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(!is_migrate_isolate_page(page
))) {
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(!is_migrate_isolate(migratetype
)))
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 #ifdef CONFIG_MEMORY_ISOLATION
927 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
932 * Move the free pages in a range to the free lists of the requested type.
933 * Note that start_page and end_pages are not aligned on a pageblock
934 * boundary. If alignment is required, use move_freepages_block()
936 int move_freepages(struct zone
*zone
,
937 struct page
*start_page
, struct page
*end_page
,
944 #ifndef CONFIG_HOLES_IN_ZONE
946 * page_zone is not safe to call in this context when
947 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
948 * anyway as we check zone boundaries in move_freepages_block().
949 * Remove at a later date when no bug reports exist related to
950 * grouping pages by mobility
952 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
955 for (page
= start_page
; page
<= end_page
;) {
956 /* Make sure we are not inadvertently changing nodes */
957 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
959 if (!pfn_valid_within(page_to_pfn(page
))) {
964 if (!PageBuddy(page
)) {
969 order
= page_order(page
);
970 list_move(&page
->lru
,
971 &zone
->free_area
[order
].free_list
[migratetype
]);
972 set_freepage_migratetype(page
, migratetype
);
974 pages_moved
+= 1 << order
;
980 int move_freepages_block(struct zone
*zone
, struct page
*page
,
983 unsigned long start_pfn
, end_pfn
;
984 struct page
*start_page
, *end_page
;
986 start_pfn
= page_to_pfn(page
);
987 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
988 start_page
= pfn_to_page(start_pfn
);
989 end_page
= start_page
+ pageblock_nr_pages
- 1;
990 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
992 /* Do not cross zone boundaries */
993 if (start_pfn
< zone
->zone_start_pfn
)
995 if (end_pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
)
998 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1001 static void change_pageblock_range(struct page
*pageblock_page
,
1002 int start_order
, int migratetype
)
1004 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1006 while (nr_pageblocks
--) {
1007 set_pageblock_migratetype(pageblock_page
, migratetype
);
1008 pageblock_page
+= pageblock_nr_pages
;
1012 /* Remove an element from the buddy allocator from the fallback list */
1013 static inline struct page
*
1014 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
1016 struct free_area
* area
;
1021 /* Find the largest possible block of pages in the other list */
1022 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
1025 migratetype
= fallbacks
[start_migratetype
][i
];
1027 /* MIGRATE_RESERVE handled later if necessary */
1028 if (migratetype
== MIGRATE_RESERVE
)
1031 area
= &(zone
->free_area
[current_order
]);
1032 if (list_empty(&area
->free_list
[migratetype
]))
1035 page
= list_entry(area
->free_list
[migratetype
].next
,
1040 * If breaking a large block of pages, move all free
1041 * pages to the preferred allocation list. If falling
1042 * back for a reclaimable kernel allocation, be more
1043 * aggressive about taking ownership of free pages
1045 * On the other hand, never change migration
1046 * type of MIGRATE_CMA pageblocks nor move CMA
1047 * pages on different free lists. We don't
1048 * want unmovable pages to be allocated from
1049 * MIGRATE_CMA areas.
1051 if (!is_migrate_cma(migratetype
) &&
1052 (unlikely(current_order
>= pageblock_order
/ 2) ||
1053 start_migratetype
== MIGRATE_RECLAIMABLE
||
1054 page_group_by_mobility_disabled
)) {
1056 pages
= move_freepages_block(zone
, page
,
1059 /* Claim the whole block if over half of it is free */
1060 if (pages
>= (1 << (pageblock_order
-1)) ||
1061 page_group_by_mobility_disabled
)
1062 set_pageblock_migratetype(page
,
1065 migratetype
= start_migratetype
;
1068 /* Remove the page from the freelists */
1069 list_del(&page
->lru
);
1070 rmv_page_order(page
);
1072 /* Take ownership for orders >= pageblock_order */
1073 if (current_order
>= pageblock_order
&&
1074 !is_migrate_cma(migratetype
))
1075 change_pageblock_range(page
, current_order
,
1078 expand(zone
, page
, order
, current_order
, area
,
1079 is_migrate_cma(migratetype
)
1080 ? migratetype
: start_migratetype
);
1082 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1083 start_migratetype
, migratetype
);
1093 * Do the hard work of removing an element from the buddy allocator.
1094 * Call me with the zone->lock already held.
1096 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1102 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1104 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1105 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1108 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1109 * is used because __rmqueue_smallest is an inline function
1110 * and we want just one call site
1113 migratetype
= MIGRATE_RESERVE
;
1118 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1123 * Obtain a specified number of elements from the buddy allocator, all under
1124 * a single hold of the lock, for efficiency. Add them to the supplied list.
1125 * Returns the number of new pages which were placed at *list.
1127 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1128 unsigned long count
, struct list_head
*list
,
1129 int migratetype
, int cold
)
1131 int mt
= migratetype
, i
;
1133 spin_lock(&zone
->lock
);
1134 for (i
= 0; i
< count
; ++i
) {
1135 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1136 if (unlikely(page
== NULL
))
1140 * Split buddy pages returned by expand() are received here
1141 * in physical page order. The page is added to the callers and
1142 * list and the list head then moves forward. From the callers
1143 * perspective, the linked list is ordered by page number in
1144 * some conditions. This is useful for IO devices that can
1145 * merge IO requests if the physical pages are ordered
1148 if (likely(cold
== 0))
1149 list_add(&page
->lru
, list
);
1151 list_add_tail(&page
->lru
, list
);
1152 if (IS_ENABLED(CONFIG_CMA
)) {
1153 mt
= get_pageblock_migratetype(page
);
1154 if (!is_migrate_cma(mt
) && !is_migrate_isolate(mt
))
1157 set_freepage_migratetype(page
, mt
);
1159 if (is_migrate_cma(mt
))
1160 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1163 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1164 spin_unlock(&zone
->lock
);
1170 * Called from the vmstat counter updater to drain pagesets of this
1171 * currently executing processor on remote nodes after they have
1174 * Note that this function must be called with the thread pinned to
1175 * a single processor.
1177 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1179 unsigned long flags
;
1182 local_irq_save(flags
);
1183 if (pcp
->count
>= pcp
->batch
)
1184 to_drain
= pcp
->batch
;
1186 to_drain
= pcp
->count
;
1188 free_pcppages_bulk(zone
, to_drain
, pcp
);
1189 pcp
->count
-= to_drain
;
1191 local_irq_restore(flags
);
1196 * Drain pages of the indicated processor.
1198 * The processor must either be the current processor and the
1199 * thread pinned to the current processor or a processor that
1202 static void drain_pages(unsigned int cpu
)
1204 unsigned long flags
;
1207 for_each_populated_zone(zone
) {
1208 struct per_cpu_pageset
*pset
;
1209 struct per_cpu_pages
*pcp
;
1211 local_irq_save(flags
);
1212 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1216 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1219 local_irq_restore(flags
);
1224 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1226 void drain_local_pages(void *arg
)
1228 drain_pages(smp_processor_id());
1232 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1234 * Note that this code is protected against sending an IPI to an offline
1235 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1236 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1237 * nothing keeps CPUs from showing up after we populated the cpumask and
1238 * before the call to on_each_cpu_mask().
1240 void drain_all_pages(void)
1243 struct per_cpu_pageset
*pcp
;
1247 * Allocate in the BSS so we wont require allocation in
1248 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1250 static cpumask_t cpus_with_pcps
;
1253 * We don't care about racing with CPU hotplug event
1254 * as offline notification will cause the notified
1255 * cpu to drain that CPU pcps and on_each_cpu_mask
1256 * disables preemption as part of its processing
1258 for_each_online_cpu(cpu
) {
1259 bool has_pcps
= false;
1260 for_each_populated_zone(zone
) {
1261 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1262 if (pcp
->pcp
.count
) {
1268 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1270 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1272 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1275 #ifdef CONFIG_HIBERNATION
1277 void mark_free_pages(struct zone
*zone
)
1279 unsigned long pfn
, max_zone_pfn
;
1280 unsigned long flags
;
1282 struct list_head
*curr
;
1284 if (!zone
->spanned_pages
)
1287 spin_lock_irqsave(&zone
->lock
, flags
);
1289 max_zone_pfn
= zone
->zone_start_pfn
+ zone
->spanned_pages
;
1290 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1291 if (pfn_valid(pfn
)) {
1292 struct page
*page
= pfn_to_page(pfn
);
1294 if (!swsusp_page_is_forbidden(page
))
1295 swsusp_unset_page_free(page
);
1298 for_each_migratetype_order(order
, t
) {
1299 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1302 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1303 for (i
= 0; i
< (1UL << order
); i
++)
1304 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1307 spin_unlock_irqrestore(&zone
->lock
, flags
);
1309 #endif /* CONFIG_PM */
1312 * Free a 0-order page
1313 * cold == 1 ? free a cold page : free a hot page
1315 void free_hot_cold_page(struct page
*page
, int cold
)
1317 struct zone
*zone
= page_zone(page
);
1318 struct per_cpu_pages
*pcp
;
1319 unsigned long flags
;
1322 if (!free_pages_prepare(page
, 0))
1325 migratetype
= get_pageblock_migratetype(page
);
1326 set_freepage_migratetype(page
, migratetype
);
1327 local_irq_save(flags
);
1328 __count_vm_event(PGFREE
);
1331 * We only track unmovable, reclaimable and movable on pcp lists.
1332 * Free ISOLATE pages back to the allocator because they are being
1333 * offlined but treat RESERVE as movable pages so we can get those
1334 * areas back if necessary. Otherwise, we may have to free
1335 * excessively into the page allocator
1337 if (migratetype
>= MIGRATE_PCPTYPES
) {
1338 if (unlikely(is_migrate_isolate(migratetype
))) {
1339 free_one_page(zone
, page
, 0, migratetype
);
1342 migratetype
= MIGRATE_MOVABLE
;
1345 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1347 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1349 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1351 if (pcp
->count
>= pcp
->high
) {
1352 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1353 pcp
->count
-= pcp
->batch
;
1357 local_irq_restore(flags
);
1361 * Free a list of 0-order pages
1363 void free_hot_cold_page_list(struct list_head
*list
, int cold
)
1365 struct page
*page
, *next
;
1367 list_for_each_entry_safe(page
, next
, list
, lru
) {
1368 trace_mm_page_free_batched(page
, cold
);
1369 free_hot_cold_page(page
, cold
);
1374 * split_page takes a non-compound higher-order page, and splits it into
1375 * n (1<<order) sub-pages: page[0..n]
1376 * Each sub-page must be freed individually.
1378 * Note: this is probably too low level an operation for use in drivers.
1379 * Please consult with lkml before using this in your driver.
1381 void split_page(struct page
*page
, unsigned int order
)
1385 VM_BUG_ON(PageCompound(page
));
1386 VM_BUG_ON(!page_count(page
));
1388 #ifdef CONFIG_KMEMCHECK
1390 * Split shadow pages too, because free(page[0]) would
1391 * otherwise free the whole shadow.
1393 if (kmemcheck_page_is_tracked(page
))
1394 split_page(virt_to_page(page
[0].shadow
), order
);
1397 for (i
= 1; i
< (1 << order
); i
++)
1398 set_page_refcounted(page
+ i
);
1401 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1403 unsigned long watermark
;
1407 BUG_ON(!PageBuddy(page
));
1409 zone
= page_zone(page
);
1410 mt
= get_pageblock_migratetype(page
);
1412 if (!is_migrate_isolate(mt
)) {
1413 /* Obey watermarks as if the page was being allocated */
1414 watermark
= low_wmark_pages(zone
) + (1 << order
);
1415 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1418 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1421 /* Remove page from free list */
1422 list_del(&page
->lru
);
1423 zone
->free_area
[order
].nr_free
--;
1424 rmv_page_order(page
);
1426 /* Set the pageblock if the isolated page is at least a pageblock */
1427 if (order
>= pageblock_order
- 1) {
1428 struct page
*endpage
= page
+ (1 << order
) - 1;
1429 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1430 int mt
= get_pageblock_migratetype(page
);
1431 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1432 set_pageblock_migratetype(page
,
1437 return 1UL << order
;
1441 * Similar to split_page except the page is already free. As this is only
1442 * being used for migration, the migratetype of the block also changes.
1443 * As this is called with interrupts disabled, the caller is responsible
1444 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1447 * Note: this is probably too low level an operation for use in drivers.
1448 * Please consult with lkml before using this in your driver.
1450 int split_free_page(struct page
*page
)
1455 order
= page_order(page
);
1457 nr_pages
= __isolate_free_page(page
, order
);
1461 /* Split into individual pages */
1462 set_page_refcounted(page
);
1463 split_page(page
, order
);
1468 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1469 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1473 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1474 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1477 unsigned long flags
;
1479 int cold
= !!(gfp_flags
& __GFP_COLD
);
1482 if (likely(order
== 0)) {
1483 struct per_cpu_pages
*pcp
;
1484 struct list_head
*list
;
1486 local_irq_save(flags
);
1487 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1488 list
= &pcp
->lists
[migratetype
];
1489 if (list_empty(list
)) {
1490 pcp
->count
+= rmqueue_bulk(zone
, 0,
1493 if (unlikely(list_empty(list
)))
1498 page
= list_entry(list
->prev
, struct page
, lru
);
1500 page
= list_entry(list
->next
, struct page
, lru
);
1502 list_del(&page
->lru
);
1505 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1507 * __GFP_NOFAIL is not to be used in new code.
1509 * All __GFP_NOFAIL callers should be fixed so that they
1510 * properly detect and handle allocation failures.
1512 * We most definitely don't want callers attempting to
1513 * allocate greater than order-1 page units with
1516 WARN_ON_ONCE(order
> 1);
1518 spin_lock_irqsave(&zone
->lock
, flags
);
1519 page
= __rmqueue(zone
, order
, migratetype
);
1520 spin_unlock(&zone
->lock
);
1523 __mod_zone_freepage_state(zone
, -(1 << order
),
1524 get_pageblock_migratetype(page
));
1527 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1528 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1529 local_irq_restore(flags
);
1531 VM_BUG_ON(bad_range(zone
, page
));
1532 if (prep_new_page(page
, order
, gfp_flags
))
1537 local_irq_restore(flags
);
1541 #ifdef CONFIG_FAIL_PAGE_ALLOC
1544 struct fault_attr attr
;
1546 u32 ignore_gfp_highmem
;
1547 u32 ignore_gfp_wait
;
1549 } fail_page_alloc
= {
1550 .attr
= FAULT_ATTR_INITIALIZER
,
1551 .ignore_gfp_wait
= 1,
1552 .ignore_gfp_highmem
= 1,
1556 static int __init
setup_fail_page_alloc(char *str
)
1558 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1560 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1562 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1564 if (order
< fail_page_alloc
.min_order
)
1566 if (gfp_mask
& __GFP_NOFAIL
)
1568 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1570 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1573 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1576 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1578 static int __init
fail_page_alloc_debugfs(void)
1580 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1583 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1584 &fail_page_alloc
.attr
);
1586 return PTR_ERR(dir
);
1588 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1589 &fail_page_alloc
.ignore_gfp_wait
))
1591 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1592 &fail_page_alloc
.ignore_gfp_highmem
))
1594 if (!debugfs_create_u32("min-order", mode
, dir
,
1595 &fail_page_alloc
.min_order
))
1600 debugfs_remove_recursive(dir
);
1605 late_initcall(fail_page_alloc_debugfs
);
1607 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1609 #else /* CONFIG_FAIL_PAGE_ALLOC */
1611 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1616 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1619 * Return true if free pages are above 'mark'. This takes into account the order
1620 * of the allocation.
1622 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1623 int classzone_idx
, int alloc_flags
, long free_pages
)
1625 /* free_pages my go negative - that's OK */
1627 long lowmem_reserve
= z
->lowmem_reserve
[classzone_idx
];
1630 free_pages
-= (1 << order
) - 1;
1631 if (alloc_flags
& ALLOC_HIGH
)
1633 if (alloc_flags
& ALLOC_HARDER
)
1636 /* If allocation can't use CMA areas don't use free CMA pages */
1637 if (!(alloc_flags
& ALLOC_CMA
))
1638 free_pages
-= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1640 if (free_pages
<= min
+ lowmem_reserve
)
1642 for (o
= 0; o
< order
; o
++) {
1643 /* At the next order, this order's pages become unavailable */
1644 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1646 /* Require fewer higher order pages to be free */
1649 if (free_pages
<= min
)
1655 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1656 int classzone_idx
, int alloc_flags
)
1658 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1659 zone_page_state(z
, NR_FREE_PAGES
));
1662 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1663 int classzone_idx
, int alloc_flags
)
1665 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1667 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1668 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1670 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1676 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1677 * skip over zones that are not allowed by the cpuset, or that have
1678 * been recently (in last second) found to be nearly full. See further
1679 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1680 * that have to skip over a lot of full or unallowed zones.
