
694 if (unlikely(migratetype
!= MIGRATE_ISOLATE
))
695 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
696 spin_unlock(&zone
->lock
);
699 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
704 trace_mm_page_free(page
, order
);
705 kmemcheck_free_shadow(page
, order
);
708 page
->mapping
= NULL
;
709 for (i
= 0; i
< (1 << order
); i
++)
710 bad
+= free_pages_check(page
+ i
);
714 if (!PageHighMem(page
)) {
715 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
716 debug_check_no_obj_freed(page_address(page
),
719 arch_free_page(page
, order
);
720 kernel_map_pages(page
, 1 << order
, 0);
725 static void __free_pages_ok(struct page
*page
, unsigned int order
)
730 if (!free_pages_prepare(page
, order
))
733 local_irq_save(flags
);
734 __count_vm_events(PGFREE
, 1 << order
);
735 migratetype
= get_pageblock_migratetype(page
);
736 set_freepage_migratetype(page
, migratetype
);
737 free_one_page(page_zone(page
), page
, order
, migratetype
);
738 local_irq_restore(flags
);
742 * Read access to zone->managed_pages is safe because it's unsigned long,
743 * but we still need to serialize writers. Currently all callers of
744 * __free_pages_bootmem() except put_page_bootmem() should only be used
745 * at boot time. So for shorter boot time, we shift the burden to
746 * put_page_bootmem() to serialize writers.
748 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
750 unsigned int nr_pages
= 1 << order
;
754 for (loop
= 0; loop
< nr_pages
; loop
++) {
755 struct page
*p
= &page
[loop
];
757 if (loop
+ 1 < nr_pages
)
759 __ClearPageReserved(p
);
760 set_page_count(p
, 0);
763 page_zone(page
)->managed_pages
+= 1 << order
;
764 set_page_refcounted(page
);
765 __free_pages(page
, order
);
769 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
770 void __init
init_cma_reserved_pageblock(struct page
*page
)
772 unsigned i
= pageblock_nr_pages
;
773 struct page
*p
= page
;
776 __ClearPageReserved(p
);
777 set_page_count(p
, 0);
780 set_page_refcounted(page
);
781 set_pageblock_migratetype(page
, MIGRATE_CMA
);
782 __free_pages(page
, pageblock_order
);
783 totalram_pages
+= pageblock_nr_pages
;
784 #ifdef CONFIG_HIGHMEM
785 if (PageHighMem(page
))
786 totalhigh_pages
+= pageblock_nr_pages
;
792 * The order of subdivision here is critical for the IO subsystem.
793 * Please do not alter this order without good reasons and regression
794 * testing. Specifically, as large blocks of memory are subdivided,
795 * the order in which smaller blocks are delivered depends on the order
796 * they're subdivided in this function. This is the primary factor
797 * influencing the order in which pages are delivered to the IO
798 * subsystem according to empirical testing, and this is also justified
799 * by considering the behavior of a buddy system containing a single
800 * large block of memory acted on by a series of small allocations.
801 * This behavior is a critical factor in sglist merging's success.
805 static inline void expand(struct zone
*zone
, struct page
*page
,
806 int low
, int high
, struct free_area
*area
,
809 unsigned long size
= 1 << high
;
815 VM_BUG_ON(bad_range(zone
, &page
[size
]));
817 #ifdef CONFIG_DEBUG_PAGEALLOC
818 if (high
< debug_guardpage_minorder()) {
820 * Mark as guard pages (or page), that will allow to
821 * merge back to allocator when buddy will be freed.
822 * Corresponding page table entries will not be touched,
823 * pages will stay not present in virtual address space
825 INIT_LIST_HEAD(&page
[size
].lru
);
826 set_page_guard_flag(&page
[size
]);
827 set_page_private(&page
[size
], high
);
828 /* Guard pages are not available for any usage */
829 __mod_zone_freepage_state(zone
, -(1 << high
),
834 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
836 set_page_order(&page
[size
], high
);
841 * This page is about to be returned from the page allocator
843 static inline int check_new_page(struct page
*page
)
845 if (unlikely(page_mapcount(page
) |
846 (page
->mapping
!= NULL
) |
847 (atomic_read(&page
->_count
) != 0) |
848 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
) |
849 (mem_cgroup_bad_page_check(page
)))) {
856 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
860 for (i
= 0; i
< (1 << order
); i
++) {
861 struct page
*p
= page
+ i
;
862 if (unlikely(check_new_page(p
)))
866 set_page_private(page
, 0);
867 set_page_refcounted(page
);
869 arch_alloc_page(page
, order
);
870 kernel_map_pages(page
, 1 << order
, 1);
872 if (gfp_flags
& __GFP_ZERO
)
873 prep_zero_page(page
, order
, gfp_flags
);
875 if (order
&& (gfp_flags
& __GFP_COMP
))
876 prep_compound_page(page
, order
);
882 * Go through the free lists for the given migratetype and remove
883 * the smallest available page from the freelists
886 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
889 unsigned int current_order
;
890 struct free_area
* area
;
893 /* Find a page of the appropriate size in the preferred list */
894 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
895 area
= &(zone
->free_area
[current_order
]);
896 if (list_empty(&area
->free_list
[migratetype
]))
899 page
= list_entry(area
->free_list
[migratetype
].next
,
901 list_del(&page
->lru
);
902 rmv_page_order(page
);
904 expand(zone
, page
, order
, current_order
, area
, migratetype
);
913 * This array describes the order lists are fallen back to when
914 * the free lists for the desirable migrate type are depleted
916 static int fallbacks
[MIGRATE_TYPES
][4] = {
917 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
918 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
920 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
921 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
923 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
925 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
926 [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 (start_pfn
< zone
->zone_start_pfn
)
993 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
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
;
1010 /* Remove an element from the buddy allocator from the fallback list */
1011 static inline struct page
*
1012 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
1014 struct free_area
* area
;
1019 /* Find the largest possible block of pages in the other list */
1020 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
1023 migratetype
= fallbacks
[start_migratetype
][i
];
1025 /* MIGRATE_RESERVE handled later if necessary */
1026 if (migratetype
== MIGRATE_RESERVE
)
1029 area
= &(zone
->free_area
[current_order
]);
1030 if (list_empty(&area
->free_list
[migratetype
]))
1033 page
= list_entry(area
->free_list
[migratetype
].next
,
1038 * If breaking a large block of pages, move all free
1039 * pages to the preferred allocation list. If falling
1040 * back for a reclaimable kernel allocation, be more
1041 * aggressive about taking ownership of free pages
1043 * On the other hand, never change migration
1044 * type of MIGRATE_CMA pageblocks nor move CMA
1045 * pages on different free lists. We don't
1046 * want unmovable pages to be allocated from
1047 * MIGRATE_CMA areas.
1049 if (!is_migrate_cma(migratetype
) &&
1050 (unlikely(current_order
>= pageblock_order
/ 2) ||
1051 start_migratetype
== MIGRATE_RECLAIMABLE
||
1052 page_group_by_mobility_disabled
)) {
1054 pages
= move_freepages_block(zone
, page
,
1057 /* Claim the whole block if over half of it is free */
1058 if (pages
>= (1 << (pageblock_order
-1)) ||
1059 page_group_by_mobility_disabled
)
1060 set_pageblock_migratetype(page
,
1063 migratetype
= start_migratetype
;
1066 /* Remove the page from the freelists */
1067 list_del(&page
->lru
);
1068 rmv_page_order(page
);
1070 /* Take ownership for orders >= pageblock_order */
1071 if (current_order
>= pageblock_order
&&
1072 !is_migrate_cma(migratetype
))
1073 change_pageblock_range(page
, current_order
,
1076 expand(zone
, page
, order
, current_order
, area
,
1077 is_migrate_cma(migratetype
)
1078 ? migratetype
: start_migratetype
);
1080 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1081 start_migratetype
, migratetype
);
1091 * Do the hard work of removing an element from the buddy allocator.
1092 * Call me with the zone->lock already held.
1094 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1100 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1102 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1103 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1106 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1107 * is used because __rmqueue_smallest is an inline function
1108 * and we want just one call site
1111 migratetype
= MIGRATE_RESERVE
;
1116 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1121 * Obtain a specified number of elements from the buddy allocator, all under
1122 * a single hold of the lock, for efficiency. Add them to the supplied list.
1123 * Returns the number of new pages which were placed at *list.
1125 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1126 unsigned long count
, struct list_head
*list
,
1127 int migratetype
, int cold
)
1129 int mt
= migratetype
, i
;
1131 spin_lock(&zone
->lock
);
1132 for (i
= 0; i
< count
; ++i
) {
1133 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1134 if (unlikely(page
== NULL
))
1138 * Split buddy pages returned by expand() are received here
1139 * in physical page order. The page is added to the callers and
1140 * list and the list head then moves forward. From the callers
1141 * perspective, the linked list is ordered by page number in
1142 * some conditions. This is useful for IO devices that can
1143 * merge IO requests if the physical pages are ordered
1146 if (likely(cold
== 0))
1147 list_add(&page
->lru
, list
);
1149 list_add_tail(&page
->lru
, list
);
1150 if (IS_ENABLED(CONFIG_CMA
)) {
1151 mt
= get_pageblock_migratetype(page
);
1152 if (!is_migrate_cma(mt
) && mt
!= MIGRATE_ISOLATE
)
1155 set_freepage_migratetype(page
, mt
);
1157 if (is_migrate_cma(mt
))
1158 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1161 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1162 spin_unlock(&zone
->lock
);
1168 * Called from the vmstat counter updater to drain pagesets of this
1169 * currently executing processor on remote nodes after they have
1172 * Note that this function must be called with the thread pinned to
1173 * a single processor.
1175 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1177 unsigned long flags
;
1180 local_irq_save(flags
);
1181 if (pcp
->count
>= pcp
->batch
)
1182 to_drain
= pcp
->batch
;
1184 to_drain
= pcp
->count
;
1186 free_pcppages_bulk(zone
, to_drain
, pcp
);
1187 pcp
->count
-= to_drain
;
1189 local_irq_restore(flags
);
1194 * Drain pages of the indicated processor.
1196 * The processor must either be the current processor and the
1197 * thread pinned to the current processor or a processor that
1200 static void drain_pages(unsigned int cpu
)
1202 unsigned long flags
;
1205 for_each_populated_zone(zone
) {
1206 struct per_cpu_pageset
*pset
;
1207 struct per_cpu_pages
*pcp
;
1209 local_irq_save(flags
);
1210 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1214 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1217 local_irq_restore(flags
);
1222 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1224 void drain_local_pages(void *arg
)
1226 drain_pages(smp_processor_id());
1230 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1232 * Note that this code is protected against sending an IPI to an offline
1233 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1234 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1235 * nothing keeps CPUs from showing up after we populated the cpumask and
1236 * before the call to on_each_cpu_mask().
1238 void drain_all_pages(void)
1241 struct per_cpu_pageset
*pcp
;
1245 * Allocate in the BSS so we wont require allocation in
1246 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1248 static cpumask_t cpus_with_pcps
;
1251 * We don't care about racing with CPU hotplug event
1252 * as offline notification will cause the notified
1253 * cpu to drain that CPU pcps and on_each_cpu_mask
1254 * disables preemption as part of its processing
1256 for_each_online_cpu(cpu
) {
1257 bool has_pcps
= false;
1258 for_each_populated_zone(zone
) {
1259 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1260 if (pcp
->pcp
.count
) {
1266 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1268 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1270 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1273 #ifdef CONFIG_HIBERNATION
1275 void mark_free_pages(struct zone
*zone
)
1277 unsigned long pfn
, max_zone_pfn
;
1278 unsigned long flags
;
1280 struct list_head
*curr
;
1282 if (!zone
->spanned_pages
)
1285 spin_lock_irqsave(&zone
->lock
, flags
);
1287 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1288 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1289 if (pfn_valid(pfn
)) {
1290 struct page
*page
= pfn_to_page(pfn
);
1292 if (!swsusp_page_is_forbidden(page
))
1293 swsusp_unset_page_free(page
);
1296 for_each_migratetype_order(order
, t
) {
1297 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1300 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1301 for (i
= 0; i
< (1UL << order
); i
++)
1302 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1305 spin_unlock_irqrestore(&zone
->lock
, flags
);
1307 #endif /* CONFIG_PM */
1310 * Free a 0-order page
1311 * cold == 1 ? free a cold page : free a hot page
1313 void free_hot_cold_page(struct page
*page
, int cold
)
1315 struct zone
*zone
= page_zone(page
);
1316 struct per_cpu_pages
*pcp
;
1317 unsigned long flags
;
1320 if (!free_pages_prepare(page
, 0))
1323 migratetype
= get_pageblock_migratetype(page
);
1324 set_freepage_migratetype(page
, migratetype
);
1325 local_irq_save(flags
);
1326 __count_vm_event(PGFREE
);
1329 * We only track unmovable, reclaimable and movable on pcp lists.
1330 * Free ISOLATE pages back to the allocator because they are being
1331 * offlined but treat RESERVE as movable pages so we can get those
1332 * areas back if necessary. Otherwise, we may have to free
1333 * excessively into the page allocator
1335 if (migratetype
>= MIGRATE_PCPTYPES
) {
1336 if (unlikely(migratetype
== MIGRATE_ISOLATE
)) {
1337 free_one_page(zone
, page
, 0, migratetype
);
1340 migratetype
= MIGRATE_MOVABLE
;
1343 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1345 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1347 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1349 if (pcp
->count
>= pcp
->high
) {
1350 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1351 pcp
->count
-= pcp
->batch
;
1355 local_irq_restore(flags
);
1359 * Free a list of 0-order pages
1361 void free_hot_cold_page_list(struct list_head
*list
, int cold
)
1363 struct page
*page
, *next
;
1365 list_for_each_entry_safe(page
, next
, list
, lru
) {
1366 trace_mm_page_free_batched(page
, cold
);
1367 free_hot_cold_page(page
, cold
);
1372 * split_page takes a non-compound higher-order page, and splits it into
1373 * n (1<<order) sub-pages: page[0..n]
1374 * Each sub-page must be freed individually.
1376 * Note: this is probably too low level an operation for use in drivers.
1377 * Please consult with lkml before using this in your driver.
1379 void split_page(struct page
*page
, unsigned int order
)
1383 VM_BUG_ON(PageCompound(page
));
1384 VM_BUG_ON(!page_count(page
));
1386 #ifdef CONFIG_KMEMCHECK
1388 * Split shadow pages too, because free(page[0]) would
1389 * otherwise free the whole shadow.
1391 if (kmemcheck_page_is_tracked(page
))
1392 split_page(virt_to_page(page
[0].shadow
), order
);
1395 for (i
= 1; i
< (1 << order
); i
++)
1396 set_page_refcounted(page
+ i
);
1399 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1401 unsigned long watermark
;
1405 BUG_ON(!PageBuddy(page
));
1407 zone
= page_zone(page
);
1408 mt
= get_pageblock_migratetype(page
);
1410 if (mt
!= MIGRATE_ISOLATE
) {
1411 /* Obey watermarks as if the page was being allocated */
1412 watermark
= low_wmark_pages(zone
) + (1 << order
);
1413 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1416 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1419 /* Remove page from free list */
1420 list_del(&page
->lru
);
1421 zone
->free_area
[order
].nr_free
--;
1422 rmv_page_order(page
);
1424 /* Set the pageblock if the isolated page is at least a pageblock */
1425 if (order
>= pageblock_order
- 1) {
1426 struct page
*endpage
= page
+ (1 << order
) - 1;
1427 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1428 int mt
= get_pageblock_migratetype(page
);
1429 if (mt
!= MIGRATE_ISOLATE
&& !is_migrate_cma(mt
))
1430 set_pageblock_migratetype(page
,
1435 return 1UL << order
;
1439 * Similar to split_page except the page is already free. As this is only
1440 * being used for migration, the migratetype of the block also changes.
1441 * As this is called with interrupts disabled, the caller is responsible
1442 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1445 * Note: this is probably too low level an operation for use in drivers.
1446 * Please consult with lkml before using this in your driver.
1448 int split_free_page(struct page
*page
)
1453 order
= page_order(page
);
1455 nr_pages
= __isolate_free_page(page
, order
);
1459 /* Split into individual pages */
1460 set_page_refcounted(page
);
1461 split_page(page
, order
);
1466 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1467 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1471 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1472 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1475 unsigned long flags
;
1477 int cold
= !!(gfp_flags
& __GFP_COLD
);
1480 if (likely(order
== 0)) {
1481 struct per_cpu_pages
*pcp
;
1482 struct list_head
*list
;
1484 local_irq_save(flags
);
1485 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1486 list
= &pcp
->lists
[migratetype
];
1487 if (list_empty(list
)) {
1488 pcp
->count
+= rmqueue_bulk(zone
, 0,
1491 if (unlikely(list_empty(list
)))
1496 page
= list_entry(list
->prev
, struct page
, lru
);
1498 page
= list_entry(list
->next
, struct page
, lru
);
1500 list_del(&page
->lru
);
1503 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1505 * __GFP_NOFAIL is not to be used in new code.
