2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/migrate.h>
60 #include <linux/page-debug-flags.h>
61 #include <linux/hugetlb.h>
62 #include <linux/sched/rt.h>
64 #include <asm/tlbflush.h>
65 #include <asm/div64.h>
68 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
69 DEFINE_PER_CPU(int, numa_node
);
70 EXPORT_PER_CPU_SYMBOL(numa_node
);
73 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
75 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
76 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
77 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
78 * defined in <linux/topology.h>.
80 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
81 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
85 * Array of node states.
87 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
88 [N_POSSIBLE
] = NODE_MASK_ALL
,
89 [N_ONLINE
] = { { [0] = 1UL } },
91 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
93 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
95 #ifdef CONFIG_MOVABLE_NODE
96 [N_MEMORY
] = { { [0] = 1UL } },
98 [N_CPU
] = { { [0] = 1UL } },
101 EXPORT_SYMBOL(node_states
);
103 unsigned long totalram_pages __read_mostly
;
104 unsigned long totalreserve_pages __read_mostly
;
106 * When calculating the number of globally allowed dirty pages, there
107 * is a certain number of per-zone reserves that should not be
108 * considered dirtyable memory. This is the sum of those reserves
109 * over all existing zones that contribute dirtyable memory.
111 unsigned long dirty_balance_reserve __read_mostly
;
113 int percpu_pagelist_fraction
;
114 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
116 #ifdef CONFIG_PM_SLEEP
118 * The following functions are used by the suspend/hibernate code to temporarily
119 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
120 * while devices are suspended. To avoid races with the suspend/hibernate code,
121 * they should always be called with pm_mutex held (gfp_allowed_mask also should
122 * only be modified with pm_mutex held, unless the suspend/hibernate code is
123 * guaranteed not to run in parallel with that modification).
126 static gfp_t saved_gfp_mask
;
128 void pm_restore_gfp_mask(void)
130 WARN_ON(!mutex_is_locked(&pm_mutex
));
131 if (saved_gfp_mask
) {
132 gfp_allowed_mask
= saved_gfp_mask
;
137 void pm_restrict_gfp_mask(void)
139 WARN_ON(!mutex_is_locked(&pm_mutex
));
140 WARN_ON(saved_gfp_mask
);
141 saved_gfp_mask
= gfp_allowed_mask
;
142 gfp_allowed_mask
&= ~GFP_IOFS
;
145 bool pm_suspended_storage(void)
147 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
151 #endif /* CONFIG_PM_SLEEP */
153 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
154 int pageblock_order __read_mostly
;
157 static void __free_pages_ok(struct page
*page
, unsigned int order
);
160 * results with 256, 32 in the lowmem_reserve sysctl:
161 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
162 * 1G machine -> (16M dma, 784M normal, 224M high)
163 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
164 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
165 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
167 * TBD: should special case ZONE_DMA32 machines here - in those we normally
168 * don't need any ZONE_NORMAL reservation
170 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
171 #ifdef CONFIG_ZONE_DMA
174 #ifdef CONFIG_ZONE_DMA32
177 #ifdef CONFIG_HIGHMEM
183 EXPORT_SYMBOL(totalram_pages
);
185 static char * const zone_names
[MAX_NR_ZONES
] = {
186 #ifdef CONFIG_ZONE_DMA
189 #ifdef CONFIG_ZONE_DMA32
193 #ifdef CONFIG_HIGHMEM
199 int min_free_kbytes
= 1024;
201 static unsigned long __meminitdata nr_kernel_pages
;
202 static unsigned long __meminitdata nr_all_pages
;
203 static unsigned long __meminitdata dma_reserve
;
205 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
206 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
207 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
208 static unsigned long __initdata required_kernelcore
;
209 static unsigned long __initdata required_movablecore
;
210 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
212 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
214 EXPORT_SYMBOL(movable_zone
);
215 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
218 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
219 int nr_online_nodes __read_mostly
= 1;
220 EXPORT_SYMBOL(nr_node_ids
);
221 EXPORT_SYMBOL(nr_online_nodes
);
224 int page_group_by_mobility_disabled __read_mostly
;
226 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
229 if (unlikely(page_group_by_mobility_disabled
))
230 migratetype
= MIGRATE_UNMOVABLE
;
232 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
233 PB_migrate
, PB_migrate_end
);
236 bool oom_killer_disabled __read_mostly
;
238 #ifdef CONFIG_DEBUG_VM
239 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
243 unsigned long pfn
= page_to_pfn(page
);
244 unsigned long sp
, start_pfn
;
247 seq
= zone_span_seqbegin(zone
);
248 start_pfn
= zone
->zone_start_pfn
;
249 sp
= zone
->spanned_pages
;
250 if (!zone_spans_pfn(zone
, pfn
))
252 } while (zone_span_seqretry(zone
, seq
));
255 pr_err("page %lu outside zone [ %lu - %lu ]\n",
256 pfn
, start_pfn
, start_pfn
+ sp
);
261 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
263 if (!pfn_valid_within(page_to_pfn(page
)))
265 if (zone
!= page_zone(page
))
271 * Temporary debugging check for pages not lying within a given zone.
273 static int bad_range(struct zone
*zone
, struct page
*page
)
275 if (page_outside_zone_boundaries(zone
, page
))
277 if (!page_is_consistent(zone
, page
))
283 static inline int bad_range(struct zone
*zone
, struct page
*page
)
289 static void bad_page(struct page
*page
)
291 static unsigned long resume
;
292 static unsigned long nr_shown
;
293 static unsigned long nr_unshown
;
295 /* Don't complain about poisoned pages */
296 if (PageHWPoison(page
)) {
297 page_mapcount_reset(page
); /* remove PageBuddy */
302 * Allow a burst of 60 reports, then keep quiet for that minute;
303 * or allow a steady drip of one report per second.
305 if (nr_shown
== 60) {
306 if (time_before(jiffies
, resume
)) {
312 "BUG: Bad page state: %lu messages suppressed\n",
319 resume
= jiffies
+ 60 * HZ
;
321 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
322 current
->comm
, page_to_pfn(page
));
328 /* Leave bad fields for debug, except PageBuddy could make trouble */
329 page_mapcount_reset(page
); /* remove PageBuddy */
330 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
334 * Higher-order pages are called "compound pages". They are structured thusly:
336 * The first PAGE_SIZE page is called the "head page".
338 * The remaining PAGE_SIZE pages are called "tail pages".
340 * All pages have PG_compound set. All tail pages have their ->first_page
341 * pointing at the head page.
343 * The first tail page's ->lru.next holds the address of the compound page's
344 * put_page() function. Its ->lru.prev holds the order of allocation.
345 * This usage means that zero-order pages may not be compound.
348 static void free_compound_page(struct page
*page
)
350 __free_pages_ok(page
, compound_order(page
));
353 void prep_compound_page(struct page
*page
, unsigned long order
)
356 int nr_pages
= 1 << order
;
358 set_compound_page_dtor(page
, free_compound_page
);
359 set_compound_order(page
, order
);
361 for (i
= 1; i
< nr_pages
; i
++) {
362 struct page
*p
= page
+ i
;
363 set_page_count(p
, 0);
364 p
->first_page
= page
;
365 /* Make sure p->first_page is always valid for PageTail() */
371 /* update __split_huge_page_refcount if you change this function */
372 static int destroy_compound_page(struct page
*page
, unsigned long order
)
375 int nr_pages
= 1 << order
;
378 if (unlikely(compound_order(page
) != order
)) {
383 __ClearPageHead(page
);
385 for (i
= 1; i
< nr_pages
; i
++) {
386 struct page
*p
= page
+ i
;
388 if (unlikely(!PageTail(p
) || (p
->first_page
!= page
))) {
398 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
403 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
404 * and __GFP_HIGHMEM from hard or soft interrupt context.
406 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
407 for (i
= 0; i
< (1 << order
); i
++)
408 clear_highpage(page
+ i
);
411 #ifdef CONFIG_DEBUG_PAGEALLOC
412 unsigned int _debug_guardpage_minorder
;
414 static int __init
debug_guardpage_minorder_setup(char *buf
)
418 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
419 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
422 _debug_guardpage_minorder
= res
;
423 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
426 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
428 static inline void set_page_guard_flag(struct page
*page
)
430 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
433 static inline void clear_page_guard_flag(struct page
*page
)
435 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
438 static inline void set_page_guard_flag(struct page
*page
) { }
439 static inline void clear_page_guard_flag(struct page
*page
) { }
442 static inline void set_page_order(struct page
*page
, int order
)
444 set_page_private(page
, order
);
445 __SetPageBuddy(page
);
448 static inline void rmv_page_order(struct page
*page
)
450 __ClearPageBuddy(page
);
451 set_page_private(page
, 0);
455 * Locate the struct page for both the matching buddy in our
456 * pair (buddy1) and the combined O(n+1) page they form (page).
458 * 1) Any buddy B1 will have an order O twin B2 which satisfies
459 * the following equation:
461 * For example, if the starting buddy (buddy2) is #8 its order
463 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
465 * 2) Any buddy B will have an order O+1 parent P which
466 * satisfies the following equation:
469 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
471 static inline unsigned long
472 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
474 return page_idx
^ (1 << order
);
478 * This function checks whether a page is free && is the buddy
479 * we can do coalesce a page and its buddy if
480 * (a) the buddy is not in a hole &&
481 * (b) the buddy is in the buddy system &&
482 * (c) a page and its buddy have the same order &&
483 * (d) a page and its buddy are in the same zone.
485 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
486 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
488 * For recording page's order, we use page_private(page).
490 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
493 if (!pfn_valid_within(page_to_pfn(buddy
)))
496 if (page_zone_id(page
) != page_zone_id(buddy
))
499 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
500 VM_BUG_ON(page_count(buddy
) != 0);
504 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
505 VM_BUG_ON(page_count(buddy
) != 0);
512 * Freeing function for a buddy system allocator.
514 * The concept of a buddy system is to maintain direct-mapped table
515 * (containing bit values) for memory blocks of various "orders".
516 * The bottom level table contains the map for the smallest allocatable
517 * units of memory (here, pages), and each level above it describes
518 * pairs of units from the levels below, hence, "buddies".
519 * At a high level, all that happens here is marking the table entry
520 * at the bottom level available, and propagating the changes upward
521 * as necessary, plus some accounting needed to play nicely with other
522 * parts of the VM system.
523 * At each level, we keep a list of pages, which are heads of continuous
524 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
525 * order is recorded in page_private(page) field.
526 * So when we are allocating or freeing one, we can derive the state of the
527 * other. That is, if we allocate a small block, and both were
528 * free, the remainder of the region must be split into blocks.
529 * If a block is freed, and its buddy is also free, then this
530 * triggers coalescing into a block of larger size.
535 static inline void __free_one_page(struct page
*page
,
536 struct zone
*zone
, unsigned int order
,
539 unsigned long page_idx
;
540 unsigned long combined_idx
;
541 unsigned long uninitialized_var(buddy_idx
);
544 VM_BUG_ON(!zone_is_initialized(zone
));
546 if (unlikely(PageCompound(page
)))
547 if (unlikely(destroy_compound_page(page
, order
)))
550 VM_BUG_ON(migratetype
== -1);
552 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
554 VM_BUG_ON(page_idx
& ((1 << order
) - 1));
555 VM_BUG_ON(bad_range(zone
, page
));
557 while (order
< MAX_ORDER
-1) {
558 buddy_idx
= __find_buddy_index(page_idx
, order
);
559 buddy
= page
+ (buddy_idx
- page_idx
);
560 if (!page_is_buddy(page
, buddy
, order
))
563 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
564 * merge with it and move up one order.
566 if (page_is_guard(buddy
)) {
567 clear_page_guard_flag(buddy
);
568 set_page_private(page
, 0);
569 __mod_zone_freepage_state(zone
, 1 << order
,
572 list_del(&buddy
->lru
);
573 zone
->free_area
[order
].nr_free
--;
574 rmv_page_order(buddy
);
576 combined_idx
= buddy_idx
& page_idx
;
577 page
= page
+ (combined_idx
- page_idx
);
578 page_idx
= combined_idx
;
581 set_page_order(page
, order
);
584 * If this is not the largest possible page, check if the buddy
585 * of the next-highest order is free. If it is, it's possible
586 * that pages are being freed that will coalesce soon. In case,
587 * that is happening, add the free page to the tail of the list
588 * so it's less likely to be used soon and more likely to be merged
589 * as a higher order page
591 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
592 struct page
*higher_page
, *higher_buddy
;
593 combined_idx
= buddy_idx
& page_idx
;
594 higher_page
= page
+ (combined_idx
- page_idx
);
595 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
596 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
597 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
598 list_add_tail(&page
->lru
,
599 &zone
->free_area
[order
].free_list
[migratetype
]);
604 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
606 zone
->free_area
[order
].nr_free
++;
609 static inline int free_pages_check(struct page
*page
)
611 if (unlikely(page_mapcount(page
) |
612 (page
->mapping
!= NULL
) |
613 (atomic_read(&page
->_count
) != 0) |
614 (page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
) |
615 (mem_cgroup_bad_page_check(page
)))) {
619 page_nid_reset_last(page
);
620 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
621 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
626 * Frees a number of pages from the PCP lists
627 * Assumes all pages on list are in same zone, and of same order.
628 * count is the number of pages to free.
630 * If the zone was previously in an "all pages pinned" state then look to
631 * see if this freeing clears that state.
633 * And clear the zone's pages_scanned counter, to hold off the "all pages are
634 * pinned" detection logic.
636 static void free_pcppages_bulk(struct zone
*zone
, int count
,
637 struct per_cpu_pages
*pcp
)
643 spin_lock(&zone
->lock
);
644 zone
->all_unreclaimable
= 0;
645 zone
->pages_scanned
= 0;
649 struct list_head
*list
;
652 * Remove pages from lists in a round-robin fashion. A
653 * batch_free count is maintained that is incremented when an
654 * empty list is encountered. This is so more pages are freed
655 * off fuller lists instead of spinning excessively around empty
660 if (++migratetype
== MIGRATE_PCPTYPES
)
662 list
= &pcp
->lists
[migratetype
];
663 } while (list_empty(list
));
665 /* This is the only non-empty list. Free them all. */
666 if (batch_free
== MIGRATE_PCPTYPES
)
667 batch_free
= to_free
;
670 int mt
; /* migratetype of the to-be-freed page */
672 page
= list_entry(list
->prev
, struct page
, lru
);
673 /* must delete as __free_one_page list manipulates */
674 list_del(&page
->lru
);
675 mt
= get_freepage_migratetype(page
);
676 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
677 __free_one_page(page
, zone
, 0, mt
);
678 trace_mm_page_pcpu_drain(page
, 0, mt
);
679 if (likely(!is_migrate_isolate_page(page
))) {
680 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1);
681 if (is_migrate_cma(mt
))
682 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
, 1);
684 } while (--to_free
&& --batch_free
&& !list_empty(list
));
686 spin_unlock(&zone
->lock
);
689 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
692 spin_lock(&zone
->lock
);
693 zone
->all_unreclaimable
= 0;
694 zone
->pages_scanned
= 0;
696 __free_one_page(page
, zone
, order
, migratetype
);
697 if (unlikely(!is_migrate_isolate(migratetype
)))
698 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
699 spin_unlock(&zone
->lock
);
702 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
707 trace_mm_page_free(page
, order
);
708 kmemcheck_free_shadow(page
, order
);
711 page
->mapping
= NULL
;
712 for (i
= 0; i
< (1 << order
); i
++)
713 bad
+= free_pages_check(page
+ i
);
717 if (!PageHighMem(page
)) {
718 debug_check_no_locks_freed(page_address(page
),PAGE_SIZE
<<order
);
719 debug_check_no_obj_freed(page_address(page
),
722 arch_free_page(page
, order
);
723 kernel_map_pages(page
, 1 << order
, 0);
728 static void __free_pages_ok(struct page
*page
, unsigned int order
)
733 if (!free_pages_prepare(page
, order
))
736 local_irq_save(flags
);
737 __count_vm_events(PGFREE
, 1 << order
);
738 migratetype
= get_pageblock_migratetype(page
);
739 set_freepage_migratetype(page
, migratetype
);
740 free_one_page(page_zone(page
), page
, order
, migratetype
);
741 local_irq_restore(flags
);
745 * Read access to zone->managed_pages is safe because it's unsigned long,
746 * but we still need to serialize writers. Currently all callers of
747 * __free_pages_bootmem() except put_page_bootmem() should only be used
748 * at boot time. So for shorter boot time, we shift the burden to
749 * put_page_bootmem() to serialize writers.
751 void __meminit
__free_pages_bootmem(struct page
*page
, unsigned int order
)
753 unsigned int nr_pages
= 1 << order
;
757 for (loop
= 0; loop
< nr_pages
; loop
++) {
758 struct page
*p
= &page
[loop
];
760 if (loop
+ 1 < nr_pages
)
762 __ClearPageReserved(p
);
763 set_page_count(p
, 0);
766 page_zone(page
)->managed_pages
+= 1 << order
;
767 set_page_refcounted(page
);
768 __free_pages(page
, order
);
772 /* Free whole pageblock and set it's migration type to MIGRATE_CMA. */
773 void __init
init_cma_reserved_pageblock(struct page
*page
)
775 unsigned i
= pageblock_nr_pages
;
776 struct page
*p
= page
;
779 __ClearPageReserved(p
);
780 set_page_count(p
, 0);
783 set_page_refcounted(page
);
784 set_pageblock_migratetype(page
, MIGRATE_CMA
);
785 __free_pages(page
, pageblock_order
);
786 totalram_pages
+= pageblock_nr_pages
;
787 #ifdef CONFIG_HIGHMEM
788 if (PageHighMem(page
))
789 totalhigh_pages
+= pageblock_nr_pages
;
795 * The order of subdivision here is critical for the IO subsystem.
796 * Please do not alter this order without good reasons and regression
797 * testing. Specifically, as large blocks of memory are subdivided,
798 * the order in which smaller blocks are delivered depends on the order
799 * they're subdivided in this function. This is the primary factor
800 * influencing the order in which pages are delivered to the IO
801 * subsystem according to empirical testing, and this is also justified
802 * by considering the behavior of a buddy system containing a single
803 * large block of memory acted on by a series of small allocations.
804 * This behavior is a critical factor in sglist merging's success.