1682 * If the zonelist cache is present in the passed in zonelist, then
1683 * returns a pointer to the allowed node mask (either the current
1684 * tasks mems_allowed, or node_states[N_MEMORY].)
1686 * If the zonelist cache is not available for this zonelist, does
1687 * nothing and returns NULL.
1689 * If the fullzones BITMAP in the zonelist cache is stale (more than
1690 * a second since last zap'd) then we zap it out (clear its bits.)
1692 * We hold off even calling zlc_setup, until after we've checked the
1693 * first zone in the zonelist, on the theory that most allocations will
1694 * be satisfied from that first zone, so best to examine that zone as
1695 * quickly as we can.
1697 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1699 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1700 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1702 zlc
= zonelist
->zlcache_ptr
;
1706 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1707 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1708 zlc
->last_full_zap
= jiffies
;
1711 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1712 &cpuset_current_mems_allowed
:
1713 &node_states
[N_MEMORY
];
1714 return allowednodes
;
1718 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1719 * if it is worth looking at further for free memory:
1720 * 1) Check that the zone isn't thought to be full (doesn't have its
1721 * bit set in the zonelist_cache fullzones BITMAP).
1722 * 2) Check that the zones node (obtained from the zonelist_cache
1723 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1724 * Return true (non-zero) if zone is worth looking at further, or
1725 * else return false (zero) if it is not.
1727 * This check -ignores- the distinction between various watermarks,
1728 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1729 * found to be full for any variation of these watermarks, it will
1730 * be considered full for up to one second by all requests, unless
1731 * we are so low on memory on all allowed nodes that we are forced
1732 * into the second scan of the zonelist.
1734 * In the second scan we ignore this zonelist cache and exactly
1735 * apply the watermarks to all zones, even it is slower to do so.
1736 * We are low on memory in the second scan, and should leave no stone
1737 * unturned looking for a free page.
1739 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1740 nodemask_t
*allowednodes
)
1742 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1743 int i
; /* index of *z in zonelist zones */
1744 int n
; /* node that zone *z is on */
1746 zlc
= zonelist
->zlcache_ptr
;
1750 i
= z
- zonelist
->_zonerefs
;
1753 /* This zone is worth trying if it is allowed but not full */
1754 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1758 * Given 'z' scanning a zonelist, set the corresponding bit in
1759 * zlc->fullzones, so that subsequent attempts to allocate a page
1760 * from that zone don't waste time re-examining it.
1762 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1764 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1765 int i
; /* index of *z in zonelist zones */
1767 zlc
= zonelist
->zlcache_ptr
;
1771 i
= z
- zonelist
->_zonerefs
;
1773 set_bit(i
, zlc
->fullzones
);
1777 * clear all zones full, called after direct reclaim makes progress so that
1778 * a zone that was recently full is not skipped over for up to a second
1780 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1782 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1784 zlc
= zonelist
->zlcache_ptr
;
1788 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1791 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1793 return node_isset(local_zone
->node
, zone
->zone_pgdat
->reclaim_nodes
);
1796 static void __paginginit
init_zone_allows_reclaim(int nid
)
1800 for_each_online_node(i
)
1801 if (node_distance(nid
, i
) <= RECLAIM_DISTANCE
)
1802 node_set(i
, NODE_DATA(nid
)->reclaim_nodes
);
1804 zone_reclaim_mode
= 1;
1807 #else /* CONFIG_NUMA */
1809 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1814 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1815 nodemask_t
*allowednodes
)
1820 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1824 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1828 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1833 static inline void init_zone_allows_reclaim(int nid
)
1836 #endif /* CONFIG_NUMA */
1839 * get_page_from_freelist goes through the zonelist trying to allocate
1842 static struct page
*
1843 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1844 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1845 struct zone
*preferred_zone
, int migratetype
)
1848 struct page
*page
= NULL
;
1851 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1852 int zlc_active
= 0; /* set if using zonelist_cache */
1853 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1855 classzone_idx
= zone_idx(preferred_zone
);
1858 * Scan zonelist, looking for a zone with enough free.
1859 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1861 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1862 high_zoneidx
, nodemask
) {
1863 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1864 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1866 if ((alloc_flags
& ALLOC_CPUSET
) &&
1867 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1870 * When allocating a page cache page for writing, we
1871 * want to get it from a zone that is within its dirty
1872 * limit, such that no single zone holds more than its
1873 * proportional share of globally allowed dirty pages.
1874 * The dirty limits take into account the zone's
1875 * lowmem reserves and high watermark so that kswapd
1876 * should be able to balance it without having to
1877 * write pages from its LRU list.
1879 * This may look like it could increase pressure on
1880 * lower zones by failing allocations in higher zones
1881 * before they are full. But the pages that do spill
1882 * over are limited as the lower zones are protected
1883 * by this very same mechanism. It should not become
1884 * a practical burden to them.
1886 * XXX: For now, allow allocations to potentially
1887 * exceed the per-zone dirty limit in the slowpath
1888 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1889 * which is important when on a NUMA setup the allowed
1890 * zones are together not big enough to reach the
1891 * global limit. The proper fix for these situations
1892 * will require awareness of zones in the
1893 * dirty-throttling and the flusher threads.
1895 if ((alloc_flags
& ALLOC_WMARK_LOW
) &&
1896 (gfp_mask
& __GFP_WRITE
) && !zone_dirty_ok(zone
))
1897 goto this_zone_full
;
1899 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1900 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1904 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1905 if (zone_watermark_ok(zone
, order
, mark
,
1906 classzone_idx
, alloc_flags
))
1909 if (IS_ENABLED(CONFIG_NUMA
) &&
1910 !did_zlc_setup
&& nr_online_nodes
> 1) {
1912 * we do zlc_setup if there are multiple nodes
1913 * and before considering the first zone allowed
1916 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1921 if (zone_reclaim_mode
== 0 ||
1922 !zone_allows_reclaim(preferred_zone
, zone
))
1923 goto this_zone_full
;
1926 * As we may have just activated ZLC, check if the first
1927 * eligible zone has failed zone_reclaim recently.
1929 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1930 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1933 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1935 case ZONE_RECLAIM_NOSCAN
:
1938 case ZONE_RECLAIM_FULL
:
1939 /* scanned but unreclaimable */
1942 /* did we reclaim enough */
1943 if (!zone_watermark_ok(zone
, order
, mark
,
1944 classzone_idx
, alloc_flags
))
1945 goto this_zone_full
;
1950 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1951 gfp_mask
, migratetype
);
1955 if (IS_ENABLED(CONFIG_NUMA
))
1956 zlc_mark_zone_full(zonelist
, z
);
1959 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && page
== NULL
&& zlc_active
)) {
1960 /* Disable zlc cache for second zonelist scan */
1967 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1968 * necessary to allocate the page. The expectation is
1969 * that the caller is taking steps that will free more
1970 * memory. The caller should avoid the page being used
1971 * for !PFMEMALLOC purposes.
1973 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
1979 * Large machines with many possible nodes should not always dump per-node
1980 * meminfo in irq context.
1982 static inline bool should_suppress_show_mem(void)
1987 ret
= in_interrupt();
1992 static DEFINE_RATELIMIT_STATE(nopage_rs
,
1993 DEFAULT_RATELIMIT_INTERVAL
,
1994 DEFAULT_RATELIMIT_BURST
);
1996 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
1998 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2000 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2001 debug_guardpage_minorder() > 0)
2005 * This documents exceptions given to allocations in certain
2006 * contexts that are allowed to allocate outside current's set
2009 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2010 if (test_thread_flag(TIF_MEMDIE
) ||
2011 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2012 filter
&= ~SHOW_MEM_FILTER_NODES
;
2013 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2014 filter
&= ~SHOW_MEM_FILTER_NODES
;
2017 struct va_format vaf
;
2020 va_start(args
, fmt
);
2025 pr_warn("%pV", &vaf
);
2030 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2031 current
->comm
, order
, gfp_mask
);
2034 if (!should_suppress_show_mem())
2039 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2040 unsigned long did_some_progress
,
2041 unsigned long pages_reclaimed
)
2043 /* Do not loop if specifically requested */
2044 if (gfp_mask
& __GFP_NORETRY
)
2047 /* Always retry if specifically requested */
2048 if (gfp_mask
& __GFP_NOFAIL
)
2052 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2053 * making forward progress without invoking OOM. Suspend also disables
2054 * storage devices so kswapd will not help. Bail if we are suspending.
2056 if (!did_some_progress
&& pm_suspended_storage())
2060 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2061 * means __GFP_NOFAIL, but that may not be true in other
2064 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2068 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2069 * specified, then we retry until we no longer reclaim any pages
2070 * (above), or we've reclaimed an order of pages at least as
2071 * large as the allocation's order. In both cases, if the
2072 * allocation still fails, we stop retrying.
2074 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2080 static inline struct page
*
2081 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2082 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2083 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2088 /* Acquire the OOM killer lock for the zones in zonelist */
2089 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
2090 schedule_timeout_uninterruptible(1);
2095 * Go through the zonelist yet one more time, keep very high watermark
2096 * here, this is only to catch a parallel oom killing, we must fail if
2097 * we're still under heavy pressure.
2099 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2100 order
, zonelist
, high_zoneidx
,
2101 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2102 preferred_zone
, migratetype
);
2106 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2107 /* The OOM killer will not help higher order allocs */
2108 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2110 /* The OOM killer does not needlessly kill tasks for lowmem */
2111 if (high_zoneidx
< ZONE_NORMAL
)
2114 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2115 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2116 * The caller should handle page allocation failure by itself if
2117 * it specifies __GFP_THISNODE.
2118 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2120 if (gfp_mask
& __GFP_THISNODE
)
2123 /* Exhausted what can be done so it's blamo time */
2124 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2127 clear_zonelist_oom(zonelist
, gfp_mask
);
2131 #ifdef CONFIG_COMPACTION
2132 /* Try memory compaction for high-order allocations before reclaim */
2133 static struct page
*
2134 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2135 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2136 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2137 int migratetype
, bool sync_migration
,
2138 bool *contended_compaction
, bool *deferred_compaction
,
2139 unsigned long *did_some_progress
)
2144 if (compaction_deferred(preferred_zone
, order
)) {
2145 *deferred_compaction
= true;
2149 current
->flags
|= PF_MEMALLOC
;
2150 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2151 nodemask
, sync_migration
,
2152 contended_compaction
);
2153 current
->flags
&= ~PF_MEMALLOC
;
2155 if (*did_some_progress
!= COMPACT_SKIPPED
) {
2158 /* Page migration frees to the PCP lists but we want merging */
2159 drain_pages(get_cpu());
2162 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2163 order
, zonelist
, high_zoneidx
,
2164 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2165 preferred_zone
, migratetype
);
2167 preferred_zone
->compact_blockskip_flush
= false;
2168 preferred_zone
->compact_considered
= 0;
2169 preferred_zone
->compact_defer_shift
= 0;
2170 if (order
>= preferred_zone
->compact_order_failed
)
2171 preferred_zone
->compact_order_failed
= order
+ 1;
2172 count_vm_event(COMPACTSUCCESS
);
2177 * It's bad if compaction run occurs and fails.
2178 * The most likely reason is that pages exist,
2179 * but not enough to satisfy watermarks.
2181 count_vm_event(COMPACTFAIL
);
2184 * As async compaction considers a subset of pageblocks, only
2185 * defer if the failure was a sync compaction failure.
2188 defer_compaction(preferred_zone
, order
);
2196 static inline struct page
*
2197 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2198 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2199 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2200 int migratetype
, bool sync_migration
,
2201 bool *contended_compaction
, bool *deferred_compaction
,
2202 unsigned long *did_some_progress
)
2206 #endif /* CONFIG_COMPACTION */
2208 /* Perform direct synchronous page reclaim */
2210 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2211 nodemask_t
*nodemask
)
2213 struct reclaim_state reclaim_state
;
2218 /* We now go into synchronous reclaim */
2219 cpuset_memory_pressure_bump();
2220 current
->flags
|= PF_MEMALLOC
;
2221 lockdep_set_current_reclaim_state(gfp_mask
);
2222 reclaim_state
.reclaimed_slab
= 0;
2223 current
->reclaim_state
= &reclaim_state
;
2225 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2227 current
->reclaim_state
= NULL
;
2228 lockdep_clear_current_reclaim_state();
2229 current
->flags
&= ~PF_MEMALLOC
;
2236 /* The really slow allocator path where we enter direct reclaim */
2237 static inline struct page
*
2238 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2239 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2240 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2241 int migratetype
, unsigned long *did_some_progress
)
2243 struct page
*page
= NULL
;
2244 bool drained
= false;
2246 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2248 if (unlikely(!(*did_some_progress
)))
2251 /* After successful reclaim, reconsider all zones for allocation */
2252 if (IS_ENABLED(CONFIG_NUMA
))
2253 zlc_clear_zones_full(zonelist
);
2256 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2257 zonelist
, high_zoneidx
,
2258 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2259 preferred_zone
, migratetype
);
2262 * If an allocation failed after direct reclaim, it could be because
2263 * pages are pinned on the per-cpu lists. Drain them and try again
2265 if (!page
&& !drained
) {
2275 * This is called in the allocator slow-path if the allocation request is of
2276 * sufficient urgency to ignore watermarks and take other desperate measures
2278 static inline struct page
*
2279 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2280 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2281 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2287 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2288 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2289 preferred_zone
, migratetype
);
2291 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2292 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2293 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2299 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
2300 enum zone_type high_zoneidx
,
2301 enum zone_type classzone_idx
)
2306 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
2307 wakeup_kswapd(zone
, order
, classzone_idx
);
2311 gfp_to_alloc_flags(gfp_t gfp_mask
)
2313 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2314 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2316 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2317 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2320 * The caller may dip into page reserves a bit more if the caller
2321 * cannot run direct reclaim, or if the caller has realtime scheduling
2322 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2323 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2325 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2329 * Not worth trying to allocate harder for
2330 * __GFP_NOMEMALLOC even if it can't schedule.
2332 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2333 alloc_flags
|= ALLOC_HARDER
;
2335 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2336 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2338 alloc_flags
&= ~ALLOC_CPUSET
;
2339 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2340 alloc_flags
|= ALLOC_HARDER
;
2342 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2343 if (gfp_mask
& __GFP_MEMALLOC
)
2344 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2345 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2346 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2347 else if (!in_interrupt() &&
2348 ((current
->flags
& PF_MEMALLOC
) ||
2349 unlikely(test_thread_flag(TIF_MEMDIE
))))
2350 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2353 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2354 alloc_flags
|= ALLOC_CMA
;
2359 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2361 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2364 static inline struct page
*
2365 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2366 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2367 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2370 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2371 struct page
*page
= NULL
;
2373 unsigned long pages_reclaimed
= 0;
2374 unsigned long did_some_progress
;
2375 bool sync_migration
= false;
2376 bool deferred_compaction
= false;
2377 bool contended_compaction
= false;
2380 * In the slowpath, we sanity check order to avoid ever trying to
2381 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2382 * be using allocators in order of preference for an area that is
2385 if (order
>= MAX_ORDER
) {
2386 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2391 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2392 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2393 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2394 * using a larger set of nodes after it has established that the
2395 * allowed per node queues are empty and that nodes are
2398 if (IS_ENABLED(CONFIG_NUMA
) &&
2399 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2403 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2404 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2405 zone_idx(preferred_zone
));
2408 * OK, we're below the kswapd watermark and have kicked background
2409 * reclaim. Now things get more complex, so set up alloc_flags according
2410 * to how we want to proceed.