1507 * All __GFP_NOFAIL callers should be fixed so that they
1508 * properly detect and handle allocation failures.
1510 * We most definitely don't want callers attempting to
1511 * allocate greater than order-1 page units with
1514 WARN_ON_ONCE(order
> 1);
1516 spin_lock_irqsave(&zone
->lock
, flags
);
1517 page
= __rmqueue(zone
, order
, migratetype
);
1518 spin_unlock(&zone
->lock
);
1521 __mod_zone_freepage_state(zone
, -(1 << order
),
1522 get_pageblock_migratetype(page
));
1525 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1526 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1527 local_irq_restore(flags
);
1529 VM_BUG_ON(bad_range(zone
, page
));
1530 if (prep_new_page(page
, order
, gfp_flags
))
1535 local_irq_restore(flags
);
1539 #ifdef CONFIG_FAIL_PAGE_ALLOC
1542 struct fault_attr attr
;
1544 u32 ignore_gfp_highmem
;
1545 u32 ignore_gfp_wait
;
1547 } fail_page_alloc
= {
1548 .attr
= FAULT_ATTR_INITIALIZER
,
1549 .ignore_gfp_wait
= 1,
1550 .ignore_gfp_highmem
= 1,
1554 static int __init
setup_fail_page_alloc(char *str
)
1556 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1558 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1560 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1562 if (order
< fail_page_alloc
.min_order
)
1564 if (gfp_mask
& __GFP_NOFAIL
)
1566 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1568 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1571 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1574 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1576 static int __init
fail_page_alloc_debugfs(void)
1578 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1581 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1582 &fail_page_alloc
.attr
);
1584 return PTR_ERR(dir
);
1586 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1587 &fail_page_alloc
.ignore_gfp_wait
))
1589 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1590 &fail_page_alloc
.ignore_gfp_highmem
))
1592 if (!debugfs_create_u32("min-order", mode
, dir
,
1593 &fail_page_alloc
.min_order
))
1598 debugfs_remove_recursive(dir
);
1603 late_initcall(fail_page_alloc_debugfs
);
1605 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1607 #else /* CONFIG_FAIL_PAGE_ALLOC */
1609 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1614 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1617 * Return true if free pages are above 'mark'. This takes into account the order
1618 * of the allocation.
1620 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1621 int classzone_idx
, int alloc_flags
, long free_pages
)
1623 /* free_pages my go negative - that's OK */
1625 long lowmem_reserve
= z
->lowmem_reserve
[classzone_idx
];
1628 free_pages
-= (1 << order
) - 1;
1629 if (alloc_flags
& ALLOC_HIGH
)
1631 if (alloc_flags
& ALLOC_HARDER
)
1634 /* If allocation can't use CMA areas don't use free CMA pages */
1635 if (!(alloc_flags
& ALLOC_CMA
))
1636 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1638 if (free_pages
<= min
+ lowmem_reserve
)
1640 for (o
= 0; o
< order
; o
++) {
1641 /* At the next order, this order's pages become unavailable */
1642 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1644 /* Require fewer higher order pages to be free */
1647 if (free_pages
<= min
)
1653 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1654 int classzone_idx
, int alloc_flags
)
1656 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1657 zone_page_state(z
, NR_FREE_PAGES
));
1660 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1661 int classzone_idx
, int alloc_flags
)
1663 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1665 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1666 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1668 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1674 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1675 * skip over zones that are not allowed by the cpuset, or that have
1676 * been recently (in last second) found to be nearly full. See further
1677 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1678 * that have to skip over a lot of full or unallowed zones.
1680 * If the zonelist cache is present in the passed in zonelist, then
1681 * returns a pointer to the allowed node mask (either the current
1682 * tasks mems_allowed, or node_states[N_MEMORY].)
1684 * If the zonelist cache is not available for this zonelist, does
1685 * nothing and returns NULL.
1687 * If the fullzones BITMAP in the zonelist cache is stale (more than
1688 * a second since last zap'd) then we zap it out (clear its bits.)
1690 * We hold off even calling zlc_setup, until after we've checked the
1691 * first zone in the zonelist, on the theory that most allocations will
1692 * be satisfied from that first zone, so best to examine that zone as
1693 * quickly as we can.
1695 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1697 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1698 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1700 zlc
= zonelist
->zlcache_ptr
;
1704 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1705 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1706 zlc
->last_full_zap
= jiffies
;
1709 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1710 &cpuset_current_mems_allowed
:
1711 &node_states
[N_MEMORY
];
1712 return allowednodes
;
1716 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1717 * if it is worth looking at further for free memory:
1718 * 1) Check that the zone isn't thought to be full (doesn't have its
1719 * bit set in the zonelist_cache fullzones BITMAP).
1720 * 2) Check that the zones node (obtained from the zonelist_cache
1721 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1722 * Return true (non-zero) if zone is worth looking at further, or
1723 * else return false (zero) if it is not.
1725 * This check -ignores- the distinction between various watermarks,
1726 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1727 * found to be full for any variation of these watermarks, it will
1728 * be considered full for up to one second by all requests, unless
1729 * we are so low on memory on all allowed nodes that we are forced
1730 * into the second scan of the zonelist.
1732 * In the second scan we ignore this zonelist cache and exactly
1733 * apply the watermarks to all zones, even it is slower to do so.
1734 * We are low on memory in the second scan, and should leave no stone
1735 * unturned looking for a free page.
1737 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1738 nodemask_t
*allowednodes
)
1740 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1741 int i
; /* index of *z in zonelist zones */
1742 int n
; /* node that zone *z is on */
1744 zlc
= zonelist
->zlcache_ptr
;
1748 i
= z
- zonelist
->_zonerefs
;
1751 /* This zone is worth trying if it is allowed but not full */
1752 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1756 * Given 'z' scanning a zonelist, set the corresponding bit in
1757 * zlc->fullzones, so that subsequent attempts to allocate a page
1758 * from that zone don't waste time re-examining it.
1760 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1762 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1763 int i
; /* index of *z in zonelist zones */
1765 zlc
= zonelist
->zlcache_ptr
;
1769 i
= z
- zonelist
->_zonerefs
;
1771 set_bit(i
, zlc
->fullzones
);
1775 * clear all zones full, called after direct reclaim makes progress so that
1776 * a zone that was recently full is not skipped over for up to a second
1778 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1780 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1782 zlc
= zonelist
->zlcache_ptr
;
1786 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1789 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1791 return node_isset(local_zone
->node
, zone
->zone_pgdat
->reclaim_nodes
);
1794 static void __paginginit
init_zone_allows_reclaim(int nid
)
1798 for_each_online_node(i
)
1799 if (node_distance(nid
, i
) <= RECLAIM_DISTANCE
)
1800 node_set(i
, NODE_DATA(nid
)->reclaim_nodes
);
1802 zone_reclaim_mode
= 1;
1805 #else /* CONFIG_NUMA */
1807 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1812 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1813 nodemask_t
*allowednodes
)
1818 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1822 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1826 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1831 static inline void init_zone_allows_reclaim(int nid
)
1834 #endif /* CONFIG_NUMA */
1837 * get_page_from_freelist goes through the zonelist trying to allocate
1840 static struct page
*
1841 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1842 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1843 struct zone
*preferred_zone
, int migratetype
)
1846 struct page
*page
= NULL
;
1849 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1850 int zlc_active
= 0; /* set if using zonelist_cache */
1851 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1853 classzone_idx
= zone_idx(preferred_zone
);
1856 * Scan zonelist, looking for a zone with enough free.
1857 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1859 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1860 high_zoneidx
, nodemask
) {
1861 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1862 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1864 if ((alloc_flags
& ALLOC_CPUSET
) &&
1865 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1868 * When allocating a page cache page for writing, we
1869 * want to get it from a zone that is within its dirty
1870 * limit, such that no single zone holds more than its
1871 * proportional share of globally allowed dirty pages.
1872 * The dirty limits take into account the zone's
1873 * lowmem reserves and high watermark so that kswapd
1874 * should be able to balance it without having to
1875 * write pages from its LRU list.
1877 * This may look like it could increase pressure on
1878 * lower zones by failing allocations in higher zones
1879 * before they are full. But the pages that do spill
1880 * over are limited as the lower zones are protected
1881 * by this very same mechanism. It should not become
1882 * a practical burden to them.
1884 * XXX: For now, allow allocations to potentially
1885 * exceed the per-zone dirty limit in the slowpath
1886 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1887 * which is important when on a NUMA setup the allowed
1888 * zones are together not big enough to reach the
1889 * global limit. The proper fix for these situations
1890 * will require awareness of zones in the
1891 * dirty-throttling and the flusher threads.
1893 if ((alloc_flags
& ALLOC_WMARK_LOW
) &&
1894 (gfp_mask
& __GFP_WRITE
) && !zone_dirty_ok(zone
))
1895 goto this_zone_full
;
1897 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1898 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1902 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1903 if (zone_watermark_ok(zone
, order
, mark
,
1904 classzone_idx
, alloc_flags
))
1907 if (IS_ENABLED(CONFIG_NUMA
) &&
1908 !did_zlc_setup
&& nr_online_nodes
> 1) {
1910 * we do zlc_setup if there are multiple nodes
1911 * and before considering the first zone allowed
1914 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1919 if (zone_reclaim_mode
== 0 ||
1920 !zone_allows_reclaim(preferred_zone
, zone
))
1921 goto this_zone_full
;
1924 * As we may have just activated ZLC, check if the first
1925 * eligible zone has failed zone_reclaim recently.
1927 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1928 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1931 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1933 case ZONE_RECLAIM_NOSCAN
:
1936 case ZONE_RECLAIM_FULL
:
1937 /* scanned but unreclaimable */
1940 /* did we reclaim enough */
1941 if (!zone_watermark_ok(zone
, order
, mark
,
1942 classzone_idx
, alloc_flags
))
1943 goto this_zone_full
;
1948 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1949 gfp_mask
, migratetype
);
1953 if (IS_ENABLED(CONFIG_NUMA
))
1954 zlc_mark_zone_full(zonelist
, z
);
1957 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && page
== NULL
&& zlc_active
)) {
1958 /* Disable zlc cache for second zonelist scan */
1965 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1966 * necessary to allocate the page. The expectation is
1967 * that the caller is taking steps that will free more
1968 * memory. The caller should avoid the page being used
1969 * for !PFMEMALLOC purposes.
1971 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
1977 * Large machines with many possible nodes should not always dump per-node
1978 * meminfo in irq context.
1980 static inline bool should_suppress_show_mem(void)
1985 ret
= in_interrupt();
1990 static DEFINE_RATELIMIT_STATE(nopage_rs
,
1991 DEFAULT_RATELIMIT_INTERVAL
,
1992 DEFAULT_RATELIMIT_BURST
);
1994 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
1996 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
1998 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
1999 debug_guardpage_minorder() > 0)
2003 * This documents exceptions given to allocations in certain
2004 * contexts that are allowed to allocate outside current's set
2007 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2008 if (test_thread_flag(TIF_MEMDIE
) ||
2009 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2010 filter
&= ~SHOW_MEM_FILTER_NODES
;
2011 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2012 filter
&= ~SHOW_MEM_FILTER_NODES
;
2015 struct va_format vaf
;
2018 va_start(args
, fmt
);
2023 pr_warn("%pV", &vaf
);
2028 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2029 current
->comm
, order
, gfp_mask
);
2032 if (!should_suppress_show_mem())
2037 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2038 unsigned long did_some_progress
,
2039 unsigned long pages_reclaimed
)
2041 /* Do not loop if specifically requested */
2042 if (gfp_mask
& __GFP_NORETRY
)
2045 /* Always retry if specifically requested */
2046 if (gfp_mask
& __GFP_NOFAIL
)
2050 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2051 * making forward progress without invoking OOM. Suspend also disables
2052 * storage devices so kswapd will not help. Bail if we are suspending.
2054 if (!did_some_progress
&& pm_suspended_storage())
2058 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2059 * means __GFP_NOFAIL, but that may not be true in other
2062 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2066 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2067 * specified, then we retry until we no longer reclaim any pages
2068 * (above), or we've reclaimed an order of pages at least as
2069 * large as the allocation's order. In both cases, if the
2070 * allocation still fails, we stop retrying.
2072 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2078 static inline struct page
*
2079 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2080 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2081 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2086 /* Acquire the OOM killer lock for the zones in zonelist */
2087 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
2088 schedule_timeout_uninterruptible(1);
2093 * Go through the zonelist yet one more time, keep very high watermark
2094 * here, this is only to catch a parallel oom killing, we must fail if
2095 * we're still under heavy pressure.
2097 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2098 order
, zonelist
, high_zoneidx
,
2099 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2100 preferred_zone
, migratetype
);
2104 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2105 /* The OOM killer will not help higher order allocs */
2106 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2108 /* The OOM killer does not needlessly kill tasks for lowmem */
2109 if (high_zoneidx
< ZONE_NORMAL
)
2112 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2113 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2114 * The caller should handle page allocation failure by itself if
2115 * it specifies __GFP_THISNODE.
2116 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2118 if (gfp_mask
& __GFP_THISNODE
)
2121 /* Exhausted what can be done so it's blamo time */
2122 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2125 clear_zonelist_oom(zonelist
, gfp_mask
);
2129 #ifdef CONFIG_COMPACTION
2130 /* Try memory compaction for high-order allocations before reclaim */
2131 static struct page
*
2132 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2133 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2134 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2135 int migratetype
, bool sync_migration
,
2136 bool *contended_compaction
, bool *deferred_compaction
,
2137 unsigned long *did_some_progress
)
2142 if (compaction_deferred(preferred_zone
, order
)) {
2143 *deferred_compaction
= true;
2147 current
->flags
|= PF_MEMALLOC
;
2148 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2149 nodemask
, sync_migration
,
2150 contended_compaction
);
2151 current
->flags
&= ~PF_MEMALLOC
;
2153 if (*did_some_progress
!= COMPACT_SKIPPED
) {
2156 /* Page migration frees to the PCP lists but we want merging */
2157 drain_pages(get_cpu());
2160 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2161 order
, zonelist
, high_zoneidx
,
2162 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2163 preferred_zone
, migratetype
);
2165 preferred_zone
->compact_blockskip_flush
= false;
2166 preferred_zone
->compact_considered
= 0;
2167 preferred_zone
->compact_defer_shift
= 0;
2168 if (order
>= preferred_zone
->compact_order_failed
)
2169 preferred_zone
->compact_order_failed
= order
+ 1;
2170 count_vm_event(COMPACTSUCCESS
);
2175 * It's bad if compaction run occurs and fails.
2176 * The most likely reason is that pages exist,
2177 * but not enough to satisfy watermarks.
2179 count_vm_event(COMPACTFAIL
);
2182 * As async compaction considers a subset of pageblocks, only
2183 * defer if the failure was a sync compaction failure.
2186 defer_compaction(preferred_zone
, order
);
2194 static inline struct page
*
2195 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2196 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2197 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2198 int migratetype
, bool sync_migration
,
2199 bool *contended_compaction
, bool *deferred_compaction
,
2200 unsigned long *did_some_progress
)
2204 #endif /* CONFIG_COMPACTION */
2206 /* Perform direct synchronous page reclaim */
2208 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2209 nodemask_t
*nodemask
)
2211 struct reclaim_state reclaim_state
;
2216 /* We now go into synchronous reclaim */
2217 cpuset_memory_pressure_bump();
2218 current
->flags
|= PF_MEMALLOC
;
2219 lockdep_set_current_reclaim_state(gfp_mask
);
2220 reclaim_state
.reclaimed_slab
= 0;
2221 current
->reclaim_state
= &reclaim_state
;
2223 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2225 current
->reclaim_state
= NULL
;
2226 lockdep_clear_current_reclaim_state();
2227 current
->flags
&= ~PF_MEMALLOC
;
2234 /* The really slow allocator path where we enter direct reclaim */
2235 static inline struct page
*
2236 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2237 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2238 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2239 int migratetype
, unsigned long *did_some_progress
)
2241 struct page
*page
= NULL
;
2242 bool drained
= false;
2244 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2246 if (unlikely(!(*did_some_progress
)))
2249 /* After successful reclaim, reconsider all zones for allocation */
2250 if (IS_ENABLED(CONFIG_NUMA
))
2251 zlc_clear_zones_full(zonelist
);
2254 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2255 zonelist
, high_zoneidx
,
2256 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2257 preferred_zone
, migratetype
);
2260 * If an allocation failed after direct reclaim, it could be because
2261 * pages are pinned on the per-cpu lists. Drain them and try again
2263 if (!page
&& !drained
) {
2273 * This is called in the allocator slow-path if the allocation request is of
2274 * sufficient urgency to ignore watermarks and take other desperate measures
2276 static inline struct page
*
2277 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2278 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2279 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2285 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2286 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2287 preferred_zone
, migratetype
);
2289 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2290 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2291 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2297 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
2298 enum zone_type high_zoneidx
,
2299 enum zone_type classzone_idx
)
2304 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
2305 wakeup_kswapd(zone
, order
, classzone_idx
);
2309 gfp_to_alloc_flags(gfp_t gfp_mask
)
2311 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2312 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2314 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2315 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2318 * The caller may dip into page reserves a bit more if the caller
2319 * cannot run direct reclaim, or if the caller has realtime scheduling
2320 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2321 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2323 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2327 * Not worth trying to allocate harder for
2328 * __GFP_NOMEMALLOC even if it can't schedule.