808 static inline void expand(struct zone
*zone
, struct page
*page
,
809 int low
, int high
, struct free_area
*area
,
812 unsigned long size
= 1 << high
;
818 VM_BUG_ON(bad_range(zone
, &page
[size
]));
820 #ifdef CONFIG_DEBUG_PAGEALLOC
821 if (high
< debug_guardpage_minorder()) {
823 * Mark as guard pages (or page), that will allow to
824 * merge back to allocator when buddy will be freed.
825 * Corresponding page table entries will not be touched,
826 * pages will stay not present in virtual address space
828 INIT_LIST_HEAD(&page
[size
].lru
);
829 set_page_guard_flag(&page
[size
]);
830 set_page_private(&page
[size
], high
);
831 /* Guard pages are not available for any usage */
832 __mod_zone_freepage_state(zone
, -(1 << high
),
837 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
839 set_page_order(&page
[size
], high
);
844 * This page is about to be returned from the page allocator
846 static inline int check_new_page(struct page
*page
)
848 if (unlikely(page_mapcount(page
) |
849 (page
->mapping
!= NULL
) |
850 (atomic_read(&page
->_count
) != 0) |
851 (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
) |
852 (mem_cgroup_bad_page_check(page
)))) {
859 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
863 for (i
= 0; i
< (1 << order
); i
++) {
864 struct page
*p
= page
+ i
;
865 if (unlikely(check_new_page(p
)))
869 set_page_private(page
, 0);
870 set_page_refcounted(page
);
872 arch_alloc_page(page
, order
);
873 kernel_map_pages(page
, 1 << order
, 1);
875 if (gfp_flags
& __GFP_ZERO
)
876 prep_zero_page(page
, order
, gfp_flags
);
878 if (order
&& (gfp_flags
& __GFP_COMP
))
879 prep_compound_page(page
, order
);
885 * Go through the free lists for the given migratetype and remove
886 * the smallest available page from the freelists
889 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
892 unsigned int current_order
;
893 struct free_area
* area
;
896 /* Find a page of the appropriate size in the preferred list */
897 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
898 area
= &(zone
->free_area
[current_order
]);
899 if (list_empty(&area
->free_list
[migratetype
]))
902 page
= list_entry(area
->free_list
[migratetype
].next
,
904 list_del(&page
->lru
);
905 rmv_page_order(page
);
907 expand(zone
, page
, order
, current_order
, area
, migratetype
);
916 * This array describes the order lists are fallen back to when
917 * the free lists for the desirable migrate type are depleted
919 static int fallbacks
[MIGRATE_TYPES
][4] = {
920 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
921 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
923 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
924 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
926 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
928 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
929 #ifdef CONFIG_MEMORY_ISOLATION
930 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
935 * Move the free pages in a range to the free lists of the requested type.
936 * Note that start_page and end_pages are not aligned on a pageblock
937 * boundary. If alignment is required, use move_freepages_block()
939 int move_freepages(struct zone
*zone
,
940 struct page
*start_page
, struct page
*end_page
,
947 #ifndef CONFIG_HOLES_IN_ZONE
949 * page_zone is not safe to call in this context when
950 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
951 * anyway as we check zone boundaries in move_freepages_block().
952 * Remove at a later date when no bug reports exist related to
953 * grouping pages by mobility
955 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
958 for (page
= start_page
; page
<= end_page
;) {
959 /* Make sure we are not inadvertently changing nodes */
960 VM_BUG_ON(page_to_nid(page
) != zone_to_nid(zone
));
962 if (!pfn_valid_within(page_to_pfn(page
))) {
967 if (!PageBuddy(page
)) {
972 order
= page_order(page
);
973 list_move(&page
->lru
,
974 &zone
->free_area
[order
].free_list
[migratetype
]);
975 set_freepage_migratetype(page
, migratetype
);
977 pages_moved
+= 1 << order
;
983 int move_freepages_block(struct zone
*zone
, struct page
*page
,
986 unsigned long start_pfn
, end_pfn
;
987 struct page
*start_page
, *end_page
;
989 start_pfn
= page_to_pfn(page
);
990 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
991 start_page
= pfn_to_page(start_pfn
);
992 end_page
= start_page
+ pageblock_nr_pages
- 1;
993 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
995 /* Do not cross zone boundaries */
996 if (!zone_spans_pfn(zone
, start_pfn
))
998 if (!zone_spans_pfn(zone
, end_pfn
))
1001 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1004 static void change_pageblock_range(struct page
*pageblock_page
,
1005 int start_order
, int migratetype
)
1007 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1009 while (nr_pageblocks
--) {
1010 set_pageblock_migratetype(pageblock_page
, migratetype
);
1011 pageblock_page
+= pageblock_nr_pages
;
1015 /* Remove an element from the buddy allocator from the fallback list */
1016 static inline struct page
*
1017 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
1019 struct free_area
* area
;
1024 /* Find the largest possible block of pages in the other list */
1025 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
1028 migratetype
= fallbacks
[start_migratetype
][i
];
1030 /* MIGRATE_RESERVE handled later if necessary */
1031 if (migratetype
== MIGRATE_RESERVE
)
1034 area
= &(zone
->free_area
[current_order
]);
1035 if (list_empty(&area
->free_list
[migratetype
]))
1038 page
= list_entry(area
->free_list
[migratetype
].next
,
1043 * If breaking a large block of pages, move all free
1044 * pages to the preferred allocation list. If falling
1045 * back for a reclaimable kernel allocation, be more
1046 * aggressive about taking ownership of free pages
1048 * On the other hand, never change migration
1049 * type of MIGRATE_CMA pageblocks nor move CMA
1050 * pages on different free lists. We don't
1051 * want unmovable pages to be allocated from
1052 * MIGRATE_CMA areas.
1054 if (!is_migrate_cma(migratetype
) &&
1055 (unlikely(current_order
>= pageblock_order
/ 2) ||
1056 start_migratetype
== MIGRATE_RECLAIMABLE
||
1057 page_group_by_mobility_disabled
)) {
1059 pages
= move_freepages_block(zone
, page
,
1062 /* Claim the whole block if over half of it is free */
1063 if (pages
>= (1 << (pageblock_order
-1)) ||
1064 page_group_by_mobility_disabled
)
1065 set_pageblock_migratetype(page
,
1068 migratetype
= start_migratetype
;
1071 /* Remove the page from the freelists */
1072 list_del(&page
->lru
);
1073 rmv_page_order(page
);
1075 /* Take ownership for orders >= pageblock_order */
1076 if (current_order
>= pageblock_order
&&
1077 !is_migrate_cma(migratetype
))
1078 change_pageblock_range(page
, current_order
,
1081 expand(zone
, page
, order
, current_order
, area
,
1082 is_migrate_cma(migratetype
)
1083 ? migratetype
: start_migratetype
);
1085 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1086 start_migratetype
, migratetype
);
1096 * Do the hard work of removing an element from the buddy allocator.
1097 * Call me with the zone->lock already held.
1099 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1105 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1107 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1108 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1111 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1112 * is used because __rmqueue_smallest is an inline function
1113 * and we want just one call site
1116 migratetype
= MIGRATE_RESERVE
;
1121 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1126 * Obtain a specified number of elements from the buddy allocator, all under
1127 * a single hold of the lock, for efficiency. Add them to the supplied list.
1128 * Returns the number of new pages which were placed at *list.
1130 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1131 unsigned long count
, struct list_head
*list
,
1132 int migratetype
, int cold
)
1134 int mt
= migratetype
, i
;
1136 spin_lock(&zone
->lock
);
1137 for (i
= 0; i
< count
; ++i
) {
1138 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1139 if (unlikely(page
== NULL
))
1143 * Split buddy pages returned by expand() are received here
1144 * in physical page order. The page is added to the callers and
1145 * list and the list head then moves forward. From the callers
1146 * perspective, the linked list is ordered by page number in
1147 * some conditions. This is useful for IO devices that can
1148 * merge IO requests if the physical pages are ordered
1151 if (likely(cold
== 0))
1152 list_add(&page
->lru
, list
);
1154 list_add_tail(&page
->lru
, list
);
1155 if (IS_ENABLED(CONFIG_CMA
)) {
1156 mt
= get_pageblock_migratetype(page
);
1157 if (!is_migrate_cma(mt
) && !is_migrate_isolate(mt
))
1160 set_freepage_migratetype(page
, mt
);
1162 if (is_migrate_cma(mt
))
1163 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1166 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1167 spin_unlock(&zone
->lock
);
1173 * Called from the vmstat counter updater to drain pagesets of this
1174 * currently executing processor on remote nodes after they have
1177 * Note that this function must be called with the thread pinned to
1178 * a single processor.
1180 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1182 unsigned long flags
;
1185 local_irq_save(flags
);
1186 if (pcp
->count
>= pcp
->batch
)
1187 to_drain
= pcp
->batch
;
1189 to_drain
= pcp
->count
;
1191 free_pcppages_bulk(zone
, to_drain
, pcp
);
1192 pcp
->count
-= to_drain
;
1194 local_irq_restore(flags
);
1199 * Drain pages of the indicated processor.
1201 * The processor must either be the current processor and the
1202 * thread pinned to the current processor or a processor that
1205 static void drain_pages(unsigned int cpu
)
1207 unsigned long flags
;
1210 for_each_populated_zone(zone
) {
1211 struct per_cpu_pageset
*pset
;
1212 struct per_cpu_pages
*pcp
;
1214 local_irq_save(flags
);
1215 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1219 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1222 local_irq_restore(flags
);
1227 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1229 void drain_local_pages(void *arg
)
1231 drain_pages(smp_processor_id());
1235 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1237 * Note that this code is protected against sending an IPI to an offline
1238 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1239 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1240 * nothing keeps CPUs from showing up after we populated the cpumask and
1241 * before the call to on_each_cpu_mask().
1243 void drain_all_pages(void)
1246 struct per_cpu_pageset
*pcp
;
1250 * Allocate in the BSS so we wont require allocation in
1251 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1253 static cpumask_t cpus_with_pcps
;
1256 * We don't care about racing with CPU hotplug event
1257 * as offline notification will cause the notified
1258 * cpu to drain that CPU pcps and on_each_cpu_mask
1259 * disables preemption as part of its processing
1261 for_each_online_cpu(cpu
) {
1262 bool has_pcps
= false;
1263 for_each_populated_zone(zone
) {
1264 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1265 if (pcp
->pcp
.count
) {
1271 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1273 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1275 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1278 #ifdef CONFIG_HIBERNATION
1280 void mark_free_pages(struct zone
*zone
)
1282 unsigned long pfn
, max_zone_pfn
;
1283 unsigned long flags
;
1285 struct list_head
*curr
;
1287 if (!zone
->spanned_pages
)
1290 spin_lock_irqsave(&zone
->lock
, flags
);
1292 max_zone_pfn
= zone_end_pfn(zone
);
1293 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1294 if (pfn_valid(pfn
)) {
1295 struct page
*page
= pfn_to_page(pfn
);
1297 if (!swsusp_page_is_forbidden(page
))
1298 swsusp_unset_page_free(page
);
1301 for_each_migratetype_order(order
, t
) {
1302 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1305 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1306 for (i
= 0; i
< (1UL << order
); i
++)
1307 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1310 spin_unlock_irqrestore(&zone
->lock
, flags
);
1312 #endif /* CONFIG_PM */
1315 * Free a 0-order page
1316 * cold == 1 ? free a cold page : free a hot page
1318 void free_hot_cold_page(struct page
*page
, int cold
)
1320 struct zone
*zone
= page_zone(page
);
1321 struct per_cpu_pages
*pcp
;
1322 unsigned long flags
;
1325 if (!free_pages_prepare(page
, 0))
1328 migratetype
= get_pageblock_migratetype(page
);
1329 set_freepage_migratetype(page
, migratetype
);
1330 local_irq_save(flags
);
1331 __count_vm_event(PGFREE
);
1334 * We only track unmovable, reclaimable and movable on pcp lists.
1335 * Free ISOLATE pages back to the allocator because they are being
1336 * offlined but treat RESERVE as movable pages so we can get those
1337 * areas back if necessary. Otherwise, we may have to free
1338 * excessively into the page allocator
1340 if (migratetype
>= MIGRATE_PCPTYPES
) {
1341 if (unlikely(is_migrate_isolate(migratetype
))) {
1342 free_one_page(zone
, page
, 0, migratetype
);
1345 migratetype
= MIGRATE_MOVABLE
;
1348 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1350 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1352 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1354 if (pcp
->count
>= pcp
->high
) {
1355 free_pcppages_bulk(zone
, pcp
->batch
, pcp
);
1356 pcp
->count
-= pcp
->batch
;
1360 local_irq_restore(flags
);
1364 * Free a list of 0-order pages
1366 void free_hot_cold_page_list(struct list_head
*list
, int cold
)
1368 struct page
*page
, *next
;
1370 list_for_each_entry_safe(page
, next
, list
, lru
) {
1371 trace_mm_page_free_batched(page
, cold
);
1372 free_hot_cold_page(page
, cold
);
1377 * split_page takes a non-compound higher-order page, and splits it into
1378 * n (1<<order) sub-pages: page[0..n]
1379 * Each sub-page must be freed individually.
1381 * Note: this is probably too low level an operation for use in drivers.
1382 * Please consult with lkml before using this in your driver.
1384 void split_page(struct page
*page
, unsigned int order
)
1388 VM_BUG_ON(PageCompound(page
));
1389 VM_BUG_ON(!page_count(page
));
1391 #ifdef CONFIG_KMEMCHECK
1393 * Split shadow pages too, because free(page[0]) would
1394 * otherwise free the whole shadow.
1396 if (kmemcheck_page_is_tracked(page
))
1397 split_page(virt_to_page(page
[0].shadow
), order
);
1400 for (i
= 1; i
< (1 << order
); i
++)
1401 set_page_refcounted(page
+ i
);
1403 EXPORT_SYMBOL_GPL(split_page
);
1405 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1407 unsigned long watermark
;
1411 BUG_ON(!PageBuddy(page
));
1413 zone
= page_zone(page
);
1414 mt
= get_pageblock_migratetype(page
);
1416 if (!is_migrate_isolate(mt
)) {
1417 /* Obey watermarks as if the page was being allocated */
1418 watermark
= low_wmark_pages(zone
) + (1 << order
);
1419 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1422 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1425 /* Remove page from free list */
1426 list_del(&page
->lru
);
1427 zone
->free_area
[order
].nr_free
--;
1428 rmv_page_order(page
);
1430 /* Set the pageblock if the isolated page is at least a pageblock */
1431 if (order
>= pageblock_order
- 1) {
1432 struct page
*endpage
= page
+ (1 << order
) - 1;
1433 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1434 int mt
= get_pageblock_migratetype(page
);
1435 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1436 set_pageblock_migratetype(page
,
1441 return 1UL << order
;
1445 * Similar to split_page except the page is already free. As this is only
1446 * being used for migration, the migratetype of the block also changes.
1447 * As this is called with interrupts disabled, the caller is responsible
1448 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1451 * Note: this is probably too low level an operation for use in drivers.
1452 * Please consult with lkml before using this in your driver.
1454 int split_free_page(struct page
*page
)
1459 order
= page_order(page
);
1461 nr_pages
= __isolate_free_page(page
, order
);
1465 /* Split into individual pages */
1466 set_page_refcounted(page
);
1467 split_page(page
, order
);
1472 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1473 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1477 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1478 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1481 unsigned long flags
;
1483 int cold
= !!(gfp_flags
& __GFP_COLD
);
1486 if (likely(order
== 0)) {
1487 struct per_cpu_pages
*pcp
;
1488 struct list_head
*list
;
1490 local_irq_save(flags
);
1491 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1492 list
= &pcp
->lists
[migratetype
];
1493 if (list_empty(list
)) {
1494 pcp
->count
+= rmqueue_bulk(zone
, 0,
1497 if (unlikely(list_empty(list
)))
1502 page
= list_entry(list
->prev
, struct page
, lru
);
1504 page
= list_entry(list
->next
, struct page
, lru
);
1506 list_del(&page
->lru
);
1509 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1511 * __GFP_NOFAIL is not to be used in new code.
1513 * All __GFP_NOFAIL callers should be fixed so that they
1514 * properly detect and handle allocation failures.
1516 * We most definitely don't want callers attempting to
1517 * allocate greater than order-1 page units with
1520 WARN_ON_ONCE(order
> 1);
1522 spin_lock_irqsave(&zone
->lock
, flags
);
1523 page
= __rmqueue(zone
, order
, migratetype
);
1524 spin_unlock(&zone
->lock
);
1527 __mod_zone_freepage_state(zone
, -(1 << order
),
1528 get_pageblock_migratetype(page
));
1531 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1532 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1533 local_irq_restore(flags
);
1535 VM_BUG_ON(bad_range(zone
, page
));
1536 if (prep_new_page(page
, order
, gfp_flags
))
1541 local_irq_restore(flags
);
1545 #ifdef CONFIG_FAIL_PAGE_ALLOC
1548 struct fault_attr attr
;
1550 u32 ignore_gfp_highmem
;
1551 u32 ignore_gfp_wait
;
1553 } fail_page_alloc
= {
1554 .attr
= FAULT_ATTR_INITIALIZER
,
1555 .ignore_gfp_wait
= 1,
1556 .ignore_gfp_highmem
= 1,
1560 static int __init
setup_fail_page_alloc(char *str
)
1562 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1564 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1566 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1568 if (order
< fail_page_alloc
.min_order
)
1570 if (gfp_mask
& __GFP_NOFAIL
)
1572 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1574 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1577 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1580 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1582 static int __init
fail_page_alloc_debugfs(void)
1584 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1587 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1588 &fail_page_alloc
.attr
);
1590 return PTR_ERR(dir
);
1592 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1593 &fail_page_alloc
.ignore_gfp_wait
))
1595 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1596 &fail_page_alloc
.ignore_gfp_highmem
))
1598 if (!debugfs_create_u32("min-order", mode
, dir
,
1599 &fail_page_alloc
.min_order
))
1604 debugfs_remove_recursive(dir
);
1609 late_initcall(fail_page_alloc_debugfs
);
1611 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1613 #else /* CONFIG_FAIL_PAGE_ALLOC */
1615 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1620 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1623 * Return true if free pages are above 'mark'. This takes into account the order
1624 * of the allocation.
