2 * linux/mm/page_alloc.c
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
7 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
8 * Swap reorganised 29.12.95, Stephen Tweedie
9 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
10 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
11 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
12 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
13 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
14 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
17 #include <linux/stddef.h>
19 #include <linux/swap.h>
20 #include <linux/interrupt.h>
21 #include <linux/pagemap.h>
22 #include <linux/jiffies.h>
23 #include <linux/bootmem.h>
24 #include <linux/memblock.h>
25 #include <linux/compiler.h>
26 #include <linux/kernel.h>
27 #include <linux/kmemcheck.h>
28 #include <linux/module.h>
29 #include <linux/suspend.h>
30 #include <linux/pagevec.h>
31 #include <linux/blkdev.h>
32 #include <linux/slab.h>
33 #include <linux/ratelimit.h>
34 #include <linux/oom.h>
35 #include <linux/notifier.h>
36 #include <linux/topology.h>
37 #include <linux/sysctl.h>
38 #include <linux/cpu.h>
39 #include <linux/cpuset.h>
40 #include <linux/memory_hotplug.h>
41 #include <linux/nodemask.h>
42 #include <linux/vmalloc.h>
43 #include <linux/vmstat.h>
44 #include <linux/mempolicy.h>
45 #include <linux/stop_machine.h>
46 #include <linux/sort.h>
47 #include <linux/pfn.h>
48 #include <linux/backing-dev.h>
49 #include <linux/fault-inject.h>
50 #include <linux/page-isolation.h>
51 #include <linux/page_cgroup.h>
52 #include <linux/debugobjects.h>
53 #include <linux/kmemleak.h>
54 #include <linux/compaction.h>
55 #include <trace/events/kmem.h>
56 #include <linux/ftrace_event.h>
57 #include <linux/memcontrol.h>
58 #include <linux/prefetch.h>
59 #include <linux/mm_inline.h>
60 #include <linux/migrate.h>
61 #include <linux/page-debug-flags.h>
62 #include <linux/hugetlb.h>
63 #include <linux/sched/rt.h>
65 #include <asm/sections.h>
66 #include <asm/tlbflush.h>
67 #include <asm/div64.h>
70 /* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
71 static DEFINE_MUTEX(pcp_batch_high_lock
);
73 #ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
74 DEFINE_PER_CPU(int, numa_node
);
75 EXPORT_PER_CPU_SYMBOL(numa_node
);
78 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
80 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
81 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
82 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
83 * defined in <linux/topology.h>.
85 DEFINE_PER_CPU(int, _numa_mem_
); /* Kernel "local memory" node */
86 EXPORT_PER_CPU_SYMBOL(_numa_mem_
);
90 * Array of node states.
92 nodemask_t node_states
[NR_NODE_STATES
] __read_mostly
= {
93 [N_POSSIBLE
] = NODE_MASK_ALL
,
94 [N_ONLINE
] = { { [0] = 1UL } },
96 [N_NORMAL_MEMORY
] = { { [0] = 1UL } },
98 [N_HIGH_MEMORY
] = { { [0] = 1UL } },
100 #ifdef CONFIG_MOVABLE_NODE
101 [N_MEMORY
] = { { [0] = 1UL } },
103 [N_CPU
] = { { [0] = 1UL } },
106 EXPORT_SYMBOL(node_states
);
108 /* Protect totalram_pages and zone->managed_pages */
109 static DEFINE_SPINLOCK(managed_page_count_lock
);
111 unsigned long totalram_pages __read_mostly
;
112 unsigned long totalreserve_pages __read_mostly
;
114 * When calculating the number of globally allowed dirty pages, there
115 * is a certain number of per-zone reserves that should not be
116 * considered dirtyable memory. This is the sum of those reserves
117 * over all existing zones that contribute dirtyable memory.
119 unsigned long dirty_balance_reserve __read_mostly
;
121 int percpu_pagelist_fraction
;
122 gfp_t gfp_allowed_mask __read_mostly
= GFP_BOOT_MASK
;
124 #ifdef CONFIG_PM_SLEEP
126 * The following functions are used by the suspend/hibernate code to temporarily
127 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
128 * while devices are suspended. To avoid races with the suspend/hibernate code,
129 * they should always be called with pm_mutex held (gfp_allowed_mask also should
130 * only be modified with pm_mutex held, unless the suspend/hibernate code is
131 * guaranteed not to run in parallel with that modification).
134 static gfp_t saved_gfp_mask
;
136 void pm_restore_gfp_mask(void)
138 WARN_ON(!mutex_is_locked(&pm_mutex
));
139 if (saved_gfp_mask
) {
140 gfp_allowed_mask
= saved_gfp_mask
;
145 void pm_restrict_gfp_mask(void)
147 WARN_ON(!mutex_is_locked(&pm_mutex
));
148 WARN_ON(saved_gfp_mask
);
149 saved_gfp_mask
= gfp_allowed_mask
;
150 gfp_allowed_mask
&= ~GFP_IOFS
;
153 bool pm_suspended_storage(void)
155 if ((gfp_allowed_mask
& GFP_IOFS
) == GFP_IOFS
)
159 #endif /* CONFIG_PM_SLEEP */
161 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
162 int pageblock_order __read_mostly
;
165 static void __free_pages_ok(struct page
*page
, unsigned int order
);
168 * results with 256, 32 in the lowmem_reserve sysctl:
169 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
170 * 1G machine -> (16M dma, 784M normal, 224M high)
171 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
172 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
173 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
175 * TBD: should special case ZONE_DMA32 machines here - in those we normally
176 * don't need any ZONE_NORMAL reservation
178 int sysctl_lowmem_reserve_ratio
[MAX_NR_ZONES
-1] = {
179 #ifdef CONFIG_ZONE_DMA
182 #ifdef CONFIG_ZONE_DMA32
185 #ifdef CONFIG_HIGHMEM
191 EXPORT_SYMBOL(totalram_pages
);
193 static char * const zone_names
[MAX_NR_ZONES
] = {
194 #ifdef CONFIG_ZONE_DMA
197 #ifdef CONFIG_ZONE_DMA32
201 #ifdef CONFIG_HIGHMEM
207 int min_free_kbytes
= 1024;
208 int user_min_free_kbytes
= -1;
210 static unsigned long __meminitdata nr_kernel_pages
;
211 static unsigned long __meminitdata nr_all_pages
;
212 static unsigned long __meminitdata dma_reserve
;
214 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
215 static unsigned long __meminitdata arch_zone_lowest_possible_pfn
[MAX_NR_ZONES
];
216 static unsigned long __meminitdata arch_zone_highest_possible_pfn
[MAX_NR_ZONES
];
217 static unsigned long __initdata required_kernelcore
;
218 static unsigned long __initdata required_movablecore
;
219 static unsigned long __meminitdata zone_movable_pfn
[MAX_NUMNODES
];
221 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
223 EXPORT_SYMBOL(movable_zone
);
224 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
227 int nr_node_ids __read_mostly
= MAX_NUMNODES
;
228 int nr_online_nodes __read_mostly
= 1;
229 EXPORT_SYMBOL(nr_node_ids
);
230 EXPORT_SYMBOL(nr_online_nodes
);
233 int page_group_by_mobility_disabled __read_mostly
;
235 void set_pageblock_migratetype(struct page
*page
, int migratetype
)
237 if (unlikely(page_group_by_mobility_disabled
&&
238 migratetype
< MIGRATE_PCPTYPES
))
239 migratetype
= MIGRATE_UNMOVABLE
;
241 set_pageblock_flags_group(page
, (unsigned long)migratetype
,
242 PB_migrate
, PB_migrate_end
);
245 bool oom_killer_disabled __read_mostly
;
247 #ifdef CONFIG_DEBUG_VM
248 static int page_outside_zone_boundaries(struct zone
*zone
, struct page
*page
)
252 unsigned long pfn
= page_to_pfn(page
);
253 unsigned long sp
, start_pfn
;
256 seq
= zone_span_seqbegin(zone
);
257 start_pfn
= zone
->zone_start_pfn
;
258 sp
= zone
->spanned_pages
;
259 if (!zone_spans_pfn(zone
, pfn
))
261 } while (zone_span_seqretry(zone
, seq
));
264 pr_err("page %lu outside zone [ %lu - %lu ]\n",
265 pfn
, start_pfn
, start_pfn
+ sp
);
270 static int page_is_consistent(struct zone
*zone
, struct page
*page
)
272 if (!pfn_valid_within(page_to_pfn(page
)))
274 if (zone
!= page_zone(page
))
280 * Temporary debugging check for pages not lying within a given zone.
282 static int bad_range(struct zone
*zone
, struct page
*page
)
284 if (page_outside_zone_boundaries(zone
, page
))
286 if (!page_is_consistent(zone
, page
))
292 static inline int bad_range(struct zone
*zone
, struct page
*page
)
298 static void bad_page(struct page
*page
, const char *reason
,
299 unsigned long bad_flags
)
301 static unsigned long resume
;
302 static unsigned long nr_shown
;
303 static unsigned long nr_unshown
;
305 /* Don't complain about poisoned pages */
306 if (PageHWPoison(page
)) {
307 page_mapcount_reset(page
); /* remove PageBuddy */
312 * Allow a burst of 60 reports, then keep quiet for that minute;
313 * or allow a steady drip of one report per second.
315 if (nr_shown
== 60) {
316 if (time_before(jiffies
, resume
)) {
322 "BUG: Bad page state: %lu messages suppressed\n",
329 resume
= jiffies
+ 60 * HZ
;
331 printk(KERN_ALERT
"BUG: Bad page state in process %s pfn:%05lx\n",
332 current
->comm
, page_to_pfn(page
));
333 dump_page_badflags(page
, reason
, bad_flags
);
338 /* Leave bad fields for debug, except PageBuddy could make trouble */
339 page_mapcount_reset(page
); /* remove PageBuddy */
340 add_taint(TAINT_BAD_PAGE
, LOCKDEP_NOW_UNRELIABLE
);
344 * Higher-order pages are called "compound pages". They are structured thusly:
346 * The first PAGE_SIZE page is called the "head page".
348 * The remaining PAGE_SIZE pages are called "tail pages".
350 * All pages have PG_compound set. All tail pages have their ->first_page
351 * pointing at the head page.
353 * The first tail page's ->lru.next holds the address of the compound page's
354 * put_page() function. Its ->lru.prev holds the order of allocation.
355 * This usage means that zero-order pages may not be compound.
358 static void free_compound_page(struct page
*page
)
360 __free_pages_ok(page
, compound_order(page
));
363 void prep_compound_page(struct page
*page
, unsigned long order
)
366 int nr_pages
= 1 << order
;
368 set_compound_page_dtor(page
, free_compound_page
);
369 set_compound_order(page
, order
);
371 for (i
= 1; i
< nr_pages
; i
++) {
372 struct page
*p
= page
+ i
;
373 set_page_count(p
, 0);
374 p
->first_page
= page
;
375 /* Make sure p->first_page is always valid for PageTail() */
381 /* update __split_huge_page_refcount if you change this function */
382 static int destroy_compound_page(struct page
*page
, unsigned long order
)
385 int nr_pages
= 1 << order
;
388 if (unlikely(compound_order(page
) != order
)) {
389 bad_page(page
, "wrong compound order", 0);
393 __ClearPageHead(page
);
395 for (i
= 1; i
< nr_pages
; i
++) {
396 struct page
*p
= page
+ i
;
398 if (unlikely(!PageTail(p
))) {
399 bad_page(page
, "PageTail not set", 0);
401 } else if (unlikely(p
->first_page
!= page
)) {
402 bad_page(page
, "first_page not consistent", 0);
411 static inline void prep_zero_page(struct page
*page
, int order
, gfp_t gfp_flags
)
416 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
417 * and __GFP_HIGHMEM from hard or soft interrupt context.
419 VM_BUG_ON((gfp_flags
& __GFP_HIGHMEM
) && in_interrupt());
420 for (i
= 0; i
< (1 << order
); i
++)
421 clear_highpage(page
+ i
);
424 #ifdef CONFIG_DEBUG_PAGEALLOC
425 unsigned int _debug_guardpage_minorder
;
427 static int __init
debug_guardpage_minorder_setup(char *buf
)
431 if (kstrtoul(buf
, 10, &res
) < 0 || res
> MAX_ORDER
/ 2) {
432 printk(KERN_ERR
"Bad debug_guardpage_minorder value\n");
435 _debug_guardpage_minorder
= res
;
436 printk(KERN_INFO
"Setting debug_guardpage_minorder to %lu\n", res
);
439 __setup("debug_guardpage_minorder=", debug_guardpage_minorder_setup
);
441 static inline void set_page_guard_flag(struct page
*page
)
443 __set_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
446 static inline void clear_page_guard_flag(struct page
*page
)
448 __clear_bit(PAGE_DEBUG_FLAG_GUARD
, &page
->debug_flags
);
451 static inline void set_page_guard_flag(struct page
*page
) { }
452 static inline void clear_page_guard_flag(struct page
*page
) { }
455 static inline void set_page_order(struct page
*page
, int order
)
457 set_page_private(page
, order
);
458 __SetPageBuddy(page
);
461 static inline void rmv_page_order(struct page
*page
)
463 __ClearPageBuddy(page
);
464 set_page_private(page
, 0);
468 * Locate the struct page for both the matching buddy in our
469 * pair (buddy1) and the combined O(n+1) page they form (page).
471 * 1) Any buddy B1 will have an order O twin B2 which satisfies
472 * the following equation:
474 * For example, if the starting buddy (buddy2) is #8 its order
476 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
478 * 2) Any buddy B will have an order O+1 parent P which
479 * satisfies the following equation:
482 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
484 static inline unsigned long
485 __find_buddy_index(unsigned long page_idx
, unsigned int order
)
487 return page_idx
^ (1 << order
);
491 * This function checks whether a page is free && is the buddy
492 * we can do coalesce a page and its buddy if
493 * (a) the buddy is not in a hole &&
494 * (b) the buddy is in the buddy system &&
495 * (c) a page and its buddy have the same order &&
496 * (d) a page and its buddy are in the same zone.
498 * For recording whether a page is in the buddy system, we set ->_mapcount
499 * PAGE_BUDDY_MAPCOUNT_VALUE.
500 * Setting, clearing, and testing _mapcount PAGE_BUDDY_MAPCOUNT_VALUE is
501 * serialized by zone->lock.
503 * For recording page's order, we use page_private(page).
505 static inline int page_is_buddy(struct page
*page
, struct page
*buddy
,
508 if (!pfn_valid_within(page_to_pfn(buddy
)))
511 if (page_zone_id(page
) != page_zone_id(buddy
))
514 if (page_is_guard(buddy
) && page_order(buddy
) == order
) {
515 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
519 if (PageBuddy(buddy
) && page_order(buddy
) == order
) {
520 VM_BUG_ON_PAGE(page_count(buddy
) != 0, buddy
);
527 * Freeing function for a buddy system allocator.
529 * The concept of a buddy system is to maintain direct-mapped table
530 * (containing bit values) for memory blocks of various "orders".
531 * The bottom level table contains the map for the smallest allocatable
532 * units of memory (here, pages), and each level above it describes
533 * pairs of units from the levels below, hence, "buddies".
534 * At a high level, all that happens here is marking the table entry
535 * at the bottom level available, and propagating the changes upward
536 * as necessary, plus some accounting needed to play nicely with other
537 * parts of the VM system.
538 * At each level, we keep a list of pages, which are heads of continuous
539 * free pages of length of (1 << order) and marked with _mapcount
540 * PAGE_BUDDY_MAPCOUNT_VALUE. Page's order is recorded in page_private(page)
542 * So when we are allocating or freeing one, we can derive the state of the
543 * other. That is, if we allocate a small block, and both were
544 * free, the remainder of the region must be split into blocks.
545 * If a block is freed, and its buddy is also free, then this
546 * triggers coalescing into a block of larger size.
551 static inline void __free_one_page(struct page
*page
,
552 struct zone
*zone
, unsigned int order
,
555 unsigned long page_idx
;
556 unsigned long combined_idx
;
557 unsigned long uninitialized_var(buddy_idx
);
560 VM_BUG_ON(!zone_is_initialized(zone
));
562 if (unlikely(PageCompound(page
)))
563 if (unlikely(destroy_compound_page(page
, order
)))
566 VM_BUG_ON(migratetype
== -1);
568 page_idx
= page_to_pfn(page
) & ((1 << MAX_ORDER
) - 1);
570 VM_BUG_ON_PAGE(page_idx
& ((1 << order
) - 1), page
);
571 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
573 while (order
< MAX_ORDER
-1) {
574 buddy_idx
= __find_buddy_index(page_idx
, order
);
575 buddy
= page
+ (buddy_idx
- page_idx
);
576 if (!page_is_buddy(page
, buddy
, order
))
579 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
580 * merge with it and move up one order.
582 if (page_is_guard(buddy
)) {
583 clear_page_guard_flag(buddy
);
584 set_page_private(page
, 0);
585 __mod_zone_freepage_state(zone
, 1 << order
,
588 list_del(&buddy
->lru
);
589 zone
->free_area
[order
].nr_free
--;
590 rmv_page_order(buddy
);
592 combined_idx
= buddy_idx
& page_idx
;
593 page
= page
+ (combined_idx
- page_idx
);
594 page_idx
= combined_idx
;
597 set_page_order(page
, order
);
600 * If this is not the largest possible page, check if the buddy
601 * of the next-highest order is free. If it is, it's possible
602 * that pages are being freed that will coalesce soon. In case,
603 * that is happening, add the free page to the tail of the list
604 * so it's less likely to be used soon and more likely to be merged
605 * as a higher order page
607 if ((order
< MAX_ORDER
-2) && pfn_valid_within(page_to_pfn(buddy
))) {
608 struct page
*higher_page
, *higher_buddy
;
609 combined_idx
= buddy_idx
& page_idx
;
610 higher_page
= page
+ (combined_idx
- page_idx
);
611 buddy_idx
= __find_buddy_index(combined_idx
, order
+ 1);
612 higher_buddy
= higher_page
+ (buddy_idx
- combined_idx
);
613 if (page_is_buddy(higher_page
, higher_buddy
, order
+ 1)) {
614 list_add_tail(&page
->lru
,
615 &zone
->free_area
[order
].free_list
[migratetype
]);
620 list_add(&page
->lru
, &zone
->free_area
[order
].free_list
[migratetype
]);
622 zone
->free_area
[order
].nr_free
++;
625 static inline int free_pages_check(struct page
*page
)
627 const char *bad_reason
= NULL
;
628 unsigned long bad_flags
= 0;
630 if (unlikely(page_mapcount(page
)))
631 bad_reason
= "nonzero mapcount";
632 if (unlikely(page
->mapping
!= NULL
))
633 bad_reason
= "non-NULL mapping";
634 if (unlikely(atomic_read(&page
->_count
) != 0))
635 bad_reason
= "nonzero _count";
636 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_FREE
)) {
637 bad_reason
= "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
638 bad_flags
= PAGE_FLAGS_CHECK_AT_FREE
;
640 if (unlikely(mem_cgroup_bad_page_check(page
)))
641 bad_reason
= "cgroup check failed";
642 if (unlikely(bad_reason
)) {
643 bad_page(page
, bad_reason
, bad_flags
);
646 page_cpupid_reset_last(page
);
647 if (page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)
648 page
->flags
&= ~PAGE_FLAGS_CHECK_AT_PREP
;
653 * Frees a number of pages from the PCP lists
654 * Assumes all pages on list are in same zone, and of same order.
655 * count is the number of pages to free.
657 * If the zone was previously in an "all pages pinned" state then look to
658 * see if this freeing clears that state.
660 * And clear the zone's pages_scanned counter, to hold off the "all pages are
661 * pinned" detection logic.
663 static void free_pcppages_bulk(struct zone
*zone
, int count
,
664 struct per_cpu_pages
*pcp
)
670 spin_lock(&zone
->lock
);
671 zone
->pages_scanned
= 0;
675 struct list_head
*list
;
678 * Remove pages from lists in a round-robin fashion. A
679 * batch_free count is maintained that is incremented when an
680 * empty list is encountered. This is so more pages are freed
681 * off fuller lists instead of spinning excessively around empty
686 if (++migratetype
== MIGRATE_PCPTYPES
)
688 list
= &pcp
->lists
[migratetype
];
689 } while (list_empty(list
));
691 /* This is the only non-empty list. Free them all. */
692 if (batch_free
== MIGRATE_PCPTYPES
)
693 batch_free
= to_free
;
696 int mt
; /* migratetype of the to-be-freed page */
698 page
= list_entry(list
->prev
, struct page
, lru
);
699 /* must delete as __free_one_page list manipulates */
700 list_del(&page
->lru
);
701 mt
= get_freepage_migratetype(page
);
702 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
703 __free_one_page(page
, zone
, 0, mt
);
704 trace_mm_page_pcpu_drain(page
, 0, mt
);
705 if (likely(!is_migrate_isolate_page(page
))) {
706 __mod_zone_page_state(zone
, NR_FREE_PAGES
, 1);
707 if (is_migrate_cma(mt
))
708 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
, 1);
710 } while (--to_free
&& --batch_free
&& !list_empty(list
));
712 spin_unlock(&zone
->lock
);
715 static void free_one_page(struct zone
*zone
, struct page
*page
, int order
,
718 spin_lock(&zone
->lock
);
719 zone
->pages_scanned
= 0;
721 __free_one_page(page
, zone
, order
, migratetype
);
722 if (unlikely(!is_migrate_isolate(migratetype
)))
723 __mod_zone_freepage_state(zone
, 1 << order
, migratetype
);
724 spin_unlock(&zone
->lock
);
727 static bool free_pages_prepare(struct page
*page
, unsigned int order
)
732 trace_mm_page_free(page
, order
);
733 kmemcheck_free_shadow(page
, order
);
736 page
->mapping
= NULL
;
737 for (i
= 0; i
< (1 << order
); i
++)
738 bad
+= free_pages_check(page
+ i
);
742 if (!PageHighMem(page
)) {
743 debug_check_no_locks_freed(page_address(page
),
745 debug_check_no_obj_freed(page_address(page
),
748 arch_free_page(page
, order
);
749 kernel_map_pages(page
, 1 << order
, 0);
754 static void __free_pages_ok(struct page
*page
, unsigned int order
)
759 if (!free_pages_prepare(page
, order
))
762 local_irq_save(flags
);
763 __count_vm_events(PGFREE
, 1 << order
);
764 migratetype
= get_pageblock_migratetype(page
);
765 set_freepage_migratetype(page
, migratetype
);
766 free_one_page(page_zone(page
), page
, order
, migratetype
);
767 local_irq_restore(flags
);
770 void __init
__free_pages_bootmem(struct page
*page
, unsigned int order
)
772 unsigned int nr_pages
= 1 << order
;
773 struct page
*p
= page
;
777 for (loop
= 0; loop
< (nr_pages
- 1); loop
++, p
++) {
779 __ClearPageReserved(p
);
780 set_page_count(p
, 0);
782 __ClearPageReserved(p
);
783 set_page_count(p
, 0);
785 page_zone(page
)->managed_pages
+= nr_pages
;
786 set_page_refcounted(page
);
787 __free_pages(page
, order
);
791 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
792 void __init
init_cma_reserved_pageblock(struct page
*page
)
794 unsigned i
= pageblock_nr_pages
;
795 struct page
*p
= page
;
798 __ClearPageReserved(p
);
799 set_page_count(p
, 0);
802 set_page_refcounted(page
);
803 set_pageblock_migratetype(page
, MIGRATE_CMA
);
804 __free_pages(page
, pageblock_order
);
805 adjust_managed_page_count(page
, pageblock_nr_pages
);
810 * The order of subdivision here is critical for the IO subsystem.