2330 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2331 alloc_flags
|= ALLOC_HARDER
;
2333 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2334 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2336 alloc_flags
&= ~ALLOC_CPUSET
;
2337 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2338 alloc_flags
|= ALLOC_HARDER
;
2340 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2341 if (gfp_mask
& __GFP_MEMALLOC
)
2342 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2343 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2344 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2345 else if (!in_interrupt() &&
2346 ((current
->flags
& PF_MEMALLOC
) ||
2347 unlikely(test_thread_flag(TIF_MEMDIE
))))
2348 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2351 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2352 alloc_flags
|= ALLOC_CMA
;
2357 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2359 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2362 static inline struct page
*
2363 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2364 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2365 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2368 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2369 struct page
*page
= NULL
;
2371 unsigned long pages_reclaimed
= 0;
2372 unsigned long did_some_progress
;
2373 bool sync_migration
= false;
2374 bool deferred_compaction
= false;
2375 bool contended_compaction
= false;
2378 * In the slowpath, we sanity check order to avoid ever trying to
2379 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2380 * be using allocators in order of preference for an area that is
2383 if (order
>= MAX_ORDER
) {
2384 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2389 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2390 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2391 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2392 * using a larger set of nodes after it has established that the
2393 * allowed per node queues are empty and that nodes are
2396 if (IS_ENABLED(CONFIG_NUMA
) &&
2397 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2401 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2402 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2403 zone_idx(preferred_zone
));
2406 * OK, we're below the kswapd watermark and have kicked background
2407 * reclaim. Now things get more complex, so set up alloc_flags according
2408 * to how we want to proceed.
2410 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2413 * Find the true preferred zone if the allocation is unconstrained by
2416 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2417 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2421 /* This is the last chance, in general, before the goto nopage. */
2422 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2423 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2424 preferred_zone
, migratetype
);
2428 /* Allocate without watermarks if the context allows */
2429 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2431 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2432 * the allocation is high priority and these type of
2433 * allocations are system rather than user orientated
2435 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2437 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2438 zonelist
, high_zoneidx
, nodemask
,
2439 preferred_zone
, migratetype
);
2445 /* Atomic allocations - we can't balance anything */
2449 /* Avoid recursion of direct reclaim */
2450 if (current
->flags
& PF_MEMALLOC
)
2453 /* Avoid allocations with no watermarks from looping endlessly */
2454 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2458 * Try direct compaction. The first pass is asynchronous. Subsequent
2459 * attempts after direct reclaim are synchronous
2461 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2462 zonelist
, high_zoneidx
,
2464 alloc_flags
, preferred_zone
,
2465 migratetype
, sync_migration
,
2466 &contended_compaction
,
2467 &deferred_compaction
,
2468 &did_some_progress
);
2471 sync_migration
= true;
2474 * If compaction is deferred for high-order allocations, it is because
2475 * sync compaction recently failed. In this is the case and the caller
2476 * requested a movable allocation that does not heavily disrupt the
2477 * system then fail the allocation instead of entering direct reclaim.
2479 if ((deferred_compaction
|| contended_compaction
) &&
2480 (gfp_mask
& __GFP_NO_KSWAPD
))
2483 /* Try direct reclaim and then allocating */
2484 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2485 zonelist
, high_zoneidx
,
2487 alloc_flags
, preferred_zone
,
2488 migratetype
, &did_some_progress
);
2493 * If we failed to make any progress reclaiming, then we are
2494 * running out of options and have to consider going OOM
2496 if (!did_some_progress
) {
2497 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2498 if (oom_killer_disabled
)
2500 /* Coredumps can quickly deplete all memory reserves */
2501 if ((current
->flags
& PF_DUMPCORE
) &&
2502 !(gfp_mask
& __GFP_NOFAIL
))
2504 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2505 zonelist
, high_zoneidx
,
2506 nodemask
, preferred_zone
,
2511 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2513 * The oom killer is not called for high-order
2514 * allocations that may fail, so if no progress
2515 * is being made, there are no other options and
2516 * retrying is unlikely to help.
2518 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2521 * The oom killer is not called for lowmem
2522 * allocations to prevent needlessly killing
2525 if (high_zoneidx
< ZONE_NORMAL
)
2533 /* Check if we should retry the allocation */
2534 pages_reclaimed
+= did_some_progress
;
2535 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2537 /* Wait for some write requests to complete then retry */
2538 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2542 * High-order allocations do not necessarily loop after
2543 * direct reclaim and reclaim/compaction depends on compaction
2544 * being called after reclaim so call directly if necessary
2546 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2547 zonelist
, high_zoneidx
,
2549 alloc_flags
, preferred_zone
,
2550 migratetype
, sync_migration
,
2551 &contended_compaction
,
2552 &deferred_compaction
,
2553 &did_some_progress
);
2559 warn_alloc_failed(gfp_mask
, order
, NULL
);
2562 if (kmemcheck_enabled
)
2563 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2569 * This is the 'heart' of the zoned buddy allocator.
2572 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2573 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2575 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2576 struct zone
*preferred_zone
;
2577 struct page
*page
= NULL
;
2578 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2579 unsigned int cpuset_mems_cookie
;
2580 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
;
2581 struct mem_cgroup
*memcg
= NULL
;
2583 gfp_mask
&= gfp_allowed_mask
;
2585 lockdep_trace_alloc(gfp_mask
);
2587 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2589 if (should_fail_alloc_page(gfp_mask
, order
))
2593 * Check the zones suitable for the gfp_mask contain at least one
2594 * valid zone. It's possible to have an empty zonelist as a result
2595 * of GFP_THISNODE and a memoryless node
2597 if (unlikely(!zonelist
->_zonerefs
->zone
))
2601 * Will only have any effect when __GFP_KMEMCG is set. This is
2602 * verified in the (always inline) callee
2604 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2608 cpuset_mems_cookie
= get_mems_allowed();
2610 /* The preferred zone is used for statistics later */
2611 first_zones_zonelist(zonelist
, high_zoneidx
,
2612 nodemask
? : &cpuset_current_mems_allowed
,
2614 if (!preferred_zone
)
2618 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2619 alloc_flags
|= ALLOC_CMA
;
2621 /* First allocation attempt */
2622 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2623 zonelist
, high_zoneidx
, alloc_flags
,
2624 preferred_zone
, migratetype
);
2625 if (unlikely(!page
))
2626 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2627 zonelist
, high_zoneidx
, nodemask
,
2628 preferred_zone
, migratetype
);
2630 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2634 * When updating a task's mems_allowed, it is possible to race with
2635 * parallel threads in such a way that an allocation can fail while
2636 * the mask is being updated. If a page allocation is about to fail,
2637 * check if the cpuset changed during allocation and if so, retry.
2639 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2642 memcg_kmem_commit_charge(page
, memcg
, order
);
2646 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2649 * Common helper functions.
2651 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2656 * __get_free_pages() returns a 32-bit address, which cannot represent
2659 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2661 page
= alloc_pages(gfp_mask
, order
);
2664 return (unsigned long) page_address(page
);
2666 EXPORT_SYMBOL(__get_free_pages
);
2668 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2670 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2672 EXPORT_SYMBOL(get_zeroed_page
);
2674 void __free_pages(struct page
*page
, unsigned int order
)
2676 if (put_page_testzero(page
)) {
2678 free_hot_cold_page(page
, 0);
2680 __free_pages_ok(page
, order
);
2684 EXPORT_SYMBOL(__free_pages
);
2686 void free_pages(unsigned long addr
, unsigned int order
)
2689 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2690 __free_pages(virt_to_page((void *)addr
), order
);
2694 EXPORT_SYMBOL(free_pages
);
2697 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2698 * pages allocated with __GFP_KMEMCG.
2700 * Those pages are accounted to a particular memcg, embedded in the
2701 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2702 * for that information only to find out that it is NULL for users who have no
2703 * interest in that whatsoever, we provide these functions.
2705 * The caller knows better which flags it relies on.
2707 void __free_memcg_kmem_pages(struct page
*page
, unsigned int order
)
2709 memcg_kmem_uncharge_pages(page
, order
);
2710 __free_pages(page
, order
);
2713 void free_memcg_kmem_pages(unsigned long addr
, unsigned int order
)
2716 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2717 __free_memcg_kmem_pages(virt_to_page((void *)addr
), order
);
2721 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2724 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2725 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2727 split_page(virt_to_page((void *)addr
), order
);
2728 while (used
< alloc_end
) {
2733 return (void *)addr
;
2737 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2738 * @size: the number of bytes to allocate
2739 * @gfp_mask: GFP flags for the allocation
2741 * This function is similar to alloc_pages(), except that it allocates the
2742 * minimum number of pages to satisfy the request. alloc_pages() can only
2743 * allocate memory in power-of-two pages.
2745 * This function is also limited by MAX_ORDER.
2747 * Memory allocated by this function must be released by free_pages_exact().
2749 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2751 unsigned int order
= get_order(size
);
2754 addr
= __get_free_pages(gfp_mask
, order
);
2755 return make_alloc_exact(addr
, order
, size
);
2757 EXPORT_SYMBOL(alloc_pages_exact
);
2760 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2762 * @nid: the preferred node ID where memory should be allocated
2763 * @size: the number of bytes to allocate
2764 * @gfp_mask: GFP flags for the allocation
2766 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2768 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2771 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2773 unsigned order
= get_order(size
);
2774 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2777 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2779 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2782 * free_pages_exact - release memory allocated via alloc_pages_exact()
2783 * @virt: the value returned by alloc_pages_exact.
2784 * @size: size of allocation, same value as passed to alloc_pages_exact().
2786 * Release the memory allocated by a previous call to alloc_pages_exact.
2788 void free_pages_exact(void *virt
, size_t size
)
2790 unsigned long addr
= (unsigned long)virt
;
2791 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2793 while (addr
< end
) {
2798 EXPORT_SYMBOL(free_pages_exact
);
2800 static unsigned int nr_free_zone_pages(int offset
)
2805 /* Just pick one node, since fallback list is circular */
2806 unsigned int sum
= 0;
2808 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2810 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2811 unsigned long size
= zone
->managed_pages
;
2812 unsigned long high
= high_wmark_pages(zone
);
2821 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2823 unsigned int nr_free_buffer_pages(void)
2825 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2827 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2830 * Amount of free RAM allocatable within all zones
2832 unsigned int nr_free_pagecache_pages(void)
2834 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2837 static inline void show_node(struct zone
*zone
)
2839 if (IS_ENABLED(CONFIG_NUMA
))
2840 printk("Node %d ", zone_to_nid(zone
));
2843 void si_meminfo(struct sysinfo
*val
)
2845 val
->totalram
= totalram_pages
;
2847 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2848 val
->bufferram
= nr_blockdev_pages();
2849 val
->totalhigh
= totalhigh_pages
;
2850 val
->freehigh
= nr_free_highpages();
2851 val
->mem_unit
= PAGE_SIZE
;
2854 EXPORT_SYMBOL(si_meminfo
);
2857 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2859 pg_data_t
*pgdat
= NODE_DATA(nid
);
2861 val
->totalram
= pgdat
->node_present_pages
;
2862 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2863 #ifdef CONFIG_HIGHMEM
2864 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
2865 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2871 val
->mem_unit
= PAGE_SIZE
;
2876 * Determine whether the node should be displayed or not, depending on whether
2877 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2879 bool skip_free_areas_node(unsigned int flags
, int nid
)
2882 unsigned int cpuset_mems_cookie
;
2884 if (!(flags
& SHOW_MEM_FILTER_NODES
))
2888 cpuset_mems_cookie
= get_mems_allowed();
2889 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
2890 } while (!put_mems_allowed(cpuset_mems_cookie
));
2895 #define K(x) ((x) << (PAGE_SHIFT-10))
2897 static void show_migration_types(unsigned char type
)
2899 static const char types
[MIGRATE_TYPES
] = {
2900 [MIGRATE_UNMOVABLE
] = 'U',
2901 [MIGRATE_RECLAIMABLE
] = 'E',
2902 [MIGRATE_MOVABLE
] = 'M',
2903 [MIGRATE_RESERVE
] = 'R',
2905 [MIGRATE_CMA
] = 'C',
2907 [MIGRATE_ISOLATE
] = 'I',
2909 char tmp
[MIGRATE_TYPES
+ 1];
2913 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
2914 if (type
& (1 << i
))
2919 printk("(%s) ", tmp
);
2923 * Show free area list (used inside shift_scroll-lock stuff)
2924 * We also calculate the percentage fragmentation. We do this by counting the
2925 * memory on each free list with the exception of the first item on the list.
2926 * Suppresses nodes that are not allowed by current's cpuset if
2927 * SHOW_MEM_FILTER_NODES is passed.
2929 void show_free_areas(unsigned int filter
)
2934 for_each_populated_zone(zone
) {
2935 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2938 printk("%s per-cpu:\n", zone
->name
);
2940 for_each_online_cpu(cpu
) {
2941 struct per_cpu_pageset
*pageset
;
2943 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2945 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2946 cpu
, pageset
->pcp
.high
,
2947 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2951 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2952 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2954 " dirty:%lu writeback:%lu unstable:%lu\n"
2955 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2956 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2958 global_page_state(NR_ACTIVE_ANON
),
2959 global_page_state(NR_INACTIVE_ANON
),
2960 global_page_state(NR_ISOLATED_ANON
),
2961 global_page_state(NR_ACTIVE_FILE
),
2962 global_page_state(NR_INACTIVE_FILE
),
2963 global_page_state(NR_ISOLATED_FILE
),
2964 global_page_state(NR_UNEVICTABLE
),
2965 global_page_state(NR_FILE_DIRTY
),
2966 global_page_state(NR_WRITEBACK
),
2967 global_page_state(NR_UNSTABLE_NFS
),
2968 global_page_state(NR_FREE_PAGES
),
2969 global_page_state(NR_SLAB_RECLAIMABLE
),
2970 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2971 global_page_state(NR_FILE_MAPPED
),
2972 global_page_state(NR_SHMEM
),
2973 global_page_state(NR_PAGETABLE
),
2974 global_page_state(NR_BOUNCE
),
2975 global_page_state(NR_FREE_CMA_PAGES
));
2977 for_each_populated_zone(zone
) {
2980 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2988 " active_anon:%lukB"
2989 " inactive_anon:%lukB"
2990 " active_file:%lukB"
2991 " inactive_file:%lukB"
2992 " unevictable:%lukB"
2993 " isolated(anon):%lukB"
2994 " isolated(file):%lukB"
3002 " slab_reclaimable:%lukB"
3003 " slab_unreclaimable:%lukB"
3004 " kernel_stack:%lukB"
3009 " writeback_tmp:%lukB"
3010 " pages_scanned:%lu"
3011 " all_unreclaimable? %s"
3014 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3015 K(min_wmark_pages(zone
)),
3016 K(low_wmark_pages(zone
)),
3017 K(high_wmark_pages(zone
)),
3018 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3019 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3020 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3021 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3022 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3023 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3024 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3025 K(zone
->present_pages
),
3026 K(zone
->managed_pages
),
3027 K(zone_page_state(zone
, NR_MLOCK
)),
3028 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3029 K(zone_page_state(zone
, NR_WRITEBACK
)),
3030 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3031 K(zone_page_state(zone
, NR_SHMEM
)),
3032 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3033 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3034 zone_page_state(zone
, NR_KERNEL_STACK
) *
3036 K(zone_page_state(zone
, NR_PAGETABLE
)),
3037 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3038 K(zone_page_state(zone
, NR_BOUNCE
)),
3039 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3040 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3041 zone
->pages_scanned
,
3042 (zone
->all_unreclaimable
? "yes" : "no")
3044 printk("lowmem_reserve[]:");
3045 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3046 printk(" %lu", zone
->lowmem_reserve
[i
]);
3050 for_each_populated_zone(zone
) {
3051 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3052 unsigned char types
[MAX_ORDER
];
3054 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3057 printk("%s: ", zone
->name
);
3059 spin_lock_irqsave(&zone
->lock
, flags
);
3060 for (order
= 0; order
< MAX_ORDER
; order
++) {
3061 struct free_area
*area
= &zone
->free_area
[order
];
3064 nr
[order
] = area
->nr_free
;
3065 total
+= nr
[order
] << order
;
3068 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3069 if (!list_empty(&area
->free_list
[type
]))
3070 types
[order
] |= 1 << type
;
3073 spin_unlock_irqrestore(&zone
->lock
, flags
);
3074 for (order
= 0; order
< MAX_ORDER
; order
++) {
3075 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3077 show_migration_types(types
[order
]);
3079 printk("= %lukB\n", K(total
));
3082 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3084 show_swap_cache_info();
3087 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3089 zoneref
->zone
= zone
;
3090 zoneref
->zone_idx
= zone_idx(zone
);
3094 * Builds allocation fallback zone lists.