1626 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1627 int classzone_idx
, int alloc_flags
, long free_pages
)
1629 /* free_pages my go negative - that's OK */
1631 long lowmem_reserve
= z
->lowmem_reserve
[classzone_idx
];
1635 free_pages
-= (1 << order
) - 1;
1636 if (alloc_flags
& ALLOC_HIGH
)
1638 if (alloc_flags
& ALLOC_HARDER
)
1641 /* If allocation can't use CMA areas don't use free CMA pages */
1642 if (!(alloc_flags
& ALLOC_CMA
))
1643 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1646 if (free_pages
- free_cma
<= min
+ lowmem_reserve
)
1648 for (o
= 0; o
< order
; o
++) {
1649 /* At the next order, this order's pages become unavailable */
1650 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1652 /* Require fewer higher order pages to be free */
1655 if (free_pages
<= min
)
1661 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1662 int classzone_idx
, int alloc_flags
)
1664 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1665 zone_page_state(z
, NR_FREE_PAGES
));
1668 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1669 int classzone_idx
, int alloc_flags
)
1671 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1673 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1674 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1676 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1682 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1683 * skip over zones that are not allowed by the cpuset, or that have
1684 * been recently (in last second) found to be nearly full. See further
1685 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1686 * that have to skip over a lot of full or unallowed zones.
1688 * If the zonelist cache is present in the passed in zonelist, then
1689 * returns a pointer to the allowed node mask (either the current
1690 * tasks mems_allowed, or node_states[N_MEMORY].)
1692 * If the zonelist cache is not available for this zonelist, does
1693 * nothing and returns NULL.
1695 * If the fullzones BITMAP in the zonelist cache is stale (more than
1696 * a second since last zap'd) then we zap it out (clear its bits.)
1698 * We hold off even calling zlc_setup, until after we've checked the
1699 * first zone in the zonelist, on the theory that most allocations will
1700 * be satisfied from that first zone, so best to examine that zone as
1701 * quickly as we can.
1703 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1705 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1706 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1708 zlc
= zonelist
->zlcache_ptr
;
1712 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1713 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1714 zlc
->last_full_zap
= jiffies
;
1717 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1718 &cpuset_current_mems_allowed
:
1719 &node_states
[N_MEMORY
];
1720 return allowednodes
;
1724 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1725 * if it is worth looking at further for free memory:
1726 * 1) Check that the zone isn't thought to be full (doesn't have its
1727 * bit set in the zonelist_cache fullzones BITMAP).
1728 * 2) Check that the zones node (obtained from the zonelist_cache
1729 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1730 * Return true (non-zero) if zone is worth looking at further, or
1731 * else return false (zero) if it is not.
1733 * This check -ignores- the distinction between various watermarks,
1734 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1735 * found to be full for any variation of these watermarks, it will
1736 * be considered full for up to one second by all requests, unless
1737 * we are so low on memory on all allowed nodes that we are forced
1738 * into the second scan of the zonelist.
1740 * In the second scan we ignore this zonelist cache and exactly
1741 * apply the watermarks to all zones, even it is slower to do so.
1742 * We are low on memory in the second scan, and should leave no stone
1743 * unturned looking for a free page.
1745 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1746 nodemask_t
*allowednodes
)
1748 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1749 int i
; /* index of *z in zonelist zones */
1750 int n
; /* node that zone *z is on */
1752 zlc
= zonelist
->zlcache_ptr
;
1756 i
= z
- zonelist
->_zonerefs
;
1759 /* This zone is worth trying if it is allowed but not full */
1760 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1764 * Given 'z' scanning a zonelist, set the corresponding bit in
1765 * zlc->fullzones, so that subsequent attempts to allocate a page
1766 * from that zone don't waste time re-examining it.
1768 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1770 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1771 int i
; /* index of *z in zonelist zones */
1773 zlc
= zonelist
->zlcache_ptr
;
1777 i
= z
- zonelist
->_zonerefs
;
1779 set_bit(i
, zlc
->fullzones
);
1783 * clear all zones full, called after direct reclaim makes progress so that
1784 * a zone that was recently full is not skipped over for up to a second
1786 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1788 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1790 zlc
= zonelist
->zlcache_ptr
;
1794 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1797 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1799 return node_isset(local_zone
->node
, zone
->zone_pgdat
->reclaim_nodes
);
1802 static void __paginginit
init_zone_allows_reclaim(int nid
)
1806 for_each_online_node(i
)
1807 if (node_distance(nid
, i
) <= RECLAIM_DISTANCE
)
1808 node_set(i
, NODE_DATA(nid
)->reclaim_nodes
);
1810 zone_reclaim_mode
= 1;
1813 #else /* CONFIG_NUMA */
1815 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1820 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1821 nodemask_t
*allowednodes
)
1826 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1830 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1834 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1839 static inline void init_zone_allows_reclaim(int nid
)
1842 #endif /* CONFIG_NUMA */
1845 * get_page_from_freelist goes through the zonelist trying to allocate
1848 static struct page
*
1849 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1850 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1851 struct zone
*preferred_zone
, int migratetype
)
1854 struct page
*page
= NULL
;
1857 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1858 int zlc_active
= 0; /* set if using zonelist_cache */
1859 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1861 classzone_idx
= zone_idx(preferred_zone
);
1864 * Scan zonelist, looking for a zone with enough free.
1865 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1867 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1868 high_zoneidx
, nodemask
) {
1869 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1870 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1872 if ((alloc_flags
& ALLOC_CPUSET
) &&
1873 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1876 * When allocating a page cache page for writing, we
1877 * want to get it from a zone that is within its dirty
1878 * limit, such that no single zone holds more than its
1879 * proportional share of globally allowed dirty pages.
1880 * The dirty limits take into account the zone's
1881 * lowmem reserves and high watermark so that kswapd
1882 * should be able to balance it without having to
1883 * write pages from its LRU list.
1885 * This may look like it could increase pressure on
1886 * lower zones by failing allocations in higher zones
1887 * before they are full. But the pages that do spill
1888 * over are limited as the lower zones are protected
1889 * by this very same mechanism. It should not become
1890 * a practical burden to them.
1892 * XXX: For now, allow allocations to potentially
1893 * exceed the per-zone dirty limit in the slowpath
1894 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1895 * which is important when on a NUMA setup the allowed
1896 * zones are together not big enough to reach the
1897 * global limit. The proper fix for these situations
1898 * will require awareness of zones in the
1899 * dirty-throttling and the flusher threads.
1901 if ((alloc_flags
& ALLOC_WMARK_LOW
) &&
1902 (gfp_mask
& __GFP_WRITE
) && !zone_dirty_ok(zone
))
1903 goto this_zone_full
;
1905 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1906 if (!(alloc_flags
& ALLOC_NO_WATERMARKS
)) {
1910 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1911 if (zone_watermark_ok(zone
, order
, mark
,
1912 classzone_idx
, alloc_flags
))
1915 if (IS_ENABLED(CONFIG_NUMA
) &&
1916 !did_zlc_setup
&& nr_online_nodes
> 1) {
1918 * we do zlc_setup if there are multiple nodes
1919 * and before considering the first zone allowed
1922 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1927 if (zone_reclaim_mode
== 0 ||
1928 !zone_allows_reclaim(preferred_zone
, zone
))
1929 goto this_zone_full
;
1932 * As we may have just activated ZLC, check if the first
1933 * eligible zone has failed zone_reclaim recently.
1935 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1936 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1939 ret
= zone_reclaim(zone
, gfp_mask
, order
);
1941 case ZONE_RECLAIM_NOSCAN
:
1944 case ZONE_RECLAIM_FULL
:
1945 /* scanned but unreclaimable */
1948 /* did we reclaim enough */
1949 if (zone_watermark_ok(zone
, order
, mark
,
1950 classzone_idx
, alloc_flags
))
1954 * Failed to reclaim enough to meet watermark.
1955 * Only mark the zone full if checking the min
1956 * watermark or if we failed to reclaim just
1957 * 1<<order pages or else the page allocator
1958 * fastpath will prematurely mark zones full
1959 * when the watermark is between the low and
1962 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
1963 ret
== ZONE_RECLAIM_SOME
)
1964 goto this_zone_full
;
1971 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
1972 gfp_mask
, migratetype
);
1976 if (IS_ENABLED(CONFIG_NUMA
))
1977 zlc_mark_zone_full(zonelist
, z
);
1980 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && page
== NULL
&& zlc_active
)) {
1981 /* Disable zlc cache for second zonelist scan */
1988 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
1989 * necessary to allocate the page. The expectation is
1990 * that the caller is taking steps that will free more
1991 * memory. The caller should avoid the page being used
1992 * for !PFMEMALLOC purposes.
1994 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2000 * Large machines with many possible nodes should not always dump per-node
2001 * meminfo in irq context.
2003 static inline bool should_suppress_show_mem(void)
2008 ret
= in_interrupt();
2013 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2014 DEFAULT_RATELIMIT_INTERVAL
,
2015 DEFAULT_RATELIMIT_BURST
);
2017 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2019 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2021 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2022 debug_guardpage_minorder() > 0)
2026 * Walking all memory to count page types is very expensive and should
2027 * be inhibited in non-blockable contexts.
2029 if (!(gfp_mask
& __GFP_WAIT
))
2030 filter
|= SHOW_MEM_FILTER_PAGE_COUNT
;
2033 * This documents exceptions given to allocations in certain
2034 * contexts that are allowed to allocate outside current's set
2037 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2038 if (test_thread_flag(TIF_MEMDIE
) ||
2039 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2040 filter
&= ~SHOW_MEM_FILTER_NODES
;
2041 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2042 filter
&= ~SHOW_MEM_FILTER_NODES
;
2045 struct va_format vaf
;
2048 va_start(args
, fmt
);
2053 pr_warn("%pV", &vaf
);
2058 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2059 current
->comm
, order
, gfp_mask
);
2062 if (!should_suppress_show_mem())
2067 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2068 unsigned long did_some_progress
,
2069 unsigned long pages_reclaimed
)
2071 /* Do not loop if specifically requested */
2072 if (gfp_mask
& __GFP_NORETRY
)
2075 /* Always retry if specifically requested */
2076 if (gfp_mask
& __GFP_NOFAIL
)
2080 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2081 * making forward progress without invoking OOM. Suspend also disables
2082 * storage devices so kswapd will not help. Bail if we are suspending.
2084 if (!did_some_progress
&& pm_suspended_storage())
2088 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2089 * means __GFP_NOFAIL, but that may not be true in other
2092 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2096 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2097 * specified, then we retry until we no longer reclaim any pages
2098 * (above), or we've reclaimed an order of pages at least as
2099 * large as the allocation's order. In both cases, if the
2100 * allocation still fails, we stop retrying.
2102 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2108 static inline struct page
*
2109 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2110 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2111 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2116 /* Acquire the OOM killer lock for the zones in zonelist */
2117 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
2118 schedule_timeout_uninterruptible(1);
2123 * PM-freezer should be notified that there might be an OOM killer on
2124 * its way to kill and wake somebody up. This is too early and we might
2125 * end up not killing anything but false positives are acceptable.
2126 * See freeze_processes.
2131 * Go through the zonelist yet one more time, keep very high watermark
2132 * here, this is only to catch a parallel oom killing, we must fail if
2133 * we're still under heavy pressure.
2135 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2136 order
, zonelist
, high_zoneidx
,
2137 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2138 preferred_zone
, migratetype
);
2142 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2143 /* The OOM killer will not help higher order allocs */
2144 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2146 /* The OOM killer does not needlessly kill tasks for lowmem */
2147 if (high_zoneidx
< ZONE_NORMAL
)
2150 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2151 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2152 * The caller should handle page allocation failure by itself if
2153 * it specifies __GFP_THISNODE.
2154 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2156 if (gfp_mask
& __GFP_THISNODE
)
2159 /* Exhausted what can be done so it's blamo time */
2160 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2163 clear_zonelist_oom(zonelist
, gfp_mask
);
2167 #ifdef CONFIG_COMPACTION
2168 /* Try memory compaction for high-order allocations before reclaim */
2169 static struct page
*
2170 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2171 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2172 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2173 int migratetype
, bool sync_migration
,
2174 bool *contended_compaction
, bool *deferred_compaction
,
2175 unsigned long *did_some_progress
)
2180 if (compaction_deferred(preferred_zone
, order
)) {
2181 *deferred_compaction
= true;
2185 current
->flags
|= PF_MEMALLOC
;
2186 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2187 nodemask
, sync_migration
,
2188 contended_compaction
);
2189 current
->flags
&= ~PF_MEMALLOC
;
2191 if (*did_some_progress
!= COMPACT_SKIPPED
) {
2194 /* Page migration frees to the PCP lists but we want merging */
2195 drain_pages(get_cpu());
2198 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2199 order
, zonelist
, high_zoneidx
,
2200 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2201 preferred_zone
, migratetype
);
2203 preferred_zone
->compact_blockskip_flush
= false;
2204 preferred_zone
->compact_considered
= 0;
2205 preferred_zone
->compact_defer_shift
= 0;
2206 if (order
>= preferred_zone
->compact_order_failed
)
2207 preferred_zone
->compact_order_failed
= order
+ 1;
2208 count_vm_event(COMPACTSUCCESS
);
2213 * It's bad if compaction run occurs and fails.
2214 * The most likely reason is that pages exist,
2215 * but not enough to satisfy watermarks.
2217 count_vm_event(COMPACTFAIL
);
2220 * As async compaction considers a subset of pageblocks, only
2221 * defer if the failure was a sync compaction failure.
2224 defer_compaction(preferred_zone
, order
);
2232 static inline struct page
*
2233 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2234 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2235 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2236 int migratetype
, bool sync_migration
,
2237 bool *contended_compaction
, bool *deferred_compaction
,
2238 unsigned long *did_some_progress
)
2242 #endif /* CONFIG_COMPACTION */
2244 /* Perform direct synchronous page reclaim */
2246 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2247 nodemask_t
*nodemask
)
2249 struct reclaim_state reclaim_state
;
2254 /* We now go into synchronous reclaim */
2255 cpuset_memory_pressure_bump();
2256 current
->flags
|= PF_MEMALLOC
;
2257 lockdep_set_current_reclaim_state(gfp_mask
);
2258 reclaim_state
.reclaimed_slab
= 0;
2259 current
->reclaim_state
= &reclaim_state
;
2261 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2263 current
->reclaim_state
= NULL
;
2264 lockdep_clear_current_reclaim_state();
2265 current
->flags
&= ~PF_MEMALLOC
;
2272 /* The really slow allocator path where we enter direct reclaim */
2273 static inline struct page
*
2274 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2275 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2276 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2277 int migratetype
, unsigned long *did_some_progress
)
2279 struct page
*page
= NULL
;
2280 bool drained
= false;
2282 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2284 if (unlikely(!(*did_some_progress
)))
2287 /* After successful reclaim, reconsider all zones for allocation */
2288 if (IS_ENABLED(CONFIG_NUMA
))
2289 zlc_clear_zones_full(zonelist
);
2292 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2293 zonelist
, high_zoneidx
,
2294 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2295 preferred_zone
, migratetype
);
2298 * If an allocation failed after direct reclaim, it could be because
2299 * pages are pinned on the per-cpu lists. Drain them and try again
2301 if (!page
&& !drained
) {
2311 * This is called in the allocator slow-path if the allocation request is of
2312 * sufficient urgency to ignore watermarks and take other desperate measures
2314 static inline struct page
*
2315 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2316 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2317 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2323 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2324 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2325 preferred_zone
, migratetype
);
2327 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2328 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2329 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2335 void wake_all_kswapd(unsigned int order
, struct zonelist
*zonelist
,
2336 enum zone_type high_zoneidx
,
2337 enum zone_type classzone_idx
)
2342 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
2343 wakeup_kswapd(zone
, order
, classzone_idx
);
2347 gfp_to_alloc_flags(gfp_t gfp_mask
)
2349 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2350 const bool atomic
= !(gfp_mask
& (__GFP_WAIT
| __GFP_NO_KSWAPD
));
2352 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2353 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2356 * The caller may dip into page reserves a bit more if the caller
2357 * cannot run direct reclaim, or if the caller has realtime scheduling
2358 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2359 * set both ALLOC_HARDER (atomic == true) and ALLOC_HIGH (__GFP_HIGH).
2361 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2365 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
2366 * if it can't schedule.
2368 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2369 alloc_flags
|= ALLOC_HARDER
;
2371 * Ignore cpuset mems for GFP_ATOMIC rather than fail, see the
2372 * comment for __cpuset_node_allowed_softwall().
2374 alloc_flags
&= ~ALLOC_CPUSET
;
2375 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2376 alloc_flags
|= ALLOC_HARDER
;
2378 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2379 if (gfp_mask
& __GFP_MEMALLOC
)
2380 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2381 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2382 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2383 else if (!in_interrupt() &&
2384 ((current
->flags
& PF_MEMALLOC
) ||
2385 unlikely(test_thread_flag(TIF_MEMDIE
))))
2386 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2389 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2390 alloc_flags
|= ALLOC_CMA
;
2395 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2397 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2400 static inline struct page
*
2401 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2402 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2403 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2406 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2407 struct page
*page
= NULL
;
2409 unsigned long pages_reclaimed
= 0;
2410 unsigned long did_some_progress
;
2411 bool sync_migration
= false;
2412 bool deferred_compaction
= false;
2413 bool contended_compaction
= false;
2416 * In the slowpath, we sanity check order to avoid ever trying to
2417 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2418 * be using allocators in order of preference for an area that is
2421 if (order
>= MAX_ORDER
) {
2422 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2427 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2428 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2429 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2430 * using a larger set of nodes after it has established that the
2431 * allowed per node queues are empty and that nodes are
2434 if (IS_ENABLED(CONFIG_NUMA
) &&
2435 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2439 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2440 wake_all_kswapd(order
, zonelist
, high_zoneidx
,
2441 zone_idx(preferred_zone
));
2444 * OK, we're below the kswapd watermark and have kicked background
2445 * reclaim. Now things get more complex, so set up alloc_flags according
2446 * to how we want to proceed.