2412 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2415 * Find the true preferred zone if the allocation is unconstrained by
2418 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2419 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2423 /* This is the last chance, in general, before the goto nopage. */
2424 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2425 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2426 preferred_zone
, migratetype
);
2430 /* Allocate without watermarks if the context allows */
2431 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2433 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2434 * the allocation is high priority and these type of
2435 * allocations are system rather than user orientated
2437 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2439 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2440 zonelist
, high_zoneidx
, nodemask
,
2441 preferred_zone
, migratetype
);
2447 /* Atomic allocations - we can't balance anything */
2451 /* Avoid recursion of direct reclaim */
2452 if (current
->flags
& PF_MEMALLOC
)
2455 /* Avoid allocations with no watermarks from looping endlessly */
2456 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2460 * Try direct compaction. The first pass is asynchronous. Subsequent
2461 * attempts after direct reclaim are synchronous
2463 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2464 zonelist
, high_zoneidx
,
2466 alloc_flags
, preferred_zone
,
2467 migratetype
, sync_migration
,
2468 &contended_compaction
,
2469 &deferred_compaction
,
2470 &did_some_progress
);
2473 sync_migration
= true;
2476 * If compaction is deferred for high-order allocations, it is because
2477 * sync compaction recently failed. In this is the case and the caller
2478 * requested a movable allocation that does not heavily disrupt the
2479 * system then fail the allocation instead of entering direct reclaim.
2481 if ((deferred_compaction
|| contended_compaction
) &&
2482 (gfp_mask
& __GFP_NO_KSWAPD
))
2485 /* Try direct reclaim and then allocating */
2486 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2487 zonelist
, high_zoneidx
,
2489 alloc_flags
, preferred_zone
,
2490 migratetype
, &did_some_progress
);
2495 * If we failed to make any progress reclaiming, then we are
2496 * running out of options and have to consider going OOM
2498 if (!did_some_progress
) {
2499 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2500 if (oom_killer_disabled
)
2502 /* Coredumps can quickly deplete all memory reserves */
2503 if ((current
->flags
& PF_DUMPCORE
) &&
2504 !(gfp_mask
& __GFP_NOFAIL
))
2506 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2507 zonelist
, high_zoneidx
,
2508 nodemask
, preferred_zone
,
2513 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2515 * The oom killer is not called for high-order
2516 * allocations that may fail, so if no progress
2517 * is being made, there are no other options and
2518 * retrying is unlikely to help.
2520 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2523 * The oom killer is not called for lowmem
2524 * allocations to prevent needlessly killing
2527 if (high_zoneidx
< ZONE_NORMAL
)
2535 /* Check if we should retry the allocation */
2536 pages_reclaimed
+= did_some_progress
;
2537 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2539 /* Wait for some write requests to complete then retry */
2540 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2544 * High-order allocations do not necessarily loop after
2545 * direct reclaim and reclaim/compaction depends on compaction
2546 * being called after reclaim so call directly if necessary
2548 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2549 zonelist
, high_zoneidx
,
2551 alloc_flags
, preferred_zone
,
2552 migratetype
, sync_migration
,
2553 &contended_compaction
,
2554 &deferred_compaction
,
2555 &did_some_progress
);
2561 warn_alloc_failed(gfp_mask
, order
, NULL
);
2564 if (kmemcheck_enabled
)
2565 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2571 * This is the 'heart' of the zoned buddy allocator.
2574 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2575 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2577 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2578 struct zone
*preferred_zone
;
2579 struct page
*page
= NULL
;
2580 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2581 unsigned int cpuset_mems_cookie
;
2582 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
;
2583 struct mem_cgroup
*memcg
= NULL
;
2585 gfp_mask
&= gfp_allowed_mask
;
2587 lockdep_trace_alloc(gfp_mask
);
2589 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2591 if (should_fail_alloc_page(gfp_mask
, order
))
2595 * Check the zones suitable for the gfp_mask contain at least one
2596 * valid zone. It's possible to have an empty zonelist as a result
2597 * of GFP_THISNODE and a memoryless node
2599 if (unlikely(!zonelist
->_zonerefs
->zone
))
2603 * Will only have any effect when __GFP_KMEMCG is set. This is
2604 * verified in the (always inline) callee
2606 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2610 cpuset_mems_cookie
= get_mems_allowed();
2612 /* The preferred zone is used for statistics later */
2613 first_zones_zonelist(zonelist
, high_zoneidx
,
2614 nodemask
? : &cpuset_current_mems_allowed
,
2616 if (!preferred_zone
)
2620 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2621 alloc_flags
|= ALLOC_CMA
;
2623 /* First allocation attempt */
2624 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2625 zonelist
, high_zoneidx
, alloc_flags
,
2626 preferred_zone
, migratetype
);
2627 if (unlikely(!page
)) {
2629 * Runtime PM, block IO and its error handling path
2630 * can deadlock because I/O on the device might not
2633 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2634 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2635 zonelist
, high_zoneidx
, nodemask
,
2636 preferred_zone
, migratetype
);
2639 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2643 * When updating a task's mems_allowed, it is possible to race with
2644 * parallel threads in such a way that an allocation can fail while
2645 * the mask is being updated. If a page allocation is about to fail,
2646 * check if the cpuset changed during allocation and if so, retry.
2648 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2651 memcg_kmem_commit_charge(page
, memcg
, order
);
2655 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2658 * Common helper functions.
2660 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2665 * __get_free_pages() returns a 32-bit address, which cannot represent
2668 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2670 page
= alloc_pages(gfp_mask
, order
);
2673 return (unsigned long) page_address(page
);
2675 EXPORT_SYMBOL(__get_free_pages
);
2677 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2679 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2681 EXPORT_SYMBOL(get_zeroed_page
);
2683 void __free_pages(struct page
*page
, unsigned int order
)
2685 if (put_page_testzero(page
)) {
2687 free_hot_cold_page(page
, 0);
2689 __free_pages_ok(page
, order
);
2693 EXPORT_SYMBOL(__free_pages
);
2695 void free_pages(unsigned long addr
, unsigned int order
)
2698 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2699 __free_pages(virt_to_page((void *)addr
), order
);
2703 EXPORT_SYMBOL(free_pages
);
2706 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2707 * pages allocated with __GFP_KMEMCG.
2709 * Those pages are accounted to a particular memcg, embedded in the
2710 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2711 * for that information only to find out that it is NULL for users who have no
2712 * interest in that whatsoever, we provide these functions.
2714 * The caller knows better which flags it relies on.
2716 void __free_memcg_kmem_pages(struct page
*page
, unsigned int order
)
2718 memcg_kmem_uncharge_pages(page
, order
);
2719 __free_pages(page
, order
);
2722 void free_memcg_kmem_pages(unsigned long addr
, unsigned int order
)
2725 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2726 __free_memcg_kmem_pages(virt_to_page((void *)addr
), order
);
2730 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2733 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2734 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2736 split_page(virt_to_page((void *)addr
), order
);
2737 while (used
< alloc_end
) {
2742 return (void *)addr
;
2746 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2747 * @size: the number of bytes to allocate
2748 * @gfp_mask: GFP flags for the allocation
2750 * This function is similar to alloc_pages(), except that it allocates the
2751 * minimum number of pages to satisfy the request. alloc_pages() can only
2752 * allocate memory in power-of-two pages.
2754 * This function is also limited by MAX_ORDER.
2756 * Memory allocated by this function must be released by free_pages_exact().
2758 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2760 unsigned int order
= get_order(size
);
2763 addr
= __get_free_pages(gfp_mask
, order
);
2764 return make_alloc_exact(addr
, order
, size
);
2766 EXPORT_SYMBOL(alloc_pages_exact
);
2769 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2771 * @nid: the preferred node ID where memory should be allocated
2772 * @size: the number of bytes to allocate
2773 * @gfp_mask: GFP flags for the allocation
2775 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2777 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2780 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2782 unsigned order
= get_order(size
);
2783 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2786 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2788 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2791 * free_pages_exact - release memory allocated via alloc_pages_exact()
2792 * @virt: the value returned by alloc_pages_exact.
2793 * @size: size of allocation, same value as passed to alloc_pages_exact().
2795 * Release the memory allocated by a previous call to alloc_pages_exact.
2797 void free_pages_exact(void *virt
, size_t size
)
2799 unsigned long addr
= (unsigned long)virt
;
2800 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2802 while (addr
< end
) {
2807 EXPORT_SYMBOL(free_pages_exact
);
2809 static unsigned int nr_free_zone_pages(int offset
)
2814 /* Just pick one node, since fallback list is circular */
2815 unsigned int sum
= 0;
2817 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2819 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2820 unsigned long size
= zone
->managed_pages
;
2821 unsigned long high
= high_wmark_pages(zone
);
2830 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2832 unsigned int nr_free_buffer_pages(void)
2834 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2836 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2839 * Amount of free RAM allocatable within all zones
2841 unsigned int nr_free_pagecache_pages(void)
2843 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2846 static inline void show_node(struct zone
*zone
)
2848 if (IS_ENABLED(CONFIG_NUMA
))
2849 printk("Node %d ", zone_to_nid(zone
));
2852 void si_meminfo(struct sysinfo
*val
)
2854 val
->totalram
= totalram_pages
;
2856 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2857 val
->bufferram
= nr_blockdev_pages();
2858 val
->totalhigh
= totalhigh_pages
;
2859 val
->freehigh
= nr_free_highpages();
2860 val
->mem_unit
= PAGE_SIZE
;
2863 EXPORT_SYMBOL(si_meminfo
);
2866 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2868 pg_data_t
*pgdat
= NODE_DATA(nid
);
2870 val
->totalram
= pgdat
->node_present_pages
;
2871 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2872 #ifdef CONFIG_HIGHMEM
2873 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
2874 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2880 val
->mem_unit
= PAGE_SIZE
;
2885 * Determine whether the node should be displayed or not, depending on whether
2886 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2888 bool skip_free_areas_node(unsigned int flags
, int nid
)
2891 unsigned int cpuset_mems_cookie
;
2893 if (!(flags
& SHOW_MEM_FILTER_NODES
))
2897 cpuset_mems_cookie
= get_mems_allowed();
2898 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
2899 } while (!put_mems_allowed(cpuset_mems_cookie
));
2904 #define K(x) ((x) << (PAGE_SHIFT-10))
2906 static void show_migration_types(unsigned char type
)
2908 static const char types
[MIGRATE_TYPES
] = {
2909 [MIGRATE_UNMOVABLE
] = 'U',
2910 [MIGRATE_RECLAIMABLE
] = 'E',
2911 [MIGRATE_MOVABLE
] = 'M',
2912 [MIGRATE_RESERVE
] = 'R',
2914 [MIGRATE_CMA
] = 'C',
2916 #ifdef CONFIG_MEMORY_ISOLATION
2917 [MIGRATE_ISOLATE
] = 'I',
2920 char tmp
[MIGRATE_TYPES
+ 1];
2924 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
2925 if (type
& (1 << i
))
2930 printk("(%s) ", tmp
);
2934 * Show free area list (used inside shift_scroll-lock stuff)
2935 * We also calculate the percentage fragmentation. We do this by counting the
2936 * memory on each free list with the exception of the first item on the list.
2937 * Suppresses nodes that are not allowed by current's cpuset if
2938 * SHOW_MEM_FILTER_NODES is passed.
2940 void show_free_areas(unsigned int filter
)
2945 for_each_populated_zone(zone
) {
2946 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2949 printk("%s per-cpu:\n", zone
->name
);
2951 for_each_online_cpu(cpu
) {
2952 struct per_cpu_pageset
*pageset
;
2954 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
2956 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2957 cpu
, pageset
->pcp
.high
,
2958 pageset
->pcp
.batch
, pageset
->pcp
.count
);
2962 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2963 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2965 " dirty:%lu writeback:%lu unstable:%lu\n"
2966 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2967 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
2969 global_page_state(NR_ACTIVE_ANON
),
2970 global_page_state(NR_INACTIVE_ANON
),
2971 global_page_state(NR_ISOLATED_ANON
),
2972 global_page_state(NR_ACTIVE_FILE
),
2973 global_page_state(NR_INACTIVE_FILE
),
2974 global_page_state(NR_ISOLATED_FILE
),
2975 global_page_state(NR_UNEVICTABLE
),
2976 global_page_state(NR_FILE_DIRTY
),
2977 global_page_state(NR_WRITEBACK
),
2978 global_page_state(NR_UNSTABLE_NFS
),
2979 global_page_state(NR_FREE_PAGES
),
2980 global_page_state(NR_SLAB_RECLAIMABLE
),
2981 global_page_state(NR_SLAB_UNRECLAIMABLE
),
2982 global_page_state(NR_FILE_MAPPED
),
2983 global_page_state(NR_SHMEM
),
2984 global_page_state(NR_PAGETABLE
),
2985 global_page_state(NR_BOUNCE
),
2986 global_page_state(NR_FREE_CMA_PAGES
));
2988 for_each_populated_zone(zone
) {
2991 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
2999 " active_anon:%lukB"
3000 " inactive_anon:%lukB"
3001 " active_file:%lukB"
3002 " inactive_file:%lukB"
3003 " unevictable:%lukB"
3004 " isolated(anon):%lukB"
3005 " isolated(file):%lukB"
3013 " slab_reclaimable:%lukB"
3014 " slab_unreclaimable:%lukB"
3015 " kernel_stack:%lukB"
3020 " writeback_tmp:%lukB"
3021 " pages_scanned:%lu"
3022 " all_unreclaimable? %s"
3025 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3026 K(min_wmark_pages(zone
)),
3027 K(low_wmark_pages(zone
)),
3028 K(high_wmark_pages(zone
)),
3029 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3030 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3031 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3032 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3033 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3034 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3035 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3036 K(zone
->present_pages
),
3037 K(zone
->managed_pages
),
3038 K(zone_page_state(zone
, NR_MLOCK
)),
3039 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3040 K(zone_page_state(zone
, NR_WRITEBACK
)),
3041 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3042 K(zone_page_state(zone
, NR_SHMEM
)),
3043 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3044 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3045 zone_page_state(zone
, NR_KERNEL_STACK
) *
3047 K(zone_page_state(zone
, NR_PAGETABLE
)),
3048 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3049 K(zone_page_state(zone
, NR_BOUNCE
)),
3050 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3051 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3052 zone
->pages_scanned
,
3053 (zone
->all_unreclaimable
? "yes" : "no")
3055 printk("lowmem_reserve[]:");
3056 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3057 printk(" %lu", zone
->lowmem_reserve
[i
]);
3061 for_each_populated_zone(zone
) {
3062 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3063 unsigned char types
[MAX_ORDER
];
3065 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3068 printk("%s: ", zone
->name
);
3070 spin_lock_irqsave(&zone
->lock
, flags
);
3071 for (order
= 0; order
< MAX_ORDER
; order
++) {
3072 struct free_area
*area
= &zone
->free_area
[order
];
3075 nr
[order
] = area
->nr_free
;
3076 total
+= nr
[order
] << order
;
3079 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3080 if (!list_empty(&area
->free_list
[type
]))
3081 types
[order
] |= 1 << type
;
3084 spin_unlock_irqrestore(&zone
->lock
, flags
);
3085 for (order
= 0; order
< MAX_ORDER
; order
++) {
3086 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3088 show_migration_types(types
[order
]);
3090 printk("= %lukB\n", K(total
));
3093 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3095 show_swap_cache_info();
3098 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3100 zoneref
->zone
= zone
;
3101 zoneref
->zone_idx
= zone_idx(zone
);
3105 * Builds allocation fallback zone lists.
3107 * Add all populated zones of a node to the zonelist.
3109 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3110 int nr_zones
, enum zone_type zone_type
)
3114 BUG_ON(zone_type
>= MAX_NR_ZONES
);
3119 zone
= pgdat
->node_zones
+ zone_type
;
3120 if (populated_zone(zone
)) {
3121 zoneref_set_zone(zone
,
3122 &zonelist
->_zonerefs
[nr_zones
++]);
3123 check_highest_zone(zone_type
);
3126 } while (zone_type
);
3133 * 0 = automatic detection of better ordering.
3134 * 1 = order by ([node] distance, -zonetype)
3135 * 2 = order by (-zonetype, [node] distance)
3137 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3138 * the same zonelist. So only NUMA can configure this param.
3140 #define ZONELIST_ORDER_DEFAULT 0
3141 #define ZONELIST_ORDER_NODE 1
3142 #define ZONELIST_ORDER_ZONE 2
3144 /* zonelist order in the kernel.
3145 * set_zonelist_order() will set this to NODE or ZONE.