3096 * Add all populated zones of a node to the zonelist.
3098 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3099 int nr_zones
, enum zone_type zone_type
)
3103 BUG_ON(zone_type
>= MAX_NR_ZONES
);
3108 zone
= pgdat
->node_zones
+ zone_type
;
3109 if (populated_zone(zone
)) {
3110 zoneref_set_zone(zone
,
3111 &zonelist
->_zonerefs
[nr_zones
++]);
3112 check_highest_zone(zone_type
);
3115 } while (zone_type
);
3122 * 0 = automatic detection of better ordering.
3123 * 1 = order by ([node] distance, -zonetype)
3124 * 2 = order by (-zonetype, [node] distance)
3126 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3127 * the same zonelist. So only NUMA can configure this param.
3129 #define ZONELIST_ORDER_DEFAULT 0
3130 #define ZONELIST_ORDER_NODE 1
3131 #define ZONELIST_ORDER_ZONE 2
3133 /* zonelist order in the kernel.
3134 * set_zonelist_order() will set this to NODE or ZONE.
3136 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3137 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3141 /* The value user specified ....changed by config */
3142 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3143 /* string for sysctl */
3144 #define NUMA_ZONELIST_ORDER_LEN 16
3145 char numa_zonelist_order
[16] = "default";
3148 * interface for configure zonelist ordering.
3149 * command line option "numa_zonelist_order"
3150 * = "[dD]efault - default, automatic configuration.
3151 * = "[nN]ode - order by node locality, then by zone within node
3152 * = "[zZ]one - order by zone, then by locality within zone
3155 static int __parse_numa_zonelist_order(char *s
)
3157 if (*s
== 'd' || *s
== 'D') {
3158 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3159 } else if (*s
== 'n' || *s
== 'N') {
3160 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3161 } else if (*s
== 'z' || *s
== 'Z') {
3162 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3165 "Ignoring invalid numa_zonelist_order value: "
3172 static __init
int setup_numa_zonelist_order(char *s
)
3179 ret
= __parse_numa_zonelist_order(s
);
3181 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3185 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3188 * sysctl handler for numa_zonelist_order
3190 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
3191 void __user
*buffer
, size_t *length
,
3194 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3196 static DEFINE_MUTEX(zl_order_mutex
);
3198 mutex_lock(&zl_order_mutex
);
3200 strcpy(saved_string
, (char*)table
->data
);
3201 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3205 int oldval
= user_zonelist_order
;
3206 if (__parse_numa_zonelist_order((char*)table
->data
)) {
3208 * bogus value. restore saved string
3210 strncpy((char*)table
->data
, saved_string
,
3211 NUMA_ZONELIST_ORDER_LEN
);
3212 user_zonelist_order
= oldval
;
3213 } else if (oldval
!= user_zonelist_order
) {
3214 mutex_lock(&zonelists_mutex
);
3215 build_all_zonelists(NULL
, NULL
);
3216 mutex_unlock(&zonelists_mutex
);
3220 mutex_unlock(&zl_order_mutex
);
3225 #define MAX_NODE_LOAD (nr_online_nodes)
3226 static int node_load
[MAX_NUMNODES
];
3229 * find_next_best_node - find the next node that should appear in a given node's fallback list
3230 * @node: node whose fallback list we're appending
3231 * @used_node_mask: nodemask_t of already used nodes
3233 * We use a number of factors to determine which is the next node that should
3234 * appear on a given node's fallback list. The node should not have appeared
3235 * already in @node's fallback list, and it should be the next closest node
3236 * according to the distance array (which contains arbitrary distance values
3237 * from each node to each node in the system), and should also prefer nodes
3238 * with no CPUs, since presumably they'll have very little allocation pressure
3239 * on them otherwise.
3240 * It returns -1 if no node is found.
3242 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3245 int min_val
= INT_MAX
;
3247 const struct cpumask
*tmp
= cpumask_of_node(0);
3249 /* Use the local node if we haven't already */
3250 if (!node_isset(node
, *used_node_mask
)) {
3251 node_set(node
, *used_node_mask
);
3255 for_each_node_state(n
, N_MEMORY
) {
3257 /* Don't want a node to appear more than once */
3258 if (node_isset(n
, *used_node_mask
))
3261 /* Use the distance array to find the distance */
3262 val
= node_distance(node
, n
);
3264 /* Penalize nodes under us ("prefer the next node") */
3267 /* Give preference to headless and unused nodes */
3268 tmp
= cpumask_of_node(n
);
3269 if (!cpumask_empty(tmp
))
3270 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3272 /* Slight preference for less loaded node */
3273 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3274 val
+= node_load
[n
];
3276 if (val
< min_val
) {
3283 node_set(best_node
, *used_node_mask
);
3290 * Build zonelists ordered by node and zones within node.
3291 * This results in maximum locality--normal zone overflows into local
3292 * DMA zone, if any--but risks exhausting DMA zone.
3294 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3297 struct zonelist
*zonelist
;
3299 zonelist
= &pgdat
->node_zonelists
[0];
3300 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3302 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3304 zonelist
->_zonerefs
[j
].zone
= NULL
;
3305 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3309 * Build gfp_thisnode zonelists
3311 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3314 struct zonelist
*zonelist
;
3316 zonelist
= &pgdat
->node_zonelists
[1];
3317 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3318 zonelist
->_zonerefs
[j
].zone
= NULL
;
3319 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3323 * Build zonelists ordered by zone and nodes within zones.
3324 * This results in conserving DMA zone[s] until all Normal memory is
3325 * exhausted, but results in overflowing to remote node while memory
3326 * may still exist in local DMA zone.
3328 static int node_order
[MAX_NUMNODES
];
3330 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3333 int zone_type
; /* needs to be signed */
3335 struct zonelist
*zonelist
;
3337 zonelist
= &pgdat
->node_zonelists
[0];
3339 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3340 for (j
= 0; j
< nr_nodes
; j
++) {
3341 node
= node_order
[j
];
3342 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3343 if (populated_zone(z
)) {
3345 &zonelist
->_zonerefs
[pos
++]);
3346 check_highest_zone(zone_type
);
3350 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3351 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3354 static int default_zonelist_order(void)
3357 unsigned long low_kmem_size
,total_size
;
3361 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3362 * If they are really small and used heavily, the system can fall
3363 * into OOM very easily.
3364 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3366 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3369 for_each_online_node(nid
) {
3370 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3371 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3372 if (populated_zone(z
)) {
3373 if (zone_type
< ZONE_NORMAL
)
3374 low_kmem_size
+= z
->present_pages
;
3375 total_size
+= z
->present_pages
;
3376 } else if (zone_type
== ZONE_NORMAL
) {
3378 * If any node has only lowmem, then node order
3379 * is preferred to allow kernel allocations
3380 * locally; otherwise, they can easily infringe
3381 * on other nodes when there is an abundance of
3382 * lowmem available to allocate from.
3384 return ZONELIST_ORDER_NODE
;
3388 if (!low_kmem_size
|| /* there are no DMA area. */
3389 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
3390 return ZONELIST_ORDER_NODE
;
3392 * look into each node's config.
3393 * If there is a node whose DMA/DMA32 memory is very big area on
3394 * local memory, NODE_ORDER may be suitable.
3396 average_size
= total_size
/
3397 (nodes_weight(node_states
[N_MEMORY
]) + 1);
3398 for_each_online_node(nid
) {
3401 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3402 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3403 if (populated_zone(z
)) {
3404 if (zone_type
< ZONE_NORMAL
)
3405 low_kmem_size
+= z
->present_pages
;
3406 total_size
+= z
->present_pages
;
3409 if (low_kmem_size
&&
3410 total_size
> average_size
&& /* ignore small node */
3411 low_kmem_size
> total_size
* 70/100)
3412 return ZONELIST_ORDER_NODE
;
3414 return ZONELIST_ORDER_ZONE
;
3417 static void set_zonelist_order(void)
3419 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3420 current_zonelist_order
= default_zonelist_order();
3422 current_zonelist_order
= user_zonelist_order
;
3425 static void build_zonelists(pg_data_t
*pgdat
)
3429 nodemask_t used_mask
;
3430 int local_node
, prev_node
;
3431 struct zonelist
*zonelist
;
3432 int order
= current_zonelist_order
;
3434 /* initialize zonelists */
3435 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3436 zonelist
= pgdat
->node_zonelists
+ i
;
3437 zonelist
->_zonerefs
[0].zone
= NULL
;
3438 zonelist
->_zonerefs
[0].zone_idx
= 0;
3441 /* NUMA-aware ordering of nodes */
3442 local_node
= pgdat
->node_id
;
3443 load
= nr_online_nodes
;
3444 prev_node
= local_node
;
3445 nodes_clear(used_mask
);
3447 memset(node_order
, 0, sizeof(node_order
));
3450 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3452 * We don't want to pressure a particular node.
3453 * So adding penalty to the first node in same
3454 * distance group to make it round-robin.
3456 if (node_distance(local_node
, node
) !=
3457 node_distance(local_node
, prev_node
))
3458 node_load
[node
] = load
;
3462 if (order
== ZONELIST_ORDER_NODE
)
3463 build_zonelists_in_node_order(pgdat
, node
);
3465 node_order
[j
++] = node
; /* remember order */
3468 if (order
== ZONELIST_ORDER_ZONE
) {
3469 /* calculate node order -- i.e., DMA last! */
3470 build_zonelists_in_zone_order(pgdat
, j
);
3473 build_thisnode_zonelists(pgdat
);
3476 /* Construct the zonelist performance cache - see further mmzone.h */
3477 static void build_zonelist_cache(pg_data_t
*pgdat
)
3479 struct zonelist
*zonelist
;
3480 struct zonelist_cache
*zlc
;
3483 zonelist
= &pgdat
->node_zonelists
[0];
3484 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3485 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3486 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3487 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3490 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3492 * Return node id of node used for "local" allocations.
3493 * I.e., first node id of first zone in arg node's generic zonelist.
3494 * Used for initializing percpu 'numa_mem', which is used primarily
3495 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3497 int local_memory_node(int node
)
3501 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3502 gfp_zone(GFP_KERNEL
),
3509 #else /* CONFIG_NUMA */
3511 static void set_zonelist_order(void)
3513 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3516 static void build_zonelists(pg_data_t
*pgdat
)
3518 int node
, local_node
;
3520 struct zonelist
*zonelist
;
3522 local_node
= pgdat
->node_id
;
3524 zonelist
= &pgdat
->node_zonelists
[0];
3525 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3528 * Now we build the zonelist so that it contains the zones
3529 * of all the other nodes.
3530 * We don't want to pressure a particular node, so when
3531 * building the zones for node N, we make sure that the
3532 * zones coming right after the local ones are those from
3533 * node N+1 (modulo N)
3535 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3536 if (!node_online(node
))
3538 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3541 for (node
= 0; node
< local_node
; node
++) {
3542 if (!node_online(node
))
3544 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3548 zonelist
->_zonerefs
[j
].zone
= NULL
;
3549 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3552 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3553 static void build_zonelist_cache(pg_data_t
*pgdat
)
3555 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3558 #endif /* CONFIG_NUMA */
3561 * Boot pageset table. One per cpu which is going to be used for all
3562 * zones and all nodes. The parameters will be set in such a way
3563 * that an item put on a list will immediately be handed over to
3564 * the buddy list. This is safe since pageset manipulation is done
3565 * with interrupts disabled.
3567 * The boot_pagesets must be kept even after bootup is complete for
3568 * unused processors and/or zones. They do play a role for bootstrapping
3569 * hotplugged processors.
3571 * zoneinfo_show() and maybe other functions do
3572 * not check if the processor is online before following the pageset pointer.
3573 * Other parts of the kernel may not check if the zone is available.
3575 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3576 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3577 static void setup_zone_pageset(struct zone
*zone
);
3580 * Global mutex to protect against size modification of zonelists
3581 * as well as to serialize pageset setup for the new populated zone.
3583 DEFINE_MUTEX(zonelists_mutex
);
3585 /* return values int ....just for stop_machine() */
3586 static int __build_all_zonelists(void *data
)
3590 pg_data_t
*self
= data
;
3593 memset(node_load
, 0, sizeof(node_load
));
3596 if (self
&& !node_online(self
->node_id
)) {
3597 build_zonelists(self
);
3598 build_zonelist_cache(self
);
3601 for_each_online_node(nid
) {
3602 pg_data_t
*pgdat
= NODE_DATA(nid
);
3604 build_zonelists(pgdat
);
3605 build_zonelist_cache(pgdat
);
3609 * Initialize the boot_pagesets that are going to be used
3610 * for bootstrapping processors. The real pagesets for
3611 * each zone will be allocated later when the per cpu
3612 * allocator is available.
3614 * boot_pagesets are used also for bootstrapping offline
3615 * cpus if the system is already booted because the pagesets
3616 * are needed to initialize allocators on a specific cpu too.
3617 * F.e. the percpu allocator needs the page allocator which
3618 * needs the percpu allocator in order to allocate its pagesets
3619 * (a chicken-egg dilemma).
3621 for_each_possible_cpu(cpu
) {
3622 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3624 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3626 * We now know the "local memory node" for each node--
3627 * i.e., the node of the first zone in the generic zonelist.
3628 * Set up numa_mem percpu variable for on-line cpus. During
3629 * boot, only the boot cpu should be on-line; we'll init the
3630 * secondary cpus' numa_mem as they come on-line. During
3631 * node/memory hotplug, we'll fixup all on-line cpus.
3633 if (cpu_online(cpu
))
3634 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3642 * Called with zonelists_mutex held always
3643 * unless system_state == SYSTEM_BOOTING.
3645 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3647 set_zonelist_order();
3649 if (system_state
== SYSTEM_BOOTING
) {
3650 __build_all_zonelists(NULL
);
3651 mminit_verify_zonelist();
3652 cpuset_init_current_mems_allowed();
3654 /* we have to stop all cpus to guarantee there is no user
3656 #ifdef CONFIG_MEMORY_HOTPLUG
3658 setup_zone_pageset(zone
);
3660 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3661 /* cpuset refresh routine should be here */
3663 vm_total_pages
= nr_free_pagecache_pages();
3665 * Disable grouping by mobility if the number of pages in the
3666 * system is too low to allow the mechanism to work. It would be
3667 * more accurate, but expensive to check per-zone. This check is
3668 * made on memory-hotadd so a system can start with mobility
3669 * disabled and enable it later
3671 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3672 page_group_by_mobility_disabled
= 1;
3674 page_group_by_mobility_disabled
= 0;
3676 printk("Built %i zonelists in %s order, mobility grouping %s. "
3677 "Total pages: %ld\n",
3679 zonelist_order_name
[current_zonelist_order
],
3680 page_group_by_mobility_disabled
? "off" : "on",
3683 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3688 * Helper functions to size the waitqueue hash table.
3689 * Essentially these want to choose hash table sizes sufficiently
3690 * large so that collisions trying to wait on pages are rare.
3691 * But in fact, the number of active page waitqueues on typical
3692 * systems is ridiculously low, less than 200. So this is even
3693 * conservative, even though it seems large.
3695 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3696 * waitqueues, i.e. the size of the waitq table given the number of pages.
3698 #define PAGES_PER_WAITQUEUE 256
3700 #ifndef CONFIG_MEMORY_HOTPLUG
3701 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3703 unsigned long size
= 1;
3705 pages
/= PAGES_PER_WAITQUEUE
;
3707 while (size
< pages
)
3711 * Once we have dozens or even hundreds of threads sleeping
3712 * on IO we've got bigger problems than wait queue collision.