2448 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2451 * Find the true preferred zone if the allocation is unconstrained by
2454 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2455 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2459 /* This is the last chance, in general, before the goto nopage. */
2460 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2461 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2462 preferred_zone
, migratetype
);
2466 /* Allocate without watermarks if the context allows */
2467 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2469 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2470 * the allocation is high priority and these type of
2471 * allocations are system rather than user orientated
2473 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2475 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2476 zonelist
, high_zoneidx
, nodemask
,
2477 preferred_zone
, migratetype
);
2483 /* Atomic allocations - we can't balance anything */
2487 /* Avoid recursion of direct reclaim */
2488 if (current
->flags
& PF_MEMALLOC
)
2491 /* Avoid allocations with no watermarks from looping endlessly */
2492 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2496 * Try direct compaction. The first pass is asynchronous. Subsequent
2497 * attempts after direct reclaim are synchronous
2499 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2500 zonelist
, high_zoneidx
,
2502 alloc_flags
, preferred_zone
,
2503 migratetype
, sync_migration
,
2504 &contended_compaction
,
2505 &deferred_compaction
,
2506 &did_some_progress
);
2509 sync_migration
= true;
2512 * If compaction is deferred for high-order allocations, it is because
2513 * sync compaction recently failed. In this is the case and the caller
2514 * requested a movable allocation that does not heavily disrupt the
2515 * system then fail the allocation instead of entering direct reclaim.
2517 if ((deferred_compaction
|| contended_compaction
) &&
2518 (gfp_mask
& __GFP_NO_KSWAPD
))
2521 /* Try direct reclaim and then allocating */
2522 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2523 zonelist
, high_zoneidx
,
2525 alloc_flags
, preferred_zone
,
2526 migratetype
, &did_some_progress
);
2531 * If we failed to make any progress reclaiming, then we are
2532 * running out of options and have to consider going OOM
2534 if (!did_some_progress
) {
2535 if ((gfp_mask
& __GFP_FS
) && !(gfp_mask
& __GFP_NORETRY
)) {
2536 if (oom_killer_disabled
)
2538 /* Coredumps can quickly deplete all memory reserves */
2539 if ((current
->flags
& PF_DUMPCORE
) &&
2540 !(gfp_mask
& __GFP_NOFAIL
))
2542 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2543 zonelist
, high_zoneidx
,
2544 nodemask
, preferred_zone
,
2549 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2551 * The oom killer is not called for high-order
2552 * allocations that may fail, so if no progress
2553 * is being made, there are no other options and
2554 * retrying is unlikely to help.
2556 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2559 * The oom killer is not called for lowmem
2560 * allocations to prevent needlessly killing
2563 if (high_zoneidx
< ZONE_NORMAL
)
2571 /* Check if we should retry the allocation */
2572 pages_reclaimed
+= did_some_progress
;
2573 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2575 /* Wait for some write requests to complete then retry */
2576 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2580 * High-order allocations do not necessarily loop after
2581 * direct reclaim and reclaim/compaction depends on compaction
2582 * being called after reclaim so call directly if necessary
2584 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2585 zonelist
, high_zoneidx
,
2587 alloc_flags
, preferred_zone
,
2588 migratetype
, sync_migration
,
2589 &contended_compaction
,
2590 &deferred_compaction
,
2591 &did_some_progress
);
2597 warn_alloc_failed(gfp_mask
, order
, NULL
);
2600 if (kmemcheck_enabled
)
2601 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2607 * This is the 'heart' of the zoned buddy allocator.
2610 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2611 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2613 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2614 struct zone
*preferred_zone
;
2615 struct page
*page
= NULL
;
2616 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2617 unsigned int cpuset_mems_cookie
;
2618 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
;
2619 struct mem_cgroup
*memcg
= NULL
;
2621 gfp_mask
&= gfp_allowed_mask
;
2623 lockdep_trace_alloc(gfp_mask
);
2625 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2627 if (should_fail_alloc_page(gfp_mask
, order
))
2631 * Check the zones suitable for the gfp_mask contain at least one
2632 * valid zone. It's possible to have an empty zonelist as a result
2633 * of GFP_THISNODE and a memoryless node
2635 if (unlikely(!zonelist
->_zonerefs
->zone
))
2639 * Will only have any effect when __GFP_KMEMCG is set. This is
2640 * verified in the (always inline) callee
2642 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2646 cpuset_mems_cookie
= get_mems_allowed();
2648 /* The preferred zone is used for statistics later */
2649 first_zones_zonelist(zonelist
, high_zoneidx
,
2650 nodemask
? : &cpuset_current_mems_allowed
,
2652 if (!preferred_zone
)
2656 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2657 alloc_flags
|= ALLOC_CMA
;
2659 /* First allocation attempt */
2660 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2661 zonelist
, high_zoneidx
, alloc_flags
,
2662 preferred_zone
, migratetype
);
2663 if (unlikely(!page
)) {
2665 * Runtime PM, block IO and its error handling path
2666 * can deadlock because I/O on the device might not
2669 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2670 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2671 zonelist
, high_zoneidx
, nodemask
,
2672 preferred_zone
, migratetype
);
2675 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2679 * When updating a task's mems_allowed, it is possible to race with
2680 * parallel threads in such a way that an allocation can fail while
2681 * the mask is being updated. If a page allocation is about to fail,
2682 * check if the cpuset changed during allocation and if so, retry.
2684 if (unlikely(!put_mems_allowed(cpuset_mems_cookie
) && !page
))
2687 memcg_kmem_commit_charge(page
, memcg
, order
);
2691 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2694 * Common helper functions.
2696 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2701 * __get_free_pages() returns a 32-bit address, which cannot represent
2704 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2706 page
= alloc_pages(gfp_mask
, order
);
2709 return (unsigned long) page_address(page
);
2711 EXPORT_SYMBOL(__get_free_pages
);
2713 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2715 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2717 EXPORT_SYMBOL(get_zeroed_page
);
2719 void __free_pages(struct page
*page
, unsigned int order
)
2721 if (put_page_testzero(page
)) {
2723 free_hot_cold_page(page
, 0);
2725 __free_pages_ok(page
, order
);
2729 EXPORT_SYMBOL(__free_pages
);
2731 void free_pages(unsigned long addr
, unsigned int order
)
2734 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2735 __free_pages(virt_to_page((void *)addr
), order
);
2739 EXPORT_SYMBOL(free_pages
);
2742 * __free_memcg_kmem_pages and free_memcg_kmem_pages will free
2743 * pages allocated with __GFP_KMEMCG.
2745 * Those pages are accounted to a particular memcg, embedded in the
2746 * corresponding page_cgroup. To avoid adding a hit in the allocator to search
2747 * for that information only to find out that it is NULL for users who have no
2748 * interest in that whatsoever, we provide these functions.
2750 * The caller knows better which flags it relies on.
2752 void __free_memcg_kmem_pages(struct page
*page
, unsigned int order
)
2754 memcg_kmem_uncharge_pages(page
, order
);
2755 __free_pages(page
, order
);
2758 void free_memcg_kmem_pages(unsigned long addr
, unsigned int order
)
2761 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2762 __free_memcg_kmem_pages(virt_to_page((void *)addr
), order
);
2766 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2769 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2770 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2772 split_page(virt_to_page((void *)addr
), order
);
2773 while (used
< alloc_end
) {
2778 return (void *)addr
;
2782 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2783 * @size: the number of bytes to allocate
2784 * @gfp_mask: GFP flags for the allocation
2786 * This function is similar to alloc_pages(), except that it allocates the
2787 * minimum number of pages to satisfy the request. alloc_pages() can only
2788 * allocate memory in power-of-two pages.
2790 * This function is also limited by MAX_ORDER.
2792 * Memory allocated by this function must be released by free_pages_exact().
2794 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2796 unsigned int order
= get_order(size
);
2799 addr
= __get_free_pages(gfp_mask
, order
);
2800 return make_alloc_exact(addr
, order
, size
);
2802 EXPORT_SYMBOL(alloc_pages_exact
);
2805 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2807 * @nid: the preferred node ID where memory should be allocated
2808 * @size: the number of bytes to allocate
2809 * @gfp_mask: GFP flags for the allocation
2811 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2813 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2816 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2818 unsigned order
= get_order(size
);
2819 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2822 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2824 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2827 * free_pages_exact - release memory allocated via alloc_pages_exact()
2828 * @virt: the value returned by alloc_pages_exact.
2829 * @size: size of allocation, same value as passed to alloc_pages_exact().
2831 * Release the memory allocated by a previous call to alloc_pages_exact.
2833 void free_pages_exact(void *virt
, size_t size
)
2835 unsigned long addr
= (unsigned long)virt
;
2836 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2838 while (addr
< end
) {
2843 EXPORT_SYMBOL(free_pages_exact
);
2846 * nr_free_zone_pages - count number of pages beyond high watermark
2847 * @offset: The zone index of the highest zone
2849 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2850 * high watermark within all zones at or below a given zone index. For each
2851 * zone, the number of pages is calculated as:
2852 * present_pages - high_pages
2854 static unsigned long nr_free_zone_pages(int offset
)
2859 /* Just pick one node, since fallback list is circular */
2860 unsigned long sum
= 0;
2862 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2864 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2865 unsigned long size
= zone
->managed_pages
;
2866 unsigned long high
= high_wmark_pages(zone
);
2875 * nr_free_buffer_pages - count number of pages beyond high watermark
2877 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2878 * watermark within ZONE_DMA and ZONE_NORMAL.
2880 unsigned long nr_free_buffer_pages(void)
2882 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2884 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2887 * nr_free_pagecache_pages - count number of pages beyond high watermark
2889 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2890 * high watermark within all zones.
2892 unsigned long nr_free_pagecache_pages(void)
2894 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2897 static inline void show_node(struct zone
*zone
)
2899 if (IS_ENABLED(CONFIG_NUMA
))
2900 printk("Node %d ", zone_to_nid(zone
));
2903 void si_meminfo(struct sysinfo
*val
)
2905 val
->totalram
= totalram_pages
;
2907 val
->freeram
= global_page_state(NR_FREE_PAGES
);
2908 val
->bufferram
= nr_blockdev_pages();
2909 val
->totalhigh
= totalhigh_pages
;
2910 val
->freehigh
= nr_free_highpages();
2911 val
->mem_unit
= PAGE_SIZE
;
2914 EXPORT_SYMBOL(si_meminfo
);
2917 void si_meminfo_node(struct sysinfo
*val
, int nid
)
2919 pg_data_t
*pgdat
= NODE_DATA(nid
);
2921 val
->totalram
= pgdat
->node_present_pages
;
2922 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
2923 #ifdef CONFIG_HIGHMEM
2924 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
2925 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
2931 val
->mem_unit
= PAGE_SIZE
;
2936 * Determine whether the node should be displayed or not, depending on whether
2937 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2939 bool skip_free_areas_node(unsigned int flags
, int nid
)
2942 unsigned int cpuset_mems_cookie
;
2944 if (!(flags
& SHOW_MEM_FILTER_NODES
))
2948 cpuset_mems_cookie
= get_mems_allowed();
2949 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
2950 } while (!put_mems_allowed(cpuset_mems_cookie
));
2955 #define K(x) ((x) << (PAGE_SHIFT-10))
2957 static void show_migration_types(unsigned char type
)
2959 static const char types
[MIGRATE_TYPES
] = {
2960 [MIGRATE_UNMOVABLE
] = 'U',
2961 [MIGRATE_RECLAIMABLE
] = 'E',
2962 [MIGRATE_MOVABLE
] = 'M',
2963 [MIGRATE_RESERVE
] = 'R',
2965 [MIGRATE_CMA
] = 'C',
2967 #ifdef CONFIG_MEMORY_ISOLATION
2968 [MIGRATE_ISOLATE
] = 'I',
2971 char tmp
[MIGRATE_TYPES
+ 1];
2975 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
2976 if (type
& (1 << i
))
2981 printk("(%s) ", tmp
);
2985 * Show free area list (used inside shift_scroll-lock stuff)
2986 * We also calculate the percentage fragmentation. We do this by counting the
2987 * memory on each free list with the exception of the first item on the list.
2988 * Suppresses nodes that are not allowed by current's cpuset if
2989 * SHOW_MEM_FILTER_NODES is passed.
2991 void show_free_areas(unsigned int filter
)
2996 for_each_populated_zone(zone
) {
2997 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3000 printk("%s per-cpu:\n", zone
->name
);
3002 for_each_online_cpu(cpu
) {
3003 struct per_cpu_pageset
*pageset
;
3005 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3007 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3008 cpu
, pageset
->pcp
.high
,
3009 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3013 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3014 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3016 " dirty:%lu writeback:%lu unstable:%lu\n"
3017 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3018 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3020 global_page_state(NR_ACTIVE_ANON
),
3021 global_page_state(NR_INACTIVE_ANON
),
3022 global_page_state(NR_ISOLATED_ANON
),
3023 global_page_state(NR_ACTIVE_FILE
),
3024 global_page_state(NR_INACTIVE_FILE
),
3025 global_page_state(NR_ISOLATED_FILE
),
3026 global_page_state(NR_UNEVICTABLE
),
3027 global_page_state(NR_FILE_DIRTY
),
3028 global_page_state(NR_WRITEBACK
),
3029 global_page_state(NR_UNSTABLE_NFS
),
3030 global_page_state(NR_FREE_PAGES
),
3031 global_page_state(NR_SLAB_RECLAIMABLE
),
3032 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3033 global_page_state(NR_FILE_MAPPED
),
3034 global_page_state(NR_SHMEM
),
3035 global_page_state(NR_PAGETABLE
),
3036 global_page_state(NR_BOUNCE
),
3037 global_page_state(NR_FREE_CMA_PAGES
));
3039 for_each_populated_zone(zone
) {
3042 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3050 " active_anon:%lukB"
3051 " inactive_anon:%lukB"
3052 " active_file:%lukB"
3053 " inactive_file:%lukB"
3054 " unevictable:%lukB"
3055 " isolated(anon):%lukB"
3056 " isolated(file):%lukB"
3064 " slab_reclaimable:%lukB"
3065 " slab_unreclaimable:%lukB"
3066 " kernel_stack:%lukB"
3071 " writeback_tmp:%lukB"
3072 " pages_scanned:%lu"
3073 " all_unreclaimable? %s"
3076 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3077 K(min_wmark_pages(zone
)),
3078 K(low_wmark_pages(zone
)),
3079 K(high_wmark_pages(zone
)),
3080 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3081 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3082 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3083 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3084 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3085 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3086 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3087 K(zone
->present_pages
),
3088 K(zone
->managed_pages
),
3089 K(zone_page_state(zone
, NR_MLOCK
)),
3090 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3091 K(zone_page_state(zone
, NR_WRITEBACK
)),
3092 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3093 K(zone_page_state(zone
, NR_SHMEM
)),
3094 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3095 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3096 zone_page_state(zone
, NR_KERNEL_STACK
) *
3098 K(zone_page_state(zone
, NR_PAGETABLE
)),
3099 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3100 K(zone_page_state(zone
, NR_BOUNCE
)),
3101 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3102 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3103 zone
->pages_scanned
,
3104 (zone
->all_unreclaimable
? "yes" : "no")
3106 printk("lowmem_reserve[]:");
3107 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3108 printk(" %lu", zone
->lowmem_reserve
[i
]);
3112 for_each_populated_zone(zone
) {
3113 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3114 unsigned char types
[MAX_ORDER
];
3116 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3119 printk("%s: ", zone
->name
);
3121 spin_lock_irqsave(&zone
->lock
, flags
);
3122 for (order
= 0; order
< MAX_ORDER
; order
++) {
3123 struct free_area
*area
= &zone
->free_area
[order
];
3126 nr
[order
] = area
->nr_free
;
3127 total
+= nr
[order
] << order
;
3130 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3131 if (!list_empty(&area
->free_list
[type
]))
3132 types
[order
] |= 1 << type
;
3135 spin_unlock_irqrestore(&zone
->lock
, flags
);
3136 for (order
= 0; order
< MAX_ORDER
; order
++) {
3137 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3139 show_migration_types(types
[order
]);
3141 printk("= %lukB\n", K(total
));
3144 hugetlb_show_meminfo();
3146 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3148 show_swap_cache_info();
3151 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3153 zoneref
->zone
= zone
;
3154 zoneref
->zone_idx
= zone_idx(zone
);
3158 * Builds allocation fallback zone lists.
3160 * Add all populated zones of a node to the zonelist.
3162 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3163 int nr_zones
, enum zone_type zone_type
)
3167 BUG_ON(zone_type
>= MAX_NR_ZONES
);
3172 zone
= pgdat
->node_zones
+ zone_type
;
3173 if (populated_zone(zone
)) {
3174 zoneref_set_zone(zone
,
3175 &zonelist
->_zonerefs
[nr_zones
++]);
3176 check_highest_zone(zone_type
);
3179 } while (zone_type
);
3186 * 0 = automatic detection of better ordering.
3187 * 1 = order by ([node] distance, -zonetype)
3188 * 2 = order by (-zonetype, [node] distance)
3190 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3191 * the same zonelist. So only NUMA can configure this param.
3193 #define ZONELIST_ORDER_DEFAULT 0
3194 #define ZONELIST_ORDER_NODE 1
3195 #define ZONELIST_ORDER_ZONE 2
3197 /* zonelist order in the kernel.
3198 * set_zonelist_order() will set this to NODE or ZONE.
3200 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3201 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3205 /* The value user specified ....changed by config */
3206 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3207 /* string for sysctl */
3208 #define NUMA_ZONELIST_ORDER_LEN 16
3209 char numa_zonelist_order
[16] = "default";
3212 * interface for configure zonelist ordering.
3213 * command line option "numa_zonelist_order"
3214 * = "[dD]efault - default, automatic configuration.