3147 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3148 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3152 /* The value user specified ....changed by config */
3153 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3154 /* string for sysctl */
3155 #define NUMA_ZONELIST_ORDER_LEN 16
3156 char numa_zonelist_order
[16] = "default";
3159 * interface for configure zonelist ordering.
3160 * command line option "numa_zonelist_order"
3161 * = "[dD]efault - default, automatic configuration.
3162 * = "[nN]ode - order by node locality, then by zone within node
3163 * = "[zZ]one - order by zone, then by locality within zone
3166 static int __parse_numa_zonelist_order(char *s
)
3168 if (*s
== 'd' || *s
== 'D') {
3169 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3170 } else if (*s
== 'n' || *s
== 'N') {
3171 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3172 } else if (*s
== 'z' || *s
== 'Z') {
3173 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3176 "Ignoring invalid numa_zonelist_order value: "
3183 static __init
int setup_numa_zonelist_order(char *s
)
3190 ret
= __parse_numa_zonelist_order(s
);
3192 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3196 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3199 * sysctl handler for numa_zonelist_order
3201 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
3202 void __user
*buffer
, size_t *length
,
3205 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3207 static DEFINE_MUTEX(zl_order_mutex
);
3209 mutex_lock(&zl_order_mutex
);
3211 strcpy(saved_string
, (char*)table
->data
);
3212 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3216 int oldval
= user_zonelist_order
;
3217 if (__parse_numa_zonelist_order((char*)table
->data
)) {
3219 * bogus value. restore saved string
3221 strncpy((char*)table
->data
, saved_string
,
3222 NUMA_ZONELIST_ORDER_LEN
);
3223 user_zonelist_order
= oldval
;
3224 } else if (oldval
!= user_zonelist_order
) {
3225 mutex_lock(&zonelists_mutex
);
3226 build_all_zonelists(NULL
, NULL
);
3227 mutex_unlock(&zonelists_mutex
);
3231 mutex_unlock(&zl_order_mutex
);
3236 #define MAX_NODE_LOAD (nr_online_nodes)
3237 static int node_load
[MAX_NUMNODES
];
3240 * find_next_best_node - find the next node that should appear in a given node's fallback list
3241 * @node: node whose fallback list we're appending
3242 * @used_node_mask: nodemask_t of already used nodes
3244 * We use a number of factors to determine which is the next node that should
3245 * appear on a given node's fallback list. The node should not have appeared
3246 * already in @node's fallback list, and it should be the next closest node
3247 * according to the distance array (which contains arbitrary distance values
3248 * from each node to each node in the system), and should also prefer nodes
3249 * with no CPUs, since presumably they'll have very little allocation pressure
3250 * on them otherwise.
3251 * It returns -1 if no node is found.
3253 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3256 int min_val
= INT_MAX
;
3258 const struct cpumask
*tmp
= cpumask_of_node(0);
3260 /* Use the local node if we haven't already */
3261 if (!node_isset(node
, *used_node_mask
)) {
3262 node_set(node
, *used_node_mask
);
3266 for_each_node_state(n
, N_MEMORY
) {
3268 /* Don't want a node to appear more than once */
3269 if (node_isset(n
, *used_node_mask
))
3272 /* Use the distance array to find the distance */
3273 val
= node_distance(node
, n
);
3275 /* Penalize nodes under us ("prefer the next node") */
3278 /* Give preference to headless and unused nodes */
3279 tmp
= cpumask_of_node(n
);
3280 if (!cpumask_empty(tmp
))
3281 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3283 /* Slight preference for less loaded node */
3284 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3285 val
+= node_load
[n
];
3287 if (val
< min_val
) {
3294 node_set(best_node
, *used_node_mask
);
3301 * Build zonelists ordered by node and zones within node.
3302 * This results in maximum locality--normal zone overflows into local
3303 * DMA zone, if any--but risks exhausting DMA zone.
3305 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3308 struct zonelist
*zonelist
;
3310 zonelist
= &pgdat
->node_zonelists
[0];
3311 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3313 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3315 zonelist
->_zonerefs
[j
].zone
= NULL
;
3316 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3320 * Build gfp_thisnode zonelists
3322 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3325 struct zonelist
*zonelist
;
3327 zonelist
= &pgdat
->node_zonelists
[1];
3328 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3329 zonelist
->_zonerefs
[j
].zone
= NULL
;
3330 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3334 * Build zonelists ordered by zone and nodes within zones.
3335 * This results in conserving DMA zone[s] until all Normal memory is
3336 * exhausted, but results in overflowing to remote node while memory
3337 * may still exist in local DMA zone.
3339 static int node_order
[MAX_NUMNODES
];
3341 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3344 int zone_type
; /* needs to be signed */
3346 struct zonelist
*zonelist
;
3348 zonelist
= &pgdat
->node_zonelists
[0];
3350 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3351 for (j
= 0; j
< nr_nodes
; j
++) {
3352 node
= node_order
[j
];
3353 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3354 if (populated_zone(z
)) {
3356 &zonelist
->_zonerefs
[pos
++]);
3357 check_highest_zone(zone_type
);
3361 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3362 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3365 static int default_zonelist_order(void)
3368 unsigned long low_kmem_size
,total_size
;
3372 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3373 * If they are really small and used heavily, the system can fall
3374 * into OOM very easily.
3375 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3377 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3380 for_each_online_node(nid
) {
3381 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3382 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3383 if (populated_zone(z
)) {
3384 if (zone_type
< ZONE_NORMAL
)
3385 low_kmem_size
+= z
->present_pages
;
3386 total_size
+= z
->present_pages
;
3387 } else if (zone_type
== ZONE_NORMAL
) {
3389 * If any node has only lowmem, then node order
3390 * is preferred to allow kernel allocations
3391 * locally; otherwise, they can easily infringe
3392 * on other nodes when there is an abundance of
3393 * lowmem available to allocate from.
3395 return ZONELIST_ORDER_NODE
;
3399 if (!low_kmem_size
|| /* there are no DMA area. */
3400 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
3401 return ZONELIST_ORDER_NODE
;
3403 * look into each node's config.
3404 * If there is a node whose DMA/DMA32 memory is very big area on
3405 * local memory, NODE_ORDER may be suitable.
3407 average_size
= total_size
/
3408 (nodes_weight(node_states
[N_MEMORY
]) + 1);
3409 for_each_online_node(nid
) {
3412 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3413 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3414 if (populated_zone(z
)) {
3415 if (zone_type
< ZONE_NORMAL
)
3416 low_kmem_size
+= z
->present_pages
;
3417 total_size
+= z
->present_pages
;
3420 if (low_kmem_size
&&
3421 total_size
> average_size
&& /* ignore small node */
3422 low_kmem_size
> total_size
* 70/100)
3423 return ZONELIST_ORDER_NODE
;
3425 return ZONELIST_ORDER_ZONE
;
3428 static void set_zonelist_order(void)
3430 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3431 current_zonelist_order
= default_zonelist_order();
3433 current_zonelist_order
= user_zonelist_order
;
3436 static void build_zonelists(pg_data_t
*pgdat
)
3440 nodemask_t used_mask
;
3441 int local_node
, prev_node
;
3442 struct zonelist
*zonelist
;
3443 int order
= current_zonelist_order
;
3445 /* initialize zonelists */
3446 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3447 zonelist
= pgdat
->node_zonelists
+ i
;
3448 zonelist
->_zonerefs
[0].zone
= NULL
;
3449 zonelist
->_zonerefs
[0].zone_idx
= 0;
3452 /* NUMA-aware ordering of nodes */
3453 local_node
= pgdat
->node_id
;
3454 load
= nr_online_nodes
;
3455 prev_node
= local_node
;
3456 nodes_clear(used_mask
);
3458 memset(node_order
, 0, sizeof(node_order
));
3461 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3463 * We don't want to pressure a particular node.
3464 * So adding penalty to the first node in same
3465 * distance group to make it round-robin.
3467 if (node_distance(local_node
, node
) !=
3468 node_distance(local_node
, prev_node
))
3469 node_load
[node
] = load
;
3473 if (order
== ZONELIST_ORDER_NODE
)
3474 build_zonelists_in_node_order(pgdat
, node
);
3476 node_order
[j
++] = node
; /* remember order */
3479 if (order
== ZONELIST_ORDER_ZONE
) {
3480 /* calculate node order -- i.e., DMA last! */
3481 build_zonelists_in_zone_order(pgdat
, j
);
3484 build_thisnode_zonelists(pgdat
);
3487 /* Construct the zonelist performance cache - see further mmzone.h */
3488 static void build_zonelist_cache(pg_data_t
*pgdat
)
3490 struct zonelist
*zonelist
;
3491 struct zonelist_cache
*zlc
;
3494 zonelist
= &pgdat
->node_zonelists
[0];
3495 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3496 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3497 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3498 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3501 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3503 * Return node id of node used for "local" allocations.
3504 * I.e., first node id of first zone in arg node's generic zonelist.
3505 * Used for initializing percpu 'numa_mem', which is used primarily
3506 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3508 int local_memory_node(int node
)
3512 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3513 gfp_zone(GFP_KERNEL
),
3520 #else /* CONFIG_NUMA */
3522 static void set_zonelist_order(void)
3524 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3527 static void build_zonelists(pg_data_t
*pgdat
)
3529 int node
, local_node
;
3531 struct zonelist
*zonelist
;
3533 local_node
= pgdat
->node_id
;
3535 zonelist
= &pgdat
->node_zonelists
[0];
3536 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3539 * Now we build the zonelist so that it contains the zones
3540 * of all the other nodes.
3541 * We don't want to pressure a particular node, so when
3542 * building the zones for node N, we make sure that the
3543 * zones coming right after the local ones are those from
3544 * node N+1 (modulo N)
3546 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3547 if (!node_online(node
))
3549 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3552 for (node
= 0; node
< local_node
; node
++) {
3553 if (!node_online(node
))
3555 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3559 zonelist
->_zonerefs
[j
].zone
= NULL
;
3560 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3563 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3564 static void build_zonelist_cache(pg_data_t
*pgdat
)
3566 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3569 #endif /* CONFIG_NUMA */
3572 * Boot pageset table. One per cpu which is going to be used for all
3573 * zones and all nodes. The parameters will be set in such a way
3574 * that an item put on a list will immediately be handed over to
3575 * the buddy list. This is safe since pageset manipulation is done
3576 * with interrupts disabled.
3578 * The boot_pagesets must be kept even after bootup is complete for
3579 * unused processors and/or zones. They do play a role for bootstrapping
3580 * hotplugged processors.
3582 * zoneinfo_show() and maybe other functions do
3583 * not check if the processor is online before following the pageset pointer.
3584 * Other parts of the kernel may not check if the zone is available.
3586 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3587 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3588 static void setup_zone_pageset(struct zone
*zone
);
3591 * Global mutex to protect against size modification of zonelists
3592 * as well as to serialize pageset setup for the new populated zone.
3594 DEFINE_MUTEX(zonelists_mutex
);
3596 /* return values int ....just for stop_machine() */
3597 static int __build_all_zonelists(void *data
)
3601 pg_data_t
*self
= data
;
3604 memset(node_load
, 0, sizeof(node_load
));
3607 if (self
&& !node_online(self
->node_id
)) {
3608 build_zonelists(self
);
3609 build_zonelist_cache(self
);
3612 for_each_online_node(nid
) {
3613 pg_data_t
*pgdat
= NODE_DATA(nid
);
3615 build_zonelists(pgdat
);
3616 build_zonelist_cache(pgdat
);
3620 * Initialize the boot_pagesets that are going to be used
3621 * for bootstrapping processors. The real pagesets for
3622 * each zone will be allocated later when the per cpu
3623 * allocator is available.
3625 * boot_pagesets are used also for bootstrapping offline
3626 * cpus if the system is already booted because the pagesets
3627 * are needed to initialize allocators on a specific cpu too.
3628 * F.e. the percpu allocator needs the page allocator which
3629 * needs the percpu allocator in order to allocate its pagesets
3630 * (a chicken-egg dilemma).
3632 for_each_possible_cpu(cpu
) {
3633 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3635 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3637 * We now know the "local memory node" for each node--
3638 * i.e., the node of the first zone in the generic zonelist.
3639 * Set up numa_mem percpu variable for on-line cpus. During
3640 * boot, only the boot cpu should be on-line; we'll init the
3641 * secondary cpus' numa_mem as they come on-line. During
3642 * node/memory hotplug, we'll fixup all on-line cpus.
3644 if (cpu_online(cpu
))
3645 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3653 * Called with zonelists_mutex held always
3654 * unless system_state == SYSTEM_BOOTING.
3656 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3658 set_zonelist_order();
3660 if (system_state
== SYSTEM_BOOTING
) {
3661 __build_all_zonelists(NULL
);
3662 mminit_verify_zonelist();
3663 cpuset_init_current_mems_allowed();
3665 /* we have to stop all cpus to guarantee there is no user
3667 #ifdef CONFIG_MEMORY_HOTPLUG
3669 setup_zone_pageset(zone
);
3671 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3672 /* cpuset refresh routine should be here */
3674 vm_total_pages
= nr_free_pagecache_pages();
3676 * Disable grouping by mobility if the number of pages in the
3677 * system is too low to allow the mechanism to work. It would be
3678 * more accurate, but expensive to check per-zone. This check is
3679 * made on memory-hotadd so a system can start with mobility
3680 * disabled and enable it later
3682 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3683 page_group_by_mobility_disabled
= 1;
3685 page_group_by_mobility_disabled
= 0;
3687 printk("Built %i zonelists in %s order, mobility grouping %s. "
3688 "Total pages: %ld\n",
3690 zonelist_order_name
[current_zonelist_order
],
3691 page_group_by_mobility_disabled
? "off" : "on",
3694 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3699 * Helper functions to size the waitqueue hash table.
3700 * Essentially these want to choose hash table sizes sufficiently
3701 * large so that collisions trying to wait on pages are rare.
3702 * But in fact, the number of active page waitqueues on typical
3703 * systems is ridiculously low, less than 200. So this is even
3704 * conservative, even though it seems large.
3706 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3707 * waitqueues, i.e. the size of the waitq table given the number of pages.
3709 #define PAGES_PER_WAITQUEUE 256
3711 #ifndef CONFIG_MEMORY_HOTPLUG
3712 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3714 unsigned long size
= 1;
3716 pages
/= PAGES_PER_WAITQUEUE
;
3718 while (size
< pages
)
3722 * Once we have dozens or even hundreds of threads sleeping
3723 * on IO we've got bigger problems than wait queue collision.
3724 * Limit the size of the wait table to a reasonable size.
3726 size
= min(size
, 4096UL);
3728 return max(size
, 4UL);
3732 * A zone's size might be changed by hot-add, so it is not possible to determine
3733 * a suitable size for its wait_table. So we use the maximum size now.
3735 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3737 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3738 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3739 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3741 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3742 * or more by the traditional way. (See above). It equals:
3744 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3745 * ia64(16K page size) : = ( 8G + 4M)byte.
3746 * powerpc (64K page size) : = (32G +16M)byte.
3748 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3755 * This is an integer logarithm so that shifts can be used later
3756 * to extract the more random high bits from the multiplicative
3757 * hash function before the remainder is taken.
3759 static inline unsigned long wait_table_bits(unsigned long size
)
3764 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3767 * Check if a pageblock contains reserved pages
3769 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3773 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3774 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3781 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3782 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3783 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3784 * higher will lead to a bigger reserve which will get freed as contiguous
3785 * blocks as reclaim kicks in
3787 static void setup_zone_migrate_reserve(struct zone
*zone
)
3789 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3791 unsigned long block_migratetype
;
3795 * Get the start pfn, end pfn and the number of blocks to reserve
3796 * We have to be careful to be aligned to pageblock_nr_pages to
3797 * make sure that we always check pfn_valid for the first page in
3800 start_pfn
= zone
->zone_start_pfn
;
3801 end_pfn
= start_pfn
+ zone
->spanned_pages
;
3802 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
3803 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3807 * Reserve blocks are generally in place to help high-order atomic
3808 * allocations that are short-lived. A min_free_kbytes value that
3809 * would result in more than 2 reserve blocks for atomic allocations
3810 * is assumed to be in place to help anti-fragmentation for the
3811 * future allocation of hugepages at runtime.