3713 * Limit the size of the wait table to a reasonable size.
3715 size
= min(size
, 4096UL);
3717 return max(size
, 4UL);
3721 * A zone's size might be changed by hot-add, so it is not possible to determine
3722 * a suitable size for its wait_table. So we use the maximum size now.
3724 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3726 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3727 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3728 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3730 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3731 * or more by the traditional way. (See above). It equals:
3733 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3734 * ia64(16K page size) : = ( 8G + 4M)byte.
3735 * powerpc (64K page size) : = (32G +16M)byte.
3737 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3744 * This is an integer logarithm so that shifts can be used later
3745 * to extract the more random high bits from the multiplicative
3746 * hash function before the remainder is taken.
3748 static inline unsigned long wait_table_bits(unsigned long size
)
3753 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3756 * Check if a pageblock contains reserved pages
3758 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3762 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3763 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3770 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3771 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3772 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3773 * higher will lead to a bigger reserve which will get freed as contiguous
3774 * blocks as reclaim kicks in
3776 static void setup_zone_migrate_reserve(struct zone
*zone
)
3778 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3780 unsigned long block_migratetype
;
3784 * Get the start pfn, end pfn and the number of blocks to reserve
3785 * We have to be careful to be aligned to pageblock_nr_pages to
3786 * make sure that we always check pfn_valid for the first page in
3789 start_pfn
= zone
->zone_start_pfn
;
3790 end_pfn
= start_pfn
+ zone
->spanned_pages
;
3791 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
3792 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3796 * Reserve blocks are generally in place to help high-order atomic
3797 * allocations that are short-lived. A min_free_kbytes value that
3798 * would result in more than 2 reserve blocks for atomic allocations
3799 * is assumed to be in place to help anti-fragmentation for the
3800 * future allocation of hugepages at runtime.
3802 reserve
= min(2, reserve
);
3804 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3805 if (!pfn_valid(pfn
))
3807 page
= pfn_to_page(pfn
);
3809 /* Watch out for overlapping nodes */
3810 if (page_to_nid(page
) != zone_to_nid(zone
))
3813 block_migratetype
= get_pageblock_migratetype(page
);
3815 /* Only test what is necessary when the reserves are not met */
3818 * Blocks with reserved pages will never free, skip
3821 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
3822 if (pageblock_is_reserved(pfn
, block_end_pfn
))
3825 /* If this block is reserved, account for it */
3826 if (block_migratetype
== MIGRATE_RESERVE
) {
3831 /* Suitable for reserving if this block is movable */
3832 if (block_migratetype
== MIGRATE_MOVABLE
) {
3833 set_pageblock_migratetype(page
,
3835 move_freepages_block(zone
, page
,
3843 * If the reserve is met and this is a previous reserved block,
3846 if (block_migratetype
== MIGRATE_RESERVE
) {
3847 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3848 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3854 * Initially all pages are reserved - free ones are freed
3855 * up by free_all_bootmem() once the early boot process is
3856 * done. Non-atomic initialization, single-pass.
3858 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3859 unsigned long start_pfn
, enum memmap_context context
)
3862 unsigned long end_pfn
= start_pfn
+ size
;
3866 if (highest_memmap_pfn
< end_pfn
- 1)
3867 highest_memmap_pfn
= end_pfn
- 1;
3869 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3870 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3872 * There can be holes in boot-time mem_map[]s
3873 * handed to this function. They do not
3874 * exist on hotplugged memory.
3876 if (context
== MEMMAP_EARLY
) {
3877 if (!early_pfn_valid(pfn
))
3879 if (!early_pfn_in_nid(pfn
, nid
))
3882 page
= pfn_to_page(pfn
);
3883 set_page_links(page
, zone
, nid
, pfn
);
3884 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3885 init_page_count(page
);
3886 reset_page_mapcount(page
);
3887 reset_page_last_nid(page
);
3888 SetPageReserved(page
);
3890 * Mark the block movable so that blocks are reserved for
3891 * movable at startup. This will force kernel allocations
3892 * to reserve their blocks rather than leaking throughout
3893 * the address space during boot when many long-lived
3894 * kernel allocations are made. Later some blocks near
3895 * the start are marked MIGRATE_RESERVE by
3896 * setup_zone_migrate_reserve()
3898 * bitmap is created for zone's valid pfn range. but memmap
3899 * can be created for invalid pages (for alignment)
3900 * check here not to call set_pageblock_migratetype() against
3903 if ((z
->zone_start_pfn
<= pfn
)
3904 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
3905 && !(pfn
& (pageblock_nr_pages
- 1)))
3906 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3908 INIT_LIST_HEAD(&page
->lru
);
3909 #ifdef WANT_PAGE_VIRTUAL
3910 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3911 if (!is_highmem_idx(zone
))
3912 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3917 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3920 for_each_migratetype_order(order
, t
) {
3921 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3922 zone
->free_area
[order
].nr_free
= 0;
3926 #ifndef __HAVE_ARCH_MEMMAP_INIT
3927 #define memmap_init(size, nid, zone, start_pfn) \
3928 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3931 static int __meminit
zone_batchsize(struct zone
*zone
)
3937 * The per-cpu-pages pools are set to around 1000th of the
3938 * size of the zone. But no more than 1/2 of a meg.
3940 * OK, so we don't know how big the cache is. So guess.
3942 batch
= zone
->managed_pages
/ 1024;
3943 if (batch
* PAGE_SIZE
> 512 * 1024)
3944 batch
= (512 * 1024) / PAGE_SIZE
;
3945 batch
/= 4; /* We effectively *= 4 below */
3950 * Clamp the batch to a 2^n - 1 value. Having a power
3951 * of 2 value was found to be more likely to have
3952 * suboptimal cache aliasing properties in some cases.
3954 * For example if 2 tasks are alternately allocating
3955 * batches of pages, one task can end up with a lot
3956 * of pages of one half of the possible page colors
3957 * and the other with pages of the other colors.
3959 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3964 /* The deferral and batching of frees should be suppressed under NOMMU
3967 * The problem is that NOMMU needs to be able to allocate large chunks
3968 * of contiguous memory as there's no hardware page translation to
3969 * assemble apparent contiguous memory from discontiguous pages.
3971 * Queueing large contiguous runs of pages for batching, however,
3972 * causes the pages to actually be freed in smaller chunks. As there
3973 * can be a significant delay between the individual batches being
3974 * recycled, this leads to the once large chunks of space being
3975 * fragmented and becoming unavailable for high-order allocations.
3981 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3983 struct per_cpu_pages
*pcp
;
3986 memset(p
, 0, sizeof(*p
));
3990 pcp
->high
= 6 * batch
;
3991 pcp
->batch
= max(1UL, 1 * batch
);
3992 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
3993 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
3997 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3998 * to the value high for the pageset p.
4001 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
4004 struct per_cpu_pages
*pcp
;
4008 pcp
->batch
= max(1UL, high
/4);
4009 if ((high
/4) > (PAGE_SHIFT
* 8))
4010 pcp
->batch
= PAGE_SHIFT
* 8;
4013 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4017 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4019 for_each_possible_cpu(cpu
) {
4020 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4022 setup_pageset(pcp
, zone_batchsize(zone
));
4024 if (percpu_pagelist_fraction
)
4025 setup_pagelist_highmark(pcp
,
4026 (zone
->managed_pages
/
4027 percpu_pagelist_fraction
));
4032 * Allocate per cpu pagesets and initialize them.
4033 * Before this call only boot pagesets were available.
4035 void __init
setup_per_cpu_pageset(void)
4039 for_each_populated_zone(zone
)
4040 setup_zone_pageset(zone
);
4043 static noinline __init_refok
4044 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4047 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4051 * The per-page waitqueue mechanism uses hashed waitqueues
4054 zone
->wait_table_hash_nr_entries
=
4055 wait_table_hash_nr_entries(zone_size_pages
);
4056 zone
->wait_table_bits
=
4057 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4058 alloc_size
= zone
->wait_table_hash_nr_entries
4059 * sizeof(wait_queue_head_t
);
4061 if (!slab_is_available()) {
4062 zone
->wait_table
= (wait_queue_head_t
*)
4063 alloc_bootmem_node_nopanic(pgdat
, alloc_size
);
4066 * This case means that a zone whose size was 0 gets new memory
4067 * via memory hot-add.
4068 * But it may be the case that a new node was hot-added. In
4069 * this case vmalloc() will not be able to use this new node's
4070 * memory - this wait_table must be initialized to use this new
4071 * node itself as well.
4072 * To use this new node's memory, further consideration will be
4075 zone
->wait_table
= vmalloc(alloc_size
);
4077 if (!zone
->wait_table
)
4080 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4081 init_waitqueue_head(zone
->wait_table
+ i
);
4086 static __meminit
void zone_pcp_init(struct zone
*zone
)
4089 * per cpu subsystem is not up at this point. The following code
4090 * relies on the ability of the linker to provide the
4091 * offset of a (static) per cpu variable into the per cpu area.
4093 zone
->pageset
= &boot_pageset
;
4095 if (zone
->present_pages
)
4096 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4097 zone
->name
, zone
->present_pages
,
4098 zone_batchsize(zone
));
4101 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4102 unsigned long zone_start_pfn
,
4104 enum memmap_context context
)
4106 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4108 ret
= zone_wait_table_init(zone
, size
);
4111 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4113 zone
->zone_start_pfn
= zone_start_pfn
;
4115 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4116 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4118 (unsigned long)zone_idx(zone
),
4119 zone_start_pfn
, (zone_start_pfn
+ size
));
4121 zone_init_free_lists(zone
);
4126 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4127 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4129 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4130 * Architectures may implement their own version but if add_active_range()
4131 * was used and there are no special requirements, this is a convenient
4134 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4136 unsigned long start_pfn
, end_pfn
;
4139 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
4140 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
4142 /* This is a memory hole */
4145 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4147 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4151 nid
= __early_pfn_to_nid(pfn
);
4154 /* just returns 0 */
4158 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4159 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4163 nid
= __early_pfn_to_nid(pfn
);
4164 if (nid
>= 0 && nid
!= node
)
4171 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4172 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4173 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4175 * If an architecture guarantees that all ranges registered with
4176 * add_active_ranges() contain no holes and may be freed, this
4177 * this function may be used instead of calling free_bootmem() manually.
4179 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4181 unsigned long start_pfn
, end_pfn
;
4184 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4185 start_pfn
= min(start_pfn
, max_low_pfn
);
4186 end_pfn
= min(end_pfn
, max_low_pfn
);
4188 if (start_pfn
< end_pfn
)
4189 free_bootmem_node(NODE_DATA(this_nid
),
4190 PFN_PHYS(start_pfn
),
4191 (end_pfn
- start_pfn
) << PAGE_SHIFT
);
4196 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4197 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4199 * If an architecture guarantees that all ranges registered with
4200 * add_active_ranges() contain no holes and may be freed, this
4201 * function may be used instead of calling memory_present() manually.
4203 void __init
sparse_memory_present_with_active_regions(int nid
)
4205 unsigned long start_pfn
, end_pfn
;
4208 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4209 memory_present(this_nid
, start_pfn
, end_pfn
);
4213 * get_pfn_range_for_nid - Return the start and end page frames for a node
4214 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4215 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4216 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4218 * It returns the start and end page frame of a node based on information
4219 * provided by an arch calling add_active_range(). If called for a node
4220 * with no available memory, a warning is printed and the start and end
4223 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4224 unsigned long *start_pfn
, unsigned long *end_pfn
)
4226 unsigned long this_start_pfn
, this_end_pfn
;
4232 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4233 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4234 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4237 if (*start_pfn
== -1UL)
4242 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4243 * assumption is made that zones within a node are ordered in monotonic
4244 * increasing memory addresses so that the "highest" populated zone is used
4246 static void __init
find_usable_zone_for_movable(void)
4249 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4250 if (zone_index
== ZONE_MOVABLE
)
4253 if (arch_zone_highest_possible_pfn
[zone_index
] >
4254 arch_zone_lowest_possible_pfn
[zone_index
])
4258 VM_BUG_ON(zone_index
== -1);
4259 movable_zone
= zone_index
;
4263 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4264 * because it is sized independent of architecture. Unlike the other zones,
4265 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4266 * in each node depending on the size of each node and how evenly kernelcore
4267 * is distributed. This helper function adjusts the zone ranges
4268 * provided by the architecture for a given node by using the end of the
4269 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4270 * zones within a node are in order of monotonic increases memory addresses
4272 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4273 unsigned long zone_type
,
4274 unsigned long node_start_pfn
,
4275 unsigned long node_end_pfn
,
4276 unsigned long *zone_start_pfn
,
4277 unsigned long *zone_end_pfn
)
4279 /* Only adjust if ZONE_MOVABLE is on this node */
4280 if (zone_movable_pfn
[nid
]) {
4281 /* Size ZONE_MOVABLE */
4282 if (zone_type
== ZONE_MOVABLE
) {
4283 *zone_start_pfn
= zone_movable_pfn
[nid
];
4284 *zone_end_pfn
= min(node_end_pfn
,
4285 arch_zone_highest_possible_pfn
[movable_zone
]);
4287 /* Adjust for ZONE_MOVABLE starting within this range */
4288 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4289 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4290 *zone_end_pfn
= zone_movable_pfn
[nid
];
4292 /* Check if this whole range is within ZONE_MOVABLE */
4293 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4294 *zone_start_pfn
= *zone_end_pfn
;
4299 * Return the number of pages a zone spans in a node, including holes
4300 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4302 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4303 unsigned long zone_type
,
4304 unsigned long *ignored
)
4306 unsigned long node_start_pfn
, node_end_pfn
;
4307 unsigned long zone_start_pfn
, zone_end_pfn
;
4309 /* Get the start and end of the node and zone */
4310 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4311 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4312 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4313 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4314 node_start_pfn
, node_end_pfn
,
4315 &zone_start_pfn
, &zone_end_pfn
);
4317 /* Check that this node has pages within the zone's required range */
4318 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4321 /* Move the zone boundaries inside the node if necessary */
4322 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4323 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4325 /* Return the spanned pages */
4326 return zone_end_pfn
- zone_start_pfn
;
4330 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4331 * then all holes in the requested range will be accounted for.
4333 unsigned long __meminit
__absent_pages_in_range(int nid
,
4334 unsigned long range_start_pfn
,
4335 unsigned long range_end_pfn
)
4337 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4338 unsigned long start_pfn
, end_pfn
;
4341 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4342 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4343 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4344 nr_absent
-= end_pfn
- start_pfn
;
4350 * absent_pages_in_range - Return number of page frames in holes within a range
4351 * @start_pfn: The start PFN to start searching for holes
4352 * @end_pfn: The end PFN to stop searching for holes
4354 * It returns the number of pages frames in memory holes within a range.
4356 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4357 unsigned long end_pfn
)
4359 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4362 /* Return the number of page frames in holes in a zone on a node */
4363 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4364 unsigned long zone_type
,
4365 unsigned long *ignored
)
4367 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4368 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4369 unsigned long node_start_pfn
, node_end_pfn
;
4370 unsigned long zone_start_pfn
, zone_end_pfn
;
4372 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4373 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4374 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4376 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4377 node_start_pfn
, node_end_pfn
,
4378 &zone_start_pfn
, &zone_end_pfn
);
4379 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4383 * sanitize_zone_movable_limit - Sanitize the zone_movable_limit array.
4385 * zone_movable_limit is initialized as 0. This function will try to get
4386 * the first ZONE_MOVABLE pfn of each node from movablemem_map, and
4387 * assigne them to zone_movable_limit.
4388 * zone_movable_limit[nid] == 0 means no limit for the node.
4390 * Note: Each range is represented as [start_pfn, end_pfn)
4392 static void __meminit
sanitize_zone_movable_limit(void)
4394 int map_pos
= 0, i
, nid
;
4395 unsigned long start_pfn
, end_pfn
;
4397 if (!movablemem_map
.nr_map
)
4400 /* Iterate all ranges from minimum to maximum */
4401 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4403 * If we have found lowest pfn of ZONE_MOVABLE of the node
4404 * specified by user, just go on to check next range.