3215 * = "[nN]ode - order by node locality, then by zone within node
3216 * = "[zZ]one - order by zone, then by locality within zone
3219 static int __parse_numa_zonelist_order(char *s
)
3221 if (*s
== 'd' || *s
== 'D') {
3222 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3223 } else if (*s
== 'n' || *s
== 'N') {
3224 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3225 } else if (*s
== 'z' || *s
== 'Z') {
3226 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3229 "Ignoring invalid numa_zonelist_order value: "
3236 static __init
int setup_numa_zonelist_order(char *s
)
3243 ret
= __parse_numa_zonelist_order(s
);
3245 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3249 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3252 * sysctl handler for numa_zonelist_order
3254 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
3255 void __user
*buffer
, size_t *length
,
3258 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3260 static DEFINE_MUTEX(zl_order_mutex
);
3262 mutex_lock(&zl_order_mutex
);
3264 strcpy(saved_string
, (char*)table
->data
);
3265 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3269 int oldval
= user_zonelist_order
;
3270 if (__parse_numa_zonelist_order((char*)table
->data
)) {
3272 * bogus value. restore saved string
3274 strncpy((char*)table
->data
, saved_string
,
3275 NUMA_ZONELIST_ORDER_LEN
);
3276 user_zonelist_order
= oldval
;
3277 } else if (oldval
!= user_zonelist_order
) {
3278 mutex_lock(&zonelists_mutex
);
3279 build_all_zonelists(NULL
, NULL
);
3280 mutex_unlock(&zonelists_mutex
);
3284 mutex_unlock(&zl_order_mutex
);
3289 #define MAX_NODE_LOAD (nr_online_nodes)
3290 static int node_load
[MAX_NUMNODES
];
3293 * find_next_best_node - find the next node that should appear in a given node's fallback list
3294 * @node: node whose fallback list we're appending
3295 * @used_node_mask: nodemask_t of already used nodes
3297 * We use a number of factors to determine which is the next node that should
3298 * appear on a given node's fallback list. The node should not have appeared
3299 * already in @node's fallback list, and it should be the next closest node
3300 * according to the distance array (which contains arbitrary distance values
3301 * from each node to each node in the system), and should also prefer nodes
3302 * with no CPUs, since presumably they'll have very little allocation pressure
3303 * on them otherwise.
3304 * It returns -1 if no node is found.
3306 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3309 int min_val
= INT_MAX
;
3310 int best_node
= NUMA_NO_NODE
;
3311 const struct cpumask
*tmp
= cpumask_of_node(0);
3313 /* Use the local node if we haven't already */
3314 if (!node_isset(node
, *used_node_mask
)) {
3315 node_set(node
, *used_node_mask
);
3319 for_each_node_state(n
, N_MEMORY
) {
3321 /* Don't want a node to appear more than once */
3322 if (node_isset(n
, *used_node_mask
))
3325 /* Use the distance array to find the distance */
3326 val
= node_distance(node
, n
);
3328 /* Penalize nodes under us ("prefer the next node") */
3331 /* Give preference to headless and unused nodes */
3332 tmp
= cpumask_of_node(n
);
3333 if (!cpumask_empty(tmp
))
3334 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3336 /* Slight preference for less loaded node */
3337 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3338 val
+= node_load
[n
];
3340 if (val
< min_val
) {
3347 node_set(best_node
, *used_node_mask
);
3354 * Build zonelists ordered by node and zones within node.
3355 * This results in maximum locality--normal zone overflows into local
3356 * DMA zone, if any--but risks exhausting DMA zone.
3358 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3361 struct zonelist
*zonelist
;
3363 zonelist
= &pgdat
->node_zonelists
[0];
3364 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3366 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3368 zonelist
->_zonerefs
[j
].zone
= NULL
;
3369 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3373 * Build gfp_thisnode zonelists
3375 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3378 struct zonelist
*zonelist
;
3380 zonelist
= &pgdat
->node_zonelists
[1];
3381 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3382 zonelist
->_zonerefs
[j
].zone
= NULL
;
3383 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3387 * Build zonelists ordered by zone and nodes within zones.
3388 * This results in conserving DMA zone[s] until all Normal memory is
3389 * exhausted, but results in overflowing to remote node while memory
3390 * may still exist in local DMA zone.
3392 static int node_order
[MAX_NUMNODES
];
3394 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3397 int zone_type
; /* needs to be signed */
3399 struct zonelist
*zonelist
;
3401 zonelist
= &pgdat
->node_zonelists
[0];
3403 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3404 for (j
= 0; j
< nr_nodes
; j
++) {
3405 node
= node_order
[j
];
3406 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3407 if (populated_zone(z
)) {
3409 &zonelist
->_zonerefs
[pos
++]);
3410 check_highest_zone(zone_type
);
3414 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3415 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3418 static int default_zonelist_order(void)
3421 unsigned long low_kmem_size
,total_size
;
3425 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3426 * If they are really small and used heavily, the system can fall
3427 * into OOM very easily.
3428 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3430 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3433 for_each_online_node(nid
) {
3434 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3435 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3436 if (populated_zone(z
)) {
3437 if (zone_type
< ZONE_NORMAL
)
3438 low_kmem_size
+= z
->present_pages
;
3439 total_size
+= z
->present_pages
;
3440 } else if (zone_type
== ZONE_NORMAL
) {
3442 * If any node has only lowmem, then node order
3443 * is preferred to allow kernel allocations
3444 * locally; otherwise, they can easily infringe
3445 * on other nodes when there is an abundance of
3446 * lowmem available to allocate from.
3448 return ZONELIST_ORDER_NODE
;
3452 if (!low_kmem_size
|| /* there are no DMA area. */
3453 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
3454 return ZONELIST_ORDER_NODE
;
3456 * look into each node's config.
3457 * If there is a node whose DMA/DMA32 memory is very big area on
3458 * local memory, NODE_ORDER may be suitable.
3460 average_size
= total_size
/
3461 (nodes_weight(node_states
[N_MEMORY
]) + 1);
3462 for_each_online_node(nid
) {
3465 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3466 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3467 if (populated_zone(z
)) {
3468 if (zone_type
< ZONE_NORMAL
)
3469 low_kmem_size
+= z
->present_pages
;
3470 total_size
+= z
->present_pages
;
3473 if (low_kmem_size
&&
3474 total_size
> average_size
&& /* ignore small node */
3475 low_kmem_size
> total_size
* 70/100)
3476 return ZONELIST_ORDER_NODE
;
3478 return ZONELIST_ORDER_ZONE
;
3481 static void set_zonelist_order(void)
3483 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3484 current_zonelist_order
= default_zonelist_order();
3486 current_zonelist_order
= user_zonelist_order
;
3489 static void build_zonelists(pg_data_t
*pgdat
)
3493 nodemask_t used_mask
;
3494 int local_node
, prev_node
;
3495 struct zonelist
*zonelist
;
3496 int order
= current_zonelist_order
;
3498 /* initialize zonelists */
3499 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3500 zonelist
= pgdat
->node_zonelists
+ i
;
3501 zonelist
->_zonerefs
[0].zone
= NULL
;
3502 zonelist
->_zonerefs
[0].zone_idx
= 0;
3505 /* NUMA-aware ordering of nodes */
3506 local_node
= pgdat
->node_id
;
3507 load
= nr_online_nodes
;
3508 prev_node
= local_node
;
3509 nodes_clear(used_mask
);
3511 memset(node_order
, 0, sizeof(node_order
));
3514 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3516 * We don't want to pressure a particular node.
3517 * So adding penalty to the first node in same
3518 * distance group to make it round-robin.
3520 if (node_distance(local_node
, node
) !=
3521 node_distance(local_node
, prev_node
))
3522 node_load
[node
] = load
;
3526 if (order
== ZONELIST_ORDER_NODE
)
3527 build_zonelists_in_node_order(pgdat
, node
);
3529 node_order
[j
++] = node
; /* remember order */
3532 if (order
== ZONELIST_ORDER_ZONE
) {
3533 /* calculate node order -- i.e., DMA last! */
3534 build_zonelists_in_zone_order(pgdat
, j
);
3537 build_thisnode_zonelists(pgdat
);
3540 /* Construct the zonelist performance cache - see further mmzone.h */
3541 static void build_zonelist_cache(pg_data_t
*pgdat
)
3543 struct zonelist
*zonelist
;
3544 struct zonelist_cache
*zlc
;
3547 zonelist
= &pgdat
->node_zonelists
[0];
3548 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3549 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3550 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3551 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3554 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3556 * Return node id of node used for "local" allocations.
3557 * I.e., first node id of first zone in arg node's generic zonelist.
3558 * Used for initializing percpu 'numa_mem', which is used primarily
3559 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3561 int local_memory_node(int node
)
3565 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3566 gfp_zone(GFP_KERNEL
),
3573 #else /* CONFIG_NUMA */
3575 static void set_zonelist_order(void)
3577 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3580 static void build_zonelists(pg_data_t
*pgdat
)
3582 int node
, local_node
;
3584 struct zonelist
*zonelist
;
3586 local_node
= pgdat
->node_id
;
3588 zonelist
= &pgdat
->node_zonelists
[0];
3589 j
= build_zonelists_node(pgdat
, zonelist
, 0, MAX_NR_ZONES
- 1);
3592 * Now we build the zonelist so that it contains the zones
3593 * of all the other nodes.
3594 * We don't want to pressure a particular node, so when
3595 * building the zones for node N, we make sure that the
3596 * zones coming right after the local ones are those from
3597 * node N+1 (modulo N)
3599 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3600 if (!node_online(node
))
3602 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3605 for (node
= 0; node
< local_node
; node
++) {
3606 if (!node_online(node
))
3608 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
,
3612 zonelist
->_zonerefs
[j
].zone
= NULL
;
3613 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3616 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3617 static void build_zonelist_cache(pg_data_t
*pgdat
)
3619 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3622 #endif /* CONFIG_NUMA */
3625 * Boot pageset table. One per cpu which is going to be used for all
3626 * zones and all nodes. The parameters will be set in such a way
3627 * that an item put on a list will immediately be handed over to
3628 * the buddy list. This is safe since pageset manipulation is done
3629 * with interrupts disabled.
3631 * The boot_pagesets must be kept even after bootup is complete for
3632 * unused processors and/or zones. They do play a role for bootstrapping
3633 * hotplugged processors.
3635 * zoneinfo_show() and maybe other functions do
3636 * not check if the processor is online before following the pageset pointer.
3637 * Other parts of the kernel may not check if the zone is available.
3639 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3640 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3641 static void setup_zone_pageset(struct zone
*zone
);
3644 * Global mutex to protect against size modification of zonelists
3645 * as well as to serialize pageset setup for the new populated zone.
3647 DEFINE_MUTEX(zonelists_mutex
);
3649 /* return values int ....just for stop_machine() */
3650 static int __build_all_zonelists(void *data
)
3654 pg_data_t
*self
= data
;
3657 memset(node_load
, 0, sizeof(node_load
));
3660 if (self
&& !node_online(self
->node_id
)) {
3661 build_zonelists(self
);
3662 build_zonelist_cache(self
);
3665 for_each_online_node(nid
) {
3666 pg_data_t
*pgdat
= NODE_DATA(nid
);
3668 build_zonelists(pgdat
);
3669 build_zonelist_cache(pgdat
);
3673 * Initialize the boot_pagesets that are going to be used
3674 * for bootstrapping processors. The real pagesets for
3675 * each zone will be allocated later when the per cpu
3676 * allocator is available.
3678 * boot_pagesets are used also for bootstrapping offline
3679 * cpus if the system is already booted because the pagesets
3680 * are needed to initialize allocators on a specific cpu too.
3681 * F.e. the percpu allocator needs the page allocator which
3682 * needs the percpu allocator in order to allocate its pagesets
3683 * (a chicken-egg dilemma).
3685 for_each_possible_cpu(cpu
) {
3686 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3688 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3690 * We now know the "local memory node" for each node--
3691 * i.e., the node of the first zone in the generic zonelist.
3692 * Set up numa_mem percpu variable for on-line cpus. During
3693 * boot, only the boot cpu should be on-line; we'll init the
3694 * secondary cpus' numa_mem as they come on-line. During
3695 * node/memory hotplug, we'll fixup all on-line cpus.
3697 if (cpu_online(cpu
))
3698 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3706 * Called with zonelists_mutex held always
3707 * unless system_state == SYSTEM_BOOTING.
3709 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3711 set_zonelist_order();
3713 if (system_state
== SYSTEM_BOOTING
) {
3714 __build_all_zonelists(NULL
);
3715 mminit_verify_zonelist();
3716 cpuset_init_current_mems_allowed();
3718 /* we have to stop all cpus to guarantee there is no user
3720 #ifdef CONFIG_MEMORY_HOTPLUG
3722 setup_zone_pageset(zone
);
3724 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3725 /* cpuset refresh routine should be here */
3727 vm_total_pages
= nr_free_pagecache_pages();
3729 * Disable grouping by mobility if the number of pages in the
3730 * system is too low to allow the mechanism to work. It would be
3731 * more accurate, but expensive to check per-zone. This check is
3732 * made on memory-hotadd so a system can start with mobility
3733 * disabled and enable it later
3735 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3736 page_group_by_mobility_disabled
= 1;
3738 page_group_by_mobility_disabled
= 0;
3740 printk("Built %i zonelists in %s order, mobility grouping %s. "
3741 "Total pages: %ld\n",
3743 zonelist_order_name
[current_zonelist_order
],
3744 page_group_by_mobility_disabled
? "off" : "on",
3747 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3752 * Helper functions to size the waitqueue hash table.
3753 * Essentially these want to choose hash table sizes sufficiently
3754 * large so that collisions trying to wait on pages are rare.
3755 * But in fact, the number of active page waitqueues on typical
3756 * systems is ridiculously low, less than 200. So this is even
3757 * conservative, even though it seems large.
3759 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3760 * waitqueues, i.e. the size of the waitq table given the number of pages.
3762 #define PAGES_PER_WAITQUEUE 256
3764 #ifndef CONFIG_MEMORY_HOTPLUG
3765 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3767 unsigned long size
= 1;
3769 pages
/= PAGES_PER_WAITQUEUE
;
3771 while (size
< pages
)
3775 * Once we have dozens or even hundreds of threads sleeping
3776 * on IO we've got bigger problems than wait queue collision.
3777 * Limit the size of the wait table to a reasonable size.
3779 size
= min(size
, 4096UL);
3781 return max(size
, 4UL);
3785 * A zone's size might be changed by hot-add, so it is not possible to determine
3786 * a suitable size for its wait_table. So we use the maximum size now.
3788 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3790 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3791 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3792 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3794 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3795 * or more by the traditional way. (See above). It equals:
3797 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3798 * ia64(16K page size) : = ( 8G + 4M)byte.
3799 * powerpc (64K page size) : = (32G +16M)byte.
3801 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3808 * This is an integer logarithm so that shifts can be used later
3809 * to extract the more random high bits from the multiplicative
3810 * hash function before the remainder is taken.
3812 static inline unsigned long wait_table_bits(unsigned long size
)
3817 #define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3820 * Check if a pageblock contains reserved pages
3822 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3826 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3827 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3834 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3835 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3836 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3837 * higher will lead to a bigger reserve which will get freed as contiguous
3838 * blocks as reclaim kicks in
3840 static void setup_zone_migrate_reserve(struct zone
*zone
)
3842 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3844 unsigned long block_migratetype
;
3848 * Get the start pfn, end pfn and the number of blocks to reserve
3849 * We have to be careful to be aligned to pageblock_nr_pages to
3850 * make sure that we always check pfn_valid for the first page in
3853 start_pfn
= zone
->zone_start_pfn
;
3854 end_pfn
= zone_end_pfn(zone
);
3855 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
3856 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3860 * Reserve blocks are generally in place to help high-order atomic
3861 * allocations that are short-lived. A min_free_kbytes value that
3862 * would result in more than 2 reserve blocks for atomic allocations
3863 * is assumed to be in place to help anti-fragmentation for the
3864 * future allocation of hugepages at runtime.
3866 reserve
= min(2, reserve
);
3868 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3869 if (!pfn_valid(pfn
))
3871 page
= pfn_to_page(pfn
);
3873 /* Watch out for overlapping nodes */
3874 if (page_to_nid(page
) != zone_to_nid(zone
))
3877 block_migratetype
= get_pageblock_migratetype(page
);
3879 /* Only test what is necessary when the reserves are not met */
3882 * Blocks with reserved pages will never free, skip
3885 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
3886 if (pageblock_is_reserved(pfn
, block_end_pfn
))
3889 /* If this block is reserved, account for it */
3890 if (block_migratetype
== MIGRATE_RESERVE
) {
3895 /* Suitable for reserving if this block is movable */
3896 if (block_migratetype
== MIGRATE_MOVABLE
) {
3897 set_pageblock_migratetype(page
,
3899 move_freepages_block(zone
, page
,
3907 * If the reserve is met and this is a previous reserved block,
3910 if (block_migratetype
== MIGRATE_RESERVE
) {
3911 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3912 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
3918 * Initially all pages are reserved - free ones are freed
3919 * up by free_all_bootmem() once the early boot process is
3920 * done. Non-atomic initialization, single-pass.
3922 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
3923 unsigned long start_pfn
, enum memmap_context context
)
3926 unsigned long end_pfn
= start_pfn
+ size
;
3930 if (highest_memmap_pfn
< end_pfn
- 1)
3931 highest_memmap_pfn
= end_pfn
- 1;
3933 z
= &NODE_DATA(nid
)->node_zones
[zone
];
3934 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3936 * There can be holes in boot-time mem_map[]s
3937 * handed to this function. They do not
3938 * exist on hotplugged memory.
3940 if (context
== MEMMAP_EARLY
) {
3941 if (!early_pfn_valid(pfn
))
3943 if (!early_pfn_in_nid(pfn
, nid
))
3946 page
= pfn_to_page(pfn
);
3947 set_page_links(page
, zone
, nid
, pfn
);
3948 mminit_verify_page_links(page
, zone
, nid
, pfn
);
3949 init_page_count(page
);
3950 page_mapcount_reset(page
);
3951 page_nid_reset_last(page
);
3952 SetPageReserved(page
);
3954 * Mark the block movable so that blocks are reserved for
3955 * movable at startup. This will force kernel allocations
3956 * to reserve their blocks rather than leaking throughout
3957 * the address space during boot when many long-lived
3958 * kernel allocations are made. Later some blocks near
3959 * the start are marked MIGRATE_RESERVE by
3960 * setup_zone_migrate_reserve()
3962 * bitmap is created for zone's valid pfn range. but memmap
3963 * can be created for invalid pages (for alignment)
3964 * check here not to call set_pageblock_migratetype() against
3967 if ((z
->zone_start_pfn
<= pfn
)
3968 && (pfn
< zone_end_pfn(z
))
3969 && !(pfn
& (pageblock_nr_pages
- 1)))
3970 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
3972 INIT_LIST_HEAD(&page
->lru
);
3973 #ifdef WANT_PAGE_VIRTUAL
3974 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3975 if (!is_highmem_idx(zone
))
3976 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
3981 static void __meminit
zone_init_free_lists(struct zone
*zone
)
3984 for_each_migratetype_order(order
, t
) {
3985 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
3986 zone
->free_area
[order
].nr_free
= 0;
3990 #ifndef __HAVE_ARCH_MEMMAP_INIT
3991 #define memmap_init(size, nid, zone, start_pfn) \
3992 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3995 static int __meminit
zone_batchsize(struct zone
*zone
)
4001 * The per-cpu-pages pools are set to around 1000th of the
4002 * size of the zone. But no more than 1/2 of a meg.