3813 reserve
= min(2, reserve
);
3815 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3816 if (!pfn_valid(pfn
))
3818 page
= pfn_to_page(pfn
);
3820 /* Watch out for overlapping nodes */
3821 if (page_to_nid(page
) != zone_to_nid(zone
))
3824 block_migratetype
= get_pageblock_migratetype(page
);
3826 /* Only test what is necessary when the reserves are not met */
3829 * Blocks with reserved pages will never free, skip
3832 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
3833 if (pageblock_is_reserved(pfn
, block_end_pfn
))
3836 /* If this block is reserved, account for it */
3837 if (block_migratetype
== MIGRATE_RESERVE
) {
3842 /* Suitable for reserving if this block is movable */
3843 if (block_migratetype
== MIGRATE_MOVABLE
) {
3844 set_pageblock_migratetype(page
,
3846 move_freepages_block(zone
, page
,
3854 * If the reserve is met and this is a previous reserved block,
3857 if (block_migratetype
== MIGRATE_RESERVE
) {
3858 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3859 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3865 * Initially all pages are reserved - free ones are freed
3866 * up by free_all_bootmem() once the early boot process is
3867 * done. Non-atomic initialization, single-pass.
3869 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3870 unsigned long start_pfn
, enum memmap_context context
)
3873 unsigned long end_pfn
= start_pfn
+ size
;
3877 if (highest_memmap_pfn
< end_pfn
- 1)
3878 highest_memmap_pfn
= end_pfn
- 1;
3880 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3881 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3883 * There can be holes in boot-time mem_map[]s
3884 * handed to this function. They do not
3885 * exist on hotplugged memory.
3887 if (context
== MEMMAP_EARLY
) {
3888 if (!early_pfn_valid(pfn
))
3890 if (!early_pfn_in_nid(pfn
, nid
))
3893 page
= pfn_to_page(pfn
);
3894 set_page_links(page
, zone
, nid
, pfn
);
3895 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3896 init_page_count(page
);
3897 reset_page_mapcount(page
);
3898 reset_page_last_nid(page
);
3899 SetPageReserved(page
);
3901 * Mark the block movable so that blocks are reserved for
3902 * movable at startup. This will force kernel allocations
3903 * to reserve their blocks rather than leaking throughout
3904 * the address space during boot when many long-lived
3905 * kernel allocations are made. Later some blocks near
3906 * the start are marked MIGRATE_RESERVE by
3907 * setup_zone_migrate_reserve()
3909 * bitmap is created for zone's valid pfn range. but memmap
3910 * can be created for invalid pages (for alignment)
3911 * check here not to call set_pageblock_migratetype() against
3914 if ((z
->zone_start_pfn
<= pfn
)
3915 && (pfn
< z
->zone_start_pfn
+ z
->spanned_pages
)
3916 && !(pfn
& (pageblock_nr_pages
- 1)))
3917 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3919 INIT_LIST_HEAD(&page
->lru
);
3920 #ifdef WANT_PAGE_VIRTUAL
3921 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3922 if (!is_highmem_idx(zone
))
3923 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3928 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3931 for_each_migratetype_order(order
, t
) {
3932 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3933 zone
->free_area
[order
].nr_free
= 0;
3937 #ifndef __HAVE_ARCH_MEMMAP_INIT
3938 #define memmap_init(size, nid, zone, start_pfn) \
3939 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3942 static int __meminit
zone_batchsize(struct zone
*zone
)
3948 * The per-cpu-pages pools are set to around 1000th of the
3949 * size of the zone. But no more than 1/2 of a meg.
3951 * OK, so we don't know how big the cache is. So guess.
3953 batch
= zone
->managed_pages
/ 1024;
3954 if (batch
* PAGE_SIZE
> 512 * 1024)
3955 batch
= (512 * 1024) / PAGE_SIZE
;
3956 batch
/= 4; /* We effectively *= 4 below */
3961 * Clamp the batch to a 2^n - 1 value. Having a power
3962 * of 2 value was found to be more likely to have
3963 * suboptimal cache aliasing properties in some cases.
3965 * For example if 2 tasks are alternately allocating
3966 * batches of pages, one task can end up with a lot
3967 * of pages of one half of the possible page colors
3968 * and the other with pages of the other colors.
3970 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
3975 /* The deferral and batching of frees should be suppressed under NOMMU
3978 * The problem is that NOMMU needs to be able to allocate large chunks
3979 * of contiguous memory as there's no hardware page translation to
3980 * assemble apparent contiguous memory from discontiguous pages.
3982 * Queueing large contiguous runs of pages for batching, however,
3983 * causes the pages to actually be freed in smaller chunks. As there
3984 * can be a significant delay between the individual batches being
3985 * recycled, this leads to the once large chunks of space being
3986 * fragmented and becoming unavailable for high-order allocations.
3992 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
3994 struct per_cpu_pages
*pcp
;
3997 memset(p
, 0, sizeof(*p
));
4001 pcp
->high
= 6 * batch
;
4002 pcp
->batch
= max(1UL, 1 * batch
);
4003 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4004 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4008 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4009 * to the value high for the pageset p.
4012 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
4015 struct per_cpu_pages
*pcp
;
4019 pcp
->batch
= max(1UL, high
/4);
4020 if ((high
/4) > (PAGE_SHIFT
* 8))
4021 pcp
->batch
= PAGE_SHIFT
* 8;
4024 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4028 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4030 for_each_possible_cpu(cpu
) {
4031 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4033 setup_pageset(pcp
, zone_batchsize(zone
));
4035 if (percpu_pagelist_fraction
)
4036 setup_pagelist_highmark(pcp
,
4037 (zone
->managed_pages
/
4038 percpu_pagelist_fraction
));
4043 * Allocate per cpu pagesets and initialize them.
4044 * Before this call only boot pagesets were available.
4046 void __init
setup_per_cpu_pageset(void)
4050 for_each_populated_zone(zone
)
4051 setup_zone_pageset(zone
);
4054 static noinline __init_refok
4055 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4058 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4062 * The per-page waitqueue mechanism uses hashed waitqueues
4065 zone
->wait_table_hash_nr_entries
=
4066 wait_table_hash_nr_entries(zone_size_pages
);
4067 zone
->wait_table_bits
=
4068 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4069 alloc_size
= zone
->wait_table_hash_nr_entries
4070 * sizeof(wait_queue_head_t
);
4072 if (!slab_is_available()) {
4073 zone
->wait_table
= (wait_queue_head_t
*)
4074 alloc_bootmem_node_nopanic(pgdat
, alloc_size
);
4077 * This case means that a zone whose size was 0 gets new memory
4078 * via memory hot-add.
4079 * But it may be the case that a new node was hot-added. In
4080 * this case vmalloc() will not be able to use this new node's
4081 * memory - this wait_table must be initialized to use this new
4082 * node itself as well.
4083 * To use this new node's memory, further consideration will be
4086 zone
->wait_table
= vmalloc(alloc_size
);
4088 if (!zone
->wait_table
)
4091 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4092 init_waitqueue_head(zone
->wait_table
+ i
);
4097 static __meminit
void zone_pcp_init(struct zone
*zone
)
4100 * per cpu subsystem is not up at this point. The following code
4101 * relies on the ability of the linker to provide the
4102 * offset of a (static) per cpu variable into the per cpu area.
4104 zone
->pageset
= &boot_pageset
;
4106 if (zone
->present_pages
)
4107 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4108 zone
->name
, zone
->present_pages
,
4109 zone_batchsize(zone
));
4112 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4113 unsigned long zone_start_pfn
,
4115 enum memmap_context context
)
4117 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4119 ret
= zone_wait_table_init(zone
, size
);
4122 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4124 zone
->zone_start_pfn
= zone_start_pfn
;
4126 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4127 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4129 (unsigned long)zone_idx(zone
),
4130 zone_start_pfn
, (zone_start_pfn
+ size
));
4132 zone_init_free_lists(zone
);
4137 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4138 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4140 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4141 * Architectures may implement their own version but if add_active_range()
4142 * was used and there are no special requirements, this is a convenient
4145 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4147 unsigned long start_pfn
, end_pfn
;
4150 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
4151 if (start_pfn
<= pfn
&& pfn
< end_pfn
)
4153 /* This is a memory hole */
4156 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4158 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4162 nid
= __early_pfn_to_nid(pfn
);
4165 /* just returns 0 */
4169 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4170 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4174 nid
= __early_pfn_to_nid(pfn
);
4175 if (nid
>= 0 && nid
!= node
)
4182 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4183 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4184 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4186 * If an architecture guarantees that all ranges registered with
4187 * add_active_ranges() contain no holes and may be freed, this
4188 * this function may be used instead of calling free_bootmem() manually.
4190 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4192 unsigned long start_pfn
, end_pfn
;
4195 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4196 start_pfn
= min(start_pfn
, max_low_pfn
);
4197 end_pfn
= min(end_pfn
, max_low_pfn
);
4199 if (start_pfn
< end_pfn
)
4200 free_bootmem_node(NODE_DATA(this_nid
),
4201 PFN_PHYS(start_pfn
),
4202 (end_pfn
- start_pfn
) << PAGE_SHIFT
);
4207 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4208 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4210 * If an architecture guarantees that all ranges registered with
4211 * add_active_ranges() contain no holes and may be freed, this
4212 * function may be used instead of calling memory_present() manually.
4214 void __init
sparse_memory_present_with_active_regions(int nid
)
4216 unsigned long start_pfn
, end_pfn
;
4219 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4220 memory_present(this_nid
, start_pfn
, end_pfn
);
4224 * get_pfn_range_for_nid - Return the start and end page frames for a node
4225 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4226 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4227 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4229 * It returns the start and end page frame of a node based on information
4230 * provided by an arch calling add_active_range(). If called for a node
4231 * with no available memory, a warning is printed and the start and end
4234 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4235 unsigned long *start_pfn
, unsigned long *end_pfn
)
4237 unsigned long this_start_pfn
, this_end_pfn
;
4243 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4244 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4245 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4248 if (*start_pfn
== -1UL)
4253 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4254 * assumption is made that zones within a node are ordered in monotonic
4255 * increasing memory addresses so that the "highest" populated zone is used
4257 static void __init
find_usable_zone_for_movable(void)
4260 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4261 if (zone_index
== ZONE_MOVABLE
)
4264 if (arch_zone_highest_possible_pfn
[zone_index
] >
4265 arch_zone_lowest_possible_pfn
[zone_index
])
4269 VM_BUG_ON(zone_index
== -1);
4270 movable_zone
= zone_index
;
4274 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4275 * because it is sized independent of architecture. Unlike the other zones,
4276 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4277 * in each node depending on the size of each node and how evenly kernelcore
4278 * is distributed. This helper function adjusts the zone ranges
4279 * provided by the architecture for a given node by using the end of the
4280 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4281 * zones within a node are in order of monotonic increases memory addresses
4283 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4284 unsigned long zone_type
,
4285 unsigned long node_start_pfn
,
4286 unsigned long node_end_pfn
,
4287 unsigned long *zone_start_pfn
,
4288 unsigned long *zone_end_pfn
)
4290 /* Only adjust if ZONE_MOVABLE is on this node */
4291 if (zone_movable_pfn
[nid
]) {
4292 /* Size ZONE_MOVABLE */
4293 if (zone_type
== ZONE_MOVABLE
) {
4294 *zone_start_pfn
= zone_movable_pfn
[nid
];
4295 *zone_end_pfn
= min(node_end_pfn
,
4296 arch_zone_highest_possible_pfn
[movable_zone
]);
4298 /* Adjust for ZONE_MOVABLE starting within this range */
4299 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4300 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4301 *zone_end_pfn
= zone_movable_pfn
[nid
];
4303 /* Check if this whole range is within ZONE_MOVABLE */
4304 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4305 *zone_start_pfn
= *zone_end_pfn
;
4310 * Return the number of pages a zone spans in a node, including holes
4311 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4313 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4314 unsigned long zone_type
,
4315 unsigned long *ignored
)
4317 unsigned long node_start_pfn
, node_end_pfn
;
4318 unsigned long zone_start_pfn
, zone_end_pfn
;
4320 /* Get the start and end of the node and zone */
4321 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4322 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4323 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4324 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4325 node_start_pfn
, node_end_pfn
,
4326 &zone_start_pfn
, &zone_end_pfn
);
4328 /* Check that this node has pages within the zone's required range */
4329 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4332 /* Move the zone boundaries inside the node if necessary */
4333 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4334 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4336 /* Return the spanned pages */
4337 return zone_end_pfn
- zone_start_pfn
;
4341 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4342 * then all holes in the requested range will be accounted for.
4344 unsigned long __meminit
__absent_pages_in_range(int nid
,
4345 unsigned long range_start_pfn
,
4346 unsigned long range_end_pfn
)
4348 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4349 unsigned long start_pfn
, end_pfn
;
4352 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4353 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4354 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4355 nr_absent
-= end_pfn
- start_pfn
;
4361 * absent_pages_in_range - Return number of page frames in holes within a range
4362 * @start_pfn: The start PFN to start searching for holes
4363 * @end_pfn: The end PFN to stop searching for holes
4365 * It returns the number of pages frames in memory holes within a range.
4367 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4368 unsigned long end_pfn
)
4370 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4373 /* Return the number of page frames in holes in a zone on a node */
4374 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4375 unsigned long zone_type
,
4376 unsigned long *ignored
)
4378 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4379 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4380 unsigned long node_start_pfn
, node_end_pfn
;
4381 unsigned long zone_start_pfn
, zone_end_pfn
;
4383 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4384 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4385 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4387 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4388 node_start_pfn
, node_end_pfn
,
4389 &zone_start_pfn
, &zone_end_pfn
);
4390 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4394 * sanitize_zone_movable_limit - Sanitize the zone_movable_limit array.
4396 * zone_movable_limit is initialized as 0. This function will try to get
4397 * the first ZONE_MOVABLE pfn of each node from movablemem_map, and
4398 * assigne them to zone_movable_limit.
4399 * zone_movable_limit[nid] == 0 means no limit for the node.
4401 * Note: Each range is represented as [start_pfn, end_pfn)
4403 static void __meminit
sanitize_zone_movable_limit(void)
4405 int map_pos
= 0, i
, nid
;
4406 unsigned long start_pfn
, end_pfn
;
4408 if (!movablemem_map
.nr_map
)
4411 /* Iterate all ranges from minimum to maximum */
4412 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4414 * If we have found lowest pfn of ZONE_MOVABLE of the node
4415 * specified by user, just go on to check next range.
4417 if (zone_movable_limit
[nid
])
4420 #ifdef CONFIG_ZONE_DMA
4421 /* Skip DMA memory. */
4422 if (start_pfn
< arch_zone_highest_possible_pfn
[ZONE_DMA
])
4423 start_pfn
= arch_zone_highest_possible_pfn
[ZONE_DMA
];
4426 #ifdef CONFIG_ZONE_DMA32
4427 /* Skip DMA32 memory. */
4428 if (start_pfn
< arch_zone_highest_possible_pfn
[ZONE_DMA32
])
4429 start_pfn
= arch_zone_highest_possible_pfn
[ZONE_DMA32
];
4432 #ifdef CONFIG_HIGHMEM
4433 /* Skip lowmem if ZONE_MOVABLE is highmem. */
4434 if (zone_movable_is_highmem() &&
4435 start_pfn
< arch_zone_lowest_possible_pfn
[ZONE_HIGHMEM
])
4436 start_pfn
= arch_zone_lowest_possible_pfn
[ZONE_HIGHMEM
];
4439 if (start_pfn
>= end_pfn
)
4442 while (map_pos
< movablemem_map
.nr_map
) {
4443 if (end_pfn
<= movablemem_map
.map
[map_pos
].start_pfn
)
4446 if (start_pfn
>= movablemem_map
.map
[map_pos
].end_pfn
) {
4452 * The start_pfn of ZONE_MOVABLE is either the minimum
4453 * pfn specified by movablemem_map, or 0, which means
4454 * the node has no ZONE_MOVABLE.