4406 if (zone_movable_limit
[nid
])
4409 #ifdef CONFIG_ZONE_DMA
4410 /* Skip DMA memory. */
4411 if (start_pfn
< arch_zone_highest_possible_pfn
[ZONE_DMA
])
4412 start_pfn
= arch_zone_highest_possible_pfn
[ZONE_DMA
];
4415 #ifdef CONFIG_ZONE_DMA32
4416 /* Skip DMA32 memory. */
4417 if (start_pfn
< arch_zone_highest_possible_pfn
[ZONE_DMA32
])
4418 start_pfn
= arch_zone_highest_possible_pfn
[ZONE_DMA32
];
4421 #ifdef CONFIG_HIGHMEM
4422 /* Skip lowmem if ZONE_MOVABLE is highmem. */
4423 if (zone_movable_is_highmem() &&
4424 start_pfn
< arch_zone_lowest_possible_pfn
[ZONE_HIGHMEM
])
4425 start_pfn
= arch_zone_lowest_possible_pfn
[ZONE_HIGHMEM
];
4428 if (start_pfn
>= end_pfn
)
4431 while (map_pos
< movablemem_map
.nr_map
) {
4432 if (end_pfn
<= movablemem_map
.map
[map_pos
].start_pfn
)
4435 if (start_pfn
>= movablemem_map
.map
[map_pos
].end_pfn
) {
4441 * The start_pfn of ZONE_MOVABLE is either the minimum
4442 * pfn specified by movablemem_map, or 0, which means
4443 * the node has no ZONE_MOVABLE.
4445 zone_movable_limit
[nid
] = max(start_pfn
,
4446 movablemem_map
.map
[map_pos
].start_pfn
);
4453 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4454 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4455 unsigned long zone_type
,
4456 unsigned long *zones_size
)
4458 return zones_size
[zone_type
];
4461 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4462 unsigned long zone_type
,
4463 unsigned long *zholes_size
)
4468 return zholes_size
[zone_type
];
4470 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4472 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4473 unsigned long *zones_size
, unsigned long *zholes_size
)
4475 unsigned long realtotalpages
, totalpages
= 0;
4478 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4479 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4481 pgdat
->node_spanned_pages
= totalpages
;
4483 realtotalpages
= totalpages
;
4484 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4486 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4488 pgdat
->node_present_pages
= realtotalpages
;
4489 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4493 #ifndef CONFIG_SPARSEMEM
4495 * Calculate the size of the zone->blockflags rounded to an unsigned long
4496 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4497 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4498 * round what is now in bits to nearest long in bits, then return it in
4501 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4503 unsigned long usemapsize
;
4505 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4506 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4507 usemapsize
= usemapsize
>> pageblock_order
;
4508 usemapsize
*= NR_PAGEBLOCK_BITS
;
4509 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4511 return usemapsize
/ 8;
4514 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4516 unsigned long zone_start_pfn
,
4517 unsigned long zonesize
)
4519 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4520 zone
->pageblock_flags
= NULL
;
4522 zone
->pageblock_flags
= alloc_bootmem_node_nopanic(pgdat
,
4526 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4527 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4528 #endif /* CONFIG_SPARSEMEM */
4530 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4532 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4533 void __init
set_pageblock_order(void)
4537 /* Check that pageblock_nr_pages has not already been setup */
4538 if (pageblock_order
)
4541 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4542 order
= HUGETLB_PAGE_ORDER
;
4544 order
= MAX_ORDER
- 1;
4547 * Assume the largest contiguous order of interest is a huge page.
4548 * This value may be variable depending on boot parameters on IA64 and
4551 pageblock_order
= order
;
4553 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4556 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4557 * is unused as pageblock_order is set at compile-time. See
4558 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4561 void __init
set_pageblock_order(void)
4565 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4567 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4568 unsigned long present_pages
)
4570 unsigned long pages
= spanned_pages
;
4573 * Provide a more accurate estimation if there are holes within
4574 * the zone and SPARSEMEM is in use. If there are holes within the
4575 * zone, each populated memory region may cost us one or two extra
4576 * memmap pages due to alignment because memmap pages for each
4577 * populated regions may not naturally algined on page boundary.
4578 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4580 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4581 IS_ENABLED(CONFIG_SPARSEMEM
))
4582 pages
= present_pages
;
4584 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4588 * Set up the zone data structures:
4589 * - mark all pages reserved
4590 * - mark all memory queues empty
4591 * - clear the memory bitmaps
4593 * NOTE: pgdat should get zeroed by caller.
4595 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4596 unsigned long *zones_size
, unsigned long *zholes_size
)
4599 int nid
= pgdat
->node_id
;
4600 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4603 pgdat_resize_init(pgdat
);
4604 #ifdef CONFIG_NUMA_BALANCING
4605 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4606 pgdat
->numabalancing_migrate_nr_pages
= 0;
4607 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4609 init_waitqueue_head(&pgdat
->kswapd_wait
);
4610 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4611 pgdat_page_cgroup_init(pgdat
);
4613 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4614 struct zone
*zone
= pgdat
->node_zones
+ j
;
4615 unsigned long size
, realsize
, freesize
, memmap_pages
;
4617 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4618 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4622 * Adjust freesize so that it accounts for how much memory
4623 * is used by this zone for memmap. This affects the watermark
4624 * and per-cpu initialisations
4626 memmap_pages
= calc_memmap_size(size
, realsize
);
4627 if (freesize
>= memmap_pages
) {
4628 freesize
-= memmap_pages
;
4631 " %s zone: %lu pages used for memmap\n",
4632 zone_names
[j
], memmap_pages
);
4635 " %s zone: %lu pages exceeds freesize %lu\n",
4636 zone_names
[j
], memmap_pages
, freesize
);
4638 /* Account for reserved pages */
4639 if (j
== 0 && freesize
> dma_reserve
) {
4640 freesize
-= dma_reserve
;
4641 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4642 zone_names
[0], dma_reserve
);
4645 if (!is_highmem_idx(j
))
4646 nr_kernel_pages
+= freesize
;
4647 /* Charge for highmem memmap if there are enough kernel pages */
4648 else if (nr_kernel_pages
> memmap_pages
* 2)
4649 nr_kernel_pages
-= memmap_pages
;
4650 nr_all_pages
+= freesize
;
4652 zone
->spanned_pages
= size
;
4653 zone
->present_pages
= freesize
;
4655 * Set an approximate value for lowmem here, it will be adjusted
4656 * when the bootmem allocator frees pages into the buddy system.
4657 * And all highmem pages will be managed by the buddy system.
4659 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4662 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4664 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4666 zone
->name
= zone_names
[j
];
4667 spin_lock_init(&zone
->lock
);
4668 spin_lock_init(&zone
->lru_lock
);
4669 zone_seqlock_init(zone
);
4670 zone
->zone_pgdat
= pgdat
;
4672 zone_pcp_init(zone
);
4673 lruvec_init(&zone
->lruvec
);
4677 set_pageblock_order();
4678 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4679 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4680 size
, MEMMAP_EARLY
);
4682 memmap_init(size
, nid
, j
, zone_start_pfn
);
4683 zone_start_pfn
+= size
;
4687 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4689 /* Skip empty nodes */
4690 if (!pgdat
->node_spanned_pages
)
4693 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4694 /* ia64 gets its own node_mem_map, before this, without bootmem */
4695 if (!pgdat
->node_mem_map
) {
4696 unsigned long size
, start
, end
;
4700 * The zone's endpoints aren't required to be MAX_ORDER
4701 * aligned but the node_mem_map endpoints must be in order
4702 * for the buddy allocator to function correctly.
4704 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4705 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
4706 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4707 size
= (end
- start
) * sizeof(struct page
);
4708 map
= alloc_remap(pgdat
->node_id
, size
);
4710 map
= alloc_bootmem_node_nopanic(pgdat
, size
);
4711 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4713 #ifndef CONFIG_NEED_MULTIPLE_NODES
4715 * With no DISCONTIG, the global mem_map is just set as node 0's
4717 if (pgdat
== NODE_DATA(0)) {
4718 mem_map
= NODE_DATA(0)->node_mem_map
;
4719 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4720 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4721 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4722 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4725 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4728 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4729 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4731 pg_data_t
*pgdat
= NODE_DATA(nid
);
4733 /* pg_data_t should be reset to zero when it's allocated */
4734 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4736 pgdat
->node_id
= nid
;
4737 pgdat
->node_start_pfn
= node_start_pfn
;
4738 init_zone_allows_reclaim(nid
);
4739 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4741 alloc_node_mem_map(pgdat
);
4742 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4743 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4744 nid
, (unsigned long)pgdat
,
4745 (unsigned long)pgdat
->node_mem_map
);
4748 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4751 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4753 #if MAX_NUMNODES > 1
4755 * Figure out the number of possible node ids.
4757 static void __init
setup_nr_node_ids(void)
4760 unsigned int highest
= 0;
4762 for_each_node_mask(node
, node_possible_map
)
4764 nr_node_ids
= highest
+ 1;
4767 static inline void setup_nr_node_ids(void)
4773 * node_map_pfn_alignment - determine the maximum internode alignment
4775 * This function should be called after node map is populated and sorted.
4776 * It calculates the maximum power of two alignment which can distinguish
4779 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4780 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4781 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4782 * shifted, 1GiB is enough and this function will indicate so.
4784 * This is used to test whether pfn -> nid mapping of the chosen memory
4785 * model has fine enough granularity to avoid incorrect mapping for the
4786 * populated node map.
4788 * Returns the determined alignment in pfn's. 0 if there is no alignment
4789 * requirement (single node).
4791 unsigned long __init
node_map_pfn_alignment(void)
4793 unsigned long accl_mask
= 0, last_end
= 0;
4794 unsigned long start
, end
, mask
;
4798 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
4799 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
4806 * Start with a mask granular enough to pin-point to the
4807 * start pfn and tick off bits one-by-one until it becomes
4808 * too coarse to separate the current node from the last.
4810 mask
= ~((1 << __ffs(start
)) - 1);
4811 while (mask
&& last_end
<= (start
& (mask
<< 1)))
4814 /* accumulate all internode masks */
4818 /* convert mask to number of pages */
4819 return ~accl_mask
+ 1;
4822 /* Find the lowest pfn for a node */
4823 static unsigned long __init
find_min_pfn_for_node(int nid
)
4825 unsigned long min_pfn
= ULONG_MAX
;
4826 unsigned long start_pfn
;
4829 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
4830 min_pfn
= min(min_pfn
, start_pfn
);
4832 if (min_pfn
== ULONG_MAX
) {
4834 "Could not find start_pfn for node %d\n", nid
);
4842 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4844 * It returns the minimum PFN based on information provided via
4845 * add_active_range().
4847 unsigned long __init
find_min_pfn_with_active_regions(void)
4849 return find_min_pfn_for_node(MAX_NUMNODES
);
4853 * early_calculate_totalpages()
4854 * Sum pages in active regions for movable zone.
4855 * Populate N_MEMORY for calculating usable_nodes.
4857 static unsigned long __init
early_calculate_totalpages(void)
4859 unsigned long totalpages
= 0;
4860 unsigned long start_pfn
, end_pfn
;
4863 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4864 unsigned long pages
= end_pfn
- start_pfn
;
4866 totalpages
+= pages
;
4868 node_set_state(nid
, N_MEMORY
);
4874 * Find the PFN the Movable zone begins in each node. Kernel memory
4875 * is spread evenly between nodes as long as the nodes have enough
4876 * memory. When they don't, some nodes will have more kernelcore than
4879 static void __init
find_zone_movable_pfns_for_nodes(void)
4882 unsigned long usable_startpfn
;
4883 unsigned long kernelcore_node
, kernelcore_remaining
;
4884 /* save the state before borrow the nodemask */
4885 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
4886 unsigned long totalpages
= early_calculate_totalpages();
4887 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
4890 * If movablecore was specified, calculate what size of
4891 * kernelcore that corresponds so that memory usable for
4892 * any allocation type is evenly spread. If both kernelcore
4893 * and movablecore are specified, then the value of kernelcore
4894 * will be used for required_kernelcore if it's greater than
4895 * what movablecore would have allowed.
4897 if (required_movablecore
) {
4898 unsigned long corepages
;
4901 * Round-up so that ZONE_MOVABLE is at least as large as what
4902 * was requested by the user
4904 required_movablecore
=
4905 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4906 corepages
= totalpages
- required_movablecore
;
4908 required_kernelcore
= max(required_kernelcore
, corepages
);
4912 * If neither kernelcore/movablecore nor movablemem_map is specified,
4913 * there is no ZONE_MOVABLE. But if movablemem_map is specified, the
4914 * start pfn of ZONE_MOVABLE has been stored in zone_movable_limit[].
4916 if (!required_kernelcore
) {
4917 if (movablemem_map
.nr_map
)
4918 memcpy(zone_movable_pfn
, zone_movable_limit
,
4919 sizeof(zone_movable_pfn
));
4923 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4924 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4927 /* Spread kernelcore memory as evenly as possible throughout nodes */
4928 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4929 for_each_node_state(nid
, N_MEMORY
) {
4930 unsigned long start_pfn
, end_pfn
;
4933 * Recalculate kernelcore_node if the division per node
4934 * now exceeds what is necessary to satisfy the requested
4935 * amount of memory for the kernel
4937 if (required_kernelcore
< kernelcore_node
)
4938 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4941 * As the map is walked, we track how much memory is usable
4942 * by the kernel using kernelcore_remaining. When it is
4943 * 0, the rest of the node is usable by ZONE_MOVABLE
4945 kernelcore_remaining
= kernelcore_node
;
4947 /* Go through each range of PFNs within this node */
4948 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4949 unsigned long size_pages
;
4952 * Find more memory for kernelcore in
4953 * [zone_movable_pfn[nid], zone_movable_limit[nid]).
4955 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
4956 if (start_pfn
>= end_pfn
)
4959 if (zone_movable_limit
[nid
]) {
4960 end_pfn
= min(end_pfn
, zone_movable_limit
[nid
]);
4961 /* No range left for kernelcore in this node */
4962 if (start_pfn
>= end_pfn
) {
4963 zone_movable_pfn
[nid
] =
4964 zone_movable_limit
[nid
];
4969 /* Account for what is only usable for kernelcore */
4970 if (start_pfn
< usable_startpfn
) {
4971 unsigned long kernel_pages
;
4972 kernel_pages
= min(end_pfn
, usable_startpfn
)
4975 kernelcore_remaining
-= min(kernel_pages
,
4976 kernelcore_remaining
);
4977 required_kernelcore
-= min(kernel_pages
,
4978 required_kernelcore
);
4980 /* Continue if range is now fully accounted */
4981 if (end_pfn
<= usable_startpfn
) {
4984 * Push zone_movable_pfn to the end so
4985 * that if we have to rebalance
4986 * kernelcore across nodes, we will
4987 * not double account here
4989 zone_movable_pfn
[nid
] = end_pfn
;
4992 start_pfn
= usable_startpfn
;
4996 * The usable PFN range for ZONE_MOVABLE is from
4997 * start_pfn->end_pfn. Calculate size_pages as the
4998 * number of pages used as kernelcore
5000 size_pages
= end_pfn
- start_pfn
;
5001 if (size_pages
> kernelcore_remaining
)
5002 size_pages
= kernelcore_remaining
;
5003 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5006 * Some kernelcore has been met, update counts and
5007 * break if the kernelcore for this node has been
5010 required_kernelcore
-= min(required_kernelcore
,
5012 kernelcore_remaining
-= size_pages
;
5013 if (!kernelcore_remaining
)
5019 * If there is still required_kernelcore, we do another pass with one
5020 * less node in the count. This will push zone_movable_pfn[nid] further
5021 * along on the nodes that still have memory until kernelcore is
5025 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5029 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5030 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5031 zone_movable_pfn
[nid
] =
5032 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5034 /* restore the node_state */
5035 node_states
[N_MEMORY
] = saved_node_state
;
5038 /* Any regular or high memory on that node ? */
5039 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5041 enum zone_type zone_type
;
5043 if (N_MEMORY
== N_NORMAL_MEMORY
)
5046 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5047 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5048 if (zone
->present_pages
) {
5049 node_set_state(nid
, N_HIGH_MEMORY
);
5050 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5051 zone_type
<= ZONE_NORMAL
)
5052 node_set_state(nid
, N_NORMAL_MEMORY
);
5059 * free_area_init_nodes - Initialise all pg_data_t and zone data
5060 * @max_zone_pfn: an array of max PFNs for each zone
5062 * This will call free_area_init_node() for each active node in the system.
5063 * Using the page ranges provided by add_active_range(), the size of each
5064 * zone in each node and their holes is calculated. If the maximum PFN
5065 * between two adjacent zones match, it is assumed that the zone is empty.
5066 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5067 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5068 * starts where the previous one ended. For example, ZONE_DMA32 starts
5069 * at arch_max_dma_pfn.