4004 * OK, so we don't know how big the cache is. So guess.
4006 batch
= zone
->managed_pages
/ 1024;
4007 if (batch
* PAGE_SIZE
> 512 * 1024)
4008 batch
= (512 * 1024) / PAGE_SIZE
;
4009 batch
/= 4; /* We effectively *= 4 below */
4014 * Clamp the batch to a 2^n - 1 value. Having a power
4015 * of 2 value was found to be more likely to have
4016 * suboptimal cache aliasing properties in some cases.
4018 * For example if 2 tasks are alternately allocating
4019 * batches of pages, one task can end up with a lot
4020 * of pages of one half of the possible page colors
4021 * and the other with pages of the other colors.
4023 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4028 /* The deferral and batching of frees should be suppressed under NOMMU
4031 * The problem is that NOMMU needs to be able to allocate large chunks
4032 * of contiguous memory as there's no hardware page translation to
4033 * assemble apparent contiguous memory from discontiguous pages.
4035 * Queueing large contiguous runs of pages for batching, however,
4036 * causes the pages to actually be freed in smaller chunks. As there
4037 * can be a significant delay between the individual batches being
4038 * recycled, this leads to the once large chunks of space being
4039 * fragmented and becoming unavailable for high-order allocations.
4045 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4047 struct per_cpu_pages
*pcp
;
4050 memset(p
, 0, sizeof(*p
));
4054 pcp
->high
= 6 * batch
;
4055 pcp
->batch
= max(1UL, 1 * batch
);
4056 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4057 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4061 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
4062 * to the value high for the pageset p.
4065 static void setup_pagelist_highmark(struct per_cpu_pageset
*p
,
4068 struct per_cpu_pages
*pcp
;
4072 pcp
->batch
= max(1UL, high
/4);
4073 if ((high
/4) > (PAGE_SHIFT
* 8))
4074 pcp
->batch
= PAGE_SHIFT
* 8;
4077 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4081 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4083 for_each_possible_cpu(cpu
) {
4084 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4086 setup_pageset(pcp
, zone_batchsize(zone
));
4088 if (percpu_pagelist_fraction
)
4089 setup_pagelist_highmark(pcp
,
4090 (zone
->managed_pages
/
4091 percpu_pagelist_fraction
));
4096 * Allocate per cpu pagesets and initialize them.
4097 * Before this call only boot pagesets were available.
4099 void __init
setup_per_cpu_pageset(void)
4103 for_each_populated_zone(zone
)
4104 setup_zone_pageset(zone
);
4107 static noinline __init_refok
4108 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4111 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4115 * The per-page waitqueue mechanism uses hashed waitqueues
4118 zone
->wait_table_hash_nr_entries
=
4119 wait_table_hash_nr_entries(zone_size_pages
);
4120 zone
->wait_table_bits
=
4121 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4122 alloc_size
= zone
->wait_table_hash_nr_entries
4123 * sizeof(wait_queue_head_t
);
4125 if (!slab_is_available()) {
4126 zone
->wait_table
= (wait_queue_head_t
*)
4127 alloc_bootmem_node_nopanic(pgdat
, alloc_size
);
4130 * This case means that a zone whose size was 0 gets new memory
4131 * via memory hot-add.
4132 * But it may be the case that a new node was hot-added. In
4133 * this case vmalloc() will not be able to use this new node's
4134 * memory - this wait_table must be initialized to use this new
4135 * node itself as well.
4136 * To use this new node's memory, further consideration will be
4139 zone
->wait_table
= vmalloc(alloc_size
);
4141 if (!zone
->wait_table
)
4144 for(i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4145 init_waitqueue_head(zone
->wait_table
+ i
);
4150 static __meminit
void zone_pcp_init(struct zone
*zone
)
4153 * per cpu subsystem is not up at this point. The following code
4154 * relies on the ability of the linker to provide the
4155 * offset of a (static) per cpu variable into the per cpu area.
4157 zone
->pageset
= &boot_pageset
;
4159 if (zone
->present_pages
)
4160 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4161 zone
->name
, zone
->present_pages
,
4162 zone_batchsize(zone
));
4165 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4166 unsigned long zone_start_pfn
,
4168 enum memmap_context context
)
4170 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4172 ret
= zone_wait_table_init(zone
, size
);
4175 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4177 zone
->zone_start_pfn
= zone_start_pfn
;
4179 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4180 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4182 (unsigned long)zone_idx(zone
),
4183 zone_start_pfn
, (zone_start_pfn
+ size
));
4185 zone_init_free_lists(zone
);
4190 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4191 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4193 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4194 * Architectures may implement their own version but if add_active_range()
4195 * was used and there are no special requirements, this is a convenient
4198 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4200 unsigned long start_pfn
, end_pfn
;
4203 * NOTE: The following SMP-unsafe globals are only used early in boot
4204 * when the kernel is running single-threaded.
4206 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4207 static int __meminitdata last_nid
;
4209 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4212 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
4213 if (start_pfn
<= pfn
&& pfn
< end_pfn
) {
4214 last_start_pfn
= start_pfn
;
4215 last_end_pfn
= end_pfn
;
4219 /* This is a memory hole */
4222 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4224 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4228 nid
= __early_pfn_to_nid(pfn
);
4231 /* just returns 0 */
4235 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4236 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4240 nid
= __early_pfn_to_nid(pfn
);
4241 if (nid
>= 0 && nid
!= node
)
4248 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
4249 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4250 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
4252 * If an architecture guarantees that all ranges registered with
4253 * add_active_ranges() contain no holes and may be freed, this
4254 * this function may be used instead of calling free_bootmem() manually.
4256 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4258 unsigned long start_pfn
, end_pfn
;
4261 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4262 start_pfn
= min(start_pfn
, max_low_pfn
);
4263 end_pfn
= min(end_pfn
, max_low_pfn
);
4265 if (start_pfn
< end_pfn
)
4266 free_bootmem_node(NODE_DATA(this_nid
),
4267 PFN_PHYS(start_pfn
),
4268 (end_pfn
- start_pfn
) << PAGE_SHIFT
);
4273 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4274 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4276 * If an architecture guarantees that all ranges registered with
4277 * add_active_ranges() contain no holes and may be freed, this
4278 * function may be used instead of calling memory_present() manually.
4280 void __init
sparse_memory_present_with_active_regions(int nid
)
4282 unsigned long start_pfn
, end_pfn
;
4285 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4286 memory_present(this_nid
, start_pfn
, end_pfn
);
4290 * get_pfn_range_for_nid - Return the start and end page frames for a node
4291 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4292 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4293 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4295 * It returns the start and end page frame of a node based on information
4296 * provided by an arch calling add_active_range(). If called for a node
4297 * with no available memory, a warning is printed and the start and end
4300 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4301 unsigned long *start_pfn
, unsigned long *end_pfn
)
4303 unsigned long this_start_pfn
, this_end_pfn
;
4309 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4310 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4311 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4314 if (*start_pfn
== -1UL)
4319 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4320 * assumption is made that zones within a node are ordered in monotonic
4321 * increasing memory addresses so that the "highest" populated zone is used
4323 static void __init
find_usable_zone_for_movable(void)
4326 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4327 if (zone_index
== ZONE_MOVABLE
)
4330 if (arch_zone_highest_possible_pfn
[zone_index
] >
4331 arch_zone_lowest_possible_pfn
[zone_index
])
4335 VM_BUG_ON(zone_index
== -1);
4336 movable_zone
= zone_index
;
4340 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4341 * because it is sized independent of architecture. Unlike the other zones,
4342 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4343 * in each node depending on the size of each node and how evenly kernelcore
4344 * is distributed. This helper function adjusts the zone ranges
4345 * provided by the architecture for a given node by using the end of the
4346 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4347 * zones within a node are in order of monotonic increases memory addresses
4349 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4350 unsigned long zone_type
,
4351 unsigned long node_start_pfn
,
4352 unsigned long node_end_pfn
,
4353 unsigned long *zone_start_pfn
,
4354 unsigned long *zone_end_pfn
)
4356 /* Only adjust if ZONE_MOVABLE is on this node */
4357 if (zone_movable_pfn
[nid
]) {
4358 /* Size ZONE_MOVABLE */
4359 if (zone_type
== ZONE_MOVABLE
) {
4360 *zone_start_pfn
= zone_movable_pfn
[nid
];
4361 *zone_end_pfn
= min(node_end_pfn
,
4362 arch_zone_highest_possible_pfn
[movable_zone
]);
4364 /* Adjust for ZONE_MOVABLE starting within this range */
4365 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4366 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4367 *zone_end_pfn
= zone_movable_pfn
[nid
];
4369 /* Check if this whole range is within ZONE_MOVABLE */
4370 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4371 *zone_start_pfn
= *zone_end_pfn
;
4376 * Return the number of pages a zone spans in a node, including holes
4377 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4379 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4380 unsigned long zone_type
,
4381 unsigned long *ignored
)
4383 unsigned long node_start_pfn
, node_end_pfn
;
4384 unsigned long zone_start_pfn
, zone_end_pfn
;
4386 /* Get the start and end of the node and zone */
4387 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4388 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4389 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4390 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4391 node_start_pfn
, node_end_pfn
,
4392 &zone_start_pfn
, &zone_end_pfn
);
4394 /* Check that this node has pages within the zone's required range */
4395 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4398 /* Move the zone boundaries inside the node if necessary */
4399 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4400 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4402 /* Return the spanned pages */
4403 return zone_end_pfn
- zone_start_pfn
;
4407 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4408 * then all holes in the requested range will be accounted for.
4410 unsigned long __meminit
__absent_pages_in_range(int nid
,
4411 unsigned long range_start_pfn
,
4412 unsigned long range_end_pfn
)
4414 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4415 unsigned long start_pfn
, end_pfn
;
4418 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4419 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4420 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4421 nr_absent
-= end_pfn
- start_pfn
;
4427 * absent_pages_in_range - Return number of page frames in holes within a range
4428 * @start_pfn: The start PFN to start searching for holes
4429 * @end_pfn: The end PFN to stop searching for holes
4431 * It returns the number of pages frames in memory holes within a range.
4433 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4434 unsigned long end_pfn
)
4436 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4439 /* Return the number of page frames in holes in a zone on a node */
4440 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4441 unsigned long zone_type
,
4442 unsigned long *ignored
)
4444 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4445 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4446 unsigned long node_start_pfn
, node_end_pfn
;
4447 unsigned long zone_start_pfn
, zone_end_pfn
;
4449 get_pfn_range_for_nid(nid
, &node_start_pfn
, &node_end_pfn
);
4450 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4451 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4453 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4454 node_start_pfn
, node_end_pfn
,
4455 &zone_start_pfn
, &zone_end_pfn
);
4456 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4459 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4460 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4461 unsigned long zone_type
,
4462 unsigned long *zones_size
)
4464 return zones_size
[zone_type
];
4467 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4468 unsigned long zone_type
,
4469 unsigned long *zholes_size
)
4474 return zholes_size
[zone_type
];
4477 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4479 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4480 unsigned long *zones_size
, unsigned long *zholes_size
)
4482 unsigned long realtotalpages
, totalpages
= 0;
4485 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4486 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4488 pgdat
->node_spanned_pages
= totalpages
;
4490 realtotalpages
= totalpages
;
4491 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4493 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4495 pgdat
->node_present_pages
= realtotalpages
;
4496 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4500 #ifndef CONFIG_SPARSEMEM
4502 * Calculate the size of the zone->blockflags rounded to an unsigned long
4503 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4504 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4505 * round what is now in bits to nearest long in bits, then return it in
4508 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4510 unsigned long usemapsize
;
4512 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4513 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4514 usemapsize
= usemapsize
>> pageblock_order
;
4515 usemapsize
*= NR_PAGEBLOCK_BITS
;
4516 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4518 return usemapsize
/ 8;
4521 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4523 unsigned long zone_start_pfn
,
4524 unsigned long zonesize
)
4526 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4527 zone
->pageblock_flags
= NULL
;
4529 zone
->pageblock_flags
= alloc_bootmem_node_nopanic(pgdat
,
4533 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4534 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4535 #endif /* CONFIG_SPARSEMEM */
4537 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4539 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4540 void __init
set_pageblock_order(void)
4544 /* Check that pageblock_nr_pages has not already been setup */
4545 if (pageblock_order
)
4548 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4549 order
= HUGETLB_PAGE_ORDER
;
4551 order
= MAX_ORDER
- 1;
4554 * Assume the largest contiguous order of interest is a huge page.
4555 * This value may be variable depending on boot parameters on IA64 and
4558 pageblock_order
= order
;
4560 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4563 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4564 * is unused as pageblock_order is set at compile-time. See
4565 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4568 void __init
set_pageblock_order(void)
4572 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4574 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4575 unsigned long present_pages
)
4577 unsigned long pages
= spanned_pages
;
4580 * Provide a more accurate estimation if there are holes within
4581 * the zone and SPARSEMEM is in use. If there are holes within the
4582 * zone, each populated memory region may cost us one or two extra
4583 * memmap pages due to alignment because memmap pages for each
4584 * populated regions may not naturally algined on page boundary.
4585 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4587 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4588 IS_ENABLED(CONFIG_SPARSEMEM
))
4589 pages
= present_pages
;
4591 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4595 * Set up the zone data structures:
4596 * - mark all pages reserved
4597 * - mark all memory queues empty
4598 * - clear the memory bitmaps
4600 * NOTE: pgdat should get zeroed by caller.
4602 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4603 unsigned long *zones_size
, unsigned long *zholes_size
)
4606 int nid
= pgdat
->node_id
;
4607 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4610 pgdat_resize_init(pgdat
);
4611 #ifdef CONFIG_NUMA_BALANCING
4612 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4613 pgdat
->numabalancing_migrate_nr_pages
= 0;
4614 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4616 init_waitqueue_head(&pgdat
->kswapd_wait
);
4617 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4618 pgdat_page_cgroup_init(pgdat
);
4620 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4621 struct zone
*zone
= pgdat
->node_zones
+ j
;
4622 unsigned long size
, realsize
, freesize
, memmap_pages
;
4624 size
= zone_spanned_pages_in_node(nid
, j
, zones_size
);
4625 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4629 * Adjust freesize so that it accounts for how much memory
4630 * is used by this zone for memmap. This affects the watermark
4631 * and per-cpu initialisations
4633 memmap_pages
= calc_memmap_size(size
, realsize
);
4634 if (freesize
>= memmap_pages
) {
4635 freesize
-= memmap_pages
;
4638 " %s zone: %lu pages used for memmap\n",
4639 zone_names
[j
], memmap_pages
);
4642 " %s zone: %lu pages exceeds freesize %lu\n",
4643 zone_names
[j
], memmap_pages
, freesize
);
4645 /* Account for reserved pages */
4646 if (j
== 0 && freesize
> dma_reserve
) {
4647 freesize
-= dma_reserve
;
4648 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4649 zone_names
[0], dma_reserve
);
4652 if (!is_highmem_idx(j
))
4653 nr_kernel_pages
+= freesize
;
4654 /* Charge for highmem memmap if there are enough kernel pages */
4655 else if (nr_kernel_pages
> memmap_pages
* 2)
4656 nr_kernel_pages
-= memmap_pages
;
4657 nr_all_pages
+= freesize
;
4659 zone
->spanned_pages
= size
;
4660 zone
->present_pages
= realsize
;
4662 * Set an approximate value for lowmem here, it will be adjusted
4663 * when the bootmem allocator frees pages into the buddy system.
4664 * And all highmem pages will be managed by the buddy system.
4666 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4669 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4671 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4673 zone
->name
= zone_names
[j
];
4674 spin_lock_init(&zone
->lock
);
4675 spin_lock_init(&zone
->lru_lock
);
4676 zone_seqlock_init(zone
);
4677 zone
->zone_pgdat
= pgdat
;
4679 zone_pcp_init(zone
);
4680 lruvec_init(&zone
->lruvec
);
4684 set_pageblock_order();
4685 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4686 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4687 size
, MEMMAP_EARLY
);
4689 memmap_init(size
, nid
, j
, zone_start_pfn
);
4690 zone_start_pfn
+= size
;
4694 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4696 /* Skip empty nodes */
4697 if (!pgdat
->node_spanned_pages
)
4700 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4701 /* ia64 gets its own node_mem_map, before this, without bootmem */
4702 if (!pgdat
->node_mem_map
) {
4703 unsigned long size
, start
, end
;
4707 * The zone's endpoints aren't required to be MAX_ORDER
4708 * aligned but the node_mem_map endpoints must be in order
4709 * for the buddy allocator to function correctly.
4711 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4712 end
= pgdat_end_pfn(pgdat
);
4713 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4714 size
= (end
- start
) * sizeof(struct page
);
4715 map
= alloc_remap(pgdat
->node_id
, size
);
4717 map
= alloc_bootmem_node_nopanic(pgdat
, size
);
4718 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4720 #ifndef CONFIG_NEED_MULTIPLE_NODES
4722 * With no DISCONTIG, the global mem_map is just set as node 0's
4724 if (pgdat
== NODE_DATA(0)) {
4725 mem_map
= NODE_DATA(0)->node_mem_map
;
4726 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4727 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4728 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4729 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4732 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4735 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4736 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4738 pg_data_t
*pgdat
= NODE_DATA(nid
);
4740 /* pg_data_t should be reset to zero when it's allocated */
4741 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4743 pgdat
->node_id
= nid
;
4744 pgdat
->node_start_pfn
= node_start_pfn
;
4745 init_zone_allows_reclaim(nid
);
4746 calculate_node_totalpages(pgdat
, zones_size
, zholes_size
);
4748 alloc_node_mem_map(pgdat
);
4749 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4750 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4751 nid
, (unsigned long)pgdat
,
4752 (unsigned long)pgdat
->node_mem_map
);
4755 free_area_init_core(pgdat
, zones_size
, zholes_size
);
4758 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4760 #if MAX_NUMNODES > 1
4762 * Figure out the number of possible node ids.