4456 zone_movable_limit
[nid
] = max(start_pfn
,
4457 movablemem_map
.map
[map_pos
].start_pfn
);
4464 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4465 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4466 unsigned long zone_type
,
4467 unsigned long *zones_size
)
4469 return zones_size
[zone_type
];
4472 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4473 unsigned long zone_type
,
4474 unsigned long *zholes_size
)
4479 return zholes_size
[zone_type
];
4481 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4483 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4484 unsigned long *zones_size
, unsigned long *zholes_size
)
4486 unsigned long realtotalpages
, totalpages
= 0;
4489 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4490 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4492 pgdat
->node_spanned_pages
= totalpages
;
4494 realtotalpages
= totalpages
;
4495 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4497 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4499 pgdat
->node_present_pages
= realtotalpages
;
4500 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4504 #ifndef CONFIG_SPARSEMEM
4506 * Calculate the size of the zone->blockflags rounded to an unsigned long
4507 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4508 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4509 * round what is now in bits to nearest long in bits, then return it in
4512 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4514 unsigned long usemapsize
;
4516 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4517 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4518 usemapsize
= usemapsize
>> pageblock_order
;
4519 usemapsize
*= NR_PAGEBLOCK_BITS
;
4520 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4522 return usemapsize
/ 8;
4525 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4527 unsigned long zone_start_pfn
,
4528 unsigned long zonesize
)
4530 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4531 zone
->pageblock_flags
= NULL
;
4533 zone
->pageblock_flags
= alloc_bootmem_node_nopanic(pgdat
,
4537 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4538 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4539 #endif /* CONFIG_SPARSEMEM */
4541 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4543 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4544 void __init
set_pageblock_order(void)
4548 /* Check that pageblock_nr_pages has not already been setup */
4549 if (pageblock_order
)
4552 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4553 order
= HUGETLB_PAGE_ORDER
;
4555 order
= MAX_ORDER
- 1;
4558 * Assume the largest contiguous order of interest is a huge page.
4559 * This value may be variable depending on boot parameters on IA64 and
4562 pageblock_order
= order
;
4564 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4567 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4568 * is unused as pageblock_order is set at compile-time. See
4569 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4572 void __init
set_pageblock_order(void)
4576 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4578 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4579 unsigned long present_pages
)
4581 unsigned long pages
= spanned_pages
;
4584 * Provide a more accurate estimation if there are holes within
4585 * the zone and SPARSEMEM is in use. If there are holes within the
4586 * zone, each populated memory region may cost us one or two extra
4587 * memmap pages due to alignment because memmap pages for each
4588 * populated regions may not naturally algined on page boundary.
4589 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4591 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4592 IS_ENABLED(CONFIG_SPARSEMEM
))
4593 pages
= present_pages
;
4595 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4599 * Set up the zone data structures:
4600 * - mark all pages reserved
4601 * - mark all memory queues empty
4602 * - clear the memory bitmaps
4604 * NOTE: pgdat should get zeroed by caller.
4606 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4607 unsigned long *zones_size
, unsigned long *zholes_size
)
4610 int nid
= pgdat
->node_id
;
4611 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4614 pgdat_resize_init(pgdat
);
4615 #ifdef CONFIG_NUMA_BALANCING
4616 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4617 pgdat
->numabalancing_migrate_nr_pages
= 0;
4618 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4620 init_waitqueue_head(&pgdat
->kswapd_wait
);
4621 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4622 pgdat_page_cgroup_init(pgdat
);
4624 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4625 struct zone
*zone
= pgdat
->node_zones
+ j
;
4626 unsigned long size
, realsize
, freesize
, memmap_pages
;
4628 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4629 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4633 * Adjust freesize so that it accounts for how much memory
4634 * is used by this zone for memmap. This affects the watermark
4635 * and per-cpu initialisations
4637 memmap_pages
= calc_memmap_size(size
, realsize
);
4638 if (freesize
>= memmap_pages
) {
4639 freesize
-= memmap_pages
;
4642 " %s zone: %lu pages used for memmap\n",
4643 zone_names
[j
], memmap_pages
);
4646 " %s zone: %lu pages exceeds freesize %lu\n",
4647 zone_names
[j
], memmap_pages
, freesize
);
4649 /* Account for reserved pages */
4650 if (j
== 0 && freesize
> dma_reserve
) {
4651 freesize
-= dma_reserve
;
4652 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4653 zone_names
[0], dma_reserve
);
4656 if (!is_highmem_idx(j
))
4657 nr_kernel_pages
+= freesize
;
4658 /* Charge for highmem memmap if there are enough kernel pages */
4659 else if (nr_kernel_pages
> memmap_pages
* 2)
4660 nr_kernel_pages
-= memmap_pages
;
4661 nr_all_pages
+= freesize
;
4663 zone
->spanned_pages
= size
;
4664 zone
->present_pages
= realsize
;
4666 * Set an approximate value for lowmem here, it will be adjusted
4667 * when the bootmem allocator frees pages into the buddy system.
4668 * And all highmem pages will be managed by the buddy system.
4670 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4673 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4675 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4677 zone
->name
= zone_names
[j
];
4678 spin_lock_init(&zone
->lock
);
4679 spin_lock_init(&zone
->lru_lock
);
4680 zone_seqlock_init(zone
);
4681 zone
->zone_pgdat
= pgdat
;
4683 zone_pcp_init(zone
);
4684 lruvec_init(&zone
->lruvec
);
4688 set_pageblock_order();
4689 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4690 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4691 size
, MEMMAP_EARLY
);
4693 memmap_init(size
, nid
, j
, zone_start_pfn
);
4694 zone_start_pfn
+= size
;
4698 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4700 /* Skip empty nodes */
4701 if (!pgdat
->node_spanned_pages
)
4704 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4705 /* ia64 gets its own node_mem_map, before this, without bootmem */
4706 if (!pgdat
->node_mem_map
) {
4707 unsigned long size
, start
, end
;
4711 * The zone's endpoints aren't required to be MAX_ORDER
4712 * aligned but the node_mem_map endpoints must be in order
4713 * for the buddy allocator to function correctly.
4715 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4716 end
= pgdat
->node_start_pfn
+ pgdat
->node_spanned_pages
;
4717 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4718 size
= (end
- start
) * sizeof(struct page
);
4719 map
= alloc_remap(pgdat
->node_id
, size
);
4721 map
= alloc_bootmem_node_nopanic(pgdat
, size
);
4722 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4724 #ifndef CONFIG_NEED_MULTIPLE_NODES
4726 * With no DISCONTIG, the global mem_map is just set as node 0's
4728 if (pgdat
== NODE_DATA(0)) {
4729 mem_map
= NODE_DATA(0)->node_mem_map
;
4730 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4731 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4732 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4733 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4736 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4739 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4740 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4742 pg_data_t
*pgdat
= NODE_DATA(nid
);
4744 /* pg_data_t should be reset to zero when it's allocated */
4745 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4747 pgdat
->node_id
= nid
;
4748 pgdat
->node_start_pfn
= node_start_pfn
;
4749 init_zone_allows_reclaim(nid
);
4750 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4752 alloc_node_mem_map(pgdat
);
4753 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4754 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4755 nid
, (unsigned long)pgdat
,
4756 (unsigned long)pgdat
->node_mem_map
);
4759 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4762 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4764 #if MAX_NUMNODES > 1
4766 * Figure out the number of possible node ids.
4768 static void __init
setup_nr_node_ids(void)
4771 unsigned int highest
= 0;
4773 for_each_node_mask(node
, node_possible_map
)
4775 nr_node_ids
= highest
+ 1;
4778 static inline void setup_nr_node_ids(void)
4784 * node_map_pfn_alignment - determine the maximum internode alignment
4786 * This function should be called after node map is populated and sorted.
4787 * It calculates the maximum power of two alignment which can distinguish
4790 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4791 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4792 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4793 * shifted, 1GiB is enough and this function will indicate so.
4795 * This is used to test whether pfn -> nid mapping of the chosen memory
4796 * model has fine enough granularity to avoid incorrect mapping for the
4797 * populated node map.
4799 * Returns the determined alignment in pfn's. 0 if there is no alignment
4800 * requirement (single node).
4802 unsigned long __init
node_map_pfn_alignment(void)
4804 unsigned long accl_mask
= 0, last_end
= 0;
4805 unsigned long start
, end
, mask
;
4809 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
4810 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
4817 * Start with a mask granular enough to pin-point to the
4818 * start pfn and tick off bits one-by-one until it becomes
4819 * too coarse to separate the current node from the last.
4821 mask
= ~((1 << __ffs(start
)) - 1);
4822 while (mask
&& last_end
<= (start
& (mask
<< 1)))
4825 /* accumulate all internode masks */
4829 /* convert mask to number of pages */
4830 return ~accl_mask
+ 1;
4833 /* Find the lowest pfn for a node */
4834 static unsigned long __init
find_min_pfn_for_node(int nid
)
4836 unsigned long min_pfn
= ULONG_MAX
;
4837 unsigned long start_pfn
;
4840 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
4841 min_pfn
= min(min_pfn
, start_pfn
);
4843 if (min_pfn
== ULONG_MAX
) {
4845 "Could not find start_pfn for node %d\n", nid
);
4853 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4855 * It returns the minimum PFN based on information provided via
4856 * add_active_range().
4858 unsigned long __init
find_min_pfn_with_active_regions(void)
4860 return find_min_pfn_for_node(MAX_NUMNODES
);
4864 * early_calculate_totalpages()
4865 * Sum pages in active regions for movable zone.
4866 * Populate N_MEMORY for calculating usable_nodes.
4868 static unsigned long __init
early_calculate_totalpages(void)
4870 unsigned long totalpages
= 0;
4871 unsigned long start_pfn
, end_pfn
;
4874 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4875 unsigned long pages
= end_pfn
- start_pfn
;
4877 totalpages
+= pages
;
4879 node_set_state(nid
, N_MEMORY
);
4885 * Find the PFN the Movable zone begins in each node. Kernel memory
4886 * is spread evenly between nodes as long as the nodes have enough
4887 * memory. When they don't, some nodes will have more kernelcore than
4890 static void __init
find_zone_movable_pfns_for_nodes(void)
4893 unsigned long usable_startpfn
;
4894 unsigned long kernelcore_node
, kernelcore_remaining
;
4895 /* save the state before borrow the nodemask */
4896 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
4897 unsigned long totalpages
= early_calculate_totalpages();
4898 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
4901 * If movablecore was specified, calculate what size of
4902 * kernelcore that corresponds so that memory usable for
4903 * any allocation type is evenly spread. If both kernelcore
4904 * and movablecore are specified, then the value of kernelcore
4905 * will be used for required_kernelcore if it's greater than
4906 * what movablecore would have allowed.
4908 if (required_movablecore
) {
4909 unsigned long corepages
;
4912 * Round-up so that ZONE_MOVABLE is at least as large as what
4913 * was requested by the user
4915 required_movablecore
=
4916 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4917 corepages
= totalpages
- required_movablecore
;
4919 required_kernelcore
= max(required_kernelcore
, corepages
);
4923 * If neither kernelcore/movablecore nor movablemem_map is specified,
4924 * there is no ZONE_MOVABLE. But if movablemem_map is specified, the
4925 * start pfn of ZONE_MOVABLE has been stored in zone_movable_limit[].
4927 if (!required_kernelcore
) {
4928 if (movablemem_map
.nr_map
)
4929 memcpy(zone_movable_pfn
, zone_movable_limit
,
4930 sizeof(zone_movable_pfn
));
4934 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4935 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4938 /* Spread kernelcore memory as evenly as possible throughout nodes */
4939 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4940 for_each_node_state(nid
, N_MEMORY
) {
4941 unsigned long start_pfn
, end_pfn
;
4944 * Recalculate kernelcore_node if the division per node
4945 * now exceeds what is necessary to satisfy the requested
4946 * amount of memory for the kernel
4948 if (required_kernelcore
< kernelcore_node
)
4949 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4952 * As the map is walked, we track how much memory is usable
4953 * by the kernel using kernelcore_remaining. When it is
4954 * 0, the rest of the node is usable by ZONE_MOVABLE
4956 kernelcore_remaining
= kernelcore_node
;
4958 /* Go through each range of PFNs within this node */
4959 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4960 unsigned long size_pages
;
4963 * Find more memory for kernelcore in
4964 * [zone_movable_pfn[nid], zone_movable_limit[nid]).
4966 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
4967 if (start_pfn
>= end_pfn
)
4970 if (zone_movable_limit
[nid
]) {
4971 end_pfn
= min(end_pfn
, zone_movable_limit
[nid
]);
4972 /* No range left for kernelcore in this node */
4973 if (start_pfn
>= end_pfn
) {
4974 zone_movable_pfn
[nid
] =
4975 zone_movable_limit
[nid
];
4980 /* Account for what is only usable for kernelcore */
4981 if (start_pfn
< usable_startpfn
) {
4982 unsigned long kernel_pages
;
4983 kernel_pages
= min(end_pfn
, usable_startpfn
)
4986 kernelcore_remaining
-= min(kernel_pages
,
4987 kernelcore_remaining
);
4988 required_kernelcore
-= min(kernel_pages
,
4989 required_kernelcore
);
4991 /* Continue if range is now fully accounted */
4992 if (end_pfn
<= usable_startpfn
) {
4995 * Push zone_movable_pfn to the end so
4996 * that if we have to rebalance
4997 * kernelcore across nodes, we will
4998 * not double account here
5000 zone_movable_pfn
[nid
] = end_pfn
;
5003 start_pfn
= usable_startpfn
;
5007 * The usable PFN range for ZONE_MOVABLE is from
5008 * start_pfn->end_pfn. Calculate size_pages as the
5009 * number of pages used as kernelcore
5011 size_pages
= end_pfn
- start_pfn
;
5012 if (size_pages
> kernelcore_remaining
)
5013 size_pages
= kernelcore_remaining
;
5014 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5017 * Some kernelcore has been met, update counts and
5018 * break if the kernelcore for this node has been
5021 required_kernelcore
-= min(required_kernelcore
,
5023 kernelcore_remaining
-= size_pages
;
5024 if (!kernelcore_remaining
)
5030 * If there is still required_kernelcore, we do another pass with one
5031 * less node in the count. This will push zone_movable_pfn[nid] further
5032 * along on the nodes that still have memory until kernelcore is
5036 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5040 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5041 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5042 zone_movable_pfn
[nid
] =
5043 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5045 /* restore the node_state */
5046 node_states
[N_MEMORY
] = saved_node_state
;
5049 /* Any regular or high memory on that node ? */
5050 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5052 enum zone_type zone_type
;
5054 if (N_MEMORY
== N_NORMAL_MEMORY
)
5057 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5058 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5059 if (zone
->present_pages
) {
5060 node_set_state(nid
, N_HIGH_MEMORY
);
5061 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5062 zone_type
<= ZONE_NORMAL
)
5063 node_set_state(nid
, N_NORMAL_MEMORY
);
5070 * free_area_init_nodes - Initialise all pg_data_t and zone data
5071 * @max_zone_pfn: an array of max PFNs for each zone
5073 * This will call free_area_init_node() for each active node in the system.
5074 * Using the page ranges provided by add_active_range(), the size of each
5075 * zone in each node and their holes is calculated. If the maximum PFN
5076 * between two adjacent zones match, it is assumed that the zone is empty.
5077 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5078 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5079 * starts where the previous one ended. For example, ZONE_DMA32 starts
5080 * at arch_max_dma_pfn.
5082 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5084 unsigned long start_pfn
, end_pfn
;
5087 /* Record where the zone boundaries are */
5088 memset(arch_zone_lowest_possible_pfn
, 0,
5089 sizeof(arch_zone_lowest_possible_pfn
));
5090 memset(arch_zone_highest_possible_pfn
, 0,
5091 sizeof(arch_zone_highest_possible_pfn
));
5092 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5093 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5094 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5095 if (i
== ZONE_MOVABLE
)
5097 arch_zone_lowest_possible_pfn
[i
] =
5098 arch_zone_highest_possible_pfn
[i
-1];
5099 arch_zone_highest_possible_pfn
[i
] =
5100 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5102 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5103 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5105 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5106 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5107 find_usable_zone_for_movable();
5108 sanitize_zone_movable_limit();
5109 find_zone_movable_pfns_for_nodes();
5111 /* Print out the zone ranges */
5112 printk("Zone ranges:\n");
5113 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5114 if (i
== ZONE_MOVABLE
)
5116 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5117 if (arch_zone_lowest_possible_pfn
[i
] ==
5118 arch_zone_highest_possible_pfn
[i
])
5119 printk(KERN_CONT
"empty\n");
5121 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5122 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5123 (arch_zone_highest_possible_pfn
[i
]
5124 << PAGE_SHIFT
) - 1);
5127 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5128 printk("Movable zone start for each node\n");
5129 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5130 if (zone_movable_pfn
[i
])
5131 printk(" Node %d: %#010lx\n", i
,
5132 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5135 /* Print out the early node map */
5136 printk("Early memory node ranges\n");
5137 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5138 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5139 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5141 /* Initialise every node */
5142 mminit_verify_pageflags_layout();
5143 setup_nr_node_ids();
5144 for_each_online_node(nid
) {
5145 pg_data_t
*pgdat
= NODE_DATA(nid
);
5146 free_area_init_node(nid
, NULL
,
5147 find_min_pfn_for_node(nid
), NULL
);
5149 /* Any memory on that node */
5150 if (pgdat
->node_present_pages
)
5151 node_set_state(nid
, N_MEMORY
);
5152 check_for_memory(pgdat
, nid
);
5156 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5158 unsigned long long coremem
;
5162 coremem
= memparse(p
, &p
);
5163 *core
= coremem
>> PAGE_SHIFT
;
5165 /* Paranoid check that UL is enough for the coremem value */
5166 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5172 * kernelcore=size sets the amount of memory for use for allocations that
5173 * cannot be reclaimed or migrated.