811 * Please do not alter this order without good reasons and regression
812 * testing. Specifically, as large blocks of memory are subdivided,
813 * the order in which smaller blocks are delivered depends on the order
814 * they're subdivided in this function. This is the primary factor
815 * influencing the order in which pages are delivered to the IO
816 * subsystem according to empirical testing, and this is also justified
817 * by considering the behavior of a buddy system containing a single
818 * large block of memory acted on by a series of small allocations.
819 * This behavior is a critical factor in sglist merging's success.
823 static inline void expand(struct zone
*zone
, struct page
*page
,
824 int low
, int high
, struct free_area
*area
,
827 unsigned long size
= 1 << high
;
833 VM_BUG_ON_PAGE(bad_range(zone
, &page
[size
]), &page
[size
]);
835 #ifdef CONFIG_DEBUG_PAGEALLOC
836 if (high
< debug_guardpage_minorder()) {
838 * Mark as guard pages (or page), that will allow to
839 * merge back to allocator when buddy will be freed.
840 * Corresponding page table entries will not be touched,
841 * pages will stay not present in virtual address space
843 INIT_LIST_HEAD(&page
[size
].lru
);
844 set_page_guard_flag(&page
[size
]);
845 set_page_private(&page
[size
], high
);
846 /* Guard pages are not available for any usage */
847 __mod_zone_freepage_state(zone
, -(1 << high
),
852 list_add(&page
[size
].lru
, &area
->free_list
[migratetype
]);
854 set_page_order(&page
[size
], high
);
859 * This page is about to be returned from the page allocator
861 static inline int check_new_page(struct page
*page
)
863 const char *bad_reason
= NULL
;
864 unsigned long bad_flags
= 0;
866 if (unlikely(page_mapcount(page
)))
867 bad_reason
= "nonzero mapcount";
868 if (unlikely(page
->mapping
!= NULL
))
869 bad_reason
= "non-NULL mapping";
870 if (unlikely(atomic_read(&page
->_count
) != 0))
871 bad_reason
= "nonzero _count";
872 if (unlikely(page
->flags
& PAGE_FLAGS_CHECK_AT_PREP
)) {
873 bad_reason
= "PAGE_FLAGS_CHECK_AT_PREP flag set";
874 bad_flags
= PAGE_FLAGS_CHECK_AT_PREP
;
876 if (unlikely(mem_cgroup_bad_page_check(page
)))
877 bad_reason
= "cgroup check failed";
878 if (unlikely(bad_reason
)) {
879 bad_page(page
, bad_reason
, bad_flags
);
885 static int prep_new_page(struct page
*page
, int order
, gfp_t gfp_flags
)
889 for (i
= 0; i
< (1 << order
); i
++) {
890 struct page
*p
= page
+ i
;
891 if (unlikely(check_new_page(p
)))
895 set_page_private(page
, 0);
896 set_page_refcounted(page
);
898 arch_alloc_page(page
, order
);
899 kernel_map_pages(page
, 1 << order
, 1);
901 if (gfp_flags
& __GFP_ZERO
)
902 prep_zero_page(page
, order
, gfp_flags
);
904 if (order
&& (gfp_flags
& __GFP_COMP
))
905 prep_compound_page(page
, order
);
911 * Go through the free lists for the given migratetype and remove
912 * the smallest available page from the freelists
915 struct page
*__rmqueue_smallest(struct zone
*zone
, unsigned int order
,
918 unsigned int current_order
;
919 struct free_area
*area
;
922 /* Find a page of the appropriate size in the preferred list */
923 for (current_order
= order
; current_order
< MAX_ORDER
; ++current_order
) {
924 area
= &(zone
->free_area
[current_order
]);
925 if (list_empty(&area
->free_list
[migratetype
]))
928 page
= list_entry(area
->free_list
[migratetype
].next
,
930 list_del(&page
->lru
);
931 rmv_page_order(page
);
933 expand(zone
, page
, order
, current_order
, area
, migratetype
);
934 set_freepage_migratetype(page
, migratetype
);
943 * This array describes the order lists are fallen back to when
944 * the free lists for the desirable migrate type are depleted
946 static int fallbacks
[MIGRATE_TYPES
][4] = {
947 [MIGRATE_UNMOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
948 [MIGRATE_RECLAIMABLE
] = { MIGRATE_UNMOVABLE
, MIGRATE_MOVABLE
, MIGRATE_RESERVE
},
950 [MIGRATE_MOVABLE
] = { MIGRATE_CMA
, MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
951 [MIGRATE_CMA
] = { MIGRATE_RESERVE
}, /* Never used */
953 [MIGRATE_MOVABLE
] = { MIGRATE_RECLAIMABLE
, MIGRATE_UNMOVABLE
, MIGRATE_RESERVE
},
955 [MIGRATE_RESERVE
] = { MIGRATE_RESERVE
}, /* Never used */
956 #ifdef CONFIG_MEMORY_ISOLATION
957 [MIGRATE_ISOLATE
] = { MIGRATE_RESERVE
}, /* Never used */
962 * Move the free pages in a range to the free lists of the requested type.
963 * Note that start_page and end_pages are not aligned on a pageblock
964 * boundary. If alignment is required, use move_freepages_block()
966 int move_freepages(struct zone
*zone
,
967 struct page
*start_page
, struct page
*end_page
,
974 #ifndef CONFIG_HOLES_IN_ZONE
976 * page_zone is not safe to call in this context when
977 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
978 * anyway as we check zone boundaries in move_freepages_block().
979 * Remove at a later date when no bug reports exist related to
980 * grouping pages by mobility
982 BUG_ON(page_zone(start_page
) != page_zone(end_page
));
985 for (page
= start_page
; page
<= end_page
;) {
986 /* Make sure we are not inadvertently changing nodes */
987 VM_BUG_ON_PAGE(page_to_nid(page
) != zone_to_nid(zone
), page
);
989 if (!pfn_valid_within(page_to_pfn(page
))) {
994 if (!PageBuddy(page
)) {
999 order
= page_order(page
);
1000 list_move(&page
->lru
,
1001 &zone
->free_area
[order
].free_list
[migratetype
]);
1002 set_freepage_migratetype(page
, migratetype
);
1004 pages_moved
+= 1 << order
;
1010 int move_freepages_block(struct zone
*zone
, struct page
*page
,
1013 unsigned long start_pfn
, end_pfn
;
1014 struct page
*start_page
, *end_page
;
1016 start_pfn
= page_to_pfn(page
);
1017 start_pfn
= start_pfn
& ~(pageblock_nr_pages
-1);
1018 start_page
= pfn_to_page(start_pfn
);
1019 end_page
= start_page
+ pageblock_nr_pages
- 1;
1020 end_pfn
= start_pfn
+ pageblock_nr_pages
- 1;
1022 /* Do not cross zone boundaries */
1023 if (!zone_spans_pfn(zone
, start_pfn
))
1025 if (!zone_spans_pfn(zone
, end_pfn
))
1028 return move_freepages(zone
, start_page
, end_page
, migratetype
);
1031 static void change_pageblock_range(struct page
*pageblock_page
,
1032 int start_order
, int migratetype
)
1034 int nr_pageblocks
= 1 << (start_order
- pageblock_order
);
1036 while (nr_pageblocks
--) {
1037 set_pageblock_migratetype(pageblock_page
, migratetype
);
1038 pageblock_page
+= pageblock_nr_pages
;
1043 * If breaking a large block of pages, move all free pages to the preferred
1044 * allocation list. If falling back for a reclaimable kernel allocation, be
1045 * more aggressive about taking ownership of free pages.
1047 * On the other hand, never change migration type of MIGRATE_CMA pageblocks
1048 * nor move CMA pages to different free lists. We don't want unmovable pages
1049 * to be allocated from MIGRATE_CMA areas.
1051 * Returns the new migratetype of the pageblock (or the same old migratetype
1052 * if it was unchanged).
1054 static int try_to_steal_freepages(struct zone
*zone
, struct page
*page
,
1055 int start_type
, int fallback_type
)
1057 int current_order
= page_order(page
);
1060 * When borrowing from MIGRATE_CMA, we need to release the excess
1061 * buddy pages to CMA itself. We also ensure the freepage_migratetype
1062 * is set to CMA so it is returned to the correct freelist in case
1063 * the page ends up being not actually allocated from the pcp lists.
1065 if (is_migrate_cma(fallback_type
))
1066 return fallback_type
;
1068 /* Take ownership for orders >= pageblock_order */
1069 if (current_order
>= pageblock_order
) {
1070 change_pageblock_range(page
, current_order
, start_type
);
1074 if (current_order
>= pageblock_order
/ 2 ||
1075 start_type
== MIGRATE_RECLAIMABLE
||
1076 page_group_by_mobility_disabled
) {
1079 pages
= move_freepages_block(zone
, page
, start_type
);
1081 /* Claim the whole block if over half of it is free */
1082 if (pages
>= (1 << (pageblock_order
-1)) ||
1083 page_group_by_mobility_disabled
) {
1085 set_pageblock_migratetype(page
, start_type
);
1091 return fallback_type
;
1094 /* Remove an element from the buddy allocator from the fallback list */
1095 static inline struct page
*
1096 __rmqueue_fallback(struct zone
*zone
, int order
, int start_migratetype
)
1098 struct free_area
*area
;
1101 int migratetype
, new_type
, i
;
1103 /* Find the largest possible block of pages in the other list */
1104 for (current_order
= MAX_ORDER
-1; current_order
>= order
;
1107 migratetype
= fallbacks
[start_migratetype
][i
];
1109 /* MIGRATE_RESERVE handled later if necessary */
1110 if (migratetype
== MIGRATE_RESERVE
)
1113 area
= &(zone
->free_area
[current_order
]);
1114 if (list_empty(&area
->free_list
[migratetype
]))
1117 page
= list_entry(area
->free_list
[migratetype
].next
,
1121 new_type
= try_to_steal_freepages(zone
, page
,
1125 /* Remove the page from the freelists */
1126 list_del(&page
->lru
);
1127 rmv_page_order(page
);
1129 expand(zone
, page
, order
, current_order
, area
,
1131 /* The freepage_migratetype may differ from pageblock's
1132 * migratetype depending on the decisions in
1133 * try_to_steal_freepages. This is OK as long as it does
1134 * not differ for MIGRATE_CMA type.
1136 set_freepage_migratetype(page
, new_type
);
1138 trace_mm_page_alloc_extfrag(page
, order
, current_order
,
1139 start_migratetype
, migratetype
, new_type
);
1149 * Do the hard work of removing an element from the buddy allocator.
1150 * Call me with the zone->lock already held.
1152 static struct page
*__rmqueue(struct zone
*zone
, unsigned int order
,
1158 page
= __rmqueue_smallest(zone
, order
, migratetype
);
1160 if (unlikely(!page
) && migratetype
!= MIGRATE_RESERVE
) {
1161 page
= __rmqueue_fallback(zone
, order
, migratetype
);
1164 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1165 * is used because __rmqueue_smallest is an inline function
1166 * and we want just one call site
1169 migratetype
= MIGRATE_RESERVE
;
1174 trace_mm_page_alloc_zone_locked(page
, order
, migratetype
);
1179 * Obtain a specified number of elements from the buddy allocator, all under
1180 * a single hold of the lock, for efficiency. Add them to the supplied list.
1181 * Returns the number of new pages which were placed at *list.
1183 static int rmqueue_bulk(struct zone
*zone
, unsigned int order
,
1184 unsigned long count
, struct list_head
*list
,
1185 int migratetype
, int cold
)
1189 spin_lock(&zone
->lock
);
1190 for (i
= 0; i
< count
; ++i
) {
1191 struct page
*page
= __rmqueue(zone
, order
, migratetype
);
1192 if (unlikely(page
== NULL
))
1196 * Split buddy pages returned by expand() are received here
1197 * in physical page order. The page is added to the callers and
1198 * list and the list head then moves forward. From the callers
1199 * perspective, the linked list is ordered by page number in
1200 * some conditions. This is useful for IO devices that can
1201 * merge IO requests if the physical pages are ordered
1204 if (likely(cold
== 0))
1205 list_add(&page
->lru
, list
);
1207 list_add_tail(&page
->lru
, list
);
1209 if (is_migrate_cma(get_freepage_migratetype(page
)))
1210 __mod_zone_page_state(zone
, NR_FREE_CMA_PAGES
,
1213 __mod_zone_page_state(zone
, NR_FREE_PAGES
, -(i
<< order
));
1214 spin_unlock(&zone
->lock
);
1220 * Called from the vmstat counter updater to drain pagesets of this
1221 * currently executing processor on remote nodes after they have
1224 * Note that this function must be called with the thread pinned to
1225 * a single processor.
1227 void drain_zone_pages(struct zone
*zone
, struct per_cpu_pages
*pcp
)
1229 unsigned long flags
;
1231 unsigned long batch
;
1233 local_irq_save(flags
);
1234 batch
= ACCESS_ONCE(pcp
->batch
);
1235 if (pcp
->count
>= batch
)
1238 to_drain
= pcp
->count
;
1240 free_pcppages_bulk(zone
, to_drain
, pcp
);
1241 pcp
->count
-= to_drain
;
1243 local_irq_restore(flags
);
1248 * Drain pages of the indicated processor.
1250 * The processor must either be the current processor and the
1251 * thread pinned to the current processor or a processor that
1254 static void drain_pages(unsigned int cpu
)
1256 unsigned long flags
;
1259 for_each_populated_zone(zone
) {
1260 struct per_cpu_pageset
*pset
;
1261 struct per_cpu_pages
*pcp
;
1263 local_irq_save(flags
);
1264 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
1268 free_pcppages_bulk(zone
, pcp
->count
, pcp
);
1271 local_irq_restore(flags
);
1276 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1278 void drain_local_pages(void *arg
)
1280 drain_pages(smp_processor_id());
1284 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
1286 * Note that this code is protected against sending an IPI to an offline
1287 * CPU but does not guarantee sending an IPI to newly hotplugged CPUs:
1288 * on_each_cpu_mask() blocks hotplug and won't talk to offlined CPUs but
1289 * nothing keeps CPUs from showing up after we populated the cpumask and
1290 * before the call to on_each_cpu_mask().
1292 void drain_all_pages(void)
1295 struct per_cpu_pageset
*pcp
;
1299 * Allocate in the BSS so we wont require allocation in
1300 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
1302 static cpumask_t cpus_with_pcps
;
1305 * We don't care about racing with CPU hotplug event
1306 * as offline notification will cause the notified
1307 * cpu to drain that CPU pcps and on_each_cpu_mask
1308 * disables preemption as part of its processing
1310 for_each_online_cpu(cpu
) {
1311 bool has_pcps
= false;
1312 for_each_populated_zone(zone
) {
1313 pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
1314 if (pcp
->pcp
.count
) {
1320 cpumask_set_cpu(cpu
, &cpus_with_pcps
);
1322 cpumask_clear_cpu(cpu
, &cpus_with_pcps
);
1324 on_each_cpu_mask(&cpus_with_pcps
, drain_local_pages
, NULL
, 1);
1327 #ifdef CONFIG_HIBERNATION
1329 void mark_free_pages(struct zone
*zone
)
1331 unsigned long pfn
, max_zone_pfn
;
1332 unsigned long flags
;
1334 struct list_head
*curr
;
1336 if (zone_is_empty(zone
))
1339 spin_lock_irqsave(&zone
->lock
, flags
);
1341 max_zone_pfn
= zone_end_pfn(zone
);
1342 for (pfn
= zone
->zone_start_pfn
; pfn
< max_zone_pfn
; pfn
++)
1343 if (pfn_valid(pfn
)) {
1344 struct page
*page
= pfn_to_page(pfn
);
1346 if (!swsusp_page_is_forbidden(page
))
1347 swsusp_unset_page_free(page
);
1350 for_each_migratetype_order(order
, t
) {
1351 list_for_each(curr
, &zone
->free_area
[order
].free_list
[t
]) {
1354 pfn
= page_to_pfn(list_entry(curr
, struct page
, lru
));
1355 for (i
= 0; i
< (1UL << order
); i
++)
1356 swsusp_set_page_free(pfn_to_page(pfn
+ i
));
1359 spin_unlock_irqrestore(&zone
->lock
, flags
);
1361 #endif /* CONFIG_PM */
1364 * Free a 0-order page
1365 * cold == 1 ? free a cold page : free a hot page
1367 void free_hot_cold_page(struct page
*page
, int cold
)
1369 struct zone
*zone
= page_zone(page
);
1370 struct per_cpu_pages
*pcp
;
1371 unsigned long flags
;
1374 if (!free_pages_prepare(page
, 0))
1377 migratetype
= get_pageblock_migratetype(page
);
1378 set_freepage_migratetype(page
, migratetype
);
1379 local_irq_save(flags
);
1380 __count_vm_event(PGFREE
);
1383 * We only track unmovable, reclaimable and movable on pcp lists.
1384 * Free ISOLATE pages back to the allocator because they are being
1385 * offlined but treat RESERVE as movable pages so we can get those
1386 * areas back if necessary. Otherwise, we may have to free
1387 * excessively into the page allocator
1389 if (migratetype
>= MIGRATE_PCPTYPES
) {
1390 if (unlikely(is_migrate_isolate(migratetype
))) {
1391 free_one_page(zone
, page
, 0, migratetype
);
1394 migratetype
= MIGRATE_MOVABLE
;
1397 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1399 list_add_tail(&page
->lru
, &pcp
->lists
[migratetype
]);
1401 list_add(&page
->lru
, &pcp
->lists
[migratetype
]);
1403 if (pcp
->count
>= pcp
->high
) {
1404 unsigned long batch
= ACCESS_ONCE(pcp
->batch
);
1405 free_pcppages_bulk(zone
, batch
, pcp
);
1406 pcp
->count
-= batch
;
1410 local_irq_restore(flags
);
1414 * Free a list of 0-order pages
1416 void free_hot_cold_page_list(struct list_head
*list
, int cold
)
1418 struct page
*page
, *next
;
1420 list_for_each_entry_safe(page
, next
, list
, lru
) {
1421 trace_mm_page_free_batched(page
, cold
);
1422 free_hot_cold_page(page
, cold
);
1427 * split_page takes a non-compound higher-order page, and splits it into
1428 * n (1<<order) sub-pages: page[0..n]
1429 * Each sub-page must be freed individually.
1431 * Note: this is probably too low level an operation for use in drivers.
1432 * Please consult with lkml before using this in your driver.
1434 void split_page(struct page
*page
, unsigned int order
)
1438 VM_BUG_ON_PAGE(PageCompound(page
), page
);
1439 VM_BUG_ON_PAGE(!page_count(page
), page
);
1441 #ifdef CONFIG_KMEMCHECK
1443 * Split shadow pages too, because free(page[0]) would
1444 * otherwise free the whole shadow.
1446 if (kmemcheck_page_is_tracked(page
))
1447 split_page(virt_to_page(page
[0].shadow
), order
);
1450 for (i
= 1; i
< (1 << order
); i
++)
1451 set_page_refcounted(page
+ i
);
1453 EXPORT_SYMBOL_GPL(split_page
);
1455 static int __isolate_free_page(struct page
*page
, unsigned int order
)
1457 unsigned long watermark
;
1461 BUG_ON(!PageBuddy(page
));
1463 zone
= page_zone(page
);
1464 mt
= get_pageblock_migratetype(page
);
1466 if (!is_migrate_isolate(mt
)) {
1467 /* Obey watermarks as if the page was being allocated */
1468 watermark
= low_wmark_pages(zone
) + (1 << order
);
1469 if (!zone_watermark_ok(zone
, 0, watermark
, 0, 0))
1472 __mod_zone_freepage_state(zone
, -(1UL << order
), mt
);
1475 /* Remove page from free list */
1476 list_del(&page
->lru
);
1477 zone
->free_area
[order
].nr_free
--;
1478 rmv_page_order(page
);
1480 /* Set the pageblock if the isolated page is at least a pageblock */
1481 if (order
>= pageblock_order
- 1) {
1482 struct page
*endpage
= page
+ (1 << order
) - 1;
1483 for (; page
< endpage
; page
+= pageblock_nr_pages
) {
1484 int mt
= get_pageblock_migratetype(page
);
1485 if (!is_migrate_isolate(mt
) && !is_migrate_cma(mt
))
1486 set_pageblock_migratetype(page
,
1491 return 1UL << order
;
1495 * Similar to split_page except the page is already free. As this is only
1496 * being used for migration, the migratetype of the block also changes.
1497 * As this is called with interrupts disabled, the caller is responsible
1498 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1501 * Note: this is probably too low level an operation for use in drivers.
1502 * Please consult with lkml before using this in your driver.
1504 int split_free_page(struct page
*page
)
1509 order
= page_order(page
);
1511 nr_pages
= __isolate_free_page(page
, order
);
1515 /* Split into individual pages */
1516 set_page_refcounted(page
);
1517 split_page(page
, order
);
1522 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1523 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1527 struct page
*buffered_rmqueue(struct zone
*preferred_zone
,
1528 struct zone
*zone
, int order
, gfp_t gfp_flags
,
1531 unsigned long flags
;
1533 int cold
= !!(gfp_flags
& __GFP_COLD
);
1536 if (likely(order
== 0)) {
1537 struct per_cpu_pages
*pcp
;
1538 struct list_head
*list
;
1540 local_irq_save(flags
);
1541 pcp
= &this_cpu_ptr(zone
->pageset
)->pcp
;
1542 list
= &pcp
->lists
[migratetype
];
1543 if (list_empty(list
)) {
1544 pcp
->count
+= rmqueue_bulk(zone
, 0,
1547 if (unlikely(list_empty(list
)))
1552 page
= list_entry(list
->prev
, struct page
, lru
);
1554 page
= list_entry(list
->next
, struct page
, lru
);
1556 list_del(&page
->lru
);
1559 if (unlikely(gfp_flags
& __GFP_NOFAIL
)) {
1561 * __GFP_NOFAIL is not to be used in new code.
1563 * All __GFP_NOFAIL callers should be fixed so that they
1564 * properly detect and handle allocation failures.