4764 void __init
setup_nr_node_ids(void)
4767 unsigned int highest
= 0;
4769 for_each_node_mask(node
, node_possible_map
)
4771 nr_node_ids
= highest
+ 1;
4776 * node_map_pfn_alignment - determine the maximum internode alignment
4778 * This function should be called after node map is populated and sorted.
4779 * It calculates the maximum power of two alignment which can distinguish
4782 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4783 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4784 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4785 * shifted, 1GiB is enough and this function will indicate so.
4787 * This is used to test whether pfn -> nid mapping of the chosen memory
4788 * model has fine enough granularity to avoid incorrect mapping for the
4789 * populated node map.
4791 * Returns the determined alignment in pfn's. 0 if there is no alignment
4792 * requirement (single node).
4794 unsigned long __init
node_map_pfn_alignment(void)
4796 unsigned long accl_mask
= 0, last_end
= 0;
4797 unsigned long start
, end
, mask
;
4801 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
4802 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
4809 * Start with a mask granular enough to pin-point to the
4810 * start pfn and tick off bits one-by-one until it becomes
4811 * too coarse to separate the current node from the last.
4813 mask
= ~((1 << __ffs(start
)) - 1);
4814 while (mask
&& last_end
<= (start
& (mask
<< 1)))
4817 /* accumulate all internode masks */
4821 /* convert mask to number of pages */
4822 return ~accl_mask
+ 1;
4825 /* Find the lowest pfn for a node */
4826 static unsigned long __init
find_min_pfn_for_node(int nid
)
4828 unsigned long min_pfn
= ULONG_MAX
;
4829 unsigned long start_pfn
;
4832 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
4833 min_pfn
= min(min_pfn
, start_pfn
);
4835 if (min_pfn
== ULONG_MAX
) {
4837 "Could not find start_pfn for node %d\n", nid
);
4845 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4847 * It returns the minimum PFN based on information provided via
4848 * add_active_range().
4850 unsigned long __init
find_min_pfn_with_active_regions(void)
4852 return find_min_pfn_for_node(MAX_NUMNODES
);
4856 * early_calculate_totalpages()
4857 * Sum pages in active regions for movable zone.
4858 * Populate N_MEMORY for calculating usable_nodes.
4860 static unsigned long __init
early_calculate_totalpages(void)
4862 unsigned long totalpages
= 0;
4863 unsigned long start_pfn
, end_pfn
;
4866 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
4867 unsigned long pages
= end_pfn
- start_pfn
;
4869 totalpages
+= pages
;
4871 node_set_state(nid
, N_MEMORY
);
4877 * Find the PFN the Movable zone begins in each node. Kernel memory
4878 * is spread evenly between nodes as long as the nodes have enough
4879 * memory. When they don't, some nodes will have more kernelcore than
4882 static void __init
find_zone_movable_pfns_for_nodes(void)
4885 unsigned long usable_startpfn
;
4886 unsigned long kernelcore_node
, kernelcore_remaining
;
4887 /* save the state before borrow the nodemask */
4888 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
4889 unsigned long totalpages
= early_calculate_totalpages();
4890 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
4893 * If movablecore was specified, calculate what size of
4894 * kernelcore that corresponds so that memory usable for
4895 * any allocation type is evenly spread. If both kernelcore
4896 * and movablecore are specified, then the value of kernelcore
4897 * will be used for required_kernelcore if it's greater than
4898 * what movablecore would have allowed.
4900 if (required_movablecore
) {
4901 unsigned long corepages
;
4904 * Round-up so that ZONE_MOVABLE is at least as large as what
4905 * was requested by the user
4907 required_movablecore
=
4908 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
4909 corepages
= totalpages
- required_movablecore
;
4911 required_kernelcore
= max(required_kernelcore
, corepages
);
4914 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4915 if (!required_kernelcore
)
4918 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4919 find_usable_zone_for_movable();
4920 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
4923 /* Spread kernelcore memory as evenly as possible throughout nodes */
4924 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4925 for_each_node_state(nid
, N_MEMORY
) {
4926 unsigned long start_pfn
, end_pfn
;
4929 * Recalculate kernelcore_node if the division per node
4930 * now exceeds what is necessary to satisfy the requested
4931 * amount of memory for the kernel
4933 if (required_kernelcore
< kernelcore_node
)
4934 kernelcore_node
= required_kernelcore
/ usable_nodes
;
4937 * As the map is walked, we track how much memory is usable
4938 * by the kernel using kernelcore_remaining. When it is
4939 * 0, the rest of the node is usable by ZONE_MOVABLE
4941 kernelcore_remaining
= kernelcore_node
;
4943 /* Go through each range of PFNs within this node */
4944 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4945 unsigned long size_pages
;
4947 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
4948 if (start_pfn
>= end_pfn
)
4951 /* Account for what is only usable for kernelcore */
4952 if (start_pfn
< usable_startpfn
) {
4953 unsigned long kernel_pages
;
4954 kernel_pages
= min(end_pfn
, usable_startpfn
)
4957 kernelcore_remaining
-= min(kernel_pages
,
4958 kernelcore_remaining
);
4959 required_kernelcore
-= min(kernel_pages
,
4960 required_kernelcore
);
4962 /* Continue if range is now fully accounted */
4963 if (end_pfn
<= usable_startpfn
) {
4966 * Push zone_movable_pfn to the end so
4967 * that if we have to rebalance
4968 * kernelcore across nodes, we will
4969 * not double account here
4971 zone_movable_pfn
[nid
] = end_pfn
;
4974 start_pfn
= usable_startpfn
;
4978 * The usable PFN range for ZONE_MOVABLE is from
4979 * start_pfn->end_pfn. Calculate size_pages as the
4980 * number of pages used as kernelcore
4982 size_pages
= end_pfn
- start_pfn
;
4983 if (size_pages
> kernelcore_remaining
)
4984 size_pages
= kernelcore_remaining
;
4985 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
4988 * Some kernelcore has been met, update counts and
4989 * break if the kernelcore for this node has been
4992 required_kernelcore
-= min(required_kernelcore
,
4994 kernelcore_remaining
-= size_pages
;
4995 if (!kernelcore_remaining
)
5001 * If there is still required_kernelcore, we do another pass with one
5002 * less node in the count. This will push zone_movable_pfn[nid] further
5003 * along on the nodes that still have memory until kernelcore is
5007 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5010 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5011 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5012 zone_movable_pfn
[nid
] =
5013 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5016 /* restore the node_state */
5017 node_states
[N_MEMORY
] = saved_node_state
;
5020 /* Any regular or high memory on that node ? */
5021 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5023 enum zone_type zone_type
;
5025 if (N_MEMORY
== N_NORMAL_MEMORY
)
5028 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5029 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5030 if (zone
->present_pages
) {
5031 node_set_state(nid
, N_HIGH_MEMORY
);
5032 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5033 zone_type
<= ZONE_NORMAL
)
5034 node_set_state(nid
, N_NORMAL_MEMORY
);
5041 * free_area_init_nodes - Initialise all pg_data_t and zone data
5042 * @max_zone_pfn: an array of max PFNs for each zone
5044 * This will call free_area_init_node() for each active node in the system.
5045 * Using the page ranges provided by add_active_range(), the size of each
5046 * zone in each node and their holes is calculated. If the maximum PFN
5047 * between two adjacent zones match, it is assumed that the zone is empty.
5048 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5049 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5050 * starts where the previous one ended. For example, ZONE_DMA32 starts
5051 * at arch_max_dma_pfn.
5053 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5055 unsigned long start_pfn
, end_pfn
;
5058 /* Record where the zone boundaries are */
5059 memset(arch_zone_lowest_possible_pfn
, 0,
5060 sizeof(arch_zone_lowest_possible_pfn
));
5061 memset(arch_zone_highest_possible_pfn
, 0,
5062 sizeof(arch_zone_highest_possible_pfn
));
5063 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5064 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5065 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5066 if (i
== ZONE_MOVABLE
)
5068 arch_zone_lowest_possible_pfn
[i
] =
5069 arch_zone_highest_possible_pfn
[i
-1];
5070 arch_zone_highest_possible_pfn
[i
] =
5071 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5073 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5074 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5076 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5077 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5078 find_zone_movable_pfns_for_nodes();
5080 /* Print out the zone ranges */
5081 printk("Zone ranges:\n");
5082 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5083 if (i
== ZONE_MOVABLE
)
5085 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5086 if (arch_zone_lowest_possible_pfn
[i
] ==
5087 arch_zone_highest_possible_pfn
[i
])
5088 printk(KERN_CONT
"empty\n");
5090 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5091 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5092 (arch_zone_highest_possible_pfn
[i
]
5093 << PAGE_SHIFT
) - 1);
5096 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5097 printk("Movable zone start for each node\n");
5098 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5099 if (zone_movable_pfn
[i
])
5100 printk(" Node %d: %#010lx\n", i
,
5101 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5104 /* Print out the early node map */
5105 printk("Early memory node ranges\n");
5106 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5107 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5108 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5110 /* Initialise every node */
5111 mminit_verify_pageflags_layout();
5112 setup_nr_node_ids();
5113 for_each_online_node(nid
) {
5114 pg_data_t
*pgdat
= NODE_DATA(nid
);
5115 free_area_init_node(nid
, NULL
,
5116 find_min_pfn_for_node(nid
), NULL
);
5118 /* Any memory on that node */
5119 if (pgdat
->node_present_pages
)
5120 node_set_state(nid
, N_MEMORY
);
5121 check_for_memory(pgdat
, nid
);
5125 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5127 unsigned long long coremem
;
5131 coremem
= memparse(p
, &p
);
5132 *core
= coremem
>> PAGE_SHIFT
;
5134 /* Paranoid check that UL is enough for the coremem value */
5135 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5141 * kernelcore=size sets the amount of memory for use for allocations that
5142 * cannot be reclaimed or migrated.
5144 static int __init
cmdline_parse_kernelcore(char *p
)
5146 return cmdline_parse_core(p
, &required_kernelcore
);
5150 * movablecore=size sets the amount of memory for use for allocations that
5151 * can be reclaimed or migrated.
5153 static int __init
cmdline_parse_movablecore(char *p
)
5155 return cmdline_parse_core(p
, &required_movablecore
);
5158 early_param("kernelcore", cmdline_parse_kernelcore
);
5159 early_param("movablecore", cmdline_parse_movablecore
);
5161 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5163 unsigned long free_reserved_area(unsigned long start
, unsigned long end
,
5164 int poison
, char *s
)
5166 unsigned long pages
, pos
;
5168 pos
= start
= PAGE_ALIGN(start
);
5170 for (pages
= 0; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5172 memset((void *)pos
, poison
, PAGE_SIZE
);
5173 free_reserved_page(virt_to_page((void *)pos
));
5177 pr_info("Freeing %s memory: %ldK (%lx - %lx)\n",
5178 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5183 #ifdef CONFIG_HIGHMEM
5184 void free_highmem_page(struct page
*page
)
5186 __free_reserved_page(page
);
5193 * set_dma_reserve - set the specified number of pages reserved in the first zone
5194 * @new_dma_reserve: The number of pages to mark reserved
5196 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5197 * In the DMA zone, a significant percentage may be consumed by kernel image
5198 * and other unfreeable allocations which can skew the watermarks badly. This
5199 * function may optionally be used to account for unfreeable pages in the
5200 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5201 * smaller per-cpu batchsize.
5203 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5205 dma_reserve
= new_dma_reserve
;
5208 void __init
free_area_init(unsigned long *zones_size
)
5210 free_area_init_node(0, zones_size
,
5211 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5214 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5215 unsigned long action
, void *hcpu
)
5217 int cpu
= (unsigned long)hcpu
;
5219 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5220 lru_add_drain_cpu(cpu
);
5224 * Spill the event counters of the dead processor
5225 * into the current processors event counters.
5226 * This artificially elevates the count of the current
5229 vm_events_fold_cpu(cpu
);
5232 * Zero the differential counters of the dead processor
5233 * so that the vm statistics are consistent.
5235 * This is only okay since the processor is dead and cannot
5236 * race with what we are doing.
5238 refresh_cpu_vm_stats(cpu
);
5243 void __init
page_alloc_init(void)
5245 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5249 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5250 * or min_free_kbytes changes.
5252 static void calculate_totalreserve_pages(void)
5254 struct pglist_data
*pgdat
;
5255 unsigned long reserve_pages
= 0;
5256 enum zone_type i
, j
;
5258 for_each_online_pgdat(pgdat
) {
5259 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5260 struct zone
*zone
= pgdat
->node_zones
+ i
;
5261 unsigned long max
= 0;
5263 /* Find valid and maximum lowmem_reserve in the zone */
5264 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5265 if (zone
->lowmem_reserve
[j
] > max
)
5266 max
= zone
->lowmem_reserve
[j
];
5269 /* we treat the high watermark as reserved pages. */
5270 max
+= high_wmark_pages(zone
);
5272 if (max
> zone
->managed_pages
)
5273 max
= zone
->managed_pages
;
5274 reserve_pages
+= max
;
5276 * Lowmem reserves are not available to
5277 * GFP_HIGHUSER page cache allocations and
5278 * kswapd tries to balance zones to their high
5279 * watermark. As a result, neither should be
5280 * regarded as dirtyable memory, to prevent a
5281 * situation where reclaim has to clean pages
5282 * in order to balance the zones.
5284 zone
->dirty_balance_reserve
= max
;
5287 dirty_balance_reserve
= reserve_pages
;
5288 totalreserve_pages
= reserve_pages
;
5292 * setup_per_zone_lowmem_reserve - called whenever
5293 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5294 * has a correct pages reserved value, so an adequate number of
5295 * pages are left in the zone after a successful __alloc_pages().
5297 static void setup_per_zone_lowmem_reserve(void)
5299 struct pglist_data
*pgdat
;
5300 enum zone_type j
, idx
;
5302 for_each_online_pgdat(pgdat
) {
5303 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5304 struct zone
*zone
= pgdat
->node_zones
+ j
;
5305 unsigned long managed_pages
= zone
->managed_pages
;
5307 zone
->lowmem_reserve
[j
] = 0;
5311 struct zone
*lower_zone
;
5315 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5316 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5318 lower_zone
= pgdat
->node_zones
+ idx
;
5319 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5320 sysctl_lowmem_reserve_ratio
[idx
];
5321 managed_pages
+= lower_zone
->managed_pages
;
5326 /* update totalreserve_pages */
5327 calculate_totalreserve_pages();
5330 static void __setup_per_zone_wmarks(void)
5332 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5333 unsigned long lowmem_pages
= 0;
5335 unsigned long flags
;
5337 /* Calculate total number of !ZONE_HIGHMEM pages */
5338 for_each_zone(zone
) {
5339 if (!is_highmem(zone
))
5340 lowmem_pages
+= zone
->managed_pages
;
5343 for_each_zone(zone
) {
5346 spin_lock_irqsave(&zone
->lock
, flags
);
5347 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5348 do_div(tmp
, lowmem_pages
);
5349 if (is_highmem(zone
)) {
5351 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5352 * need highmem pages, so cap pages_min to a small
5355 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5356 * deltas controls asynch page reclaim, and so should
5357 * not be capped for highmem.
5359 unsigned long min_pages
;
5361 min_pages
= zone
->managed_pages
/ 1024;
5362 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5363 zone
->watermark
[WMARK_MIN
] = min_pages
;
5366 * If it's a lowmem zone, reserve a number of pages
5367 * proportionate to the zone's size.
5369 zone
->watermark
[WMARK_MIN
] = tmp
;
5372 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5373 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5375 setup_zone_migrate_reserve(zone
);
5376 spin_unlock_irqrestore(&zone
->lock
, flags
);
5379 /* update totalreserve_pages */
5380 calculate_totalreserve_pages();
5384 * setup_per_zone_wmarks - called when min_free_kbytes changes
5385 * or when memory is hot-{added|removed}
5387 * Ensures that the watermark[min,low,high] values for each zone are set
5388 * correctly with respect to min_free_kbytes.
5390 void setup_per_zone_wmarks(void)
5392 mutex_lock(&zonelists_mutex
);
5393 __setup_per_zone_wmarks();
5394 mutex_unlock(&zonelists_mutex
);
5398 * The inactive anon list should be small enough that the VM never has to
5399 * do too much work, but large enough that each inactive page has a chance
5400 * to be referenced again before it is swapped out.
5402 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5403 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5404 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5405 * the anonymous pages are kept on the inactive list.
5408 * memory ratio inactive anon
5409 * -------------------------------------
5418 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5420 unsigned int gb
, ratio
;
5422 /* Zone size in gigabytes */
5423 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5425 ratio
= int_sqrt(10 * gb
);
5429 zone
->inactive_ratio
= ratio
;
5432 static void __meminit
setup_per_zone_inactive_ratio(void)
5437 calculate_zone_inactive_ratio(zone
);
5441 * Initialise min_free_kbytes.
5443 * For small machines we want it small (128k min). For large machines
5444 * we want it large (64MB max). But it is not linear, because network
5445 * bandwidth does not increase linearly with machine size. We use
5447 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5448 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5464 int __meminit
init_per_zone_wmark_min(void)
5466 unsigned long lowmem_kbytes
;
5468 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5470 min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5471 if (min_free_kbytes
< 128)
5472 min_free_kbytes
= 128;
5473 if (min_free_kbytes
> 65536)
5474 min_free_kbytes
= 65536;
5475 setup_per_zone_wmarks();
5476 refresh_zone_stat_thresholds();
5477 setup_per_zone_lowmem_reserve();
5478 setup_per_zone_inactive_ratio();
5481 module_init(init_per_zone_wmark_min
)
5484 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5485 * that we can call two helper functions whenever min_free_kbytes
5488 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5489 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5491 proc_dointvec(table
, write
, buffer
, length
, ppos
);
5493 setup_per_zone_wmarks();
5498 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5499 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5504 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5509 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5510 sysctl_min_unmapped_ratio
) / 100;
5514 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5515 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5520 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5525 zone
->min_slab_pages
= (zone
->managed_pages
*
5526 sysctl_min_slab_ratio
) / 100;
5532 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5533 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5534 * whenever sysctl_lowmem_reserve_ratio changes.
5536 * The reserve ratio obviously has absolutely no relation with the
5537 * minimum watermarks. The lowmem reserve ratio can only make sense
5538 * if in function of the boot time zone sizes.