5071 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5073 unsigned long start_pfn
, end_pfn
;
5076 /* Record where the zone boundaries are */
5077 memset(arch_zone_lowest_possible_pfn
, 0,
5078 sizeof(arch_zone_lowest_possible_pfn
));
5079 memset(arch_zone_highest_possible_pfn
, 0,
5080 sizeof(arch_zone_highest_possible_pfn
));
5081 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5082 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5083 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5084 if (i
== ZONE_MOVABLE
)
5086 arch_zone_lowest_possible_pfn
[i
] =
5087 arch_zone_highest_possible_pfn
[i
-1];
5088 arch_zone_highest_possible_pfn
[i
] =
5089 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5091 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5092 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5094 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5095 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5096 find_usable_zone_for_movable();
5097 sanitize_zone_movable_limit();
5098 find_zone_movable_pfns_for_nodes();
5100 /* Print out the zone ranges */
5101 printk("Zone ranges:\n");
5102 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5103 if (i
== ZONE_MOVABLE
)
5105 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5106 if (arch_zone_lowest_possible_pfn
[i
] ==
5107 arch_zone_highest_possible_pfn
[i
])
5108 printk(KERN_CONT
"empty\n");
5110 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5111 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5112 (arch_zone_highest_possible_pfn
[i
]
5113 << PAGE_SHIFT
) - 1);
5116 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5117 printk("Movable zone start for each node\n");
5118 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5119 if (zone_movable_pfn
[i
])
5120 printk(" Node %d: %#010lx\n", i
,
5121 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5124 /* Print out the early node map */
5125 printk("Early memory node ranges\n");
5126 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5127 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5128 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5130 /* Initialise every node */
5131 mminit_verify_pageflags_layout();
5132 setup_nr_node_ids();
5133 for_each_online_node(nid
) {
5134 pg_data_t
*pgdat
= NODE_DATA(nid
);
5135 free_area_init_node(nid
, NULL
,
5136 find_min_pfn_for_node(nid
), NULL
);
5138 /* Any memory on that node */
5139 if (pgdat
->node_present_pages
)
5140 node_set_state(nid
, N_MEMORY
);
5141 check_for_memory(pgdat
, nid
);
5145 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5147 unsigned long long coremem
;
5151 coremem
= memparse(p
, &p
);
5152 *core
= coremem
>> PAGE_SHIFT
;
5154 /* Paranoid check that UL is enough for the coremem value */
5155 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5161 * kernelcore=size sets the amount of memory for use for allocations that
5162 * cannot be reclaimed or migrated.
5164 static int __init
cmdline_parse_kernelcore(char *p
)
5166 return cmdline_parse_core(p
, &required_kernelcore
);
5170 * movablecore=size sets the amount of memory for use for allocations that
5171 * can be reclaimed or migrated.
5173 static int __init
cmdline_parse_movablecore(char *p
)
5175 return cmdline_parse_core(p
, &required_movablecore
);
5178 early_param("kernelcore", cmdline_parse_kernelcore
);
5179 early_param("movablecore", cmdline_parse_movablecore
);
5182 * movablemem_map_overlap() - Check if a range overlaps movablemem_map.map[].
5183 * @start_pfn: start pfn of the range to be checked
5184 * @end_pfn: end pfn of the range to be checked (exclusive)
5186 * This function checks if a given memory range [start_pfn, end_pfn) overlaps
5187 * the movablemem_map.map[] array.
5189 * Return: index of the first overlapped element in movablemem_map.map[]
5190 * or -1 if they don't overlap each other.
5192 int __init
movablemem_map_overlap(unsigned long start_pfn
,
5193 unsigned long end_pfn
)
5197 if (!movablemem_map
.nr_map
)
5200 for (overlap
= 0; overlap
< movablemem_map
.nr_map
; overlap
++)
5201 if (start_pfn
< movablemem_map
.map
[overlap
].end_pfn
)
5204 if (overlap
== movablemem_map
.nr_map
||
5205 end_pfn
<= movablemem_map
.map
[overlap
].start_pfn
)
5212 * insert_movablemem_map - Insert a memory range in to movablemem_map.map.
5213 * @start_pfn: start pfn of the range
5214 * @end_pfn: end pfn of the range
5216 * This function will also merge the overlapped ranges, and sort the array
5217 * by start_pfn in monotonic increasing order.
5219 void __init
insert_movablemem_map(unsigned long start_pfn
,
5220 unsigned long end_pfn
)
5225 * pos will be at the 1st overlapped range, or the position
5226 * where the element should be inserted.
5228 for (pos
= 0; pos
< movablemem_map
.nr_map
; pos
++)
5229 if (start_pfn
<= movablemem_map
.map
[pos
].end_pfn
)
5232 /* If there is no overlapped range, just insert the element. */
5233 if (pos
== movablemem_map
.nr_map
||
5234 end_pfn
< movablemem_map
.map
[pos
].start_pfn
) {
5236 * If pos is not the end of array, we need to move all
5237 * the rest elements backward.
5239 if (pos
< movablemem_map
.nr_map
)
5240 memmove(&movablemem_map
.map
[pos
+1],
5241 &movablemem_map
.map
[pos
],
5242 sizeof(struct movablemem_entry
) *
5243 (movablemem_map
.nr_map
- pos
));
5244 movablemem_map
.map
[pos
].start_pfn
= start_pfn
;
5245 movablemem_map
.map
[pos
].end_pfn
= end_pfn
;
5246 movablemem_map
.nr_map
++;
5250 /* overlap will be at the last overlapped range */
5251 for (overlap
= pos
+ 1; overlap
< movablemem_map
.nr_map
; overlap
++)
5252 if (end_pfn
< movablemem_map
.map
[overlap
].start_pfn
)
5256 * If there are more ranges overlapped, we need to merge them,
5257 * and move the rest elements forward.
5260 movablemem_map
.map
[pos
].start_pfn
= min(start_pfn
,
5261 movablemem_map
.map
[pos
].start_pfn
);
5262 movablemem_map
.map
[pos
].end_pfn
= max(end_pfn
,
5263 movablemem_map
.map
[overlap
].end_pfn
);
5265 if (pos
!= overlap
&& overlap
+ 1 != movablemem_map
.nr_map
)
5266 memmove(&movablemem_map
.map
[pos
+1],
5267 &movablemem_map
.map
[overlap
+1],
5268 sizeof(struct movablemem_entry
) *
5269 (movablemem_map
.nr_map
- overlap
- 1));
5271 movablemem_map
.nr_map
-= overlap
- pos
;
5275 * movablemem_map_add_region - Add a memory range into movablemem_map.
5276 * @start: physical start address of range
5277 * @end: physical end address of range
5279 * This function transform the physical address into pfn, and then add the
5280 * range into movablemem_map by calling insert_movablemem_map().
5282 static void __init
movablemem_map_add_region(u64 start
, u64 size
)
5284 unsigned long start_pfn
, end_pfn
;
5286 /* In case size == 0 or start + size overflows */
5287 if (start
+ size
<= start
)
5290 if (movablemem_map
.nr_map
>= ARRAY_SIZE(movablemem_map
.map
)) {
5291 pr_err("movablemem_map: too many entries;"
5292 " ignoring [mem %#010llx-%#010llx]\n",
5293 (unsigned long long) start
,
5294 (unsigned long long) (start
+ size
- 1));
5298 start_pfn
= PFN_DOWN(start
);
5299 end_pfn
= PFN_UP(start
+ size
);
5300 insert_movablemem_map(start_pfn
, end_pfn
);
5304 * cmdline_parse_movablemem_map - Parse boot option movablemem_map.
5305 * @p: The boot option of the following format:
5306 * movablemem_map=nn[KMG]@ss[KMG]
5308 * This option sets the memory range [ss, ss+nn) to be used as movable memory.
5310 * Return: 0 on success or -EINVAL on failure.
5312 static int __init
cmdline_parse_movablemem_map(char *p
)
5315 u64 start_at
, mem_size
;
5320 if (!strcmp(p
, "acpi"))
5321 movablemem_map
.acpi
= true;
5324 * If user decide to use info from BIOS, all the other user specified
5325 * ranges will be ingored.
5327 if (movablemem_map
.acpi
) {
5328 if (movablemem_map
.nr_map
) {
5329 memset(movablemem_map
.map
, 0,
5330 sizeof(struct movablemem_entry
)
5331 * movablemem_map
.nr_map
);
5332 movablemem_map
.nr_map
= 0;
5338 mem_size
= memparse(p
, &p
);
5344 start_at
= memparse(p
, &p
);
5345 if (p
== oldp
|| *p
!= '\0')
5348 movablemem_map_add_region(start_at
, mem_size
);
5354 early_param("movablemem_map", cmdline_parse_movablemem_map
);
5356 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5359 * set_dma_reserve - set the specified number of pages reserved in the first zone
5360 * @new_dma_reserve: The number of pages to mark reserved
5362 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5363 * In the DMA zone, a significant percentage may be consumed by kernel image
5364 * and other unfreeable allocations which can skew the watermarks badly. This
5365 * function may optionally be used to account for unfreeable pages in the
5366 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5367 * smaller per-cpu batchsize.
5369 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5371 dma_reserve
= new_dma_reserve
;
5374 void __init
free_area_init(unsigned long *zones_size
)
5376 free_area_init_node(0, zones_size
,
5377 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5380 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5381 unsigned long action
, void *hcpu
)
5383 int cpu
= (unsigned long)hcpu
;
5385 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5386 lru_add_drain_cpu(cpu
);
5390 * Spill the event counters of the dead processor
5391 * into the current processors event counters.
5392 * This artificially elevates the count of the current
5395 vm_events_fold_cpu(cpu
);
5398 * Zero the differential counters of the dead processor
5399 * so that the vm statistics are consistent.
5401 * This is only okay since the processor is dead and cannot
5402 * race with what we are doing.
5404 refresh_cpu_vm_stats(cpu
);
5409 void __init
page_alloc_init(void)
5411 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5415 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5416 * or min_free_kbytes changes.
5418 static void calculate_totalreserve_pages(void)
5420 struct pglist_data
*pgdat
;
5421 unsigned long reserve_pages
= 0;
5422 enum zone_type i
, j
;
5424 for_each_online_pgdat(pgdat
) {
5425 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5426 struct zone
*zone
= pgdat
->node_zones
+ i
;
5427 unsigned long max
= 0;
5429 /* Find valid and maximum lowmem_reserve in the zone */
5430 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5431 if (zone
->lowmem_reserve
[j
] > max
)
5432 max
= zone
->lowmem_reserve
[j
];
5435 /* we treat the high watermark as reserved pages. */
5436 max
+= high_wmark_pages(zone
);
5438 if (max
> zone
->managed_pages
)
5439 max
= zone
->managed_pages
;
5440 reserve_pages
+= max
;
5442 * Lowmem reserves are not available to
5443 * GFP_HIGHUSER page cache allocations and
5444 * kswapd tries to balance zones to their high
5445 * watermark. As a result, neither should be
5446 * regarded as dirtyable memory, to prevent a
5447 * situation where reclaim has to clean pages
5448 * in order to balance the zones.
5450 zone
->dirty_balance_reserve
= max
;
5453 dirty_balance_reserve
= reserve_pages
;
5454 totalreserve_pages
= reserve_pages
;
5458 * setup_per_zone_lowmem_reserve - called whenever
5459 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5460 * has a correct pages reserved value, so an adequate number of
5461 * pages are left in the zone after a successful __alloc_pages().
5463 static void setup_per_zone_lowmem_reserve(void)
5465 struct pglist_data
*pgdat
;
5466 enum zone_type j
, idx
;
5468 for_each_online_pgdat(pgdat
) {
5469 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5470 struct zone
*zone
= pgdat
->node_zones
+ j
;
5471 unsigned long managed_pages
= zone
->managed_pages
;
5473 zone
->lowmem_reserve
[j
] = 0;
5477 struct zone
*lower_zone
;
5481 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5482 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5484 lower_zone
= pgdat
->node_zones
+ idx
;
5485 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5486 sysctl_lowmem_reserve_ratio
[idx
];
5487 managed_pages
+= lower_zone
->managed_pages
;
5492 /* update totalreserve_pages */
5493 calculate_totalreserve_pages();
5496 static void __setup_per_zone_wmarks(void)
5498 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5499 unsigned long lowmem_pages
= 0;
5501 unsigned long flags
;
5503 /* Calculate total number of !ZONE_HIGHMEM pages */
5504 for_each_zone(zone
) {
5505 if (!is_highmem(zone
))
5506 lowmem_pages
+= zone
->managed_pages
;
5509 for_each_zone(zone
) {
5512 spin_lock_irqsave(&zone
->lock
, flags
);
5513 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5514 do_div(tmp
, lowmem_pages
);
5515 if (is_highmem(zone
)) {
5517 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5518 * need highmem pages, so cap pages_min to a small
5521 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5522 * deltas controls asynch page reclaim, and so should
5523 * not be capped for highmem.
5525 unsigned long min_pages
;
5527 min_pages
= zone
->managed_pages
/ 1024;
5528 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5529 zone
->watermark
[WMARK_MIN
] = min_pages
;
5532 * If it's a lowmem zone, reserve a number of pages
5533 * proportionate to the zone's size.
5535 zone
->watermark
[WMARK_MIN
] = tmp
;
5538 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5539 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5541 setup_zone_migrate_reserve(zone
);
5542 spin_unlock_irqrestore(&zone
->lock
, flags
);
5545 /* update totalreserve_pages */
5546 calculate_totalreserve_pages();
5550 * setup_per_zone_wmarks - called when min_free_kbytes changes
5551 * or when memory is hot-{added|removed}
5553 * Ensures that the watermark[min,low,high] values for each zone are set
5554 * correctly with respect to min_free_kbytes.
5556 void setup_per_zone_wmarks(void)
5558 mutex_lock(&zonelists_mutex
);
5559 __setup_per_zone_wmarks();
5560 mutex_unlock(&zonelists_mutex
);
5564 * The inactive anon list should be small enough that the VM never has to
5565 * do too much work, but large enough that each inactive page has a chance
5566 * to be referenced again before it is swapped out.
5568 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5569 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5570 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5571 * the anonymous pages are kept on the inactive list.
5574 * memory ratio inactive anon
5575 * -------------------------------------
5584 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5586 unsigned int gb
, ratio
;
5588 /* Zone size in gigabytes */
5589 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5591 ratio
= int_sqrt(10 * gb
);
5595 zone
->inactive_ratio
= ratio
;
5598 static void __meminit
setup_per_zone_inactive_ratio(void)
5603 calculate_zone_inactive_ratio(zone
);
5607 * Initialise min_free_kbytes.
5609 * For small machines we want it small (128k min). For large machines
5610 * we want it large (64MB max). But it is not linear, because network
5611 * bandwidth does not increase linearly with machine size. We use
5613 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5614 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5630 int __meminit
init_per_zone_wmark_min(void)
5632 unsigned long lowmem_kbytes
;
5634 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5636 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5637 if (min_free_kbytes
< 128)
5638 min_free_kbytes
= 128;
5639 if (min_free_kbytes
> 65536)
5640 min_free_kbytes
= 65536;
5641 setup_per_zone_wmarks();
5642 refresh_zone_stat_thresholds();
5643 setup_per_zone_lowmem_reserve();
5644 setup_per_zone_inactive_ratio();
5647 module_init(init_per_zone_wmark_min
)
5650 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5651 * that we can call two helper functions whenever min_free_kbytes
5654 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5655 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5657 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5659 setup_per_zone_wmarks();
5664 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5665 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5670 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5675 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5676 sysctl_min_unmapped_ratio
) / 100;
5680 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5681 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5686 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5691 zone
->min_slab_pages
= (zone
->managed_pages
*
5692 sysctl_min_slab_ratio
) / 100;
5698 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5699 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5700 * whenever sysctl_lowmem_reserve_ratio changes.
5702 * The reserve ratio obviously has absolutely no relation with the
5703 * minimum watermarks. The lowmem reserve ratio can only make sense
5704 * if in function of the boot time zone sizes.
5706 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5707 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5709 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5710 setup_per_zone_lowmem_reserve();
5715 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5716 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5717 * can have before it gets flushed back to buddy allocator.