1566 * We most definitely don't want callers attempting to
1567 * allocate greater than order-1 page units with
1570 WARN_ON_ONCE(order
> 1);
1572 spin_lock_irqsave(&zone
->lock
, flags
);
1573 page
= __rmqueue(zone
, order
, migratetype
);
1574 spin_unlock(&zone
->lock
);
1577 __mod_zone_freepage_state(zone
, -(1 << order
),
1578 get_freepage_migratetype(page
));
1581 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
, -(1 << order
));
1583 __count_zone_vm_events(PGALLOC
, zone
, 1 << order
);
1584 zone_statistics(preferred_zone
, zone
, gfp_flags
);
1585 local_irq_restore(flags
);
1587 VM_BUG_ON_PAGE(bad_range(zone
, page
), page
);
1588 if (prep_new_page(page
, order
, gfp_flags
))
1593 local_irq_restore(flags
);
1597 #ifdef CONFIG_FAIL_PAGE_ALLOC
1600 struct fault_attr attr
;
1602 u32 ignore_gfp_highmem
;
1603 u32 ignore_gfp_wait
;
1605 } fail_page_alloc
= {
1606 .attr
= FAULT_ATTR_INITIALIZER
,
1607 .ignore_gfp_wait
= 1,
1608 .ignore_gfp_highmem
= 1,
1612 static int __init
setup_fail_page_alloc(char *str
)
1614 return setup_fault_attr(&fail_page_alloc
.attr
, str
);
1616 __setup("fail_page_alloc=", setup_fail_page_alloc
);
1618 static bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1620 if (order
< fail_page_alloc
.min_order
)
1622 if (gfp_mask
& __GFP_NOFAIL
)
1624 if (fail_page_alloc
.ignore_gfp_highmem
&& (gfp_mask
& __GFP_HIGHMEM
))
1626 if (fail_page_alloc
.ignore_gfp_wait
&& (gfp_mask
& __GFP_WAIT
))
1629 return should_fail(&fail_page_alloc
.attr
, 1 << order
);
1632 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1634 static int __init
fail_page_alloc_debugfs(void)
1636 umode_t mode
= S_IFREG
| S_IRUSR
| S_IWUSR
;
1639 dir
= fault_create_debugfs_attr("fail_page_alloc", NULL
,
1640 &fail_page_alloc
.attr
);
1642 return PTR_ERR(dir
);
1644 if (!debugfs_create_bool("ignore-gfp-wait", mode
, dir
,
1645 &fail_page_alloc
.ignore_gfp_wait
))
1647 if (!debugfs_create_bool("ignore-gfp-highmem", mode
, dir
,
1648 &fail_page_alloc
.ignore_gfp_highmem
))
1650 if (!debugfs_create_u32("min-order", mode
, dir
,
1651 &fail_page_alloc
.min_order
))
1656 debugfs_remove_recursive(dir
);
1661 late_initcall(fail_page_alloc_debugfs
);
1663 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1665 #else /* CONFIG_FAIL_PAGE_ALLOC */
1667 static inline bool should_fail_alloc_page(gfp_t gfp_mask
, unsigned int order
)
1672 #endif /* CONFIG_FAIL_PAGE_ALLOC */
1675 * Return true if free pages are above 'mark'. This takes into account the order
1676 * of the allocation.
1678 static bool __zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1679 int classzone_idx
, int alloc_flags
, long free_pages
)
1681 /* free_pages my go negative - that's OK */
1683 long lowmem_reserve
= z
->lowmem_reserve
[classzone_idx
];
1687 free_pages
-= (1 << order
) - 1;
1688 if (alloc_flags
& ALLOC_HIGH
)
1690 if (alloc_flags
& ALLOC_HARDER
)
1693 /* If allocation can't use CMA areas don't use free CMA pages */
1694 if (!(alloc_flags
& ALLOC_CMA
))
1695 free_cma
= zone_page_state(z
, NR_FREE_CMA_PAGES
);
1698 if (free_pages
- free_cma
<= min
+ lowmem_reserve
)
1700 for (o
= 0; o
< order
; o
++) {
1701 /* At the next order, this order's pages become unavailable */
1702 free_pages
-= z
->free_area
[o
].nr_free
<< o
;
1704 /* Require fewer higher order pages to be free */
1707 if (free_pages
<= min
)
1713 bool zone_watermark_ok(struct zone
*z
, int order
, unsigned long mark
,
1714 int classzone_idx
, int alloc_flags
)
1716 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1717 zone_page_state(z
, NR_FREE_PAGES
));
1720 bool zone_watermark_ok_safe(struct zone
*z
, int order
, unsigned long mark
,
1721 int classzone_idx
, int alloc_flags
)
1723 long free_pages
= zone_page_state(z
, NR_FREE_PAGES
);
1725 if (z
->percpu_drift_mark
&& free_pages
< z
->percpu_drift_mark
)
1726 free_pages
= zone_page_state_snapshot(z
, NR_FREE_PAGES
);
1728 return __zone_watermark_ok(z
, order
, mark
, classzone_idx
, alloc_flags
,
1734 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1735 * skip over zones that are not allowed by the cpuset, or that have
1736 * been recently (in last second) found to be nearly full. See further
1737 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1738 * that have to skip over a lot of full or unallowed zones.
1740 * If the zonelist cache is present in the passed zonelist, then
1741 * returns a pointer to the allowed node mask (either the current
1742 * tasks mems_allowed, or node_states[N_MEMORY].)
1744 * If the zonelist cache is not available for this zonelist, does
1745 * nothing and returns NULL.
1747 * If the fullzones BITMAP in the zonelist cache is stale (more than
1748 * a second since last zap'd) then we zap it out (clear its bits.)
1750 * We hold off even calling zlc_setup, until after we've checked the
1751 * first zone in the zonelist, on the theory that most allocations will
1752 * be satisfied from that first zone, so best to examine that zone as
1753 * quickly as we can.
1755 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1757 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1758 nodemask_t
*allowednodes
; /* zonelist_cache approximation */
1760 zlc
= zonelist
->zlcache_ptr
;
1764 if (time_after(jiffies
, zlc
->last_full_zap
+ HZ
)) {
1765 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1766 zlc
->last_full_zap
= jiffies
;
1769 allowednodes
= !in_interrupt() && (alloc_flags
& ALLOC_CPUSET
) ?
1770 &cpuset_current_mems_allowed
:
1771 &node_states
[N_MEMORY
];
1772 return allowednodes
;
1776 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1777 * if it is worth looking at further for free memory:
1778 * 1) Check that the zone isn't thought to be full (doesn't have its
1779 * bit set in the zonelist_cache fullzones BITMAP).
1780 * 2) Check that the zones node (obtained from the zonelist_cache
1781 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1782 * Return true (non-zero) if zone is worth looking at further, or
1783 * else return false (zero) if it is not.
1785 * This check -ignores- the distinction between various watermarks,
1786 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1787 * found to be full for any variation of these watermarks, it will
1788 * be considered full for up to one second by all requests, unless
1789 * we are so low on memory on all allowed nodes that we are forced
1790 * into the second scan of the zonelist.
1792 * In the second scan we ignore this zonelist cache and exactly
1793 * apply the watermarks to all zones, even it is slower to do so.
1794 * We are low on memory in the second scan, and should leave no stone
1795 * unturned looking for a free page.
1797 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1798 nodemask_t
*allowednodes
)
1800 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1801 int i
; /* index of *z in zonelist zones */
1802 int n
; /* node that zone *z is on */
1804 zlc
= zonelist
->zlcache_ptr
;
1808 i
= z
- zonelist
->_zonerefs
;
1811 /* This zone is worth trying if it is allowed but not full */
1812 return node_isset(n
, *allowednodes
) && !test_bit(i
, zlc
->fullzones
);
1816 * Given 'z' scanning a zonelist, set the corresponding bit in
1817 * zlc->fullzones, so that subsequent attempts to allocate a page
1818 * from that zone don't waste time re-examining it.
1820 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1822 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1823 int i
; /* index of *z in zonelist zones */
1825 zlc
= zonelist
->zlcache_ptr
;
1829 i
= z
- zonelist
->_zonerefs
;
1831 set_bit(i
, zlc
->fullzones
);
1835 * clear all zones full, called after direct reclaim makes progress so that
1836 * a zone that was recently full is not skipped over for up to a second
1838 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1840 struct zonelist_cache
*zlc
; /* cached zonelist speedup info */
1842 zlc
= zonelist
->zlcache_ptr
;
1846 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
1849 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1851 return local_zone
->node
== zone
->node
;
1854 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1856 return node_distance(zone_to_nid(local_zone
), zone_to_nid(zone
)) <
1860 #else /* CONFIG_NUMA */
1862 static nodemask_t
*zlc_setup(struct zonelist
*zonelist
, int alloc_flags
)
1867 static int zlc_zone_worth_trying(struct zonelist
*zonelist
, struct zoneref
*z
,
1868 nodemask_t
*allowednodes
)
1873 static void zlc_mark_zone_full(struct zonelist
*zonelist
, struct zoneref
*z
)
1877 static void zlc_clear_zones_full(struct zonelist
*zonelist
)
1881 static bool zone_local(struct zone
*local_zone
, struct zone
*zone
)
1886 static bool zone_allows_reclaim(struct zone
*local_zone
, struct zone
*zone
)
1891 #endif /* CONFIG_NUMA */
1894 * get_page_from_freelist goes through the zonelist trying to allocate
1897 static struct page
*
1898 get_page_from_freelist(gfp_t gfp_mask
, nodemask_t
*nodemask
, unsigned int order
,
1899 struct zonelist
*zonelist
, int high_zoneidx
, int alloc_flags
,
1900 struct zone
*preferred_zone
, int migratetype
)
1903 struct page
*page
= NULL
;
1906 nodemask_t
*allowednodes
= NULL
;/* zonelist_cache approximation */
1907 int zlc_active
= 0; /* set if using zonelist_cache */
1908 int did_zlc_setup
= 0; /* just call zlc_setup() one time */
1910 classzone_idx
= zone_idx(preferred_zone
);
1913 * Scan zonelist, looking for a zone with enough free.
1914 * See also __cpuset_node_allowed_softwall() comment in kernel/cpuset.c.
1916 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
1917 high_zoneidx
, nodemask
) {
1920 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1921 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
1923 if ((alloc_flags
& ALLOC_CPUSET
) &&
1924 !cpuset_zone_allowed_softwall(zone
, gfp_mask
))
1926 BUILD_BUG_ON(ALLOC_NO_WATERMARKS
< NR_WMARK
);
1927 if (unlikely(alloc_flags
& ALLOC_NO_WATERMARKS
))
1930 * Distribute pages in proportion to the individual
1931 * zone size to ensure fair page aging. The zone a
1932 * page was allocated in should have no effect on the
1933 * time the page has in memory before being reclaimed.
1935 if (alloc_flags
& ALLOC_FAIR
) {
1936 if (!zone_local(preferred_zone
, zone
))
1938 if (zone_page_state(zone
, NR_ALLOC_BATCH
) <= 0)
1942 * When allocating a page cache page for writing, we
1943 * want to get it from a zone that is within its dirty
1944 * limit, such that no single zone holds more than its
1945 * proportional share of globally allowed dirty pages.
1946 * The dirty limits take into account the zone's
1947 * lowmem reserves and high watermark so that kswapd
1948 * should be able to balance it without having to
1949 * write pages from its LRU list.
1951 * This may look like it could increase pressure on
1952 * lower zones by failing allocations in higher zones
1953 * before they are full. But the pages that do spill
1954 * over are limited as the lower zones are protected
1955 * by this very same mechanism. It should not become
1956 * a practical burden to them.
1958 * XXX: For now, allow allocations to potentially
1959 * exceed the per-zone dirty limit in the slowpath
1960 * (ALLOC_WMARK_LOW unset) before going into reclaim,
1961 * which is important when on a NUMA setup the allowed
1962 * zones are together not big enough to reach the
1963 * global limit. The proper fix for these situations
1964 * will require awareness of zones in the
1965 * dirty-throttling and the flusher threads.
1967 if ((alloc_flags
& ALLOC_WMARK_LOW
) &&
1968 (gfp_mask
& __GFP_WRITE
) && !zone_dirty_ok(zone
))
1969 goto this_zone_full
;
1971 mark
= zone
->watermark
[alloc_flags
& ALLOC_WMARK_MASK
];
1972 if (!zone_watermark_ok(zone
, order
, mark
,
1973 classzone_idx
, alloc_flags
)) {
1976 if (IS_ENABLED(CONFIG_NUMA
) &&
1977 !did_zlc_setup
&& nr_online_nodes
> 1) {
1979 * we do zlc_setup if there are multiple nodes
1980 * and before considering the first zone allowed
1983 allowednodes
= zlc_setup(zonelist
, alloc_flags
);
1988 if (zone_reclaim_mode
== 0 ||
1989 !zone_allows_reclaim(preferred_zone
, zone
))
1990 goto this_zone_full
;
1993 * As we may have just activated ZLC, check if the first
1994 * eligible zone has failed zone_reclaim recently.
1996 if (IS_ENABLED(CONFIG_NUMA
) && zlc_active
&&
1997 !zlc_zone_worth_trying(zonelist
, z
, allowednodes
))
2000 ret
= zone_reclaim(zone
, gfp_mask
, order
);
2002 case ZONE_RECLAIM_NOSCAN
:
2005 case ZONE_RECLAIM_FULL
:
2006 /* scanned but unreclaimable */
2009 /* did we reclaim enough */
2010 if (zone_watermark_ok(zone
, order
, mark
,
2011 classzone_idx
, alloc_flags
))
2015 * Failed to reclaim enough to meet watermark.
2016 * Only mark the zone full if checking the min
2017 * watermark or if we failed to reclaim just
2018 * 1<<order pages or else the page allocator
2019 * fastpath will prematurely mark zones full
2020 * when the watermark is between the low and
2023 if (((alloc_flags
& ALLOC_WMARK_MASK
) == ALLOC_WMARK_MIN
) ||
2024 ret
== ZONE_RECLAIM_SOME
)
2025 goto this_zone_full
;
2032 page
= buffered_rmqueue(preferred_zone
, zone
, order
,
2033 gfp_mask
, migratetype
);
2037 if (IS_ENABLED(CONFIG_NUMA
))
2038 zlc_mark_zone_full(zonelist
, z
);
2041 if (unlikely(IS_ENABLED(CONFIG_NUMA
) && page
== NULL
&& zlc_active
)) {
2042 /* Disable zlc cache for second zonelist scan */
2049 * page->pfmemalloc is set when ALLOC_NO_WATERMARKS was
2050 * necessary to allocate the page. The expectation is
2051 * that the caller is taking steps that will free more
2052 * memory. The caller should avoid the page being used
2053 * for !PFMEMALLOC purposes.
2055 page
->pfmemalloc
= !!(alloc_flags
& ALLOC_NO_WATERMARKS
);
2061 * Large machines with many possible nodes should not always dump per-node
2062 * meminfo in irq context.
2064 static inline bool should_suppress_show_mem(void)
2069 ret
= in_interrupt();
2074 static DEFINE_RATELIMIT_STATE(nopage_rs
,
2075 DEFAULT_RATELIMIT_INTERVAL
,
2076 DEFAULT_RATELIMIT_BURST
);
2078 void warn_alloc_failed(gfp_t gfp_mask
, int order
, const char *fmt
, ...)
2080 unsigned int filter
= SHOW_MEM_FILTER_NODES
;
2082 if ((gfp_mask
& __GFP_NOWARN
) || !__ratelimit(&nopage_rs
) ||
2083 debug_guardpage_minorder() > 0)
2087 * This documents exceptions given to allocations in certain
2088 * contexts that are allowed to allocate outside current's set
2091 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2092 if (test_thread_flag(TIF_MEMDIE
) ||
2093 (current
->flags
& (PF_MEMALLOC
| PF_EXITING
)))
2094 filter
&= ~SHOW_MEM_FILTER_NODES
;
2095 if (in_interrupt() || !(gfp_mask
& __GFP_WAIT
))
2096 filter
&= ~SHOW_MEM_FILTER_NODES
;
2099 struct va_format vaf
;
2102 va_start(args
, fmt
);
2107 pr_warn("%pV", &vaf
);
2112 pr_warn("%s: page allocation failure: order:%d, mode:0x%x\n",
2113 current
->comm
, order
, gfp_mask
);
2116 if (!should_suppress_show_mem())
2121 should_alloc_retry(gfp_t gfp_mask
, unsigned int order
,
2122 unsigned long did_some_progress
,
2123 unsigned long pages_reclaimed
)
2125 /* Do not loop if specifically requested */
2126 if (gfp_mask
& __GFP_NORETRY
)
2129 /* Always retry if specifically requested */
2130 if (gfp_mask
& __GFP_NOFAIL
)
2134 * Suspend converts GFP_KERNEL to __GFP_WAIT which can prevent reclaim
2135 * making forward progress without invoking OOM. Suspend also disables
2136 * storage devices so kswapd will not help. Bail if we are suspending.
2138 if (!did_some_progress
&& pm_suspended_storage())
2142 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
2143 * means __GFP_NOFAIL, but that may not be true in other
2146 if (order
<= PAGE_ALLOC_COSTLY_ORDER
)
2150 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
2151 * specified, then we retry until we no longer reclaim any pages
2152 * (above), or we've reclaimed an order of pages at least as
2153 * large as the allocation's order. In both cases, if the
2154 * allocation still fails, we stop retrying.
2156 if (gfp_mask
& __GFP_REPEAT
&& pages_reclaimed
< (1 << order
))
2162 static inline struct page
*
2163 __alloc_pages_may_oom(gfp_t gfp_mask
, unsigned int order
,
2164 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2165 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2170 /* Acquire the OOM killer lock for the zones in zonelist */
2171 if (!try_set_zonelist_oom(zonelist
, gfp_mask
)) {
2172 schedule_timeout_uninterruptible(1);
2177 * Go through the zonelist yet one more time, keep very high watermark
2178 * here, this is only to catch a parallel oom killing, we must fail if
2179 * we're still under heavy pressure.
2181 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
,
2182 order
, zonelist
, high_zoneidx
,
2183 ALLOC_WMARK_HIGH
|ALLOC_CPUSET
,
2184 preferred_zone
, migratetype
);
2188 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2189 /* The OOM killer will not help higher order allocs */
2190 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2192 /* The OOM killer does not needlessly kill tasks for lowmem */
2193 if (high_zoneidx
< ZONE_NORMAL
)
2196 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
2197 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
2198 * The caller should handle page allocation failure by itself if
2199 * it specifies __GFP_THISNODE.
2200 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
2202 if (gfp_mask
& __GFP_THISNODE
)
2205 /* Exhausted what can be done so it's blamo time */
2206 out_of_memory(zonelist
, gfp_mask
, order
, nodemask
, false);
2209 clear_zonelist_oom(zonelist
, gfp_mask
);
2213 #ifdef CONFIG_COMPACTION
2214 /* Try memory compaction for high-order allocations before reclaim */
2215 static struct page
*
2216 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2217 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2218 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2219 int migratetype
, bool sync_migration
,
2220 bool *contended_compaction
, bool *deferred_compaction
,
2221 unsigned long *did_some_progress
)
2226 if (compaction_deferred(preferred_zone
, order
)) {
2227 *deferred_compaction
= true;
2231 current
->flags
|= PF_MEMALLOC
;
2232 *did_some_progress
= try_to_compact_pages(zonelist
, order
, gfp_mask
,
2233 nodemask
, sync_migration
,
2234 contended_compaction
);
2235 current
->flags
&= ~PF_MEMALLOC
;
2237 if (*did_some_progress
!= COMPACT_SKIPPED
) {
2240 /* Page migration frees to the PCP lists but we want merging */
2241 drain_pages(get_cpu());
2244 page
= get_page_from_freelist(gfp_mask
, nodemask
,
2245 order
, zonelist
, high_zoneidx
,
2246 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2247 preferred_zone
, migratetype
);
2249 preferred_zone
->compact_blockskip_flush
= false;
2250 compaction_defer_reset(preferred_zone
, order
, true);
2251 count_vm_event(COMPACTSUCCESS
);
2256 * It's bad if compaction run occurs and fails.
2257 * The most likely reason is that pages exist,
2258 * but not enough to satisfy watermarks.
2260 count_vm_event(COMPACTFAIL
);
2263 * As async compaction considers a subset of pageblocks, only
2264 * defer if the failure was a sync compaction failure.
2267 defer_compaction(preferred_zone
, order
);
2275 static inline struct page
*
2276 __alloc_pages_direct_compact(gfp_t gfp_mask
, unsigned int order
,
2277 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2278 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2279 int migratetype
, bool sync_migration
,
2280 bool *contended_compaction
, bool *deferred_compaction
,
2281 unsigned long *did_some_progress
)
2285 #endif /* CONFIG_COMPACTION */
2287 /* Perform direct synchronous page reclaim */
2289 __perform_reclaim(gfp_t gfp_mask
, unsigned int order
, struct zonelist
*zonelist
,
2290 nodemask_t
*nodemask
)
2292 struct reclaim_state reclaim_state
;
2297 /* We now go into synchronous reclaim */
2298 cpuset_memory_pressure_bump();
2299 current
->flags
|= PF_MEMALLOC
;
2300 lockdep_set_current_reclaim_state(gfp_mask
);
2301 reclaim_state
.reclaimed_slab
= 0;
2302 current
->reclaim_state
= &reclaim_state
;
2304 progress
= try_to_free_pages(zonelist
, order
, gfp_mask
, nodemask
);
2306 current
->reclaim_state
= NULL
;
2307 lockdep_clear_current_reclaim_state();
2308 current
->flags
&= ~PF_MEMALLOC
;
2315 /* The really slow allocator path where we enter direct reclaim */
2316 static inline struct page
*
2317 __alloc_pages_direct_reclaim(gfp_t gfp_mask
, unsigned int order
,
2318 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2319 nodemask_t
*nodemask
, int alloc_flags
, struct zone
*preferred_zone
,
2320 int migratetype
, unsigned long *did_some_progress
)
2322 struct page
*page
= NULL
;
2323 bool drained
= false;
2325 *did_some_progress
= __perform_reclaim(gfp_mask
, order
, zonelist
,
2327 if (unlikely(!(*did_some_progress
)))
2330 /* After successful reclaim, reconsider all zones for allocation */
2331 if (IS_ENABLED(CONFIG_NUMA
))
2332 zlc_clear_zones_full(zonelist
);
2335 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2336 zonelist
, high_zoneidx
,
2337 alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2338 preferred_zone
, migratetype
);
2341 * If an allocation failed after direct reclaim, it could be because
2342 * pages are pinned on the per-cpu lists. Drain them and try again
2344 if (!page
&& !drained
) {
2354 * This is called in the allocator slow-path if the allocation request is of
2355 * sufficient urgency to ignore watermarks and take other desperate measures
2357 static inline struct page
*
2358 __alloc_pages_high_priority(gfp_t gfp_mask
, unsigned int order
,
2359 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2360 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2366 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
,
2367 zonelist
, high_zoneidx
, ALLOC_NO_WATERMARKS
,
2368 preferred_zone
, migratetype
);
2370 if (!page
&& gfp_mask
& __GFP_NOFAIL
)
2371 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2372 } while (!page
&& (gfp_mask
& __GFP_NOFAIL
));
2377 static void reset_alloc_batches(struct zonelist
*zonelist
,
2378 enum zone_type high_zoneidx
,
2379 struct zone
*preferred_zone
)
2384 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
) {
2386 * Only reset the batches of zones that were actually
2387 * considered in the fairness pass, we don't want to
2388 * trash fairness information for zones that are not
2389 * actually part of this zonelist's round-robin cycle.