5540 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5541 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5543 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5544 setup_per_zone_lowmem_reserve();
5549 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5550 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5551 * can have before it gets flushed back to buddy allocator.
5554 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5555 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5561 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5562 if (!write
|| (ret
< 0))
5564 for_each_populated_zone(zone
) {
5565 for_each_possible_cpu(cpu
) {
5567 high
= zone
->managed_pages
/ percpu_pagelist_fraction
;
5568 setup_pagelist_highmark(
5569 per_cpu_ptr(zone
->pageset
, cpu
), high
);
5575 int hashdist
= HASHDIST_DEFAULT
;
5578 static int __init
set_hashdist(char *str
)
5582 hashdist
= simple_strtoul(str
, &str
, 0);
5585 __setup("hashdist=", set_hashdist
);
5589 * allocate a large system hash table from bootmem
5590 * - it is assumed that the hash table must contain an exact power-of-2
5591 * quantity of entries
5592 * - limit is the number of hash buckets, not the total allocation size
5594 void *__init
alloc_large_system_hash(const char *tablename
,
5595 unsigned long bucketsize
,
5596 unsigned long numentries
,
5599 unsigned int *_hash_shift
,
5600 unsigned int *_hash_mask
,
5601 unsigned long low_limit
,
5602 unsigned long high_limit
)
5604 unsigned long long max
= high_limit
;
5605 unsigned long log2qty
, size
;
5608 /* allow the kernel cmdline to have a say */
5610 /* round applicable memory size up to nearest megabyte */
5611 numentries
= nr_kernel_pages
;
5612 numentries
+= (1UL << (20 - PAGE_SHIFT
)) - 1;
5613 numentries
>>= 20 - PAGE_SHIFT
;
5614 numentries
<<= 20 - PAGE_SHIFT
;
5616 /* limit to 1 bucket per 2^scale bytes of low memory */
5617 if (scale
> PAGE_SHIFT
)
5618 numentries
>>= (scale
- PAGE_SHIFT
);
5620 numentries
<<= (PAGE_SHIFT
- scale
);
5622 /* Make sure we've got at least a 0-order allocation.. */
5623 if (unlikely(flags
& HASH_SMALL
)) {
5624 /* Makes no sense without HASH_EARLY */
5625 WARN_ON(!(flags
& HASH_EARLY
));
5626 if (!(numentries
>> *_hash_shift
)) {
5627 numentries
= 1UL << *_hash_shift
;
5628 BUG_ON(!numentries
);
5630 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5631 numentries
= PAGE_SIZE
/ bucketsize
;
5633 numentries
= roundup_pow_of_two(numentries
);
5635 /* limit allocation size to 1/16 total memory by default */
5637 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5638 do_div(max
, bucketsize
);
5640 max
= min(max
, 0x80000000ULL
);
5642 if (numentries
< low_limit
)
5643 numentries
= low_limit
;
5644 if (numentries
> max
)
5647 log2qty
= ilog2(numentries
);
5650 size
= bucketsize
<< log2qty
;
5651 if (flags
& HASH_EARLY
)
5652 table
= alloc_bootmem_nopanic(size
);
5654 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5657 * If bucketsize is not a power-of-two, we may free
5658 * some pages at the end of hash table which
5659 * alloc_pages_exact() automatically does
5661 if (get_order(size
) < MAX_ORDER
) {
5662 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5663 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5666 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5669 panic("Failed to allocate %s hash table\n", tablename
);
5671 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5674 ilog2(size
) - PAGE_SHIFT
,
5678 *_hash_shift
= log2qty
;
5680 *_hash_mask
= (1 << log2qty
) - 1;
5685 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5686 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5689 #ifdef CONFIG_SPARSEMEM
5690 return __pfn_to_section(pfn
)->pageblock_flags
;
5692 return zone
->pageblock_flags
;
5693 #endif /* CONFIG_SPARSEMEM */
5696 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5698 #ifdef CONFIG_SPARSEMEM
5699 pfn
&= (PAGES_PER_SECTION
-1);
5700 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5702 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
5703 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
5704 #endif /* CONFIG_SPARSEMEM */
5708 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5709 * @page: The page within the block of interest
5710 * @start_bitidx: The first bit of interest to retrieve
5711 * @end_bitidx: The last bit of interest
5712 * returns pageblock_bits flags
5714 unsigned long get_pageblock_flags_group(struct page
*page
,
5715 int start_bitidx
, int end_bitidx
)
5718 unsigned long *bitmap
;
5719 unsigned long pfn
, bitidx
;
5720 unsigned long flags
= 0;
5721 unsigned long value
= 1;
5723 zone
= page_zone(page
);
5724 pfn
= page_to_pfn(page
);
5725 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5726 bitidx
= pfn_to_bitidx(zone
, pfn
);
5728 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5729 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
5736 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5737 * @page: The page within the block of interest
5738 * @start_bitidx: The first bit of interest
5739 * @end_bitidx: The last bit of interest
5740 * @flags: The flags to set
5742 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
5743 int start_bitidx
, int end_bitidx
)
5746 unsigned long *bitmap
;
5747 unsigned long pfn
, bitidx
;
5748 unsigned long value
= 1;
5750 zone
= page_zone(page
);
5751 pfn
= page_to_pfn(page
);
5752 bitmap
= get_pageblock_bitmap(zone
, pfn
);
5753 bitidx
= pfn_to_bitidx(zone
, pfn
);
5754 VM_BUG_ON(!zone_spans_pfn(zone
, pfn
));
5756 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
5758 __set_bit(bitidx
+ start_bitidx
, bitmap
);
5760 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
5764 * This function checks whether pageblock includes unmovable pages or not.
5765 * If @count is not zero, it is okay to include less @count unmovable pages
5767 * PageLRU check wihtout isolation or lru_lock could race so that
5768 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
5769 * expect this function should be exact.
5771 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
5772 bool skip_hwpoisoned_pages
)
5774 unsigned long pfn
, iter
, found
;
5778 * For avoiding noise data, lru_add_drain_all() should be called
5779 * If ZONE_MOVABLE, the zone never contains unmovable pages
5781 if (zone_idx(zone
) == ZONE_MOVABLE
)
5783 mt
= get_pageblock_migratetype(page
);
5784 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
5787 pfn
= page_to_pfn(page
);
5788 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
5789 unsigned long check
= pfn
+ iter
;
5791 if (!pfn_valid_within(check
))
5794 page
= pfn_to_page(check
);
5796 * We can't use page_count without pin a page
5797 * because another CPU can free compound page.
5798 * This check already skips compound tails of THP
5799 * because their page->_count is zero at all time.
5801 if (!atomic_read(&page
->_count
)) {
5802 if (PageBuddy(page
))
5803 iter
+= (1 << page_order(page
)) - 1;
5808 * The HWPoisoned page may be not in buddy system, and
5809 * page_count() is not 0.
5811 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
5817 * If there are RECLAIMABLE pages, we need to check it.
5818 * But now, memory offline itself doesn't call shrink_slab()
5819 * and it still to be fixed.
5822 * If the page is not RAM, page_count()should be 0.
5823 * we don't need more check. This is an _used_ not-movable page.
5825 * The problematic thing here is PG_reserved pages. PG_reserved
5826 * is set to both of a memory hole page and a _used_ kernel
5835 bool is_pageblock_removable_nolock(struct page
*page
)
5841 * We have to be careful here because we are iterating over memory
5842 * sections which are not zone aware so we might end up outside of
5843 * the zone but still within the section.
5844 * We have to take care about the node as well. If the node is offline
5845 * its NODE_DATA will be NULL - see page_zone.
5847 if (!node_online(page_to_nid(page
)))
5850 zone
= page_zone(page
);
5851 pfn
= page_to_pfn(page
);
5852 if (!zone_spans_pfn(zone
, pfn
))
5855 return !has_unmovable_pages(zone
, page
, 0, true);
5860 static unsigned long pfn_max_align_down(unsigned long pfn
)
5862 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
5863 pageblock_nr_pages
) - 1);
5866 static unsigned long pfn_max_align_up(unsigned long pfn
)
5868 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
5869 pageblock_nr_pages
));
5872 /* [start, end) must belong to a single zone. */
5873 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
5874 unsigned long start
, unsigned long end
)
5876 /* This function is based on compact_zone() from compaction.c. */
5877 unsigned long nr_reclaimed
;
5878 unsigned long pfn
= start
;
5879 unsigned int tries
= 0;
5884 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
5885 if (fatal_signal_pending(current
)) {
5890 if (list_empty(&cc
->migratepages
)) {
5891 cc
->nr_migratepages
= 0;
5892 pfn
= isolate_migratepages_range(cc
->zone
, cc
,
5899 } else if (++tries
== 5) {
5900 ret
= ret
< 0 ? ret
: -EBUSY
;
5904 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
5906 cc
->nr_migratepages
-= nr_reclaimed
;
5908 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
5909 0, MIGRATE_SYNC
, MR_CMA
);
5912 putback_movable_pages(&cc
->migratepages
);
5919 * alloc_contig_range() -- tries to allocate given range of pages
5920 * @start: start PFN to allocate
5921 * @end: one-past-the-last PFN to allocate
5922 * @migratetype: migratetype of the underlaying pageblocks (either
5923 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
5924 * in range must have the same migratetype and it must
5925 * be either of the two.
5927 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
5928 * aligned, however it's the caller's responsibility to guarantee that
5929 * we are the only thread that changes migrate type of pageblocks the
5932 * The PFN range must belong to a single zone.
5934 * Returns zero on success or negative error code. On success all
5935 * pages which PFN is in [start, end) are allocated for the caller and
5936 * need to be freed with free_contig_range().
5938 int alloc_contig_range(unsigned long start
, unsigned long end
,
5939 unsigned migratetype
)
5941 unsigned long outer_start
, outer_end
;
5944 struct compact_control cc
= {
5945 .nr_migratepages
= 0,
5947 .zone
= page_zone(pfn_to_page(start
)),
5949 .ignore_skip_hint
= true,
5951 INIT_LIST_HEAD(&cc
.migratepages
);
5954 * What we do here is we mark all pageblocks in range as
5955 * MIGRATE_ISOLATE. Because pageblock and max order pages may
5956 * have different sizes, and due to the way page allocator
5957 * work, we align the range to biggest of the two pages so
5958 * that page allocator won't try to merge buddies from
5959 * different pageblocks and change MIGRATE_ISOLATE to some
5960 * other migration type.
5962 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
5963 * migrate the pages from an unaligned range (ie. pages that
5964 * we are interested in). This will put all the pages in
5965 * range back to page allocator as MIGRATE_ISOLATE.
5967 * When this is done, we take the pages in range from page
5968 * allocator removing them from the buddy system. This way
5969 * page allocator will never consider using them.
5971 * This lets us mark the pageblocks back as
5972 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
5973 * aligned range but not in the unaligned, original range are
5974 * put back to page allocator so that buddy can use them.
5977 ret
= start_isolate_page_range(pfn_max_align_down(start
),
5978 pfn_max_align_up(end
), migratetype
,
5983 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
5988 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
5989 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
5990 * more, all pages in [start, end) are free in page allocator.
5991 * What we are going to do is to allocate all pages from
5992 * [start, end) (that is remove them from page allocator).
5994 * The only problem is that pages at the beginning and at the
5995 * end of interesting range may be not aligned with pages that
5996 * page allocator holds, ie. they can be part of higher order
5997 * pages. Because of this, we reserve the bigger range and
5998 * once this is done free the pages we are not interested in.
6000 * We don't have to hold zone->lock here because the pages are
6001 * isolated thus they won't get removed from buddy.
6004 lru_add_drain_all();
6008 outer_start
= start
;
6009 while (!PageBuddy(pfn_to_page(outer_start
))) {
6010 if (++order
>= MAX_ORDER
) {
6014 outer_start
&= ~0UL << order
;
6017 /* Make sure the range is really isolated. */
6018 if (test_pages_isolated(outer_start
, end
, false)) {
6019 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6026 /* Grab isolated pages from freelists. */
6027 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6033 /* Free head and tail (if any) */
6034 if (start
!= outer_start
)
6035 free_contig_range(outer_start
, start
- outer_start
);
6036 if (end
!= outer_end
)
6037 free_contig_range(end
, outer_end
- end
);
6040 undo_isolate_page_range(pfn_max_align_down(start
),
6041 pfn_max_align_up(end
), migratetype
);
6045 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6047 unsigned int count
= 0;
6049 for (; nr_pages
--; pfn
++) {
6050 struct page
*page
= pfn_to_page(pfn
);
6052 count
+= page_count(page
) != 1;
6055 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6059 #ifdef CONFIG_MEMORY_HOTPLUG
6060 static int __meminit
__zone_pcp_update(void *data
)
6062 struct zone
*zone
= data
;
6064 unsigned long batch
= zone_batchsize(zone
), flags
;
6066 for_each_possible_cpu(cpu
) {
6067 struct per_cpu_pageset
*pset
;
6068 struct per_cpu_pages
*pcp
;
6070 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6073 local_irq_save(flags
);
6075 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
6076 drain_zonestat(zone
, pset
);
6077 setup_pageset(pset
, batch
);
6078 local_irq_restore(flags
);
6083 void __meminit
zone_pcp_update(struct zone
*zone
)
6085 stop_machine(__zone_pcp_update
, zone
, NULL
);
6089 void zone_pcp_reset(struct zone
*zone
)
6091 unsigned long flags
;
6093 struct per_cpu_pageset
*pset
;
6095 /* avoid races with drain_pages() */
6096 local_irq_save(flags
);
6097 if (zone
->pageset
!= &boot_pageset
) {
6098 for_each_online_cpu(cpu
) {
6099 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6100 drain_zonestat(zone
, pset
);
6102 free_percpu(zone
->pageset
);
6103 zone
->pageset
= &boot_pageset
;
6105 local_irq_restore(flags
);
6108 #ifdef CONFIG_MEMORY_HOTREMOVE
6110 * All pages in the range must be isolated before calling this.
6113 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6119 unsigned long flags
;
6120 /* find the first valid pfn */
6121 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6126 zone
= page_zone(pfn_to_page(pfn
));
6127 spin_lock_irqsave(&zone
->lock
, flags
);
6129 while (pfn
< end_pfn
) {
6130 if (!pfn_valid(pfn
)) {
6134 page
= pfn_to_page(pfn
);
6136 * The HWPoisoned page may be not in buddy system, and
6137 * page_count() is not 0.
6139 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6141 SetPageReserved(page
);
6145 BUG_ON(page_count(page
));
6146 BUG_ON(!PageBuddy(page
));
6147 order
= page_order(page
);
6148 #ifdef CONFIG_DEBUG_VM
6149 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6150 pfn
, 1 << order
, end_pfn
);
6152 list_del(&page
->lru
);
6153 rmv_page_order(page
);
6154 zone
->free_area
[order
].nr_free
--;
6155 #ifdef CONFIG_HIGHMEM
6156 if (PageHighMem(page
))
6157 totalhigh_pages
-= 1 << order
;
6159 for (i
= 0; i
< (1 << order
); i
++)
6160 SetPageReserved((page
+i
));
6161 pfn
+= (1 << order
);
6163 spin_unlock_irqrestore(&zone
->lock
, flags
);
6167 #ifdef CONFIG_MEMORY_FAILURE
6168 bool is_free_buddy_page(struct page
*page
)
6170 struct zone
*zone
= page_zone(page
);
6171 unsigned long pfn
= page_to_pfn(page
);
6172 unsigned long flags
;
6175 spin_lock_irqsave(&zone
->lock
, flags
);
6176 for (order
= 0; order
< MAX_ORDER
; order
++) {
6177 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6179 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6182 spin_unlock_irqrestore(&zone
->lock
, flags
);
6184 return order
< MAX_ORDER
;
6188 static const struct trace_print_flags pageflag_names
[] = {
6189 {1UL << PG_locked
, "locked" },
6190 {1UL << PG_error
, "error" },
6191 {1UL << PG_referenced
, "referenced" },
6192 {1UL << PG_uptodate
, "uptodate" },
6193 {1UL << PG_dirty
, "dirty" },
6194 {1UL << PG_lru
, "lru" },
6195 {1UL << PG_active
, "active" },
6196 {1UL << PG_slab
, "slab" },
6197 {1UL << PG_owner_priv_1
, "owner_priv_1" },
6198 {1UL << PG_arch_1
, "arch_1" },
6199 {1UL << PG_reserved
, "reserved" },
6200 {1UL << PG_private
, "private" },
6201 {1UL << PG_private_2
, "private_2" },
6202 {1UL << PG_writeback
, "writeback" },
6203 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6204 {1UL << PG_head
, "head" },
6205 {1UL << PG_tail
, "tail" },
6207 {1UL << PG_compound
, "compound" },
6209 {1UL << PG_swapcache
, "swapcache" },
6210 {1UL << PG_mappedtodisk
, "mappedtodisk" },
6211 {1UL << PG_reclaim
, "reclaim" },
6212 {1UL << PG_swapbacked
, "swapbacked" },
6213 {1UL << PG_unevictable
, "unevictable" },
6215 {1UL << PG_mlocked
, "mlocked" },
6217 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6218 {1UL << PG_uncached
, "uncached" },
6220 #ifdef CONFIG_MEMORY_FAILURE
6221 {1UL << PG_hwpoison
, "hwpoison" },
6223 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6224 {1UL << PG_compound_lock
, "compound_lock" },
6228 static void dump_page_flags(unsigned long flags
)
6230 const char *delim
= "";
6234 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names
) != __NR_PAGEFLAGS
);
6236 printk(KERN_ALERT
"page flags: %#lx(", flags
);
6238 /* remove zone id */
6239 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
6241 for (i
= 0; i
< ARRAY_SIZE(pageflag_names
) && flags
; i
++) {
6243 mask
= pageflag_names
[i
].mask
;
6244 if ((flags
& mask
) != mask
)
6248 printk("%s%s", delim
, pageflag_names
[i
].name
);
6252 /* check for left over flags */
6254 printk("%s%#lx", delim
, flags
);
6259 void dump_page(struct page
*page
)
6262 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6263 page
, atomic_read(&page
->_count
), page_mapcount(page
),
6264 page
->mapping
, page
->index
);
6265 dump_page_flags(page
->flags
);
6266 mem_cgroup_print_bad_page(page
);