5720 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5721 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5727 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5728 if (!write
|| (ret
< 0))
5730 for_each_populated_zone(zone
) {
5731 for_each_possible_cpu(cpu
) {
5733 high
= zone
->managed_pages
/ percpu_pagelist_fraction
;
5734 setup_pagelist_highmark(
5735 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5741 int hashdist
= HASHDIST_DEFAULT
;
5744 static int __init
set_hashdist(char *str
)
5748 hashdist
= simple_strtoul(str
, &str
, 0);
5751 __setup("hashdist=", set_hashdist
);
5755 * allocate a large system hash table from bootmem
5756 * - it is assumed that the hash table must contain an exact power-of-2
5757 * quantity of entries
5758 * - limit is the number of hash buckets, not the total allocation size
5760 void *__init
alloc_large_system_hash(const char *tablename
,
5761 unsigned long bucketsize
,
5762 unsigned long numentries
,
5765 unsigned int *_hash_shift
,
5766 unsigned int *_hash_mask
,
5767 unsigned long low_limit
,
5768 unsigned long high_limit
)
5770 unsigned long long max
= high_limit
;
5771 unsigned long log2qty
, size
;
5774 /* allow the kernel cmdline to have a say */
5776 /* round applicable memory size up to nearest megabyte */
5777 numentries
= nr_kernel_pages
;
5778 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5779 numentries
>>= 20 - PAGE_SHIFT
;
5780 numentries
<<= 20 - PAGE_SHIFT
;
5782 /* limit to 1 bucket per 2^scale bytes of low memory */
5783 if (scale
> PAGE_SHIFT
)
5784 numentries
>>= (scale
- PAGE_SHIFT
);
5786 numentries
<<= (PAGE_SHIFT
- scale
);
5788 /* Make sure we've got at least a 0-order allocation.. */
5789 if (unlikely(flags
& HASH_SMALL
)) {
5790 /* Makes no sense without HASH_EARLY */
5791 WARN_ON(!(flags
& HASH_EARLY
));
5792 if (!(numentries
>> *_hash_shift
)) {
5793 numentries
= 1UL << *_hash_shift
;
5794 BUG_ON(!numentries
);
5796 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5797 numentries
= PAGE_SIZE
/ bucketsize
;
5799 numentries
= roundup_pow_of_two(numentries
);
5801 /* limit allocation size to 1/16 total memory by default */
5803 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5804 do_div(max
, bucketsize
);
5806 max
= min(max
, 0x80000000ULL
);
5808 if (numentries
< low_limit
)
5809 numentries
= low_limit
;
5810 if (numentries
> max
)
5813 log2qty
= ilog2(numentries
);
5816 size
= bucketsize
<< log2qty
;
5817 if (flags
& HASH_EARLY
)
5818 table
= alloc_bootmem_nopanic(size
);
5820 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5823 * If bucketsize is not a power-of-two, we may free
5824 * some pages at the end of hash table which
5825 * alloc_pages_exact() automatically does
5827 if (get_order(size
) < MAX_ORDER
) {
5828 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5829 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5832 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5835 panic("Failed to allocate %s hash table\n", tablename
);
5837 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5840 ilog2(size
) - PAGE_SHIFT
,
5844 *_hash_shift
= log2qty
;
5846 *_hash_mask
= (1 << log2qty
) - 1;
5851 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5852 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5855 #ifdef CONFIG_SPARSEMEM
5856 return __pfn_to_section(pfn
)->pageblock_flags
;
5858 return zone
->pageblock_flags
;
5859 #endif /* CONFIG_SPARSEMEM */
5862 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5864 #ifdef CONFIG_SPARSEMEM
5865 pfn
&= (PAGES_PER_SECTION
-1);
5866 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5868 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
5869 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5870 #endif /* CONFIG_SPARSEMEM */
5874 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5875 * @page: The page within the block of interest
5876 * @start_bitidx: The first bit of interest to retrieve
5877 * @end_bitidx: The last bit of interest
5878 * returns pageblock_bits flags
5880 unsigned long get_pageblock_flags_group(struct page
*page
,
5881 int start_bitidx
, int end_bitidx
)
5884 unsigned long *bitmap
;
5885 unsigned long pfn
, bitidx
;
5886 unsigned long flags
= 0;
5887 unsigned long value
= 1;
5889 zone
= page_zone(page
);
5890 pfn
= page_to_pfn(page
);
5891 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5892 bitidx
= pfn_to_bitidx(zone
, pfn
);
5894 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5895 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5902 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5903 * @page: The page within the block of interest
5904 * @start_bitidx: The first bit of interest
5905 * @end_bitidx: The last bit of interest
5906 * @flags: The flags to set
5908 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5909 int start_bitidx
, int end_bitidx
)
5912 unsigned long *bitmap
;
5913 unsigned long pfn
, bitidx
;
5914 unsigned long value
= 1;
5916 zone
= page_zone(page
);
5917 pfn
= page_to_pfn(page
);
5918 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5919 bitidx
= pfn_to_bitidx(zone
, pfn
);
5920 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
5921 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
5923 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5925 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5927 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5931 * This function checks whether pageblock includes unmovable pages or not.
5932 * If @count is not zero, it is okay to include less @count unmovable pages
5934 * PageLRU check wihtout isolation or lru_lock could race so that
5935 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5936 * expect this function should be exact.
5938 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
5939 bool skip_hwpoisoned_pages
)
5941 unsigned long pfn
, iter
, found
;
5945 * For avoiding noise data, lru_add_drain_all() should be called
5946 * If ZONE_MOVABLE, the zone never contains unmovable pages
5948 if (zone_idx(zone
) == ZONE_MOVABLE
)
5950 mt
= get_pageblock_migratetype(page
);
5951 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
5954 pfn
= page_to_pfn(page
);
5955 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5956 unsigned long check
= pfn
+ iter
;
5958 if (!pfn_valid_within(check
))
5961 page
= pfn_to_page(check
);
5963 * We can't use page_count without pin a page
5964 * because another CPU can free compound page.
5965 * This check already skips compound tails of THP
5966 * because their page->_count is zero at all time.
5968 if (!atomic_read(&page
->_count
)) {
5969 if (PageBuddy(page
))
5970 iter
+= (1 << page_order(page
)) - 1;
5975 * The HWPoisoned page may be not in buddy system, and
5976 * page_count() is not 0.
5978 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
5984 * If there are RECLAIMABLE pages, we need to check it.
5985 * But now, memory offline itself doesn't call shrink_slab()
5986 * and it still to be fixed.
5989 * If the page is not RAM, page_count()should be 0.
5990 * we don't need more check. This is an _used_ not-movable page.
5992 * The problematic thing here is PG_reserved pages. PG_reserved
5993 * is set to both of a memory hole page and a _used_ kernel
6002 bool is_pageblock_removable_nolock(struct page
*page
)
6008 * We have to be careful here because we are iterating over memory
6009 * sections which are not zone aware so we might end up outside of
6010 * the zone but still within the section.
6011 * We have to take care about the node as well. If the node is offline
6012 * its NODE_DATA will be NULL - see page_zone.
6014 if (!node_online(page_to_nid(page
)))
6017 zone
= page_zone(page
);
6018 pfn
= page_to_pfn(page
);
6019 if (zone
->zone_start_pfn
> pfn
||
6020 zone
->zone_start_pfn
+ zone
->spanned_pages
<= pfn
)
6023 return !has_unmovable_pages(zone
, page
, 0, true);
6028 static unsigned long pfn_max_align_down(unsigned long pfn
)
6030 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6031 pageblock_nr_pages
) - 1);
6034 static unsigned long pfn_max_align_up(unsigned long pfn
)
6036 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6037 pageblock_nr_pages
));
6040 /* [start, end) must belong to a single zone. */
6041 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6042 unsigned long start
, unsigned long end
)
6044 /* This function is based on compact_zone() from compaction.c. */
6045 unsigned long nr_reclaimed
;
6046 unsigned long pfn
= start
;
6047 unsigned int tries
= 0;
6052 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6053 if (fatal_signal_pending(current
)) {
6058 if (list_empty(&cc
->migratepages
)) {
6059 cc
->nr_migratepages
= 0;
6060 pfn
= isolate_migratepages_range(cc
->zone
, cc
,
6067 } else if (++tries
== 5) {
6068 ret
= ret
< 0 ? ret
: -EBUSY
;
6072 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6074 cc
->nr_migratepages
-= nr_reclaimed
;
6076 ret
= migrate_pages(&cc
->migratepages
,
6077 alloc_migrate_target
,
6078 0, false, MIGRATE_SYNC
,
6082 putback_movable_pages(&cc
->migratepages
);
6089 * alloc_contig_range() -- tries to allocate given range of pages
6090 * @start: start PFN to allocate
6091 * @end: one-past-the-last PFN to allocate
6092 * @migratetype: migratetype of the underlaying pageblocks (either
6093 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6094 * in range must have the same migratetype and it must
6095 * be either of the two.
6097 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6098 * aligned, however it's the caller's responsibility to guarantee that
6099 * we are the only thread that changes migrate type of pageblocks the
6102 * The PFN range must belong to a single zone.
6104 * Returns zero on success or negative error code. On success all
6105 * pages which PFN is in [start, end) are allocated for the caller and
6106 * need to be freed with free_contig_range().
6108 int alloc_contig_range(unsigned long start
, unsigned long end
,
6109 unsigned migratetype
)
6111 unsigned long outer_start
, outer_end
;
6114 struct compact_control cc
= {
6115 .nr_migratepages
= 0,
6117 .zone
= page_zone(pfn_to_page(start
)),
6119 .ignore_skip_hint
= true,
6121 INIT_LIST_HEAD(&cc
.migratepages
);
6124 * What we do here is we mark all pageblocks in range as
6125 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6126 * have different sizes, and due to the way page allocator
6127 * work, we align the range to biggest of the two pages so
6128 * that page allocator won't try to merge buddies from
6129 * different pageblocks and change MIGRATE_ISOLATE to some
6130 * other migration type.
6132 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6133 * migrate the pages from an unaligned range (ie. pages that
6134 * we are interested in). This will put all the pages in
6135 * range back to page allocator as MIGRATE_ISOLATE.
6137 * When this is done, we take the pages in range from page
6138 * allocator removing them from the buddy system. This way
6139 * page allocator will never consider using them.
6141 * This lets us mark the pageblocks back as
6142 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6143 * aligned range but not in the unaligned, original range are
6144 * put back to page allocator so that buddy can use them.
6147 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6148 pfn_max_align_up(end
), migratetype
,
6153 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6158 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6159 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6160 * more, all pages in [start, end) are free in page allocator.
6161 * What we are going to do is to allocate all pages from
6162 * [start, end) (that is remove them from page allocator).
6164 * The only problem is that pages at the beginning and at the
6165 * end of interesting range may be not aligned with pages that
6166 * page allocator holds, ie. they can be part of higher order
6167 * pages. Because of this, we reserve the bigger range and
6168 * once this is done free the pages we are not interested in.
6170 * We don't have to hold zone->lock here because the pages are
6171 * isolated thus they won't get removed from buddy.
6174 lru_add_drain_all();
6178 outer_start
= start
;
6179 while (!PageBuddy(pfn_to_page(outer_start
))) {
6180 if (++order
>= MAX_ORDER
) {
6184 outer_start
&= ~0UL << order
;
6187 /* Make sure the range is really isolated. */
6188 if (test_pages_isolated(outer_start
, end
, false)) {
6189 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6196 /* Grab isolated pages from freelists. */
6197 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6203 /* Free head and tail (if any) */
6204 if (start
!= outer_start
)
6205 free_contig_range(outer_start
, start
- outer_start
);
6206 if (end
!= outer_end
)
6207 free_contig_range(end
, outer_end
- end
);
6210 undo_isolate_page_range(pfn_max_align_down(start
),
6211 pfn_max_align_up(end
), migratetype
);
6215 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6217 unsigned int count
= 0;
6219 for (; nr_pages
--; pfn
++) {
6220 struct page
*page
= pfn_to_page(pfn
);
6222 count
+= page_count(page
) != 1;
6225 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6229 #ifdef CONFIG_MEMORY_HOTPLUG
6230 static int __meminit
__zone_pcp_update(void *data
)
6232 struct zone
*zone
= data
;
6234 unsigned long batch
= zone_batchsize(zone
), flags
;
6236 for_each_possible_cpu(cpu
) {
6237 struct per_cpu_pageset
*pset
;
6238 struct per_cpu_pages
*pcp
;
6240 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6243 local_irq_save(flags
);
6245 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
6246 drain_zonestat(zone
, pset
);
6247 setup_pageset(pset
, batch
);
6248 local_irq_restore(flags
);
6253 void __meminit
zone_pcp_update(struct zone
*zone
)
6255 stop_machine(__zone_pcp_update
, zone
, NULL
);
6259 void zone_pcp_reset(struct zone
*zone
)
6261 unsigned long flags
;
6263 struct per_cpu_pageset
*pset
;
6265 /* avoid races with drain_pages() */
6266 local_irq_save(flags
);
6267 if (zone
->pageset
!= &boot_pageset
) {
6268 for_each_online_cpu(cpu
) {
6269 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6270 drain_zonestat(zone
, pset
);
6272 free_percpu(zone
->pageset
);
6273 zone
->pageset
= &boot_pageset
;
6275 local_irq_restore(flags
);
6278 #ifdef CONFIG_MEMORY_HOTREMOVE
6280 * All pages in the range must be isolated before calling this.
6283 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6289 unsigned long flags
;
6290 /* find the first valid pfn */
6291 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6296 zone
= page_zone(pfn_to_page(pfn
));
6297 spin_lock_irqsave(&zone
->lock
, flags
);
6299 while (pfn
< end_pfn
) {
6300 if (!pfn_valid(pfn
)) {
6304 page
= pfn_to_page(pfn
);
6306 * The HWPoisoned page may be not in buddy system, and
6307 * page_count() is not 0.
6309 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6311 SetPageReserved(page
);
6315 BUG_ON(page_count(page
));
6316 BUG_ON(!PageBuddy(page
));
6317 order
= page_order(page
);
6318 #ifdef CONFIG_DEBUG_VM
6319 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6320 pfn
, 1 << order
, end_pfn
);
6322 list_del(&page
->lru
);
6323 rmv_page_order(page
);
6324 zone
->free_area
[order
].nr_free
--;
6325 for (i
= 0; i
< (1 << order
); i
++)
6326 SetPageReserved((page
+i
));
6327 pfn
+= (1 << order
);
6329 spin_unlock_irqrestore(&zone
->lock
, flags
);
6333 #ifdef CONFIG_MEMORY_FAILURE
6334 bool is_free_buddy_page(struct page
*page
)
6336 struct zone
*zone
= page_zone(page
);
6337 unsigned long pfn
= page_to_pfn(page
);
6338 unsigned long flags
;
6341 spin_lock_irqsave(&zone
->lock
, flags
);
6342 for (order
= 0; order
< MAX_ORDER
; order
++) {
6343 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6345 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6348 spin_unlock_irqrestore(&zone
->lock
, flags
);
6350 return order
< MAX_ORDER
;
6354 static const struct trace_print_flags pageflag_names
[] = {
6355 {1UL << PG_locked
, "locked" },
6356 {1UL << PG_error
, "error" },
6357 {1UL << PG_referenced
, "referenced" },
6358 {1UL << PG_uptodate
, "uptodate" },
6359 {1UL << PG_dirty
, "dirty" },
6360 {1UL << PG_lru
, "lru" },
6361 {1UL << PG_active
, "active" },
6362 {1UL << PG_slab
, "slab" },
6363 {1UL << PG_owner_priv_1
, "owner_priv_1" },
6364 {1UL << PG_arch_1
, "arch_1" },
6365 {1UL << PG_reserved
, "reserved" },
6366 {1UL << PG_private
, "private" },
6367 {1UL << PG_private_2
, "private_2" },
6368 {1UL << PG_writeback
, "writeback" },
6369 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6370 {1UL << PG_head
, "head" },
6371 {1UL << PG_tail
, "tail" },
6373 {1UL << PG_compound
, "compound" },
6375 {1UL << PG_swapcache
, "swapcache" },
6376 {1UL << PG_mappedtodisk
, "mappedtodisk" },
6377 {1UL << PG_reclaim
, "reclaim" },
6378 {1UL << PG_swapbacked
, "swapbacked" },
6379 {1UL << PG_unevictable
, "unevictable" },
6381 {1UL << PG_mlocked
, "mlocked" },
6383 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6384 {1UL << PG_uncached
, "uncached" },
6386 #ifdef CONFIG_MEMORY_FAILURE
6387 {1UL << PG_hwpoison
, "hwpoison" },
6389 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6390 {1UL << PG_compound_lock
, "compound_lock" },
6394 static void dump_page_flags(unsigned long flags
)
6396 const char *delim
= "";
6400 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names
) != __NR_PAGEFLAGS
);
6402 printk(KERN_ALERT
"page flags: %#lx(", flags
);
6404 /* remove zone id */
6405 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
6407 for (i
= 0; i
< ARRAY_SIZE(pageflag_names
) && flags
; i
++) {
6409 mask
= pageflag_names
[i
].mask
;
6410 if ((flags
& mask
) != mask
)
6414 printk("%s%s", delim
, pageflag_names
[i
].name
);
6418 /* check for left over flags */
6420 printk("%s%#lx", delim
, flags
);
6425 void dump_page(struct page
*page
)
6428 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6429 page
, atomic_read(&page
->_count
), page_mapcount(page
),
6430 page
->mapping
, page
->index
);
6431 dump_page_flags(page
->flags
);
6432 mem_cgroup_print_bad_page(page
);