2391 if (!zone_local(preferred_zone
, zone
))
2393 mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
2394 high_wmark_pages(zone
) - low_wmark_pages(zone
) -
2395 atomic_long_read(&zone
->vm_stat
[NR_ALLOC_BATCH
]));
2399 static void wake_all_kswapds(unsigned int order
,
2400 struct zonelist
*zonelist
,
2401 enum zone_type high_zoneidx
,
2402 struct zone
*preferred_zone
)
2407 for_each_zone_zonelist(zone
, z
, zonelist
, high_zoneidx
)
2408 wakeup_kswapd(zone
, order
, zone_idx(preferred_zone
));
2412 gfp_to_alloc_flags(gfp_t gfp_mask
)
2414 int alloc_flags
= ALLOC_WMARK_MIN
| ALLOC_CPUSET
;
2415 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2417 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2418 BUILD_BUG_ON(__GFP_HIGH
!= (__force gfp_t
) ALLOC_HIGH
);
2421 * The caller may dip into page reserves a bit more if the caller
2422 * cannot run direct reclaim, or if the caller has realtime scheduling
2423 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2424 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2426 alloc_flags
|= (__force
int) (gfp_mask
& __GFP_HIGH
);
2430 * Not worth trying to allocate harder for
2431 * __GFP_NOMEMALLOC even if it can't schedule.
2433 if (!(gfp_mask
& __GFP_NOMEMALLOC
))
2434 alloc_flags
|= ALLOC_HARDER
;
2436 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2437 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2439 alloc_flags
&= ~ALLOC_CPUSET
;
2440 } else if (unlikely(rt_task(current
)) && !in_interrupt())
2441 alloc_flags
|= ALLOC_HARDER
;
2443 if (likely(!(gfp_mask
& __GFP_NOMEMALLOC
))) {
2444 if (gfp_mask
& __GFP_MEMALLOC
)
2445 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2446 else if (in_serving_softirq() && (current
->flags
& PF_MEMALLOC
))
2447 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2448 else if (!in_interrupt() &&
2449 ((current
->flags
& PF_MEMALLOC
) ||
2450 unlikely(test_thread_flag(TIF_MEMDIE
))))
2451 alloc_flags
|= ALLOC_NO_WATERMARKS
;
2454 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2455 alloc_flags
|= ALLOC_CMA
;
2460 bool gfp_pfmemalloc_allowed(gfp_t gfp_mask
)
2462 return !!(gfp_to_alloc_flags(gfp_mask
) & ALLOC_NO_WATERMARKS
);
2465 static inline struct page
*
2466 __alloc_pages_slowpath(gfp_t gfp_mask
, unsigned int order
,
2467 struct zonelist
*zonelist
, enum zone_type high_zoneidx
,
2468 nodemask_t
*nodemask
, struct zone
*preferred_zone
,
2471 const gfp_t wait
= gfp_mask
& __GFP_WAIT
;
2472 struct page
*page
= NULL
;
2474 unsigned long pages_reclaimed
= 0;
2475 unsigned long did_some_progress
;
2476 bool sync_migration
= false;
2477 bool deferred_compaction
= false;
2478 bool contended_compaction
= false;
2481 * In the slowpath, we sanity check order to avoid ever trying to
2482 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2483 * be using allocators in order of preference for an area that is
2486 if (order
>= MAX_ORDER
) {
2487 WARN_ON_ONCE(!(gfp_mask
& __GFP_NOWARN
));
2492 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2493 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2494 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2495 * using a larger set of nodes after it has established that the
2496 * allowed per node queues are empty and that nodes are
2499 if (IS_ENABLED(CONFIG_NUMA
) &&
2500 (gfp_mask
& GFP_THISNODE
) == GFP_THISNODE
)
2504 if (!(gfp_mask
& __GFP_NO_KSWAPD
))
2505 wake_all_kswapds(order
, zonelist
, high_zoneidx
, preferred_zone
);
2508 * OK, we're below the kswapd watermark and have kicked background
2509 * reclaim. Now things get more complex, so set up alloc_flags according
2510 * to how we want to proceed.
2512 alloc_flags
= gfp_to_alloc_flags(gfp_mask
);
2515 * Find the true preferred zone if the allocation is unconstrained by
2518 if (!(alloc_flags
& ALLOC_CPUSET
) && !nodemask
)
2519 first_zones_zonelist(zonelist
, high_zoneidx
, NULL
,
2523 /* This is the last chance, in general, before the goto nopage. */
2524 page
= get_page_from_freelist(gfp_mask
, nodemask
, order
, zonelist
,
2525 high_zoneidx
, alloc_flags
& ~ALLOC_NO_WATERMARKS
,
2526 preferred_zone
, migratetype
);
2530 /* Allocate without watermarks if the context allows */
2531 if (alloc_flags
& ALLOC_NO_WATERMARKS
) {
2533 * Ignore mempolicies if ALLOC_NO_WATERMARKS on the grounds
2534 * the allocation is high priority and these type of
2535 * allocations are system rather than user orientated
2537 zonelist
= node_zonelist(numa_node_id(), gfp_mask
);
2539 page
= __alloc_pages_high_priority(gfp_mask
, order
,
2540 zonelist
, high_zoneidx
, nodemask
,
2541 preferred_zone
, migratetype
);
2547 /* Atomic allocations - we can't balance anything */
2550 * All existing users of the deprecated __GFP_NOFAIL are
2551 * blockable, so warn of any new users that actually allow this
2552 * type of allocation to fail.
2554 WARN_ON_ONCE(gfp_mask
& __GFP_NOFAIL
);
2558 /* Avoid recursion of direct reclaim */
2559 if (current
->flags
& PF_MEMALLOC
)
2562 /* Avoid allocations with no watermarks from looping endlessly */
2563 if (test_thread_flag(TIF_MEMDIE
) && !(gfp_mask
& __GFP_NOFAIL
))
2567 * Try direct compaction. The first pass is asynchronous. Subsequent
2568 * attempts after direct reclaim are synchronous
2570 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2571 zonelist
, high_zoneidx
,
2573 alloc_flags
, preferred_zone
,
2574 migratetype
, sync_migration
,
2575 &contended_compaction
,
2576 &deferred_compaction
,
2577 &did_some_progress
);
2580 sync_migration
= true;
2583 * If compaction is deferred for high-order allocations, it is because
2584 * sync compaction recently failed. In this is the case and the caller
2585 * requested a movable allocation that does not heavily disrupt the
2586 * system then fail the allocation instead of entering direct reclaim.
2588 if ((deferred_compaction
|| contended_compaction
) &&
2589 (gfp_mask
& __GFP_NO_KSWAPD
))
2592 /* Try direct reclaim and then allocating */
2593 page
= __alloc_pages_direct_reclaim(gfp_mask
, order
,
2594 zonelist
, high_zoneidx
,
2596 alloc_flags
, preferred_zone
,
2597 migratetype
, &did_some_progress
);
2602 * If we failed to make any progress reclaiming, then we are
2603 * running out of options and have to consider going OOM
2605 if (!did_some_progress
) {
2606 if (oom_gfp_allowed(gfp_mask
)) {
2607 if (oom_killer_disabled
)
2609 /* Coredumps can quickly deplete all memory reserves */
2610 if ((current
->flags
& PF_DUMPCORE
) &&
2611 !(gfp_mask
& __GFP_NOFAIL
))
2613 page
= __alloc_pages_may_oom(gfp_mask
, order
,
2614 zonelist
, high_zoneidx
,
2615 nodemask
, preferred_zone
,
2620 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2622 * The oom killer is not called for high-order
2623 * allocations that may fail, so if no progress
2624 * is being made, there are no other options and
2625 * retrying is unlikely to help.
2627 if (order
> PAGE_ALLOC_COSTLY_ORDER
)
2630 * The oom killer is not called for lowmem
2631 * allocations to prevent needlessly killing
2634 if (high_zoneidx
< ZONE_NORMAL
)
2642 /* Check if we should retry the allocation */
2643 pages_reclaimed
+= did_some_progress
;
2644 if (should_alloc_retry(gfp_mask
, order
, did_some_progress
,
2646 /* Wait for some write requests to complete then retry */
2647 wait_iff_congested(preferred_zone
, BLK_RW_ASYNC
, HZ
/50);
2651 * High-order allocations do not necessarily loop after
2652 * direct reclaim and reclaim/compaction depends on compaction
2653 * being called after reclaim so call directly if necessary
2655 page
= __alloc_pages_direct_compact(gfp_mask
, order
,
2656 zonelist
, high_zoneidx
,
2658 alloc_flags
, preferred_zone
,
2659 migratetype
, sync_migration
,
2660 &contended_compaction
,
2661 &deferred_compaction
,
2662 &did_some_progress
);
2668 warn_alloc_failed(gfp_mask
, order
, NULL
);
2671 if (kmemcheck_enabled
)
2672 kmemcheck_pagealloc_alloc(page
, order
, gfp_mask
);
2678 * This is the 'heart' of the zoned buddy allocator.
2681 __alloc_pages_nodemask(gfp_t gfp_mask
, unsigned int order
,
2682 struct zonelist
*zonelist
, nodemask_t
*nodemask
)
2684 enum zone_type high_zoneidx
= gfp_zone(gfp_mask
);
2685 struct zone
*preferred_zone
;
2686 struct page
*page
= NULL
;
2687 int migratetype
= allocflags_to_migratetype(gfp_mask
);
2688 unsigned int cpuset_mems_cookie
;
2689 int alloc_flags
= ALLOC_WMARK_LOW
|ALLOC_CPUSET
|ALLOC_FAIR
;
2691 gfp_mask
&= gfp_allowed_mask
;
2693 lockdep_trace_alloc(gfp_mask
);
2695 might_sleep_if(gfp_mask
& __GFP_WAIT
);
2697 if (should_fail_alloc_page(gfp_mask
, order
))
2701 * Check the zones suitable for the gfp_mask contain at least one
2702 * valid zone. It's possible to have an empty zonelist as a result
2703 * of GFP_THISNODE and a memoryless node
2705 if (unlikely(!zonelist
->_zonerefs
->zone
))
2709 cpuset_mems_cookie
= read_mems_allowed_begin();
2711 /* The preferred zone is used for statistics later */
2712 first_zones_zonelist(zonelist
, high_zoneidx
,
2713 nodemask
? : &cpuset_current_mems_allowed
,
2715 if (!preferred_zone
)
2719 if (allocflags_to_migratetype(gfp_mask
) == MIGRATE_MOVABLE
)
2720 alloc_flags
|= ALLOC_CMA
;
2723 /* First allocation attempt */
2724 page
= get_page_from_freelist(gfp_mask
|__GFP_HARDWALL
, nodemask
, order
,
2725 zonelist
, high_zoneidx
, alloc_flags
,
2726 preferred_zone
, migratetype
);
2727 if (unlikely(!page
)) {
2729 * The first pass makes sure allocations are spread
2730 * fairly within the local node. However, the local
2731 * node might have free pages left after the fairness
2732 * batches are exhausted, and remote zones haven't
2733 * even been considered yet. Try once more without
2734 * fairness, and include remote zones now, before
2735 * entering the slowpath and waking kswapd: prefer
2736 * spilling to a remote zone over swapping locally.
2738 if (alloc_flags
& ALLOC_FAIR
) {
2739 reset_alloc_batches(zonelist
, high_zoneidx
,
2741 alloc_flags
&= ~ALLOC_FAIR
;
2745 * Runtime PM, block IO and its error handling path
2746 * can deadlock because I/O on the device might not
2749 gfp_mask
= memalloc_noio_flags(gfp_mask
);
2750 page
= __alloc_pages_slowpath(gfp_mask
, order
,
2751 zonelist
, high_zoneidx
, nodemask
,
2752 preferred_zone
, migratetype
);
2755 trace_mm_page_alloc(page
, order
, gfp_mask
, migratetype
);
2759 * When updating a task's mems_allowed, it is possible to race with
2760 * parallel threads in such a way that an allocation can fail while
2761 * the mask is being updated. If a page allocation is about to fail,
2762 * check if the cpuset changed during allocation and if so, retry.
2764 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
2769 EXPORT_SYMBOL(__alloc_pages_nodemask
);
2772 * Common helper functions.
2774 unsigned long __get_free_pages(gfp_t gfp_mask
, unsigned int order
)
2779 * __get_free_pages() returns a 32-bit address, which cannot represent
2782 VM_BUG_ON((gfp_mask
& __GFP_HIGHMEM
) != 0);
2784 page
= alloc_pages(gfp_mask
, order
);
2787 return (unsigned long) page_address(page
);
2789 EXPORT_SYMBOL(__get_free_pages
);
2791 unsigned long get_zeroed_page(gfp_t gfp_mask
)
2793 return __get_free_pages(gfp_mask
| __GFP_ZERO
, 0);
2795 EXPORT_SYMBOL(get_zeroed_page
);
2797 void __free_pages(struct page
*page
, unsigned int order
)
2799 if (put_page_testzero(page
)) {
2801 free_hot_cold_page(page
, 0);
2803 __free_pages_ok(page
, order
);
2807 EXPORT_SYMBOL(__free_pages
);
2809 void free_pages(unsigned long addr
, unsigned int order
)
2812 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2813 __free_pages(virt_to_page((void *)addr
), order
);
2817 EXPORT_SYMBOL(free_pages
);
2820 * alloc_kmem_pages charges newly allocated pages to the kmem resource counter
2821 * of the current memory cgroup.
2823 * It should be used when the caller would like to use kmalloc, but since the
2824 * allocation is large, it has to fall back to the page allocator.
2826 struct page
*alloc_kmem_pages(gfp_t gfp_mask
, unsigned int order
)
2829 struct mem_cgroup
*memcg
= NULL
;
2831 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2833 page
= alloc_pages(gfp_mask
, order
);
2834 memcg_kmem_commit_charge(page
, memcg
, order
);
2838 struct page
*alloc_kmem_pages_node(int nid
, gfp_t gfp_mask
, unsigned int order
)
2841 struct mem_cgroup
*memcg
= NULL
;
2843 if (!memcg_kmem_newpage_charge(gfp_mask
, &memcg
, order
))
2845 page
= alloc_pages_node(nid
, gfp_mask
, order
);
2846 memcg_kmem_commit_charge(page
, memcg
, order
);
2851 * __free_kmem_pages and free_kmem_pages will free pages allocated with
2854 void __free_kmem_pages(struct page
*page
, unsigned int order
)
2856 memcg_kmem_uncharge_pages(page
, order
);
2857 __free_pages(page
, order
);
2860 void free_kmem_pages(unsigned long addr
, unsigned int order
)
2863 VM_BUG_ON(!virt_addr_valid((void *)addr
));
2864 __free_kmem_pages(virt_to_page((void *)addr
), order
);
2868 static void *make_alloc_exact(unsigned long addr
, unsigned order
, size_t size
)
2871 unsigned long alloc_end
= addr
+ (PAGE_SIZE
<< order
);
2872 unsigned long used
= addr
+ PAGE_ALIGN(size
);
2874 split_page(virt_to_page((void *)addr
), order
);
2875 while (used
< alloc_end
) {
2880 return (void *)addr
;
2884 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2885 * @size: the number of bytes to allocate
2886 * @gfp_mask: GFP flags for the allocation
2888 * This function is similar to alloc_pages(), except that it allocates the
2889 * minimum number of pages to satisfy the request. alloc_pages() can only
2890 * allocate memory in power-of-two pages.
2892 * This function is also limited by MAX_ORDER.
2894 * Memory allocated by this function must be released by free_pages_exact().
2896 void *alloc_pages_exact(size_t size
, gfp_t gfp_mask
)
2898 unsigned int order
= get_order(size
);
2901 addr
= __get_free_pages(gfp_mask
, order
);
2902 return make_alloc_exact(addr
, order
, size
);
2904 EXPORT_SYMBOL(alloc_pages_exact
);
2907 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2909 * @nid: the preferred node ID where memory should be allocated
2910 * @size: the number of bytes to allocate
2911 * @gfp_mask: GFP flags for the allocation
2913 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2915 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2918 void *alloc_pages_exact_nid(int nid
, size_t size
, gfp_t gfp_mask
)
2920 unsigned order
= get_order(size
);
2921 struct page
*p
= alloc_pages_node(nid
, gfp_mask
, order
);
2924 return make_alloc_exact((unsigned long)page_address(p
), order
, size
);
2926 EXPORT_SYMBOL(alloc_pages_exact_nid
);
2929 * free_pages_exact - release memory allocated via alloc_pages_exact()
2930 * @virt: the value returned by alloc_pages_exact.
2931 * @size: size of allocation, same value as passed to alloc_pages_exact().
2933 * Release the memory allocated by a previous call to alloc_pages_exact.
2935 void free_pages_exact(void *virt
, size_t size
)
2937 unsigned long addr
= (unsigned long)virt
;
2938 unsigned long end
= addr
+ PAGE_ALIGN(size
);
2940 while (addr
< end
) {
2945 EXPORT_SYMBOL(free_pages_exact
);
2948 * nr_free_zone_pages - count number of pages beyond high watermark
2949 * @offset: The zone index of the highest zone
2951 * nr_free_zone_pages() counts the number of counts pages which are beyond the
2952 * high watermark within all zones at or below a given zone index. For each
2953 * zone, the number of pages is calculated as:
2954 * managed_pages - high_pages
2956 static unsigned long nr_free_zone_pages(int offset
)
2961 /* Just pick one node, since fallback list is circular */
2962 unsigned long sum
= 0;
2964 struct zonelist
*zonelist
= node_zonelist(numa_node_id(), GFP_KERNEL
);
2966 for_each_zone_zonelist(zone
, z
, zonelist
, offset
) {
2967 unsigned long size
= zone
->managed_pages
;
2968 unsigned long high
= high_wmark_pages(zone
);
2977 * nr_free_buffer_pages - count number of pages beyond high watermark
2979 * nr_free_buffer_pages() counts the number of pages which are beyond the high
2980 * watermark within ZONE_DMA and ZONE_NORMAL.
2982 unsigned long nr_free_buffer_pages(void)
2984 return nr_free_zone_pages(gfp_zone(GFP_USER
));
2986 EXPORT_SYMBOL_GPL(nr_free_buffer_pages
);
2989 * nr_free_pagecache_pages - count number of pages beyond high watermark
2991 * nr_free_pagecache_pages() counts the number of pages which are beyond the
2992 * high watermark within all zones.
2994 unsigned long nr_free_pagecache_pages(void)
2996 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE
));
2999 static inline void show_node(struct zone
*zone
)
3001 if (IS_ENABLED(CONFIG_NUMA
))
3002 printk("Node %d ", zone_to_nid(zone
));
3005 void si_meminfo(struct sysinfo
*val
)
3007 val
->totalram
= totalram_pages
;
3009 val
->freeram
= global_page_state(NR_FREE_PAGES
);
3010 val
->bufferram
= nr_blockdev_pages();
3011 val
->totalhigh
= totalhigh_pages
;
3012 val
->freehigh
= nr_free_highpages();
3013 val
->mem_unit
= PAGE_SIZE
;
3016 EXPORT_SYMBOL(si_meminfo
);
3019 void si_meminfo_node(struct sysinfo
*val
, int nid
)
3021 int zone_type
; /* needs to be signed */
3022 unsigned long managed_pages
= 0;
3023 pg_data_t
*pgdat
= NODE_DATA(nid
);
3025 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++)
3026 managed_pages
+= pgdat
->node_zones
[zone_type
].managed_pages
;
3027 val
->totalram
= managed_pages
;
3028 val
->freeram
= node_page_state(nid
, NR_FREE_PAGES
);
3029 #ifdef CONFIG_HIGHMEM
3030 val
->totalhigh
= pgdat
->node_zones
[ZONE_HIGHMEM
].managed_pages
;
3031 val
->freehigh
= zone_page_state(&pgdat
->node_zones
[ZONE_HIGHMEM
],
3037 val
->mem_unit
= PAGE_SIZE
;
3042 * Determine whether the node should be displayed or not, depending on whether
3043 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
3045 bool skip_free_areas_node(unsigned int flags
, int nid
)
3048 unsigned int cpuset_mems_cookie
;
3050 if (!(flags
& SHOW_MEM_FILTER_NODES
))
3054 cpuset_mems_cookie
= read_mems_allowed_begin();
3055 ret
= !node_isset(nid
, cpuset_current_mems_allowed
);
3056 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
3061 #define K(x) ((x) << (PAGE_SHIFT-10))
3063 static void show_migration_types(unsigned char type
)
3065 static const char types
[MIGRATE_TYPES
] = {
3066 [MIGRATE_UNMOVABLE
] = 'U',
3067 [MIGRATE_RECLAIMABLE
] = 'E',
3068 [MIGRATE_MOVABLE
] = 'M',
3069 [MIGRATE_RESERVE
] = 'R',
3071 [MIGRATE_CMA
] = 'C',
3073 #ifdef CONFIG_MEMORY_ISOLATION
3074 [MIGRATE_ISOLATE
] = 'I',
3077 char tmp
[MIGRATE_TYPES
+ 1];
3081 for (i
= 0; i
< MIGRATE_TYPES
; i
++) {
3082 if (type
& (1 << i
))
3087 printk("(%s) ", tmp
);
3091 * Show free area list (used inside shift_scroll-lock stuff)
3092 * We also calculate the percentage fragmentation. We do this by counting the
3093 * memory on each free list with the exception of the first item on the list.
3094 * Suppresses nodes that are not allowed by current's cpuset if
3095 * SHOW_MEM_FILTER_NODES is passed.