5175 static int __init
cmdline_parse_kernelcore(char *p
)
5177 return cmdline_parse_core(p
, &required_kernelcore
);
5181 * movablecore=size sets the amount of memory for use for allocations that
5182 * can be reclaimed or migrated.
5184 static int __init
cmdline_parse_movablecore(char *p
)
5186 return cmdline_parse_core(p
, &required_movablecore
);
5189 early_param("kernelcore", cmdline_parse_kernelcore
);
5190 early_param("movablecore", cmdline_parse_movablecore
);
5193 * movablemem_map_overlap() - Check if a range overlaps movablemem_map.map[].
5194 * @start_pfn: start pfn of the range to be checked
5195 * @end_pfn: end pfn of the range to be checked (exclusive)
5197 * This function checks if a given memory range [start_pfn, end_pfn) overlaps
5198 * the movablemem_map.map[] array.
5200 * Return: index of the first overlapped element in movablemem_map.map[]
5201 * or -1 if they don't overlap each other.
5203 int __init
movablemem_map_overlap(unsigned long start_pfn
,
5204 unsigned long end_pfn
)
5208 if (!movablemem_map
.nr_map
)
5211 for (overlap
= 0; overlap
< movablemem_map
.nr_map
; overlap
++)
5212 if (start_pfn
< movablemem_map
.map
[overlap
].end_pfn
)
5215 if (overlap
== movablemem_map
.nr_map
||
5216 end_pfn
<= movablemem_map
.map
[overlap
].start_pfn
)
5223 * insert_movablemem_map - Insert a memory range in to movablemem_map.map.
5224 * @start_pfn: start pfn of the range
5225 * @end_pfn: end pfn of the range
5227 * This function will also merge the overlapped ranges, and sort the array
5228 * by start_pfn in monotonic increasing order.
5230 void __init
insert_movablemem_map(unsigned long start_pfn
,
5231 unsigned long end_pfn
)
5236 * pos will be at the 1st overlapped range, or the position
5237 * where the element should be inserted.
5239 for (pos
= 0; pos
< movablemem_map
.nr_map
; pos
++)
5240 if (start_pfn
<= movablemem_map
.map
[pos
].end_pfn
)
5243 /* If there is no overlapped range, just insert the element. */
5244 if (pos
== movablemem_map
.nr_map
||
5245 end_pfn
< movablemem_map
.map
[pos
].start_pfn
) {
5247 * If pos is not the end of array, we need to move all
5248 * the rest elements backward.
5250 if (pos
< movablemem_map
.nr_map
)
5251 memmove(&movablemem_map
.map
[pos
+1],
5252 &movablemem_map
.map
[pos
],
5253 sizeof(struct movablemem_entry
) *
5254 (movablemem_map
.nr_map
- pos
));
5255 movablemem_map
.map
[pos
].start_pfn
= start_pfn
;
5256 movablemem_map
.map
[pos
].end_pfn
= end_pfn
;
5257 movablemem_map
.nr_map
++;
5261 /* overlap will be at the last overlapped range */
5262 for (overlap
= pos
+ 1; overlap
< movablemem_map
.nr_map
; overlap
++)
5263 if (end_pfn
< movablemem_map
.map
[overlap
].start_pfn
)
5267 * If there are more ranges overlapped, we need to merge them,
5268 * and move the rest elements forward.
5271 movablemem_map
.map
[pos
].start_pfn
= min(start_pfn
,
5272 movablemem_map
.map
[pos
].start_pfn
);
5273 movablemem_map
.map
[pos
].end_pfn
= max(end_pfn
,
5274 movablemem_map
.map
[overlap
].end_pfn
);
5276 if (pos
!= overlap
&& overlap
+ 1 != movablemem_map
.nr_map
)
5277 memmove(&movablemem_map
.map
[pos
+1],
5278 &movablemem_map
.map
[overlap
+1],
5279 sizeof(struct movablemem_entry
) *
5280 (movablemem_map
.nr_map
- overlap
- 1));
5282 movablemem_map
.nr_map
-= overlap
- pos
;
5286 * movablemem_map_add_region - Add a memory range into movablemem_map.
5287 * @start: physical start address of range
5288 * @end: physical end address of range
5290 * This function transform the physical address into pfn, and then add the
5291 * range into movablemem_map by calling insert_movablemem_map().
5293 static void __init
movablemem_map_add_region(u64 start
, u64 size
)
5295 unsigned long start_pfn
, end_pfn
;
5297 /* In case size == 0 or start + size overflows */
5298 if (start
+ size
<= start
)
5301 if (movablemem_map
.nr_map
>= ARRAY_SIZE(movablemem_map
.map
)) {
5302 pr_err("movablemem_map: too many entries;"
5303 " ignoring [mem %#010llx-%#010llx]\n",
5304 (unsigned long long) start
,
5305 (unsigned long long) (start
+ size
- 1));
5309 start_pfn
= PFN_DOWN(start
);
5310 end_pfn
= PFN_UP(start
+ size
);
5311 insert_movablemem_map(start_pfn
, end_pfn
);
5315 * cmdline_parse_movablemem_map - Parse boot option movablemem_map.
5316 * @p: The boot option of the following format:
5317 * movablemem_map=nn[KMG]@ss[KMG]
5319 * This option sets the memory range [ss, ss+nn) to be used as movable memory.
5321 * Return: 0 on success or -EINVAL on failure.
5323 static int __init
cmdline_parse_movablemem_map(char *p
)
5326 u64 start_at
, mem_size
;
5331 if (!strcmp(p
, "acpi"))
5332 movablemem_map
.acpi
= true;
5335 * If user decide to use info from BIOS, all the other user specified
5336 * ranges will be ingored.
5338 if (movablemem_map
.acpi
) {
5339 if (movablemem_map
.nr_map
) {
5340 memset(movablemem_map
.map
, 0,
5341 sizeof(struct movablemem_entry
)
5342 * movablemem_map
.nr_map
);
5343 movablemem_map
.nr_map
= 0;
5349 mem_size
= memparse(p
, &p
);
5355 start_at
= memparse(p
, &p
);
5356 if (p
== oldp
|| *p
!= '\0')
5359 movablemem_map_add_region(start_at
, mem_size
);
5365 early_param("movablemem_map", cmdline_parse_movablemem_map
);
5367 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5370 * set_dma_reserve - set the specified number of pages reserved in the first zone
5371 * @new_dma_reserve: The number of pages to mark reserved
5373 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5374 * In the DMA zone, a significant percentage may be consumed by kernel image
5375 * and other unfreeable allocations which can skew the watermarks badly. This
5376 * function may optionally be used to account for unfreeable pages in the
5377 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5378 * smaller per-cpu batchsize.
5380 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5382 dma_reserve
= new_dma_reserve
;
5385 void __init
free_area_init(unsigned long *zones_size
)
5387 free_area_init_node(0, zones_size
,
5388 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5391 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5392 unsigned long action
, void *hcpu
)
5394 int cpu
= (unsigned long)hcpu
;
5396 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5397 lru_add_drain_cpu(cpu
);
5401 * Spill the event counters of the dead processor
5402 * into the current processors event counters.
5403 * This artificially elevates the count of the current
5406 vm_events_fold_cpu(cpu
);
5409 * Zero the differential counters of the dead processor
5410 * so that the vm statistics are consistent.
5412 * This is only okay since the processor is dead and cannot
5413 * race with what we are doing.
5415 refresh_cpu_vm_stats(cpu
);
5420 void __init
page_alloc_init(void)
5422 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5426 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5427 * or min_free_kbytes changes.
5429 static void calculate_totalreserve_pages(void)
5431 struct pglist_data
*pgdat
;
5432 unsigned long reserve_pages
= 0;
5433 enum zone_type i
, j
;
5435 for_each_online_pgdat(pgdat
) {
5436 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5437 struct zone
*zone
= pgdat
->node_zones
+ i
;
5438 unsigned long max
= 0;
5440 /* Find valid and maximum lowmem_reserve in the zone */
5441 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5442 if (zone
->lowmem_reserve
[j
] > max
)
5443 max
= zone
->lowmem_reserve
[j
];
5446 /* we treat the high watermark as reserved pages. */
5447 max
+= high_wmark_pages(zone
);
5449 if (max
> zone
->managed_pages
)
5450 max
= zone
->managed_pages
;
5451 reserve_pages
+= max
;
5453 * Lowmem reserves are not available to
5454 * GFP_HIGHUSER page cache allocations and
5455 * kswapd tries to balance zones to their high
5456 * watermark. As a result, neither should be
5457 * regarded as dirtyable memory, to prevent a
5458 * situation where reclaim has to clean pages
5459 * in order to balance the zones.
5461 zone
->dirty_balance_reserve
= max
;
5464 dirty_balance_reserve
= reserve_pages
;
5465 totalreserve_pages
= reserve_pages
;
5469 * setup_per_zone_lowmem_reserve - called whenever
5470 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5471 * has a correct pages reserved value, so an adequate number of
5472 * pages are left in the zone after a successful __alloc_pages().
5474 static void setup_per_zone_lowmem_reserve(void)
5476 struct pglist_data
*pgdat
;
5477 enum zone_type j
, idx
;
5479 for_each_online_pgdat(pgdat
) {
5480 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5481 struct zone
*zone
= pgdat
->node_zones
+ j
;
5482 unsigned long managed_pages
= zone
->managed_pages
;
5484 zone
->lowmem_reserve
[j
] = 0;
5488 struct zone
*lower_zone
;
5492 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5493 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5495 lower_zone
= pgdat
->node_zones
+ idx
;
5496 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5497 sysctl_lowmem_reserve_ratio
[idx
];
5498 managed_pages
+= lower_zone
->managed_pages
;
5503 /* update totalreserve_pages */
5504 calculate_totalreserve_pages();
5507 static void __setup_per_zone_wmarks(void)
5509 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5510 unsigned long lowmem_pages
= 0;
5512 unsigned long flags
;
5514 /* Calculate total number of !ZONE_HIGHMEM pages */
5515 for_each_zone(zone
) {
5516 if (!is_highmem(zone
))
5517 lowmem_pages
+= zone
->managed_pages
;
5520 for_each_zone(zone
) {
5523 spin_lock_irqsave(&zone
->lock
, flags
);
5524 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5525 do_div(tmp
, lowmem_pages
);
5526 if (is_highmem(zone
)) {
5528 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5529 * need highmem pages, so cap pages_min to a small
5532 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5533 * deltas controls asynch page reclaim, and so should
5534 * not be capped for highmem.
5536 unsigned long min_pages
;
5538 min_pages
= zone
->managed_pages
/ 1024;
5539 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5540 zone
->watermark
[WMARK_MIN
] = min_pages
;
5543 * If it's a lowmem zone, reserve a number of pages
5544 * proportionate to the zone's size.
5546 zone
->watermark
[WMARK_MIN
] = tmp
;
5549 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5550 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5552 setup_zone_migrate_reserve(zone
);
5553 spin_unlock_irqrestore(&zone
->lock
, flags
);
5556 /* update totalreserve_pages */
5557 calculate_totalreserve_pages();
5561 * setup_per_zone_wmarks - called when min_free_kbytes changes
5562 * or when memory is hot-{added|removed}
5564 * Ensures that the watermark[min,low,high] values for each zone are set
5565 * correctly with respect to min_free_kbytes.
5567 void setup_per_zone_wmarks(void)
5569 mutex_lock(&zonelists_mutex
);
5570 __setup_per_zone_wmarks();
5571 mutex_unlock(&zonelists_mutex
);
5575 * The inactive anon list should be small enough that the VM never has to
5576 * do too much work, but large enough that each inactive page has a chance
5577 * to be referenced again before it is swapped out.
5579 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5580 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5581 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5582 * the anonymous pages are kept on the inactive list.
5585 * memory ratio inactive anon
5586 * -------------------------------------
5595 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5597 unsigned int gb
, ratio
;
5599 /* Zone size in gigabytes */
5600 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5602 ratio
= int_sqrt(10 * gb
);
5606 zone
->inactive_ratio
= ratio
;
5609 static void __meminit
setup_per_zone_inactive_ratio(void)
5614 calculate_zone_inactive_ratio(zone
);
5618 * Initialise min_free_kbytes.
5620 * For small machines we want it small (128k min). For large machines
5621 * we want it large (64MB max). But it is not linear, because network
5622 * bandwidth does not increase linearly with machine size. We use
5624 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5625 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5641 int __meminit
init_per_zone_wmark_min(void)
5643 unsigned long lowmem_kbytes
;
5645 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5647 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5648 if (min_free_kbytes
< 128)
5649 min_free_kbytes
= 128;
5650 if (min_free_kbytes
> 65536)
5651 min_free_kbytes
= 65536;
5652 setup_per_zone_wmarks();
5653 refresh_zone_stat_thresholds();
5654 setup_per_zone_lowmem_reserve();
5655 setup_per_zone_inactive_ratio();
5658 module_init(init_per_zone_wmark_min
)
5661 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5662 * that we can call two helper functions whenever min_free_kbytes
5665 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5666 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5668 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5670 setup_per_zone_wmarks();
5675 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5676 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5681 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5686 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5687 sysctl_min_unmapped_ratio
) / 100;
5691 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5692 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5697 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5702 zone
->min_slab_pages
= (zone
->managed_pages
*
5703 sysctl_min_slab_ratio
) / 100;
5709 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5710 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5711 * whenever sysctl_lowmem_reserve_ratio changes.
5713 * The reserve ratio obviously has absolutely no relation with the
5714 * minimum watermarks. The lowmem reserve ratio can only make sense
5715 * if in function of the boot time zone sizes.
5717 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5718 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5720 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5721 setup_per_zone_lowmem_reserve();
5726 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5727 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5728 * can have before it gets flushed back to buddy allocator.