3097 void show_free_areas(unsigned int filter
)
3102 for_each_populated_zone(zone
) {
3103 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3106 printk("%s per-cpu:\n", zone
->name
);
3108 for_each_online_cpu(cpu
) {
3109 struct per_cpu_pageset
*pageset
;
3111 pageset
= per_cpu_ptr(zone
->pageset
, cpu
);
3113 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
3114 cpu
, pageset
->pcp
.high
,
3115 pageset
->pcp
.batch
, pageset
->pcp
.count
);
3119 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
3120 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
3122 " dirty:%lu writeback:%lu unstable:%lu\n"
3123 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
3124 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n"
3126 global_page_state(NR_ACTIVE_ANON
),
3127 global_page_state(NR_INACTIVE_ANON
),
3128 global_page_state(NR_ISOLATED_ANON
),
3129 global_page_state(NR_ACTIVE_FILE
),
3130 global_page_state(NR_INACTIVE_FILE
),
3131 global_page_state(NR_ISOLATED_FILE
),
3132 global_page_state(NR_UNEVICTABLE
),
3133 global_page_state(NR_FILE_DIRTY
),
3134 global_page_state(NR_WRITEBACK
),
3135 global_page_state(NR_UNSTABLE_NFS
),
3136 global_page_state(NR_FREE_PAGES
),
3137 global_page_state(NR_SLAB_RECLAIMABLE
),
3138 global_page_state(NR_SLAB_UNRECLAIMABLE
),
3139 global_page_state(NR_FILE_MAPPED
),
3140 global_page_state(NR_SHMEM
),
3141 global_page_state(NR_PAGETABLE
),
3142 global_page_state(NR_BOUNCE
),
3143 global_page_state(NR_FREE_CMA_PAGES
));
3145 for_each_populated_zone(zone
) {
3148 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3156 " active_anon:%lukB"
3157 " inactive_anon:%lukB"
3158 " active_file:%lukB"
3159 " inactive_file:%lukB"
3160 " unevictable:%lukB"
3161 " isolated(anon):%lukB"
3162 " isolated(file):%lukB"
3170 " slab_reclaimable:%lukB"
3171 " slab_unreclaimable:%lukB"
3172 " kernel_stack:%lukB"
3177 " writeback_tmp:%lukB"
3178 " pages_scanned:%lu"
3179 " all_unreclaimable? %s"
3182 K(zone_page_state(zone
, NR_FREE_PAGES
)),
3183 K(min_wmark_pages(zone
)),
3184 K(low_wmark_pages(zone
)),
3185 K(high_wmark_pages(zone
)),
3186 K(zone_page_state(zone
, NR_ACTIVE_ANON
)),
3187 K(zone_page_state(zone
, NR_INACTIVE_ANON
)),
3188 K(zone_page_state(zone
, NR_ACTIVE_FILE
)),
3189 K(zone_page_state(zone
, NR_INACTIVE_FILE
)),
3190 K(zone_page_state(zone
, NR_UNEVICTABLE
)),
3191 K(zone_page_state(zone
, NR_ISOLATED_ANON
)),
3192 K(zone_page_state(zone
, NR_ISOLATED_FILE
)),
3193 K(zone
->present_pages
),
3194 K(zone
->managed_pages
),
3195 K(zone_page_state(zone
, NR_MLOCK
)),
3196 K(zone_page_state(zone
, NR_FILE_DIRTY
)),
3197 K(zone_page_state(zone
, NR_WRITEBACK
)),
3198 K(zone_page_state(zone
, NR_FILE_MAPPED
)),
3199 K(zone_page_state(zone
, NR_SHMEM
)),
3200 K(zone_page_state(zone
, NR_SLAB_RECLAIMABLE
)),
3201 K(zone_page_state(zone
, NR_SLAB_UNRECLAIMABLE
)),
3202 zone_page_state(zone
, NR_KERNEL_STACK
) *
3204 K(zone_page_state(zone
, NR_PAGETABLE
)),
3205 K(zone_page_state(zone
, NR_UNSTABLE_NFS
)),
3206 K(zone_page_state(zone
, NR_BOUNCE
)),
3207 K(zone_page_state(zone
, NR_FREE_CMA_PAGES
)),
3208 K(zone_page_state(zone
, NR_WRITEBACK_TEMP
)),
3209 zone
->pages_scanned
,
3210 (!zone_reclaimable(zone
) ? "yes" : "no")
3212 printk("lowmem_reserve[]:");
3213 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
3214 printk(" %lu", zone
->lowmem_reserve
[i
]);
3218 for_each_populated_zone(zone
) {
3219 unsigned long nr
[MAX_ORDER
], flags
, order
, total
= 0;
3220 unsigned char types
[MAX_ORDER
];
3222 if (skip_free_areas_node(filter
, zone_to_nid(zone
)))
3225 printk("%s: ", zone
->name
);
3227 spin_lock_irqsave(&zone
->lock
, flags
);
3228 for (order
= 0; order
< MAX_ORDER
; order
++) {
3229 struct free_area
*area
= &zone
->free_area
[order
];
3232 nr
[order
] = area
->nr_free
;
3233 total
+= nr
[order
] << order
;
3236 for (type
= 0; type
< MIGRATE_TYPES
; type
++) {
3237 if (!list_empty(&area
->free_list
[type
]))
3238 types
[order
] |= 1 << type
;
3241 spin_unlock_irqrestore(&zone
->lock
, flags
);
3242 for (order
= 0; order
< MAX_ORDER
; order
++) {
3243 printk("%lu*%lukB ", nr
[order
], K(1UL) << order
);
3245 show_migration_types(types
[order
]);
3247 printk("= %lukB\n", K(total
));
3250 hugetlb_show_meminfo();
3252 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES
));
3254 show_swap_cache_info();
3257 static void zoneref_set_zone(struct zone
*zone
, struct zoneref
*zoneref
)
3259 zoneref
->zone
= zone
;
3260 zoneref
->zone_idx
= zone_idx(zone
);
3264 * Builds allocation fallback zone lists.
3266 * Add all populated zones of a node to the zonelist.
3268 static int build_zonelists_node(pg_data_t
*pgdat
, struct zonelist
*zonelist
,
3272 enum zone_type zone_type
= MAX_NR_ZONES
;
3276 zone
= pgdat
->node_zones
+ zone_type
;
3277 if (populated_zone(zone
)) {
3278 zoneref_set_zone(zone
,
3279 &zonelist
->_zonerefs
[nr_zones
++]);
3280 check_highest_zone(zone_type
);
3282 } while (zone_type
);
3290 * 0 = automatic detection of better ordering.
3291 * 1 = order by ([node] distance, -zonetype)
3292 * 2 = order by (-zonetype, [node] distance)
3294 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
3295 * the same zonelist. So only NUMA can configure this param.
3297 #define ZONELIST_ORDER_DEFAULT 0
3298 #define ZONELIST_ORDER_NODE 1
3299 #define ZONELIST_ORDER_ZONE 2
3301 /* zonelist order in the kernel.
3302 * set_zonelist_order() will set this to NODE or ZONE.
3304 static int current_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3305 static char zonelist_order_name
[3][8] = {"Default", "Node", "Zone"};
3309 /* The value user specified ....changed by config */
3310 static int user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3311 /* string for sysctl */
3312 #define NUMA_ZONELIST_ORDER_LEN 16
3313 char numa_zonelist_order
[16] = "default";
3316 * interface for configure zonelist ordering.
3317 * command line option "numa_zonelist_order"
3318 * = "[dD]efault - default, automatic configuration.
3319 * = "[nN]ode - order by node locality, then by zone within node
3320 * = "[zZ]one - order by zone, then by locality within zone
3323 static int __parse_numa_zonelist_order(char *s
)
3325 if (*s
== 'd' || *s
== 'D') {
3326 user_zonelist_order
= ZONELIST_ORDER_DEFAULT
;
3327 } else if (*s
== 'n' || *s
== 'N') {
3328 user_zonelist_order
= ZONELIST_ORDER_NODE
;
3329 } else if (*s
== 'z' || *s
== 'Z') {
3330 user_zonelist_order
= ZONELIST_ORDER_ZONE
;
3333 "Ignoring invalid numa_zonelist_order value: "
3340 static __init
int setup_numa_zonelist_order(char *s
)
3347 ret
= __parse_numa_zonelist_order(s
);
3349 strlcpy(numa_zonelist_order
, s
, NUMA_ZONELIST_ORDER_LEN
);
3353 early_param("numa_zonelist_order", setup_numa_zonelist_order
);
3356 * sysctl handler for numa_zonelist_order
3358 int numa_zonelist_order_handler(ctl_table
*table
, int write
,
3359 void __user
*buffer
, size_t *length
,
3362 char saved_string
[NUMA_ZONELIST_ORDER_LEN
];
3364 static DEFINE_MUTEX(zl_order_mutex
);
3366 mutex_lock(&zl_order_mutex
);
3368 if (strlen((char *)table
->data
) >= NUMA_ZONELIST_ORDER_LEN
) {
3372 strcpy(saved_string
, (char *)table
->data
);
3374 ret
= proc_dostring(table
, write
, buffer
, length
, ppos
);
3378 int oldval
= user_zonelist_order
;
3380 ret
= __parse_numa_zonelist_order((char *)table
->data
);
3383 * bogus value. restore saved string
3385 strncpy((char *)table
->data
, saved_string
,
3386 NUMA_ZONELIST_ORDER_LEN
);
3387 user_zonelist_order
= oldval
;
3388 } else if (oldval
!= user_zonelist_order
) {
3389 mutex_lock(&zonelists_mutex
);
3390 build_all_zonelists(NULL
, NULL
);
3391 mutex_unlock(&zonelists_mutex
);
3395 mutex_unlock(&zl_order_mutex
);
3400 #define MAX_NODE_LOAD (nr_online_nodes)
3401 static int node_load
[MAX_NUMNODES
];
3404 * find_next_best_node - find the next node that should appear in a given node's fallback list
3405 * @node: node whose fallback list we're appending
3406 * @used_node_mask: nodemask_t of already used nodes
3408 * We use a number of factors to determine which is the next node that should
3409 * appear on a given node's fallback list. The node should not have appeared
3410 * already in @node's fallback list, and it should be the next closest node
3411 * according to the distance array (which contains arbitrary distance values
3412 * from each node to each node in the system), and should also prefer nodes
3413 * with no CPUs, since presumably they'll have very little allocation pressure
3414 * on them otherwise.
3415 * It returns -1 if no node is found.
3417 static int find_next_best_node(int node
, nodemask_t
*used_node_mask
)
3420 int min_val
= INT_MAX
;
3421 int best_node
= NUMA_NO_NODE
;
3422 const struct cpumask
*tmp
= cpumask_of_node(0);
3424 /* Use the local node if we haven't already */
3425 if (!node_isset(node
, *used_node_mask
)) {
3426 node_set(node
, *used_node_mask
);
3430 for_each_node_state(n
, N_MEMORY
) {
3432 /* Don't want a node to appear more than once */
3433 if (node_isset(n
, *used_node_mask
))
3436 /* Use the distance array to find the distance */
3437 val
= node_distance(node
, n
);
3439 /* Penalize nodes under us ("prefer the next node") */
3442 /* Give preference to headless and unused nodes */
3443 tmp
= cpumask_of_node(n
);
3444 if (!cpumask_empty(tmp
))
3445 val
+= PENALTY_FOR_NODE_WITH_CPUS
;
3447 /* Slight preference for less loaded node */
3448 val
*= (MAX_NODE_LOAD
*MAX_NUMNODES
);
3449 val
+= node_load
[n
];
3451 if (val
< min_val
) {
3458 node_set(best_node
, *used_node_mask
);
3465 * Build zonelists ordered by node and zones within node.
3466 * This results in maximum locality--normal zone overflows into local
3467 * DMA zone, if any--but risks exhausting DMA zone.
3469 static void build_zonelists_in_node_order(pg_data_t
*pgdat
, int node
)
3472 struct zonelist
*zonelist
;
3474 zonelist
= &pgdat
->node_zonelists
[0];
3475 for (j
= 0; zonelist
->_zonerefs
[j
].zone
!= NULL
; j
++)
3477 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3478 zonelist
->_zonerefs
[j
].zone
= NULL
;
3479 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3483 * Build gfp_thisnode zonelists
3485 static void build_thisnode_zonelists(pg_data_t
*pgdat
)
3488 struct zonelist
*zonelist
;
3490 zonelist
= &pgdat
->node_zonelists
[1];
3491 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3492 zonelist
->_zonerefs
[j
].zone
= NULL
;
3493 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3497 * Build zonelists ordered by zone and nodes within zones.
3498 * This results in conserving DMA zone[s] until all Normal memory is
3499 * exhausted, but results in overflowing to remote node while memory
3500 * may still exist in local DMA zone.
3502 static int node_order
[MAX_NUMNODES
];
3504 static void build_zonelists_in_zone_order(pg_data_t
*pgdat
, int nr_nodes
)
3507 int zone_type
; /* needs to be signed */
3509 struct zonelist
*zonelist
;
3511 zonelist
= &pgdat
->node_zonelists
[0];
3513 for (zone_type
= MAX_NR_ZONES
- 1; zone_type
>= 0; zone_type
--) {
3514 for (j
= 0; j
< nr_nodes
; j
++) {
3515 node
= node_order
[j
];
3516 z
= &NODE_DATA(node
)->node_zones
[zone_type
];
3517 if (populated_zone(z
)) {
3519 &zonelist
->_zonerefs
[pos
++]);
3520 check_highest_zone(zone_type
);
3524 zonelist
->_zonerefs
[pos
].zone
= NULL
;
3525 zonelist
->_zonerefs
[pos
].zone_idx
= 0;
3528 static int default_zonelist_order(void)
3531 unsigned long low_kmem_size
, total_size
;
3535 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
3536 * If they are really small and used heavily, the system can fall
3537 * into OOM very easily.
3538 * This function detect ZONE_DMA/DMA32 size and configures zone order.
3540 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
3543 for_each_online_node(nid
) {
3544 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3545 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3546 if (populated_zone(z
)) {
3547 if (zone_type
< ZONE_NORMAL
)
3548 low_kmem_size
+= z
->managed_pages
;
3549 total_size
+= z
->managed_pages
;
3550 } else if (zone_type
== ZONE_NORMAL
) {
3552 * If any node has only lowmem, then node order
3553 * is preferred to allow kernel allocations
3554 * locally; otherwise, they can easily infringe
3555 * on other nodes when there is an abundance of
3556 * lowmem available to allocate from.
3558 return ZONELIST_ORDER_NODE
;
3562 if (!low_kmem_size
|| /* there are no DMA area. */
3563 low_kmem_size
> total_size
/2) /* DMA/DMA32 is big. */
3564 return ZONELIST_ORDER_NODE
;
3566 * look into each node's config.
3567 * If there is a node whose DMA/DMA32 memory is very big area on
3568 * local memory, NODE_ORDER may be suitable.
3570 average_size
= total_size
/
3571 (nodes_weight(node_states
[N_MEMORY
]) + 1);
3572 for_each_online_node(nid
) {
3575 for (zone_type
= 0; zone_type
< MAX_NR_ZONES
; zone_type
++) {
3576 z
= &NODE_DATA(nid
)->node_zones
[zone_type
];
3577 if (populated_zone(z
)) {
3578 if (zone_type
< ZONE_NORMAL
)
3579 low_kmem_size
+= z
->present_pages
;
3580 total_size
+= z
->present_pages
;
3583 if (low_kmem_size
&&
3584 total_size
> average_size
&& /* ignore small node */
3585 low_kmem_size
> total_size
* 70/100)
3586 return ZONELIST_ORDER_NODE
;
3588 return ZONELIST_ORDER_ZONE
;
3591 static void set_zonelist_order(void)
3593 if (user_zonelist_order
== ZONELIST_ORDER_DEFAULT
)
3594 current_zonelist_order
= default_zonelist_order();
3596 current_zonelist_order
= user_zonelist_order
;
3599 static void build_zonelists(pg_data_t
*pgdat
)
3603 nodemask_t used_mask
;
3604 int local_node
, prev_node
;
3605 struct zonelist
*zonelist
;
3606 int order
= current_zonelist_order
;
3608 /* initialize zonelists */
3609 for (i
= 0; i
< MAX_ZONELISTS
; i
++) {
3610 zonelist
= pgdat
->node_zonelists
+ i
;
3611 zonelist
->_zonerefs
[0].zone
= NULL
;
3612 zonelist
->_zonerefs
[0].zone_idx
= 0;
3615 /* NUMA-aware ordering of nodes */
3616 local_node
= pgdat
->node_id
;
3617 load
= nr_online_nodes
;
3618 prev_node
= local_node
;
3619 nodes_clear(used_mask
);
3621 memset(node_order
, 0, sizeof(node_order
));
3624 while ((node
= find_next_best_node(local_node
, &used_mask
)) >= 0) {
3626 * We don't want to pressure a particular node.
3627 * So adding penalty to the first node in same
3628 * distance group to make it round-robin.
3630 if (node_distance(local_node
, node
) !=
3631 node_distance(local_node
, prev_node
))
3632 node_load
[node
] = load
;
3636 if (order
== ZONELIST_ORDER_NODE
)
3637 build_zonelists_in_node_order(pgdat
, node
);
3639 node_order
[j
++] = node
; /* remember order */
3642 if (order
== ZONELIST_ORDER_ZONE
) {
3643 /* calculate node order -- i.e., DMA last! */
3644 build_zonelists_in_zone_order(pgdat
, j
);
3647 build_thisnode_zonelists(pgdat
);
3650 /* Construct the zonelist performance cache - see further mmzone.h */
3651 static void build_zonelist_cache(pg_data_t
*pgdat
)
3653 struct zonelist
*zonelist
;
3654 struct zonelist_cache
*zlc
;
3657 zonelist
= &pgdat
->node_zonelists
[0];
3658 zonelist
->zlcache_ptr
= zlc
= &zonelist
->zlcache
;
3659 bitmap_zero(zlc
->fullzones
, MAX_ZONES_PER_ZONELIST
);
3660 for (z
= zonelist
->_zonerefs
; z
->zone
; z
++)
3661 zlc
->z_to_n
[z
- zonelist
->_zonerefs
] = zonelist_node_idx(z
);
3664 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3666 * Return node id of node used for "local" allocations.
3667 * I.e., first node id of first zone in arg node's generic zonelist.
3668 * Used for initializing percpu 'numa_mem', which is used primarily
3669 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3671 int local_memory_node(int node
)
3675 (void)first_zones_zonelist(node_zonelist(node
, GFP_KERNEL
),
3676 gfp_zone(GFP_KERNEL
),
3683 #else /* CONFIG_NUMA */
3685 static void set_zonelist_order(void)
3687 current_zonelist_order
= ZONELIST_ORDER_ZONE
;
3690 static void build_zonelists(pg_data_t
*pgdat
)
3692 int node
, local_node
;
3694 struct zonelist
*zonelist
;
3696 local_node
= pgdat
->node_id
;
3698 zonelist
= &pgdat
->node_zonelists
[0];
3699 j
= build_zonelists_node(pgdat
, zonelist
, 0);
3702 * Now we build the zonelist so that it contains the zones
3703 * of all the other nodes.
3704 * We don't want to pressure a particular node, so when
3705 * building the zones for node N, we make sure that the
3706 * zones coming right after the local ones are those from
3707 * node N+1 (modulo N)
3709 for (node
= local_node
+ 1; node
< MAX_NUMNODES
; node
++) {
3710 if (!node_online(node
))
3712 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3714 for (node
= 0; node
< local_node
; node
++) {
3715 if (!node_online(node
))
3717 j
= build_zonelists_node(NODE_DATA(node
), zonelist
, j
);
3720 zonelist
->_zonerefs
[j
].zone
= NULL
;
3721 zonelist
->_zonerefs
[j
].zone_idx
= 0;
3724 /* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3725 static void build_zonelist_cache(pg_data_t
*pgdat
)
3727 pgdat
->node_zonelists
[0].zlcache_ptr
= NULL
;
3730 #endif /* CONFIG_NUMA */
3733 * Boot pageset table. One per cpu which is going to be used for all
3734 * zones and all nodes. The parameters will be set in such a way
3735 * that an item put on a list will immediately be handed over to
3736 * the buddy list. This is safe since pageset manipulation is done
3737 * with interrupts disabled.
3739 * The boot_pagesets must be kept even after bootup is complete for
3740 * unused processors and/or zones. They do play a role for bootstrapping
3741 * hotplugged processors.
3743 * zoneinfo_show() and maybe other functions do
3744 * not check if the processor is online before following the pageset pointer.
3745 * Other parts of the kernel may not check if the zone is available.
3747 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
);
3748 static DEFINE_PER_CPU(struct per_cpu_pageset
, boot_pageset
);
3749 static void setup_zone_pageset(struct zone
*zone
);
3752 * Global mutex to protect against size modification of zonelists
3753 * as well as to serialize pageset setup for the new populated zone.
3755 DEFINE_MUTEX(zonelists_mutex
);
3757 /* return values int ....just for stop_machine() */
3758 static int __build_all_zonelists(void *data
)
3762 pg_data_t
*self
= data
;
3765 memset(node_load
, 0, sizeof(node_load
));
3768 if (self
&& !node_online(self
->node_id
)) {
3769 build_zonelists(self
);
3770 build_zonelist_cache(self
);
3773 for_each_online_node(nid
) {
3774 pg_data_t
*pgdat
= NODE_DATA(nid
);
3776 build_zonelists(pgdat
);
3777 build_zonelist_cache(pgdat
);
3781 * Initialize the boot_pagesets that are going to be used
3782 * for bootstrapping processors. The real pagesets for
3783 * each zone will be allocated later when the per cpu
3784 * allocator is available.
3786 * boot_pagesets are used also for bootstrapping offline
3787 * cpus if the system is already booted because the pagesets
3788 * are needed to initialize allocators on a specific cpu too.
3789 * F.e. the percpu allocator needs the page allocator which
3790 * needs the percpu allocator in order to allocate its pagesets
3791 * (a chicken-egg dilemma).
3793 for_each_possible_cpu(cpu
) {
3794 setup_pageset(&per_cpu(boot_pageset
, cpu
), 0);
3796 #ifdef CONFIG_HAVE_MEMORYLESS_NODES
3798 * We now know the "local memory node" for each node--
3799 * i.e., the node of the first zone in the generic zonelist.
3800 * Set up numa_mem percpu variable for on-line cpus. During
3801 * boot, only the boot cpu should be on-line; we'll init the
3802 * secondary cpus' numa_mem as they come on-line. During
3803 * node/memory hotplug, we'll fixup all on-line cpus.
3805 if (cpu_online(cpu
))
3806 set_cpu_numa_mem(cpu
, local_memory_node(cpu_to_node(cpu
)));
3814 * Called with zonelists_mutex held always
3815 * unless system_state == SYSTEM_BOOTING.
3817 void __ref
build_all_zonelists(pg_data_t
*pgdat
, struct zone
*zone
)
3819 set_zonelist_order();
3821 if (system_state
== SYSTEM_BOOTING
) {
3822 __build_all_zonelists(NULL
);
3823 mminit_verify_zonelist();
3824 cpuset_init_current_mems_allowed();
3826 #ifdef CONFIG_MEMORY_HOTPLUG
3828 setup_zone_pageset(zone
);
3830 /* we have to stop all cpus to guarantee there is no user
3832 stop_machine(__build_all_zonelists
, pgdat
, NULL
);
3833 /* cpuset refresh routine should be here */
3835 vm_total_pages
= nr_free_pagecache_pages();
3837 * Disable grouping by mobility if the number of pages in the
3838 * system is too low to allow the mechanism to work. It would be
3839 * more accurate, but expensive to check per-zone. This check is
3840 * made on memory-hotadd so a system can start with mobility
3841 * disabled and enable it later
3843 if (vm_total_pages
< (pageblock_nr_pages
* MIGRATE_TYPES
))
3844 page_group_by_mobility_disabled
= 1;
3846 page_group_by_mobility_disabled
= 0;
3848 printk("Built %i zonelists in %s order, mobility grouping %s. "
3849 "Total pages: %ld\n",
3851 zonelist_order_name
[current_zonelist_order
],
3852 page_group_by_mobility_disabled
? "off" : "on",
3855 printk("Policy zone: %s\n", zone_names
[policy_zone
]);
3860 * Helper functions to size the waitqueue hash table.
3861 * Essentially these want to choose hash table sizes sufficiently
3862 * large so that collisions trying to wait on pages are rare.
3863 * But in fact, the number of active page waitqueues on typical
3864 * systems is ridiculously low, less than 200. So this is even
3865 * conservative, even though it seems large.
3867 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3868 * waitqueues, i.e. the size of the waitq table given the number of pages.
3870 #define PAGES_PER_WAITQUEUE 256
3872 #ifndef CONFIG_MEMORY_HOTPLUG
3873 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3875 unsigned long size
= 1;
3877 pages
/= PAGES_PER_WAITQUEUE
;
3879 while (size
< pages
)
3883 * Once we have dozens or even hundreds of threads sleeping
3884 * on IO we've got bigger problems than wait queue collision.
3885 * Limit the size of the wait table to a reasonable size.
3887 size
= min(size
, 4096UL);
3889 return max(size
, 4UL);
3893 * A zone's size might be changed by hot-add, so it is not possible to determine
3894 * a suitable size for its wait_table. So we use the maximum size now.
3896 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3898 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3899 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3900 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3902 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3903 * or more by the traditional way. (See above). It equals:
3905 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3906 * ia64(16K page size) : = ( 8G + 4M)byte.
3907 * powerpc (64K page size) : = (32G +16M)byte.
3909 static inline unsigned long wait_table_hash_nr_entries(unsigned long pages
)
3916 * This is an integer logarithm so that shifts can be used later
3917 * to extract the more random high bits from the multiplicative
3918 * hash function before the remainder is taken.
3920 static inline unsigned long wait_table_bits(unsigned long size
)
3926 * Check if a pageblock contains reserved pages
3928 static int pageblock_is_reserved(unsigned long start_pfn
, unsigned long end_pfn
)
3932 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
3933 if (!pfn_valid_within(pfn
) || PageReserved(pfn_to_page(pfn
)))
3940 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3941 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3942 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3943 * higher will lead to a bigger reserve which will get freed as contiguous
3944 * blocks as reclaim kicks in
3946 static void setup_zone_migrate_reserve(struct zone
*zone
)
3948 unsigned long start_pfn
, pfn
, end_pfn
, block_end_pfn
;
3950 unsigned long block_migratetype
;
3955 * Get the start pfn, end pfn and the number of blocks to reserve
3956 * We have to be careful to be aligned to pageblock_nr_pages to
3957 * make sure that we always check pfn_valid for the first page in
3960 start_pfn
= zone
->zone_start_pfn
;
3961 end_pfn
= zone_end_pfn(zone
);
3962 start_pfn
= roundup(start_pfn
, pageblock_nr_pages
);
3963 reserve
= roundup(min_wmark_pages(zone
), pageblock_nr_pages
) >>
3967 * Reserve blocks are generally in place to help high-order atomic
3968 * allocations that are short-lived. A min_free_kbytes value that
3969 * would result in more than 2 reserve blocks for atomic allocations
3970 * is assumed to be in place to help anti-fragmentation for the
3971 * future allocation of hugepages at runtime.
3973 reserve
= min(2, reserve
);
3974 old_reserve
= zone
->nr_migrate_reserve_block
;
3976 /* When memory hot-add, we almost always need to do nothing */
3977 if (reserve
== old_reserve
)
3979 zone
->nr_migrate_reserve_block
= reserve
;
3981 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= pageblock_nr_pages
) {
3982 if (!pfn_valid(pfn
))
3984 page
= pfn_to_page(pfn
);
3986 /* Watch out for overlapping nodes */
3987 if (page_to_nid(page
) != zone_to_nid(zone
))
3990 block_migratetype
= get_pageblock_migratetype(page
);
3992 /* Only test what is necessary when the reserves are not met */
3995 * Blocks with reserved pages will never free, skip
3998 block_end_pfn
= min(pfn
+ pageblock_nr_pages
, end_pfn
);
3999 if (pageblock_is_reserved(pfn
, block_end_pfn
))
4002 /* If this block is reserved, account for it */
4003 if (block_migratetype
== MIGRATE_RESERVE
) {
4008 /* Suitable for reserving if this block is movable */
4009 if (block_migratetype
== MIGRATE_MOVABLE
) {
4010 set_pageblock_migratetype(page
,
4012 move_freepages_block(zone
, page
,
4017 } else if (!old_reserve
) {
4019 * At boot time we don't need to scan the whole zone
4020 * for turning off MIGRATE_RESERVE.
4026 * If the reserve is met and this is a previous reserved block,
4029 if (block_migratetype
== MIGRATE_RESERVE
) {
4030 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4031 move_freepages_block(zone
, page
, MIGRATE_MOVABLE
);
4037 * Initially all pages are reserved - free ones are freed
4038 * up by free_all_bootmem() once the early boot process is
4039 * done. Non-atomic initialization, single-pass.
4041 void __meminit
memmap_init_zone(unsigned long size
, int nid
, unsigned long zone
,
4042 unsigned long start_pfn
, enum memmap_context context
)
4045 unsigned long end_pfn
= start_pfn
+ size
;
4049 if (highest_memmap_pfn
< end_pfn
- 1)
4050 highest_memmap_pfn
= end_pfn
- 1;
4052 z
= &NODE_DATA(nid
)->node_zones
[zone
];
4053 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++) {
4055 * There can be holes in boot-time mem_map[]s
4056 * handed to this function. They do not
4057 * exist on hotplugged memory.
4059 if (context
== MEMMAP_EARLY
) {
4060 if (!early_pfn_valid(pfn
))
4062 if (!early_pfn_in_nid(pfn
, nid
))
4065 page
= pfn_to_page(pfn
);
4066 set_page_links(page
, zone
, nid
, pfn
);
4067 mminit_verify_page_links(page
, zone
, nid
, pfn
);
4068 init_page_count(page
);
4069 page_mapcount_reset(page
);
4070 page_cpupid_reset_last(page
);
4071 SetPageReserved(page
);
4073 * Mark the block movable so that blocks are reserved for
4074 * movable at startup. This will force kernel allocations
4075 * to reserve their blocks rather than leaking throughout
4076 * the address space during boot when many long-lived
4077 * kernel allocations are made. Later some blocks near
4078 * the start are marked MIGRATE_RESERVE by
4079 * setup_zone_migrate_reserve()
4081 * bitmap is created for zone's valid pfn range. but memmap
4082 * can be created for invalid pages (for alignment)
4083 * check here not to call set_pageblock_migratetype() against
4086 if ((z
->zone_start_pfn
<= pfn
)
4087 && (pfn
< zone_end_pfn(z
))
4088 && !(pfn
& (pageblock_nr_pages
- 1)))
4089 set_pageblock_migratetype(page
, MIGRATE_MOVABLE
);
4091 INIT_LIST_HEAD(&page
->lru
);
4092 #ifdef WANT_PAGE_VIRTUAL
4093 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
4094 if (!is_highmem_idx(zone
))
4095 set_page_address(page
, __va(pfn
<< PAGE_SHIFT
));
4100 static void __meminit
zone_init_free_lists(struct zone
*zone
)
4103 for_each_migratetype_order(order
, t
) {
4104 INIT_LIST_HEAD(&zone
->free_area
[order
].free_list
[t
]);
4105 zone
->free_area
[order
].nr_free
= 0;
4109 #ifndef __HAVE_ARCH_MEMMAP_INIT
4110 #define memmap_init(size, nid, zone, start_pfn) \
4111 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
4114 static int __meminit
zone_batchsize(struct zone
*zone
)
4120 * The per-cpu-pages pools are set to around 1000th of the
4121 * size of the zone. But no more than 1/2 of a meg.
4123 * OK, so we don't know how big the cache is. So guess.
4125 batch
= zone
->managed_pages
/ 1024;
4126 if (batch
* PAGE_SIZE
> 512 * 1024)
4127 batch
= (512 * 1024) / PAGE_SIZE
;
4128 batch
/= 4; /* We effectively *= 4 below */
4133 * Clamp the batch to a 2^n - 1 value. Having a power
4134 * of 2 value was found to be more likely to have
4135 * suboptimal cache aliasing properties in some cases.
4137 * For example if 2 tasks are alternately allocating
4138 * batches of pages, one task can end up with a lot
4139 * of pages of one half of the possible page colors
4140 * and the other with pages of the other colors.
4142 batch
= rounddown_pow_of_two(batch
+ batch
/2) - 1;
4147 /* The deferral and batching of frees should be suppressed under NOMMU
4150 * The problem is that NOMMU needs to be able to allocate large chunks
4151 * of contiguous memory as there's no hardware page translation to
4152 * assemble apparent contiguous memory from discontiguous pages.
4154 * Queueing large contiguous runs of pages for batching, however,
4155 * causes the pages to actually be freed in smaller chunks. As there
4156 * can be a significant delay between the individual batches being
4157 * recycled, this leads to the once large chunks of space being
4158 * fragmented and becoming unavailable for high-order allocations.
4165 * pcp->high and pcp->batch values are related and dependent on one another:
4166 * ->batch must never be higher then ->high.
4167 * The following function updates them in a safe manner without read side
4170 * Any new users of pcp->batch and pcp->high should ensure they can cope with
4171 * those fields changing asynchronously (acording the the above rule).
4173 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
4174 * outside of boot time (or some other assurance that no concurrent updaters
4177 static void pageset_update(struct per_cpu_pages
*pcp
, unsigned long high
,
4178 unsigned long batch
)
4180 /* start with a fail safe value for batch */
4184 /* Update high, then batch, in order */
4191 /* a companion to pageset_set_high() */
4192 static void pageset_set_batch(struct per_cpu_pageset
*p
, unsigned long batch
)
4194 pageset_update(&p
->pcp
, 6 * batch
, max(1UL, 1 * batch
));
4197 static void pageset_init(struct per_cpu_pageset
*p
)
4199 struct per_cpu_pages
*pcp
;
4202 memset(p
, 0, sizeof(*p
));
4206 for (migratetype
= 0; migratetype
< MIGRATE_PCPTYPES
; migratetype
++)
4207 INIT_LIST_HEAD(&pcp
->lists
[migratetype
]);
4210 static void setup_pageset(struct per_cpu_pageset
*p
, unsigned long batch
)
4213 pageset_set_batch(p
, batch
);
4217 * pageset_set_high() sets the high water mark for hot per_cpu_pagelist
4218 * to the value high for the pageset p.
4220 static void pageset_set_high(struct per_cpu_pageset
*p
,
4223 unsigned long batch
= max(1UL, high
/ 4);
4224 if ((high
/ 4) > (PAGE_SHIFT
* 8))
4225 batch
= PAGE_SHIFT
* 8;
4227 pageset_update(&p
->pcp
, high
, batch
);
4230 static void __meminit
pageset_set_high_and_batch(struct zone
*zone
,
4231 struct per_cpu_pageset
*pcp
)
4233 if (percpu_pagelist_fraction
)
4234 pageset_set_high(pcp
,
4235 (zone
->managed_pages
/
4236 percpu_pagelist_fraction
));
4238 pageset_set_batch(pcp
, zone_batchsize(zone
));
4241 static void __meminit
zone_pageset_init(struct zone
*zone
, int cpu
)
4243 struct per_cpu_pageset
*pcp
= per_cpu_ptr(zone
->pageset
, cpu
);
4246 pageset_set_high_and_batch(zone
, pcp
);
4249 static void __meminit
setup_zone_pageset(struct zone
*zone
)
4252 zone
->pageset
= alloc_percpu(struct per_cpu_pageset
);
4253 for_each_possible_cpu(cpu
)
4254 zone_pageset_init(zone
, cpu
);
4258 * Allocate per cpu pagesets and initialize them.
4259 * Before this call only boot pagesets were available.
4261 void __init
setup_per_cpu_pageset(void)
4265 for_each_populated_zone(zone
)
4266 setup_zone_pageset(zone
);
4269 static noinline __init_refok
4270 int zone_wait_table_init(struct zone
*zone
, unsigned long zone_size_pages
)
4276 * The per-page waitqueue mechanism uses hashed waitqueues
4279 zone
->wait_table_hash_nr_entries
=
4280 wait_table_hash_nr_entries(zone_size_pages
);
4281 zone
->wait_table_bits
=
4282 wait_table_bits(zone
->wait_table_hash_nr_entries
);
4283 alloc_size
= zone
->wait_table_hash_nr_entries
4284 * sizeof(wait_queue_head_t
);
4286 if (!slab_is_available()) {
4287 zone
->wait_table
= (wait_queue_head_t
*)
4288 memblock_virt_alloc_node_nopanic(
4289 alloc_size
, zone
->zone_pgdat
->node_id
);
4292 * This case means that a zone whose size was 0 gets new memory
4293 * via memory hot-add.
4294 * But it may be the case that a new node was hot-added. In
4295 * this case vmalloc() will not be able to use this new node's
4296 * memory - this wait_table must be initialized to use this new
4297 * node itself as well.
4298 * To use this new node's memory, further consideration will be
4301 zone
->wait_table
= vmalloc(alloc_size
);
4303 if (!zone
->wait_table
)
4306 for (i
= 0; i
< zone
->wait_table_hash_nr_entries
; ++i
)
4307 init_waitqueue_head(zone
->wait_table
+ i
);
4312 static __meminit
void zone_pcp_init(struct zone
*zone
)
4315 * per cpu subsystem is not up at this point. The following code
4316 * relies on the ability of the linker to provide the
4317 * offset of a (static) per cpu variable into the per cpu area.
4319 zone
->pageset
= &boot_pageset
;
4321 if (populated_zone(zone
))
4322 printk(KERN_DEBUG
" %s zone: %lu pages, LIFO batch:%u\n",
4323 zone
->name
, zone
->present_pages
,
4324 zone_batchsize(zone
));
4327 int __meminit
init_currently_empty_zone(struct zone
*zone
,
4328 unsigned long zone_start_pfn
,
4330 enum memmap_context context
)
4332 struct pglist_data
*pgdat
= zone
->zone_pgdat
;
4334 ret
= zone_wait_table_init(zone
, size
);
4337 pgdat
->nr_zones
= zone_idx(zone
) + 1;
4339 zone
->zone_start_pfn
= zone_start_pfn
;
4341 mminit_dprintk(MMINIT_TRACE
, "memmap_init",
4342 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
4344 (unsigned long)zone_idx(zone
),
4345 zone_start_pfn
, (zone_start_pfn
+ size
));
4347 zone_init_free_lists(zone
);
4352 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4353 #ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
4355 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
4356 * Architectures may implement their own version but if add_active_range()
4357 * was used and there are no special requirements, this is a convenient
4360 int __meminit
__early_pfn_to_nid(unsigned long pfn
)
4362 unsigned long start_pfn
, end_pfn
;
4365 * NOTE: The following SMP-unsafe globals are only used early in boot
4366 * when the kernel is running single-threaded.
4368 static unsigned long __meminitdata last_start_pfn
, last_end_pfn
;
4369 static int __meminitdata last_nid
;
4371 if (last_start_pfn
<= pfn
&& pfn
< last_end_pfn
)
4374 nid
= memblock_search_pfn_nid(pfn
, &start_pfn
, &end_pfn
);
4376 last_start_pfn
= start_pfn
;
4377 last_end_pfn
= end_pfn
;
4383 #endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
4385 int __meminit
early_pfn_to_nid(unsigned long pfn
)
4389 nid
= __early_pfn_to_nid(pfn
);
4392 /* just returns 0 */
4396 #ifdef CONFIG_NODES_SPAN_OTHER_NODES
4397 bool __meminit
early_pfn_in_nid(unsigned long pfn
, int node
)
4401 nid
= __early_pfn_to_nid(pfn
);
4402 if (nid
>= 0 && nid
!= node
)
4409 * free_bootmem_with_active_regions - Call memblock_free_early_nid for each active range
4410 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
4411 * @max_low_pfn: The highest PFN that will be passed to memblock_free_early_nid
4413 * If an architecture guarantees that all ranges registered with
4414 * add_active_ranges() contain no holes and may be freed, this
4415 * this function may be used instead of calling memblock_free_early_nid()
4418 void __init
free_bootmem_with_active_regions(int nid
, unsigned long max_low_pfn
)
4420 unsigned long start_pfn
, end_pfn
;
4423 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
) {
4424 start_pfn
= min(start_pfn
, max_low_pfn
);
4425 end_pfn
= min(end_pfn
, max_low_pfn
);
4427 if (start_pfn
< end_pfn
)
4428 memblock_free_early_nid(PFN_PHYS(start_pfn
),
4429 (end_pfn
- start_pfn
) << PAGE_SHIFT
,
4435 * sparse_memory_present_with_active_regions - Call memory_present for each active range
4436 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
4438 * If an architecture guarantees that all ranges registered with
4439 * add_active_ranges() contain no holes and may be freed, this
4440 * function may be used instead of calling memory_present() manually.
4442 void __init
sparse_memory_present_with_active_regions(int nid
)
4444 unsigned long start_pfn
, end_pfn
;
4447 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, &this_nid
)
4448 memory_present(this_nid
, start_pfn
, end_pfn
);
4452 * get_pfn_range_for_nid - Return the start and end page frames for a node
4453 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
4454 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
4455 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
4457 * It returns the start and end page frame of a node based on information
4458 * provided by an arch calling add_active_range(). If called for a node
4459 * with no available memory, a warning is printed and the start and end
4462 void __meminit
get_pfn_range_for_nid(unsigned int nid
,
4463 unsigned long *start_pfn
, unsigned long *end_pfn
)
4465 unsigned long this_start_pfn
, this_end_pfn
;
4471 for_each_mem_pfn_range(i
, nid
, &this_start_pfn
, &this_end_pfn
, NULL
) {
4472 *start_pfn
= min(*start_pfn
, this_start_pfn
);
4473 *end_pfn
= max(*end_pfn
, this_end_pfn
);
4476 if (*start_pfn
== -1UL)
4481 * This finds a zone that can be used for ZONE_MOVABLE pages. The
4482 * assumption is made that zones within a node are ordered in monotonic
4483 * increasing memory addresses so that the "highest" populated zone is used
4485 static void __init
find_usable_zone_for_movable(void)
4488 for (zone_index
= MAX_NR_ZONES
- 1; zone_index
>= 0; zone_index
--) {
4489 if (zone_index
== ZONE_MOVABLE
)
4492 if (arch_zone_highest_possible_pfn
[zone_index
] >
4493 arch_zone_lowest_possible_pfn
[zone_index
])
4497 VM_BUG_ON(zone_index
== -1);
4498 movable_zone
= zone_index
;
4502 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4503 * because it is sized independent of architecture. Unlike the other zones,
4504 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4505 * in each node depending on the size of each node and how evenly kernelcore
4506 * is distributed. This helper function adjusts the zone ranges
4507 * provided by the architecture for a given node by using the end of the
4508 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4509 * zones within a node are in order of monotonic increases memory addresses
4511 static void __meminit
adjust_zone_range_for_zone_movable(int nid
,
4512 unsigned long zone_type
,
4513 unsigned long node_start_pfn
,
4514 unsigned long node_end_pfn
,
4515 unsigned long *zone_start_pfn
,
4516 unsigned long *zone_end_pfn
)
4518 /* Only adjust if ZONE_MOVABLE is on this node */
4519 if (zone_movable_pfn
[nid
]) {
4520 /* Size ZONE_MOVABLE */
4521 if (zone_type
== ZONE_MOVABLE
) {
4522 *zone_start_pfn
= zone_movable_pfn
[nid
];
4523 *zone_end_pfn
= min(node_end_pfn
,
4524 arch_zone_highest_possible_pfn
[movable_zone
]);
4526 /* Adjust for ZONE_MOVABLE starting within this range */
4527 } else if (*zone_start_pfn
< zone_movable_pfn
[nid
] &&
4528 *zone_end_pfn
> zone_movable_pfn
[nid
]) {
4529 *zone_end_pfn
= zone_movable_pfn
[nid
];
4531 /* Check if this whole range is within ZONE_MOVABLE */
4532 } else if (*zone_start_pfn
>= zone_movable_pfn
[nid
])
4533 *zone_start_pfn
= *zone_end_pfn
;
4538 * Return the number of pages a zone spans in a node, including holes
4539 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4541 static unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4542 unsigned long zone_type
,
4543 unsigned long node_start_pfn
,
4544 unsigned long node_end_pfn
,
4545 unsigned long *ignored
)
4547 unsigned long zone_start_pfn
, zone_end_pfn
;
4549 /* Get the start and end of the zone */
4550 zone_start_pfn
= arch_zone_lowest_possible_pfn
[zone_type
];
4551 zone_end_pfn
= arch_zone_highest_possible_pfn
[zone_type
];
4552 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4553 node_start_pfn
, node_end_pfn
,
4554 &zone_start_pfn
, &zone_end_pfn
);
4556 /* Check that this node has pages within the zone's required range */
4557 if (zone_end_pfn
< node_start_pfn
|| zone_start_pfn
> node_end_pfn
)
4560 /* Move the zone boundaries inside the node if necessary */
4561 zone_end_pfn
= min(zone_end_pfn
, node_end_pfn
);
4562 zone_start_pfn
= max(zone_start_pfn
, node_start_pfn
);
4564 /* Return the spanned pages */
4565 return zone_end_pfn
- zone_start_pfn
;
4569 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4570 * then all holes in the requested range will be accounted for.
4572 unsigned long __meminit
__absent_pages_in_range(int nid
,
4573 unsigned long range_start_pfn
,
4574 unsigned long range_end_pfn
)
4576 unsigned long nr_absent
= range_end_pfn
- range_start_pfn
;
4577 unsigned long start_pfn
, end_pfn
;
4580 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
4581 start_pfn
= clamp(start_pfn
, range_start_pfn
, range_end_pfn
);
4582 end_pfn
= clamp(end_pfn
, range_start_pfn
, range_end_pfn
);
4583 nr_absent
-= end_pfn
- start_pfn
;
4589 * absent_pages_in_range - Return number of page frames in holes within a range
4590 * @start_pfn: The start PFN to start searching for holes
4591 * @end_pfn: The end PFN to stop searching for holes
4593 * It returns the number of pages frames in memory holes within a range.
4595 unsigned long __init
absent_pages_in_range(unsigned long start_pfn
,
4596 unsigned long end_pfn
)
4598 return __absent_pages_in_range(MAX_NUMNODES
, start_pfn
, end_pfn
);
4601 /* Return the number of page frames in holes in a zone on a node */
4602 static unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4603 unsigned long zone_type
,
4604 unsigned long node_start_pfn
,
4605 unsigned long node_end_pfn
,
4606 unsigned long *ignored
)
4608 unsigned long zone_low
= arch_zone_lowest_possible_pfn
[zone_type
];
4609 unsigned long zone_high
= arch_zone_highest_possible_pfn
[zone_type
];
4610 unsigned long zone_start_pfn
, zone_end_pfn
;
4612 zone_start_pfn
= clamp(node_start_pfn
, zone_low
, zone_high
);
4613 zone_end_pfn
= clamp(node_end_pfn
, zone_low
, zone_high
);
4615 adjust_zone_range_for_zone_movable(nid
, zone_type
,
4616 node_start_pfn
, node_end_pfn
,
4617 &zone_start_pfn
, &zone_end_pfn
);
4618 return __absent_pages_in_range(nid
, zone_start_pfn
, zone_end_pfn
);
4621 #else /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4622 static inline unsigned long __meminit
zone_spanned_pages_in_node(int nid
,
4623 unsigned long zone_type
,
4624 unsigned long node_start_pfn
,
4625 unsigned long node_end_pfn
,
4626 unsigned long *zones_size
)
4628 return zones_size
[zone_type
];
4631 static inline unsigned long __meminit
zone_absent_pages_in_node(int nid
,
4632 unsigned long zone_type
,
4633 unsigned long node_start_pfn
,
4634 unsigned long node_end_pfn
,
4635 unsigned long *zholes_size
)
4640 return zholes_size
[zone_type
];
4643 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4645 static void __meminit
calculate_node_totalpages(struct pglist_data
*pgdat
,
4646 unsigned long node_start_pfn
,
4647 unsigned long node_end_pfn
,
4648 unsigned long *zones_size
,
4649 unsigned long *zholes_size
)
4651 unsigned long realtotalpages
, totalpages
= 0;
4654 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4655 totalpages
+= zone_spanned_pages_in_node(pgdat
->node_id
, i
,
4659 pgdat
->node_spanned_pages
= totalpages
;
4661 realtotalpages
= totalpages
;
4662 for (i
= 0; i
< MAX_NR_ZONES
; i
++)
4664 zone_absent_pages_in_node(pgdat
->node_id
, i
,
4665 node_start_pfn
, node_end_pfn
,
4667 pgdat
->node_present_pages
= realtotalpages
;
4668 printk(KERN_DEBUG
"On node %d totalpages: %lu\n", pgdat
->node_id
,
4672 #ifndef CONFIG_SPARSEMEM
4674 * Calculate the size of the zone->blockflags rounded to an unsigned long
4675 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4676 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4677 * round what is now in bits to nearest long in bits, then return it in
4680 static unsigned long __init
usemap_size(unsigned long zone_start_pfn
, unsigned long zonesize
)
4682 unsigned long usemapsize
;
4684 zonesize
+= zone_start_pfn
& (pageblock_nr_pages
-1);
4685 usemapsize
= roundup(zonesize
, pageblock_nr_pages
);
4686 usemapsize
= usemapsize
>> pageblock_order
;
4687 usemapsize
*= NR_PAGEBLOCK_BITS
;
4688 usemapsize
= roundup(usemapsize
, 8 * sizeof(unsigned long));
4690 return usemapsize
/ 8;
4693 static void __init
setup_usemap(struct pglist_data
*pgdat
,
4695 unsigned long zone_start_pfn
,
4696 unsigned long zonesize
)
4698 unsigned long usemapsize
= usemap_size(zone_start_pfn
, zonesize
);
4699 zone
->pageblock_flags
= NULL
;
4701 zone
->pageblock_flags
=
4702 memblock_virt_alloc_node_nopanic(usemapsize
,
4706 static inline void setup_usemap(struct pglist_data
*pgdat
, struct zone
*zone
,
4707 unsigned long zone_start_pfn
, unsigned long zonesize
) {}
4708 #endif /* CONFIG_SPARSEMEM */
4710 #ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4712 /* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4713 void __paginginit
set_pageblock_order(void)
4717 /* Check that pageblock_nr_pages has not already been setup */
4718 if (pageblock_order
)
4721 if (HPAGE_SHIFT
> PAGE_SHIFT
)
4722 order
= HUGETLB_PAGE_ORDER
;
4724 order
= MAX_ORDER
- 1;
4727 * Assume the largest contiguous order of interest is a huge page.