5731 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5732 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5738 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5739 if (!write
|| (ret
< 0))
5741 for_each_populated_zone(zone
) {
5742 for_each_possible_cpu(cpu
) {
5744 high
= zone
->managed_pages
/ percpu_pagelist_fraction
;
5745 setup_pagelist_highmark(
5746 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5752 int hashdist
= HASHDIST_DEFAULT
;
5755 static int __init
set_hashdist(char *str
)
5759 hashdist
= simple_strtoul(str
, &str
, 0);
5762 __setup("hashdist=", set_hashdist
);
5766 * allocate a large system hash table from bootmem
5767 * - it is assumed that the hash table must contain an exact power-of-2
5768 * quantity of entries
5769 * - limit is the number of hash buckets, not the total allocation size
5771 void *__init
alloc_large_system_hash(const char *tablename
,
5772 unsigned long bucketsize
,
5773 unsigned long numentries
,
5776 unsigned int *_hash_shift
,
5777 unsigned int *_hash_mask
,
5778 unsigned long low_limit
,
5779 unsigned long high_limit
)
5781 unsigned long long max
= high_limit
;
5782 unsigned long log2qty
, size
;
5785 /* allow the kernel cmdline to have a say */
5787 /* round applicable memory size up to nearest megabyte */
5788 numentries
= nr_kernel_pages
;
5789 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5790 numentries
>>= 20 - PAGE_SHIFT
;
5791 numentries
<<= 20 - PAGE_SHIFT
;
5793 /* limit to 1 bucket per 2^scale bytes of low memory */
5794 if (scale
> PAGE_SHIFT
)
5795 numentries
>>= (scale
- PAGE_SHIFT
);
5797 numentries
<<= (PAGE_SHIFT
- scale
);
5799 /* Make sure we've got at least a 0-order allocation.. */
5800 if (unlikely(flags
& HASH_SMALL
)) {
5801 /* Makes no sense without HASH_EARLY */
5802 WARN_ON(!(flags
& HASH_EARLY
));
5803 if (!(numentries
>> *_hash_shift
)) {
5804 numentries
= 1UL << *_hash_shift
;
5805 BUG_ON(!numentries
);
5807 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5808 numentries
= PAGE_SIZE
/ bucketsize
;
5810 numentries
= roundup_pow_of_two(numentries
);
5812 /* limit allocation size to 1/16 total memory by default */
5814 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5815 do_div(max
, bucketsize
);
5817 max
= min(max
, 0x80000000ULL
);
5819 if (numentries
< low_limit
)
5820 numentries
= low_limit
;
5821 if (numentries
> max
)
5824 log2qty
= ilog2(numentries
);
5827 size
= bucketsize
<< log2qty
;
5828 if (flags
& HASH_EARLY
)
5829 table
= alloc_bootmem_nopanic(size
);
5831 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5834 * If bucketsize is not a power-of-two, we may free
5835 * some pages at the end of hash table which
5836 * alloc_pages_exact() automatically does
5838 if (get_order(size
) < MAX_ORDER
) {
5839 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5840 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5843 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5846 panic("Failed to allocate %s hash table\n", tablename
);
5848 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5851 ilog2(size
) - PAGE_SHIFT
,
5855 *_hash_shift
= log2qty
;
5857 *_hash_mask
= (1 << log2qty
) - 1;
5862 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5863 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5866 #ifdef CONFIG_SPARSEMEM
5867 return __pfn_to_section(pfn
)->pageblock_flags
;
5869 return zone
->pageblock_flags
;
5870 #endif /* CONFIG_SPARSEMEM */
5873 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5875 #ifdef CONFIG_SPARSEMEM
5876 pfn
&= (PAGES_PER_SECTION
-1);
5877 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5879 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
5880 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5881 #endif /* CONFIG_SPARSEMEM */
5885 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5886 * @page: The page within the block of interest
5887 * @start_bitidx: The first bit of interest to retrieve
5888 * @end_bitidx: The last bit of interest
5889 * returns pageblock_bits flags
5891 unsigned long get_pageblock_flags_group(struct page
*page
,
5892 int start_bitidx
, int end_bitidx
)
5895 unsigned long *bitmap
;
5896 unsigned long pfn
, bitidx
;
5897 unsigned long flags
= 0;
5898 unsigned long value
= 1;
5900 zone
= page_zone(page
);
5901 pfn
= page_to_pfn(page
);
5902 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5903 bitidx
= pfn_to_bitidx(zone
, pfn
);
5905 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5906 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5913 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5914 * @page: The page within the block of interest
5915 * @start_bitidx: The first bit of interest
5916 * @end_bitidx: The last bit of interest
5917 * @flags: The flags to set
5919 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5920 int start_bitidx
, int end_bitidx
)
5923 unsigned long *bitmap
;
5924 unsigned long pfn
, bitidx
;
5925 unsigned long value
= 1;
5927 zone
= page_zone(page
);
5928 pfn
= page_to_pfn(page
);
5929 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5930 bitidx
= pfn_to_bitidx(zone
, pfn
);
5931 VM_BUG_ON(pfn
< zone
->zone_start_pfn
);
5932 VM_BUG_ON(pfn
>= zone
->zone_start_pfn
+ zone
->spanned_pages
);
5934 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5936 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5938 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5942 * This function checks whether pageblock includes unmovable pages or not.
5943 * If @count is not zero, it is okay to include less @count unmovable pages
5945 * PageLRU check wihtout isolation or lru_lock could race so that
5946 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5947 * expect this function should be exact.
5949 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
5950 bool skip_hwpoisoned_pages
)
5952 unsigned long pfn
, iter
, found
;
5956 * For avoiding noise data, lru_add_drain_all() should be called
5957 * If ZONE_MOVABLE, the zone never contains unmovable pages
5959 if (zone_idx(zone
) == ZONE_MOVABLE
)
5961 mt
= get_pageblock_migratetype(page
);
5962 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
5965 pfn
= page_to_pfn(page
);
5966 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5967 unsigned long check
= pfn
+ iter
;
5969 if (!pfn_valid_within(check
))
5972 page
= pfn_to_page(check
);
5974 * We can't use page_count without pin a page
5975 * because another CPU can free compound page.
5976 * This check already skips compound tails of THP
5977 * because their page->_count is zero at all time.
5979 if (!atomic_read(&page
->_count
)) {
5980 if (PageBuddy(page
))
5981 iter
+= (1 << page_order(page
)) - 1;
5986 * The HWPoisoned page may be not in buddy system, and
5987 * page_count() is not 0.
5989 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
5995 * If there are RECLAIMABLE pages, we need to check it.
5996 * But now, memory offline itself doesn't call shrink_slab()
5997 * and it still to be fixed.
6000 * If the page is not RAM, page_count()should be 0.
6001 * we don't need more check. This is an _used_ not-movable page.
6003 * The problematic thing here is PG_reserved pages. PG_reserved
6004 * is set to both of a memory hole page and a _used_ kernel
6013 bool is_pageblock_removable_nolock(struct page
*page
)
6019 * We have to be careful here because we are iterating over memory
6020 * sections which are not zone aware so we might end up outside of
6021 * the zone but still within the section.
6022 * We have to take care about the node as well. If the node is offline
6023 * its NODE_DATA will be NULL - see page_zone.
6025 if (!node_online(page_to_nid(page
)))
6028 zone
= page_zone(page
);
6029 pfn
= page_to_pfn(page
);
6030 if (zone
->zone_start_pfn
> pfn
||
6031 zone
->zone_start_pfn
+ zone
->spanned_pages
<= pfn
)
6034 return !has_unmovable_pages(zone
, page
, 0, true);
6039 static unsigned long pfn_max_align_down(unsigned long pfn
)
6041 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6042 pageblock_nr_pages
) - 1);
6045 static unsigned long pfn_max_align_up(unsigned long pfn
)
6047 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6048 pageblock_nr_pages
));
6051 /* [start, end) must belong to a single zone. */
6052 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6053 unsigned long start
, unsigned long end
)
6055 /* This function is based on compact_zone() from compaction.c. */
6056 unsigned long nr_reclaimed
;
6057 unsigned long pfn
= start
;
6058 unsigned int tries
= 0;
6063 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6064 if (fatal_signal_pending(current
)) {
6069 if (list_empty(&cc
->migratepages
)) {
6070 cc
->nr_migratepages
= 0;
6071 pfn
= isolate_migratepages_range(cc
->zone
, cc
,
6078 } else if (++tries
== 5) {
6079 ret
= ret
< 0 ? ret
: -EBUSY
;
6083 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6085 cc
->nr_migratepages
-= nr_reclaimed
;
6087 ret
= migrate_pages(&cc
->migratepages
,
6088 alloc_migrate_target
,
6089 0, false, MIGRATE_SYNC
,
6093 putback_movable_pages(&cc
->migratepages
);
6100 * alloc_contig_range() -- tries to allocate given range of pages
6101 * @start: start PFN to allocate
6102 * @end: one-past-the-last PFN to allocate
6103 * @migratetype: migratetype of the underlaying pageblocks (either
6104 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6105 * in range must have the same migratetype and it must
6106 * be either of the two.
6108 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6109 * aligned, however it's the caller's responsibility to guarantee that
6110 * we are the only thread that changes migrate type of pageblocks the
6113 * The PFN range must belong to a single zone.
6115 * Returns zero on success or negative error code. On success all
6116 * pages which PFN is in [start, end) are allocated for the caller and
6117 * need to be freed with free_contig_range().
6119 int alloc_contig_range(unsigned long start
, unsigned long end
,
6120 unsigned migratetype
)
6122 unsigned long outer_start
, outer_end
;
6125 struct compact_control cc
= {
6126 .nr_migratepages
= 0,
6128 .zone
= page_zone(pfn_to_page(start
)),
6130 .ignore_skip_hint
= true,
6132 INIT_LIST_HEAD(&cc
.migratepages
);
6135 * What we do here is we mark all pageblocks in range as
6136 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6137 * have different sizes, and due to the way page allocator
6138 * work, we align the range to biggest of the two pages so
6139 * that page allocator won't try to merge buddies from
6140 * different pageblocks and change MIGRATE_ISOLATE to some
6141 * other migration type.
6143 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6144 * migrate the pages from an unaligned range (ie. pages that
6145 * we are interested in). This will put all the pages in
6146 * range back to page allocator as MIGRATE_ISOLATE.
6148 * When this is done, we take the pages in range from page
6149 * allocator removing them from the buddy system. This way
6150 * page allocator will never consider using them.
6152 * This lets us mark the pageblocks back as
6153 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6154 * aligned range but not in the unaligned, original range are
6155 * put back to page allocator so that buddy can use them.
6158 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6159 pfn_max_align_up(end
), migratetype
,
6164 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6169 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6170 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6171 * more, all pages in [start, end) are free in page allocator.
6172 * What we are going to do is to allocate all pages from
6173 * [start, end) (that is remove them from page allocator).
6175 * The only problem is that pages at the beginning and at the
6176 * end of interesting range may be not aligned with pages that
6177 * page allocator holds, ie. they can be part of higher order
6178 * pages. Because of this, we reserve the bigger range and
6179 * once this is done free the pages we are not interested in.
6181 * We don't have to hold zone->lock here because the pages are
6182 * isolated thus they won't get removed from buddy.
6185 lru_add_drain_all();
6189 outer_start
= start
;
6190 while (!PageBuddy(pfn_to_page(outer_start
))) {
6191 if (++order
>= MAX_ORDER
) {
6195 outer_start
&= ~0UL << order
;
6198 /* Make sure the range is really isolated. */
6199 if (test_pages_isolated(outer_start
, end
, false)) {
6200 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6207 /* Grab isolated pages from freelists. */
6208 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6214 /* Free head and tail (if any) */
6215 if (start
!= outer_start
)
6216 free_contig_range(outer_start
, start
- outer_start
);
6217 if (end
!= outer_end
)
6218 free_contig_range(end
, outer_end
- end
);
6221 undo_isolate_page_range(pfn_max_align_down(start
),
6222 pfn_max_align_up(end
), migratetype
);
6226 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6228 unsigned int count
= 0;
6230 for (; nr_pages
--; pfn
++) {
6231 struct page
*page
= pfn_to_page(pfn
);
6233 count
+= page_count(page
) != 1;
6236 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6240 #ifdef CONFIG_MEMORY_HOTPLUG
6241 static int __meminit
__zone_pcp_update(void *data
)
6243 struct zone
*zone
= data
;
6245 unsigned long batch
= zone_batchsize(zone
), flags
;
6247 for_each_possible_cpu(cpu
) {
6248 struct per_cpu_pageset
*pset
;
6249 struct per_cpu_pages
*pcp
;
6251 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6254 local_irq_save(flags
);
6256 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
6257 drain_zonestat(zone
, pset
);
6258 setup_pageset(pset
, batch
);
6259 local_irq_restore(flags
);
6264 void __meminit
zone_pcp_update(struct zone
*zone
)
6266 stop_machine(__zone_pcp_update
, zone
, NULL
);
6270 void zone_pcp_reset(struct zone
*zone
)
6272 unsigned long flags
;
6274 struct per_cpu_pageset
*pset
;
6276 /* avoid races with drain_pages() */
6277 local_irq_save(flags
);
6278 if (zone
->pageset
!= &boot_pageset
) {
6279 for_each_online_cpu(cpu
) {
6280 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6281 drain_zonestat(zone
, pset
);
6283 free_percpu(zone
->pageset
);
6284 zone
->pageset
= &boot_pageset
;
6286 local_irq_restore(flags
);
6289 #ifdef CONFIG_MEMORY_HOTREMOVE
6291 * All pages in the range must be isolated before calling this.
6294 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6300 unsigned long flags
;
6301 /* find the first valid pfn */
6302 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6307 zone
= page_zone(pfn_to_page(pfn
));
6308 spin_lock_irqsave(&zone
->lock
, flags
);
6310 while (pfn
< end_pfn
) {
6311 if (!pfn_valid(pfn
)) {
6315 page
= pfn_to_page(pfn
);
6317 * The HWPoisoned page may be not in buddy system, and
6318 * page_count() is not 0.
6320 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6322 SetPageReserved(page
);
6326 BUG_ON(page_count(page
));
6327 BUG_ON(!PageBuddy(page
));
6328 order
= page_order(page
);
6329 #ifdef CONFIG_DEBUG_VM
6330 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6331 pfn
, 1 << order
, end_pfn
);
6333 list_del(&page
->lru
);
6334 rmv_page_order(page
);
6335 zone
->free_area
[order
].nr_free
--;
6336 for (i
= 0; i
< (1 << order
); i
++)
6337 SetPageReserved((page
+i
));
6338 pfn
+= (1 << order
);
6340 spin_unlock_irqrestore(&zone
->lock
, flags
);
6344 #ifdef CONFIG_MEMORY_FAILURE
6345 bool is_free_buddy_page(struct page
*page
)
6347 struct zone
*zone
= page_zone(page
);
6348 unsigned long pfn
= page_to_pfn(page
);
6349 unsigned long flags
;
6352 spin_lock_irqsave(&zone
->lock
, flags
);
6353 for (order
= 0; order
< MAX_ORDER
; order
++) {
6354 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6356 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6359 spin_unlock_irqrestore(&zone
->lock
, flags
);
6361 return order
< MAX_ORDER
;
6365 static const struct trace_print_flags pageflag_names
[] = {
6366 {1UL << PG_locked
, "locked" },
6367 {1UL << PG_error
, "error" },
6368 {1UL << PG_referenced
, "referenced" },
6369 {1UL << PG_uptodate
, "uptodate" },
6370 {1UL << PG_dirty
, "dirty" },
6371 {1UL << PG_lru
, "lru" },
6372 {1UL << PG_active
, "active" },
6373 {1UL << PG_slab
, "slab" },
6374 {1UL << PG_owner_priv_1
, "owner_priv_1" },
6375 {1UL << PG_arch_1
, "arch_1" },
6376 {1UL << PG_reserved
, "reserved" },
6377 {1UL << PG_private
, "private" },
6378 {1UL << PG_private_2
, "private_2" },
6379 {1UL << PG_writeback
, "writeback" },
6380 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6381 {1UL << PG_head
, "head" },
6382 {1UL << PG_tail
, "tail" },
6384 {1UL << PG_compound
, "compound" },
6386 {1UL << PG_swapcache
, "swapcache" },
6387 {1UL << PG_mappedtodisk
, "mappedtodisk" },
6388 {1UL << PG_reclaim
, "reclaim" },
6389 {1UL << PG_swapbacked
, "swapbacked" },
6390 {1UL << PG_unevictable
, "unevictable" },
6392 {1UL << PG_mlocked
, "mlocked" },
6394 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6395 {1UL << PG_uncached
, "uncached" },
6397 #ifdef CONFIG_MEMORY_FAILURE
6398 {1UL << PG_hwpoison
, "hwpoison" },
6400 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6401 {1UL << PG_compound_lock
, "compound_lock" },
6405 static void dump_page_flags(unsigned long flags
)
6407 const char *delim
= "";
6411 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names
) != __NR_PAGEFLAGS
);
6413 printk(KERN_ALERT
"page flags: %#lx(", flags
);
6415 /* remove zone id */
6416 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
6418 for (i
= 0; i
< ARRAY_SIZE(pageflag_names
) && flags
; i
++) {
6420 mask
= pageflag_names
[i
].mask
;
6421 if ((flags
& mask
) != mask
)
6425 printk("%s%s", delim
, pageflag_names
[i
].name
);
6429 /* check for left over flags */
6431 printk("%s%#lx", delim
, flags
);
6436 void dump_page(struct page
*page
)
6439 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6440 page
, atomic_read(&page
->_count
), page_mapcount(page
),
6441 page
->mapping
, page
->index
);
6442 dump_page_flags(page
->flags
);
6443 mem_cgroup_print_bad_page(page
);