4728 * This value may be variable depending on boot parameters on IA64 and
4731 pageblock_order
= order
;
4733 #else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4736 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4737 * is unused as pageblock_order is set at compile-time. See
4738 * include/linux/pageblock-flags.h for the values of pageblock_order based on
4741 void __paginginit
set_pageblock_order(void)
4745 #endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4747 static unsigned long __paginginit
calc_memmap_size(unsigned long spanned_pages
,
4748 unsigned long present_pages
)
4750 unsigned long pages
= spanned_pages
;
4753 * Provide a more accurate estimation if there are holes within
4754 * the zone and SPARSEMEM is in use. If there are holes within the
4755 * zone, each populated memory region may cost us one or two extra
4756 * memmap pages due to alignment because memmap pages for each
4757 * populated regions may not naturally algined on page boundary.
4758 * So the (present_pages >> 4) heuristic is a tradeoff for that.
4760 if (spanned_pages
> present_pages
+ (present_pages
>> 4) &&
4761 IS_ENABLED(CONFIG_SPARSEMEM
))
4762 pages
= present_pages
;
4764 return PAGE_ALIGN(pages
* sizeof(struct page
)) >> PAGE_SHIFT
;
4768 * Set up the zone data structures:
4769 * - mark all pages reserved
4770 * - mark all memory queues empty
4771 * - clear the memory bitmaps
4773 * NOTE: pgdat should get zeroed by caller.
4775 static void __paginginit
free_area_init_core(struct pglist_data
*pgdat
,
4776 unsigned long node_start_pfn
, unsigned long node_end_pfn
,
4777 unsigned long *zones_size
, unsigned long *zholes_size
)
4780 int nid
= pgdat
->node_id
;
4781 unsigned long zone_start_pfn
= pgdat
->node_start_pfn
;
4784 pgdat_resize_init(pgdat
);
4785 #ifdef CONFIG_NUMA_BALANCING
4786 spin_lock_init(&pgdat
->numabalancing_migrate_lock
);
4787 pgdat
->numabalancing_migrate_nr_pages
= 0;
4788 pgdat
->numabalancing_migrate_next_window
= jiffies
;
4790 init_waitqueue_head(&pgdat
->kswapd_wait
);
4791 init_waitqueue_head(&pgdat
->pfmemalloc_wait
);
4792 pgdat_page_cgroup_init(pgdat
);
4794 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
4795 struct zone
*zone
= pgdat
->node_zones
+ j
;
4796 unsigned long size
, realsize
, freesize
, memmap_pages
;
4798 size
= zone_spanned_pages_in_node(nid
, j
, node_start_pfn
,
4799 node_end_pfn
, zones_size
);
4800 realsize
= freesize
= size
- zone_absent_pages_in_node(nid
, j
,
4806 * Adjust freesize so that it accounts for how much memory
4807 * is used by this zone for memmap. This affects the watermark
4808 * and per-cpu initialisations
4810 memmap_pages
= calc_memmap_size(size
, realsize
);
4811 if (freesize
>= memmap_pages
) {
4812 freesize
-= memmap_pages
;
4815 " %s zone: %lu pages used for memmap\n",
4816 zone_names
[j
], memmap_pages
);
4819 " %s zone: %lu pages exceeds freesize %lu\n",
4820 zone_names
[j
], memmap_pages
, freesize
);
4822 /* Account for reserved pages */
4823 if (j
== 0 && freesize
> dma_reserve
) {
4824 freesize
-= dma_reserve
;
4825 printk(KERN_DEBUG
" %s zone: %lu pages reserved\n",
4826 zone_names
[0], dma_reserve
);
4829 if (!is_highmem_idx(j
))
4830 nr_kernel_pages
+= freesize
;
4831 /* Charge for highmem memmap if there are enough kernel pages */
4832 else if (nr_kernel_pages
> memmap_pages
* 2)
4833 nr_kernel_pages
-= memmap_pages
;
4834 nr_all_pages
+= freesize
;
4836 zone
->spanned_pages
= size
;
4837 zone
->present_pages
= realsize
;
4839 * Set an approximate value for lowmem here, it will be adjusted
4840 * when the bootmem allocator frees pages into the buddy system.
4841 * And all highmem pages will be managed by the buddy system.
4843 zone
->managed_pages
= is_highmem_idx(j
) ? realsize
: freesize
;
4846 zone
->min_unmapped_pages
= (freesize
*sysctl_min_unmapped_ratio
)
4848 zone
->min_slab_pages
= (freesize
* sysctl_min_slab_ratio
) / 100;
4850 zone
->name
= zone_names
[j
];
4851 spin_lock_init(&zone
->lock
);
4852 spin_lock_init(&zone
->lru_lock
);
4853 zone_seqlock_init(zone
);
4854 zone
->zone_pgdat
= pgdat
;
4855 zone_pcp_init(zone
);
4857 /* For bootup, initialized properly in watermark setup */
4858 mod_zone_page_state(zone
, NR_ALLOC_BATCH
, zone
->managed_pages
);
4860 lruvec_init(&zone
->lruvec
);
4864 set_pageblock_order();
4865 setup_usemap(pgdat
, zone
, zone_start_pfn
, size
);
4866 ret
= init_currently_empty_zone(zone
, zone_start_pfn
,
4867 size
, MEMMAP_EARLY
);
4869 memmap_init(size
, nid
, j
, zone_start_pfn
);
4870 zone_start_pfn
+= size
;
4874 static void __init_refok
alloc_node_mem_map(struct pglist_data
*pgdat
)
4876 /* Skip empty nodes */
4877 if (!pgdat
->node_spanned_pages
)
4880 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4881 /* ia64 gets its own node_mem_map, before this, without bootmem */
4882 if (!pgdat
->node_mem_map
) {
4883 unsigned long size
, start
, end
;
4887 * The zone's endpoints aren't required to be MAX_ORDER
4888 * aligned but the node_mem_map endpoints must be in order
4889 * for the buddy allocator to function correctly.
4891 start
= pgdat
->node_start_pfn
& ~(MAX_ORDER_NR_PAGES
- 1);
4892 end
= pgdat_end_pfn(pgdat
);
4893 end
= ALIGN(end
, MAX_ORDER_NR_PAGES
);
4894 size
= (end
- start
) * sizeof(struct page
);
4895 map
= alloc_remap(pgdat
->node_id
, size
);
4897 map
= memblock_virt_alloc_node_nopanic(size
,
4899 pgdat
->node_mem_map
= map
+ (pgdat
->node_start_pfn
- start
);
4901 #ifndef CONFIG_NEED_MULTIPLE_NODES
4903 * With no DISCONTIG, the global mem_map is just set as node 0's
4905 if (pgdat
== NODE_DATA(0)) {
4906 mem_map
= NODE_DATA(0)->node_mem_map
;
4907 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4908 if (page_to_pfn(mem_map
) != pgdat
->node_start_pfn
)
4909 mem_map
-= (pgdat
->node_start_pfn
- ARCH_PFN_OFFSET
);
4910 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
4913 #endif /* CONFIG_FLAT_NODE_MEM_MAP */
4916 void __paginginit
free_area_init_node(int nid
, unsigned long *zones_size
,
4917 unsigned long node_start_pfn
, unsigned long *zholes_size
)
4919 pg_data_t
*pgdat
= NODE_DATA(nid
);
4920 unsigned long start_pfn
= 0;
4921 unsigned long end_pfn
= 0;
4923 /* pg_data_t should be reset to zero when it's allocated */
4924 WARN_ON(pgdat
->nr_zones
|| pgdat
->classzone_idx
);
4926 pgdat
->node_id
= nid
;
4927 pgdat
->node_start_pfn
= node_start_pfn
;
4928 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4929 get_pfn_range_for_nid(nid
, &start_pfn
, &end_pfn
);
4931 calculate_node_totalpages(pgdat
, start_pfn
, end_pfn
,
4932 zones_size
, zholes_size
);
4934 alloc_node_mem_map(pgdat
);
4935 #ifdef CONFIG_FLAT_NODE_MEM_MAP
4936 printk(KERN_DEBUG
"free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4937 nid
, (unsigned long)pgdat
,
4938 (unsigned long)pgdat
->node_mem_map
);
4941 free_area_init_core(pgdat
, start_pfn
, end_pfn
,
4942 zones_size
, zholes_size
);
4945 #ifdef CONFIG_HAVE_MEMBLOCK_NODE_MAP
4947 #if MAX_NUMNODES > 1
4949 * Figure out the number of possible node ids.
4951 void __init
setup_nr_node_ids(void)
4954 unsigned int highest
= 0;
4956 for_each_node_mask(node
, node_possible_map
)
4958 nr_node_ids
= highest
+ 1;
4963 * node_map_pfn_alignment - determine the maximum internode alignment
4965 * This function should be called after node map is populated and sorted.
4966 * It calculates the maximum power of two alignment which can distinguish
4969 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4970 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4971 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4972 * shifted, 1GiB is enough and this function will indicate so.
4974 * This is used to test whether pfn -> nid mapping of the chosen memory
4975 * model has fine enough granularity to avoid incorrect mapping for the
4976 * populated node map.
4978 * Returns the determined alignment in pfn's. 0 if there is no alignment
4979 * requirement (single node).
4981 unsigned long __init
node_map_pfn_alignment(void)
4983 unsigned long accl_mask
= 0, last_end
= 0;
4984 unsigned long start
, end
, mask
;
4988 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start
, &end
, &nid
) {
4989 if (!start
|| last_nid
< 0 || last_nid
== nid
) {
4996 * Start with a mask granular enough to pin-point to the
4997 * start pfn and tick off bits one-by-one until it becomes
4998 * too coarse to separate the current node from the last.
5000 mask
= ~((1 << __ffs(start
)) - 1);
5001 while (mask
&& last_end
<= (start
& (mask
<< 1)))
5004 /* accumulate all internode masks */
5008 /* convert mask to number of pages */
5009 return ~accl_mask
+ 1;
5012 /* Find the lowest pfn for a node */
5013 static unsigned long __init
find_min_pfn_for_node(int nid
)
5015 unsigned long min_pfn
= ULONG_MAX
;
5016 unsigned long start_pfn
;
5019 for_each_mem_pfn_range(i
, nid
, &start_pfn
, NULL
, NULL
)
5020 min_pfn
= min(min_pfn
, start_pfn
);
5022 if (min_pfn
== ULONG_MAX
) {
5024 "Could not find start_pfn for node %d\n", nid
);
5032 * find_min_pfn_with_active_regions - Find the minimum PFN registered
5034 * It returns the minimum PFN based on information provided via
5035 * add_active_range().
5037 unsigned long __init
find_min_pfn_with_active_regions(void)
5039 return find_min_pfn_for_node(MAX_NUMNODES
);
5043 * early_calculate_totalpages()
5044 * Sum pages in active regions for movable zone.
5045 * Populate N_MEMORY for calculating usable_nodes.
5047 static unsigned long __init
early_calculate_totalpages(void)
5049 unsigned long totalpages
= 0;
5050 unsigned long start_pfn
, end_pfn
;
5053 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
) {
5054 unsigned long pages
= end_pfn
- start_pfn
;
5056 totalpages
+= pages
;
5058 node_set_state(nid
, N_MEMORY
);
5064 * Find the PFN the Movable zone begins in each node. Kernel memory
5065 * is spread evenly between nodes as long as the nodes have enough
5066 * memory. When they don't, some nodes will have more kernelcore than
5069 static void __init
find_zone_movable_pfns_for_nodes(void)
5072 unsigned long usable_startpfn
;
5073 unsigned long kernelcore_node
, kernelcore_remaining
;
5074 /* save the state before borrow the nodemask */
5075 nodemask_t saved_node_state
= node_states
[N_MEMORY
];
5076 unsigned long totalpages
= early_calculate_totalpages();
5077 int usable_nodes
= nodes_weight(node_states
[N_MEMORY
]);
5078 struct memblock_region
*r
;
5080 /* Need to find movable_zone earlier when movable_node is specified. */
5081 find_usable_zone_for_movable();
5084 * If movable_node is specified, ignore kernelcore and movablecore
5087 if (movable_node_is_enabled()) {
5088 for_each_memblock(memory
, r
) {
5089 if (!memblock_is_hotpluggable(r
))
5094 usable_startpfn
= PFN_DOWN(r
->base
);
5095 zone_movable_pfn
[nid
] = zone_movable_pfn
[nid
] ?
5096 min(usable_startpfn
, zone_movable_pfn
[nid
]) :
5104 * If movablecore=nn[KMG] was specified, calculate what size of
5105 * kernelcore that corresponds so that memory usable for
5106 * any allocation type is evenly spread. If both kernelcore
5107 * and movablecore are specified, then the value of kernelcore
5108 * will be used for required_kernelcore if it's greater than
5109 * what movablecore would have allowed.
5111 if (required_movablecore
) {
5112 unsigned long corepages
;
5115 * Round-up so that ZONE_MOVABLE is at least as large as what
5116 * was requested by the user
5118 required_movablecore
=
5119 roundup(required_movablecore
, MAX_ORDER_NR_PAGES
);
5120 corepages
= totalpages
- required_movablecore
;
5122 required_kernelcore
= max(required_kernelcore
, corepages
);
5125 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
5126 if (!required_kernelcore
)
5129 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
5130 usable_startpfn
= arch_zone_lowest_possible_pfn
[movable_zone
];
5133 /* Spread kernelcore memory as evenly as possible throughout nodes */
5134 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5135 for_each_node_state(nid
, N_MEMORY
) {
5136 unsigned long start_pfn
, end_pfn
;
5139 * Recalculate kernelcore_node if the division per node
5140 * now exceeds what is necessary to satisfy the requested
5141 * amount of memory for the kernel
5143 if (required_kernelcore
< kernelcore_node
)
5144 kernelcore_node
= required_kernelcore
/ usable_nodes
;
5147 * As the map is walked, we track how much memory is usable
5148 * by the kernel using kernelcore_remaining. When it is
5149 * 0, the rest of the node is usable by ZONE_MOVABLE
5151 kernelcore_remaining
= kernelcore_node
;
5153 /* Go through each range of PFNs within this node */
5154 for_each_mem_pfn_range(i
, nid
, &start_pfn
, &end_pfn
, NULL
) {
5155 unsigned long size_pages
;
5157 start_pfn
= max(start_pfn
, zone_movable_pfn
[nid
]);
5158 if (start_pfn
>= end_pfn
)
5161 /* Account for what is only usable for kernelcore */
5162 if (start_pfn
< usable_startpfn
) {
5163 unsigned long kernel_pages
;
5164 kernel_pages
= min(end_pfn
, usable_startpfn
)
5167 kernelcore_remaining
-= min(kernel_pages
,
5168 kernelcore_remaining
);
5169 required_kernelcore
-= min(kernel_pages
,
5170 required_kernelcore
);
5172 /* Continue if range is now fully accounted */
5173 if (end_pfn
<= usable_startpfn
) {
5176 * Push zone_movable_pfn to the end so
5177 * that if we have to rebalance
5178 * kernelcore across nodes, we will
5179 * not double account here
5181 zone_movable_pfn
[nid
] = end_pfn
;
5184 start_pfn
= usable_startpfn
;
5188 * The usable PFN range for ZONE_MOVABLE is from
5189 * start_pfn->end_pfn. Calculate size_pages as the
5190 * number of pages used as kernelcore
5192 size_pages
= end_pfn
- start_pfn
;
5193 if (size_pages
> kernelcore_remaining
)
5194 size_pages
= kernelcore_remaining
;
5195 zone_movable_pfn
[nid
] = start_pfn
+ size_pages
;
5198 * Some kernelcore has been met, update counts and
5199 * break if the kernelcore for this node has been
5202 required_kernelcore
-= min(required_kernelcore
,
5204 kernelcore_remaining
-= size_pages
;
5205 if (!kernelcore_remaining
)
5211 * If there is still required_kernelcore, we do another pass with one
5212 * less node in the count. This will push zone_movable_pfn[nid] further
5213 * along on the nodes that still have memory until kernelcore is
5217 if (usable_nodes
&& required_kernelcore
> usable_nodes
)
5221 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
5222 for (nid
= 0; nid
< MAX_NUMNODES
; nid
++)
5223 zone_movable_pfn
[nid
] =
5224 roundup(zone_movable_pfn
[nid
], MAX_ORDER_NR_PAGES
);
5227 /* restore the node_state */
5228 node_states
[N_MEMORY
] = saved_node_state
;
5231 /* Any regular or high memory on that node ? */
5232 static void check_for_memory(pg_data_t
*pgdat
, int nid
)
5234 enum zone_type zone_type
;
5236 if (N_MEMORY
== N_NORMAL_MEMORY
)
5239 for (zone_type
= 0; zone_type
<= ZONE_MOVABLE
- 1; zone_type
++) {
5240 struct zone
*zone
= &pgdat
->node_zones
[zone_type
];
5241 if (populated_zone(zone
)) {
5242 node_set_state(nid
, N_HIGH_MEMORY
);
5243 if (N_NORMAL_MEMORY
!= N_HIGH_MEMORY
&&
5244 zone_type
<= ZONE_NORMAL
)
5245 node_set_state(nid
, N_NORMAL_MEMORY
);
5252 * free_area_init_nodes - Initialise all pg_data_t and zone data
5253 * @max_zone_pfn: an array of max PFNs for each zone
5255 * This will call free_area_init_node() for each active node in the system.
5256 * Using the page ranges provided by add_active_range(), the size of each
5257 * zone in each node and their holes is calculated. If the maximum PFN
5258 * between two adjacent zones match, it is assumed that the zone is empty.
5259 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
5260 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
5261 * starts where the previous one ended. For example, ZONE_DMA32 starts
5262 * at arch_max_dma_pfn.
5264 void __init
free_area_init_nodes(unsigned long *max_zone_pfn
)
5266 unsigned long start_pfn
, end_pfn
;
5269 /* Record where the zone boundaries are */
5270 memset(arch_zone_lowest_possible_pfn
, 0,
5271 sizeof(arch_zone_lowest_possible_pfn
));
5272 memset(arch_zone_highest_possible_pfn
, 0,
5273 sizeof(arch_zone_highest_possible_pfn
));
5274 arch_zone_lowest_possible_pfn
[0] = find_min_pfn_with_active_regions();
5275 arch_zone_highest_possible_pfn
[0] = max_zone_pfn
[0];
5276 for (i
= 1; i
< MAX_NR_ZONES
; i
++) {
5277 if (i
== ZONE_MOVABLE
)
5279 arch_zone_lowest_possible_pfn
[i
] =
5280 arch_zone_highest_possible_pfn
[i
-1];
5281 arch_zone_highest_possible_pfn
[i
] =
5282 max(max_zone_pfn
[i
], arch_zone_lowest_possible_pfn
[i
]);
5284 arch_zone_lowest_possible_pfn
[ZONE_MOVABLE
] = 0;
5285 arch_zone_highest_possible_pfn
[ZONE_MOVABLE
] = 0;
5287 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
5288 memset(zone_movable_pfn
, 0, sizeof(zone_movable_pfn
));
5289 find_zone_movable_pfns_for_nodes();
5291 /* Print out the zone ranges */
5292 printk("Zone ranges:\n");
5293 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5294 if (i
== ZONE_MOVABLE
)
5296 printk(KERN_CONT
" %-8s ", zone_names
[i
]);
5297 if (arch_zone_lowest_possible_pfn
[i
] ==
5298 arch_zone_highest_possible_pfn
[i
])
5299 printk(KERN_CONT
"empty\n");
5301 printk(KERN_CONT
"[mem %0#10lx-%0#10lx]\n",
5302 arch_zone_lowest_possible_pfn
[i
] << PAGE_SHIFT
,
5303 (arch_zone_highest_possible_pfn
[i
]
5304 << PAGE_SHIFT
) - 1);
5307 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
5308 printk("Movable zone start for each node\n");
5309 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
5310 if (zone_movable_pfn
[i
])
5311 printk(" Node %d: %#010lx\n", i
,
5312 zone_movable_pfn
[i
] << PAGE_SHIFT
);
5315 /* Print out the early node map */
5316 printk("Early memory node ranges\n");
5317 for_each_mem_pfn_range(i
, MAX_NUMNODES
, &start_pfn
, &end_pfn
, &nid
)
5318 printk(" node %3d: [mem %#010lx-%#010lx]\n", nid
,
5319 start_pfn
<< PAGE_SHIFT
, (end_pfn
<< PAGE_SHIFT
) - 1);
5321 /* Initialise every node */
5322 mminit_verify_pageflags_layout();
5323 setup_nr_node_ids();
5324 for_each_online_node(nid
) {
5325 pg_data_t
*pgdat
= NODE_DATA(nid
);
5326 free_area_init_node(nid
, NULL
,
5327 find_min_pfn_for_node(nid
), NULL
);
5329 /* Any memory on that node */
5330 if (pgdat
->node_present_pages
)
5331 node_set_state(nid
, N_MEMORY
);
5332 check_for_memory(pgdat
, nid
);
5336 static int __init
cmdline_parse_core(char *p
, unsigned long *core
)
5338 unsigned long long coremem
;
5342 coremem
= memparse(p
, &p
);
5343 *core
= coremem
>> PAGE_SHIFT
;
5345 /* Paranoid check that UL is enough for the coremem value */
5346 WARN_ON((coremem
>> PAGE_SHIFT
) > ULONG_MAX
);
5352 * kernelcore=size sets the amount of memory for use for allocations that
5353 * cannot be reclaimed or migrated.
5355 static int __init
cmdline_parse_kernelcore(char *p
)
5357 return cmdline_parse_core(p
, &required_kernelcore
);
5361 * movablecore=size sets the amount of memory for use for allocations that
5362 * can be reclaimed or migrated.
5364 static int __init
cmdline_parse_movablecore(char *p
)
5366 return cmdline_parse_core(p
, &required_movablecore
);
5369 early_param("kernelcore", cmdline_parse_kernelcore
);
5370 early_param("movablecore", cmdline_parse_movablecore
);
5372 #endif /* CONFIG_HAVE_MEMBLOCK_NODE_MAP */
5374 void adjust_managed_page_count(struct page
*page
, long count
)
5376 spin_lock(&managed_page_count_lock
);
5377 page_zone(page
)->managed_pages
+= count
;
5378 totalram_pages
+= count
;
5379 #ifdef CONFIG_HIGHMEM
5380 if (PageHighMem(page
))
5381 totalhigh_pages
+= count
;
5383 spin_unlock(&managed_page_count_lock
);
5385 EXPORT_SYMBOL(adjust_managed_page_count
);
5387 unsigned long free_reserved_area(void *start
, void *end
, int poison
, char *s
)
5390 unsigned long pages
= 0;
5392 start
= (void *)PAGE_ALIGN((unsigned long)start
);
5393 end
= (void *)((unsigned long)end
& PAGE_MASK
);
5394 for (pos
= start
; pos
< end
; pos
+= PAGE_SIZE
, pages
++) {
5395 if ((unsigned int)poison
<= 0xFF)
5396 memset(pos
, poison
, PAGE_SIZE
);
5397 free_reserved_page(virt_to_page(pos
));
5401 pr_info("Freeing %s memory: %ldK (%p - %p)\n",
5402 s
, pages
<< (PAGE_SHIFT
- 10), start
, end
);
5406 EXPORT_SYMBOL(free_reserved_area
);
5408 #ifdef CONFIG_HIGHMEM
5409 void free_highmem_page(struct page
*page
)
5411 __free_reserved_page(page
);
5413 page_zone(page
)->managed_pages
++;
5419 void __init
mem_init_print_info(const char *str
)
5421 unsigned long physpages
, codesize
, datasize
, rosize
, bss_size
;
5422 unsigned long init_code_size
, init_data_size
;
5424 physpages
= get_num_physpages();
5425 codesize
= _etext
- _stext
;
5426 datasize
= _edata
- _sdata
;
5427 rosize
= __end_rodata
- __start_rodata
;
5428 bss_size
= __bss_stop
- __bss_start
;
5429 init_data_size
= __init_end
- __init_begin
;
5430 init_code_size
= _einittext
- _sinittext
;
5433 * Detect special cases and adjust section sizes accordingly:
5434 * 1) .init.* may be embedded into .data sections
5435 * 2) .init.text.* may be out of [__init_begin, __init_end],
5436 * please refer to arch/tile/kernel/vmlinux.lds.S.
5437 * 3) .rodata.* may be embedded into .text or .data sections.
5439 #define adj_init_size(start, end, size, pos, adj) \
5441 if (start <= pos && pos < end && size > adj) \
5445 adj_init_size(__init_begin
, __init_end
, init_data_size
,
5446 _sinittext
, init_code_size
);
5447 adj_init_size(_stext
, _etext
, codesize
, _sinittext
, init_code_size
);
5448 adj_init_size(_sdata
, _edata
, datasize
, __init_begin
, init_data_size
);
5449 adj_init_size(_stext
, _etext
, codesize
, __start_rodata
, rosize
);
5450 adj_init_size(_sdata
, _edata
, datasize
, __start_rodata
, rosize
);
5452 #undef adj_init_size
5454 printk("Memory: %luK/%luK available "
5455 "(%luK kernel code, %luK rwdata, %luK rodata, "
5456 "%luK init, %luK bss, %luK reserved"
5457 #ifdef CONFIG_HIGHMEM
5461 nr_free_pages() << (PAGE_SHIFT
-10), physpages
<< (PAGE_SHIFT
-10),
5462 codesize
>> 10, datasize
>> 10, rosize
>> 10,
5463 (init_data_size
+ init_code_size
) >> 10, bss_size
>> 10,
5464 (physpages
- totalram_pages
) << (PAGE_SHIFT
-10),
5465 #ifdef CONFIG_HIGHMEM
5466 totalhigh_pages
<< (PAGE_SHIFT
-10),
5468 str
? ", " : "", str
? str
: "");
5472 * set_dma_reserve - set the specified number of pages reserved in the first zone
5473 * @new_dma_reserve: The number of pages to mark reserved
5475 * The per-cpu batchsize and zone watermarks are determined by present_pages.
5476 * In the DMA zone, a significant percentage may be consumed by kernel image
5477 * and other unfreeable allocations which can skew the watermarks badly. This
5478 * function may optionally be used to account for unfreeable pages in the
5479 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
5480 * smaller per-cpu batchsize.
5482 void __init
set_dma_reserve(unsigned long new_dma_reserve
)
5484 dma_reserve
= new_dma_reserve
;
5487 void __init
free_area_init(unsigned long *zones_size
)
5489 free_area_init_node(0, zones_size
,
5490 __pa(PAGE_OFFSET
) >> PAGE_SHIFT
, NULL
);
5493 static int page_alloc_cpu_notify(struct notifier_block
*self
,
5494 unsigned long action
, void *hcpu
)
5496 int cpu
= (unsigned long)hcpu
;
5498 if (action
== CPU_DEAD
|| action
== CPU_DEAD_FROZEN
) {
5499 lru_add_drain_cpu(cpu
);
5503 * Spill the event counters of the dead processor
5504 * into the current processors event counters.
5505 * This artificially elevates the count of the current
5508 vm_events_fold_cpu(cpu
);
5511 * Zero the differential counters of the dead processor
5512 * so that the vm statistics are consistent.
5514 * This is only okay since the processor is dead and cannot
5515 * race with what we are doing.
5517 cpu_vm_stats_fold(cpu
);
5522 void __init
page_alloc_init(void)
5524 hotcpu_notifier(page_alloc_cpu_notify
, 0);
5528 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5529 * or min_free_kbytes changes.
5531 static void calculate_totalreserve_pages(void)
5533 struct pglist_data
*pgdat
;
5534 unsigned long reserve_pages
= 0;
5535 enum zone_type i
, j
;
5537 for_each_online_pgdat(pgdat
) {
5538 for (i
= 0; i
< MAX_NR_ZONES
; i
++) {
5539 struct zone
*zone
= pgdat
->node_zones
+ i
;
5540 unsigned long max
= 0;
5542 /* Find valid and maximum lowmem_reserve in the zone */
5543 for (j
= i
; j
< MAX_NR_ZONES
; j
++) {
5544 if (zone
->lowmem_reserve
[j
] > max
)
5545 max
= zone
->lowmem_reserve
[j
];
5548 /* we treat the high watermark as reserved pages. */
5549 max
+= high_wmark_pages(zone
);
5551 if (max
> zone
->managed_pages
)
5552 max
= zone
->managed_pages
;
5553 reserve_pages
+= max
;
5555 * Lowmem reserves are not available to
5556 * GFP_HIGHUSER page cache allocations and
5557 * kswapd tries to balance zones to their high
5558 * watermark. As a result, neither should be
5559 * regarded as dirtyable memory, to prevent a
5560 * situation where reclaim has to clean pages
5561 * in order to balance the zones.
5563 zone
->dirty_balance_reserve
= max
;
5566 dirty_balance_reserve
= reserve_pages
;
5567 totalreserve_pages
= reserve_pages
;
5571 * setup_per_zone_lowmem_reserve - called whenever
5572 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5573 * has a correct pages reserved value, so an adequate number of
5574 * pages are left in the zone after a successful __alloc_pages().
5576 static void setup_per_zone_lowmem_reserve(void)
5578 struct pglist_data
*pgdat
;
5579 enum zone_type j
, idx
;
5581 for_each_online_pgdat(pgdat
) {
5582 for (j
= 0; j
< MAX_NR_ZONES
; j
++) {
5583 struct zone
*zone
= pgdat
->node_zones
+ j
;
5584 unsigned long managed_pages
= zone
->managed_pages
;
5586 zone
->lowmem_reserve
[j
] = 0;
5590 struct zone
*lower_zone
;
5594 if (sysctl_lowmem_reserve_ratio
[idx
] < 1)
5595 sysctl_lowmem_reserve_ratio
[idx
] = 1;
5597 lower_zone
= pgdat
->node_zones
+ idx
;
5598 lower_zone
->lowmem_reserve
[j
] = managed_pages
/
5599 sysctl_lowmem_reserve_ratio
[idx
];
5600 managed_pages
+= lower_zone
->managed_pages
;
5605 /* update totalreserve_pages */
5606 calculate_totalreserve_pages();
5609 static void __setup_per_zone_wmarks(void)
5611 unsigned long pages_min
= min_free_kbytes
>> (PAGE_SHIFT
- 10);
5612 unsigned long lowmem_pages
= 0;
5614 unsigned long flags
;
5616 /* Calculate total number of !ZONE_HIGHMEM pages */
5617 for_each_zone(zone
) {
5618 if (!is_highmem(zone
))
5619 lowmem_pages
+= zone
->managed_pages
;
5622 for_each_zone(zone
) {
5625 spin_lock_irqsave(&zone
->lock
, flags
);
5626 tmp
= (u64
)pages_min
* zone
->managed_pages
;
5627 do_div(tmp
, lowmem_pages
);
5628 if (is_highmem(zone
)) {
5630 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5631 * need highmem pages, so cap pages_min to a small
5634 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5635 * deltas controls asynch page reclaim, and so should
5636 * not be capped for highmem.
5638 unsigned long min_pages
;
5640 min_pages
= zone
->managed_pages
/ 1024;
5641 min_pages
= clamp(min_pages
, SWAP_CLUSTER_MAX
, 128UL);
5642 zone
->watermark
[WMARK_MIN
] = min_pages
;
5645 * If it's a lowmem zone, reserve a number of pages
5646 * proportionate to the zone's size.
5648 zone
->watermark
[WMARK_MIN
] = tmp
;
5651 zone
->watermark
[WMARK_LOW
] = min_wmark_pages(zone
) + (tmp
>> 2);
5652 zone
->watermark
[WMARK_HIGH
] = min_wmark_pages(zone
) + (tmp
>> 1);
5654 __mod_zone_page_state(zone
, NR_ALLOC_BATCH
,
5655 high_wmark_pages(zone
) -
5656 low_wmark_pages(zone
) -
5657 zone_page_state(zone
, NR_ALLOC_BATCH
));
5659 setup_zone_migrate_reserve(zone
);
5660 spin_unlock_irqrestore(&zone
->lock
, flags
);
5663 /* update totalreserve_pages */
5664 calculate_totalreserve_pages();
5668 * setup_per_zone_wmarks - called when min_free_kbytes changes
5669 * or when memory is hot-{added|removed}
5671 * Ensures that the watermark[min,low,high] values for each zone are set
5672 * correctly with respect to min_free_kbytes.
5674 void setup_per_zone_wmarks(void)
5676 mutex_lock(&zonelists_mutex
);
5677 __setup_per_zone_wmarks();
5678 mutex_unlock(&zonelists_mutex
);
5682 * The inactive anon list should be small enough that the VM never has to
5683 * do too much work, but large enough that each inactive page has a chance
5684 * to be referenced again before it is swapped out.
5686 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5687 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5688 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5689 * the anonymous pages are kept on the inactive list.
5692 * memory ratio inactive anon
5693 * -------------------------------------
5702 static void __meminit
calculate_zone_inactive_ratio(struct zone
*zone
)
5704 unsigned int gb
, ratio
;
5706 /* Zone size in gigabytes */
5707 gb
= zone
->managed_pages
>> (30 - PAGE_SHIFT
);
5709 ratio
= int_sqrt(10 * gb
);
5713 zone
->inactive_ratio
= ratio
;
5716 static void __meminit
setup_per_zone_inactive_ratio(void)
5721 calculate_zone_inactive_ratio(zone
);
5725 * Initialise min_free_kbytes.
5727 * For small machines we want it small (128k min). For large machines
5728 * we want it large (64MB max). But it is not linear, because network
5729 * bandwidth does not increase linearly with machine size. We use
5731 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5732 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5748 int __meminit
init_per_zone_wmark_min(void)
5750 unsigned long lowmem_kbytes
;
5751 int new_min_free_kbytes
;
5753 lowmem_kbytes
= nr_free_buffer_pages() * (PAGE_SIZE
>> 10);
5754 new_min_free_kbytes
= int_sqrt(lowmem_kbytes
* 16);
5756 if (new_min_free_kbytes
> user_min_free_kbytes
) {
5757 min_free_kbytes
= new_min_free_kbytes
;
5758 if (min_free_kbytes
< 128)
5759 min_free_kbytes
= 128;
5760 if (min_free_kbytes
> 65536)
5761 min_free_kbytes
= 65536;
5763 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
5764 new_min_free_kbytes
, user_min_free_kbytes
);
5766 setup_per_zone_wmarks();
5767 refresh_zone_stat_thresholds();
5768 setup_per_zone_lowmem_reserve();
5769 setup_per_zone_inactive_ratio();
5772 module_init(init_per_zone_wmark_min
)
5775 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5776 * that we can call two helper functions whenever min_free_kbytes
5779 int min_free_kbytes_sysctl_handler(ctl_table
*table
, int write
,
5780 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5784 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5789 user_min_free_kbytes
= min_free_kbytes
;
5790 setup_per_zone_wmarks();
5796 int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table
*table
, int write
,
5797 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5802 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5807 zone
->min_unmapped_pages
= (zone
->managed_pages
*
5808 sysctl_min_unmapped_ratio
) / 100;
5812 int sysctl_min_slab_ratio_sysctl_handler(ctl_table
*table
, int write
,
5813 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5818 rc
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5823 zone
->min_slab_pages
= (zone
->managed_pages
*
5824 sysctl_min_slab_ratio
) / 100;
5830 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5831 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5832 * whenever sysctl_lowmem_reserve_ratio changes.
5834 * The reserve ratio obviously has absolutely no relation with the
5835 * minimum watermarks. The lowmem reserve ratio can only make sense
5836 * if in function of the boot time zone sizes.
5838 int lowmem_reserve_ratio_sysctl_handler(ctl_table
*table
, int write
,
5839 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5841 proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5842 setup_per_zone_lowmem_reserve();
5847 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5848 * cpu. It is the fraction of total pages in each zone that a hot per cpu
5849 * pagelist can have before it gets flushed back to buddy allocator.
5851 int percpu_pagelist_fraction_sysctl_handler(ctl_table
*table
, int write
,
5852 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
5858 ret
= proc_dointvec_minmax(table
, write
, buffer
, length
, ppos
);
5859 if (!write
|| (ret
< 0))
5862 mutex_lock(&pcp_batch_high_lock
);
5863 for_each_populated_zone(zone
) {
5865 high
= zone
->managed_pages
/ percpu_pagelist_fraction
;
5866 for_each_possible_cpu(cpu
)
5867 pageset_set_high(per_cpu_ptr(zone
->pageset
, cpu
),
5870 mutex_unlock(&pcp_batch_high_lock
);
5874 int hashdist
= HASHDIST_DEFAULT
;
5877 static int __init
set_hashdist(char *str
)
5881 hashdist
= simple_strtoul(str
, &str
, 0);
5884 __setup("hashdist=", set_hashdist
);
5888 * allocate a large system hash table from bootmem
5889 * - it is assumed that the hash table must contain an exact power-of-2
5890 * quantity of entries
5891 * - limit is the number of hash buckets, not the total allocation size
5893 void *__init
alloc_large_system_hash(const char *tablename
,
5894 unsigned long bucketsize
,
5895 unsigned long numentries
,
5898 unsigned int *_hash_shift
,
5899 unsigned int *_hash_mask
,
5900 unsigned long low_limit
,
5901 unsigned long high_limit
)
5903 unsigned long long max
= high_limit
;
5904 unsigned long log2qty
, size
;
5907 /* allow the kernel cmdline to have a say */
5909 /* round applicable memory size up to nearest megabyte */
5910 numentries
= nr_kernel_pages
;
5912 /* It isn't necessary when PAGE_SIZE >= 1MB */
5913 if (PAGE_SHIFT
< 20)
5914 numentries
= round_up(numentries
, (1<<20)/PAGE_SIZE
);
5916 /* limit to 1 bucket per 2^scale bytes of low memory */
5917 if (scale
> PAGE_SHIFT
)
5918 numentries
>>= (scale
- PAGE_SHIFT
);
5920 numentries
<<= (PAGE_SHIFT
- scale
);
5922 /* Make sure we've got at least a 0-order allocation.. */
5923 if (unlikely(flags
& HASH_SMALL
)) {
5924 /* Makes no sense without HASH_EARLY */
5925 WARN_ON(!(flags
& HASH_EARLY
));
5926 if (!(numentries
>> *_hash_shift
)) {
5927 numentries
= 1UL << *_hash_shift
;
5928 BUG_ON(!numentries
);
5930 } else if (unlikely((numentries
* bucketsize
) < PAGE_SIZE
))
5931 numentries
= PAGE_SIZE
/ bucketsize
;
5933 numentries
= roundup_pow_of_two(numentries
);
5935 /* limit allocation size to 1/16 total memory by default */
5937 max
= ((unsigned long long)nr_all_pages
<< PAGE_SHIFT
) >> 4;
5938 do_div(max
, bucketsize
);
5940 max
= min(max
, 0x80000000ULL
);
5942 if (numentries
< low_limit
)
5943 numentries
= low_limit
;
5944 if (numentries
> max
)
5947 log2qty
= ilog2(numentries
);
5950 size
= bucketsize
<< log2qty
;
5951 if (flags
& HASH_EARLY
)
5952 table
= memblock_virt_alloc_nopanic(size
, 0);
5954 table
= __vmalloc(size
, GFP_ATOMIC
, PAGE_KERNEL
);
5957 * If bucketsize is not a power-of-two, we may free
5958 * some pages at the end of hash table which
5959 * alloc_pages_exact() automatically does
5961 if (get_order(size
) < MAX_ORDER
) {
5962 table
= alloc_pages_exact(size
, GFP_ATOMIC
);
5963 kmemleak_alloc(table
, size
, 1, GFP_ATOMIC
);
5966 } while (!table
&& size
> PAGE_SIZE
&& --log2qty
);
5969 panic("Failed to allocate %s hash table\n", tablename
);
5971 printk(KERN_INFO
"%s hash table entries: %ld (order: %d, %lu bytes)\n",
5974 ilog2(size
) - PAGE_SHIFT
,
5978 *_hash_shift
= log2qty
;
5980 *_hash_mask
= (1 << log2qty
) - 1;
5985 /* Return a pointer to the bitmap storing bits affecting a block of pages */
5986 static inline unsigned long *get_pageblock_bitmap(struct zone
*zone
,
5989 #ifdef CONFIG_SPARSEMEM
5990 return __pfn_to_section(pfn
)->pageblock_flags
;
5992 return zone
->pageblock_flags
;
5993 #endif /* CONFIG_SPARSEMEM */
5996 static inline int pfn_to_bitidx(struct zone
*zone
, unsigned long pfn
)
5998 #ifdef CONFIG_SPARSEMEM
5999 pfn
&= (PAGES_PER_SECTION
-1);
6000 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6002 pfn
= pfn
- round_down(zone
->zone_start_pfn
, pageblock_nr_pages
);
6003 return (pfn
>> pageblock_order
) * NR_PAGEBLOCK_BITS
;
6004 #endif /* CONFIG_SPARSEMEM */
6008 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
6009 * @page: The page within the block of interest
6010 * @start_bitidx: The first bit of interest to retrieve
6011 * @end_bitidx: The last bit of interest
6012 * returns pageblock_bits flags
6014 unsigned long get_pageblock_flags_group(struct page
*page
,
6015 int start_bitidx
, int end_bitidx
)
6018 unsigned long *bitmap
;
6019 unsigned long pfn
, bitidx
;
6020 unsigned long flags
= 0;
6021 unsigned long value
= 1;
6023 zone
= page_zone(page
);
6024 pfn
= page_to_pfn(page
);
6025 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6026 bitidx
= pfn_to_bitidx(zone
, pfn
);
6028 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
6029 if (test_bit(bitidx
+ start_bitidx
, bitmap
))
6036 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
6037 * @page: The page within the block of interest
6038 * @start_bitidx: The first bit of interest
6039 * @end_bitidx: The last bit of interest
6040 * @flags: The flags to set
6042 void set_pageblock_flags_group(struct page
*page
, unsigned long flags
,
6043 int start_bitidx
, int end_bitidx
)
6046 unsigned long *bitmap
;
6047 unsigned long pfn
, bitidx
;
6048 unsigned long value
= 1;
6050 zone
= page_zone(page
);
6051 pfn
= page_to_pfn(page
);
6052 bitmap
= get_pageblock_bitmap(zone
, pfn
);
6053 bitidx
= pfn_to_bitidx(zone
, pfn
);
6054 VM_BUG_ON_PAGE(!zone_spans_pfn(zone
, pfn
), page
);
6056 for (; start_bitidx
<= end_bitidx
; start_bitidx
++, value
<<= 1)
6058 __set_bit(bitidx
+ start_bitidx
, bitmap
);
6060 __clear_bit(bitidx
+ start_bitidx
, bitmap
);
6064 * This function checks whether pageblock includes unmovable pages or not.
6065 * If @count is not zero, it is okay to include less @count unmovable pages
6067 * PageLRU check without isolation or lru_lock could race so that
6068 * MIGRATE_MOVABLE block might include unmovable pages. It means you can't
6069 * expect this function should be exact.
6071 bool has_unmovable_pages(struct zone
*zone
, struct page
*page
, int count
,
6072 bool skip_hwpoisoned_pages
)
6074 unsigned long pfn
, iter
, found
;
6078 * For avoiding noise data, lru_add_drain_all() should be called
6079 * If ZONE_MOVABLE, the zone never contains unmovable pages
6081 if (zone_idx(zone
) == ZONE_MOVABLE
)
6083 mt
= get_pageblock_migratetype(page
);
6084 if (mt
== MIGRATE_MOVABLE
|| is_migrate_cma(mt
))
6087 pfn
= page_to_pfn(page
);
6088 for (found
= 0, iter
= 0; iter
< pageblock_nr_pages
; iter
++) {
6089 unsigned long check
= pfn
+ iter
;
6091 if (!pfn_valid_within(check
))
6094 page
= pfn_to_page(check
);
6097 * Hugepages are not in LRU lists, but they're movable.
6098 * We need not scan over tail pages bacause we don't
6099 * handle each tail page individually in migration.
6101 if (PageHuge(page
)) {
6102 iter
= round_up(iter
+ 1, 1<<compound_order(page
)) - 1;
6107 * We can't use page_count without pin a page
6108 * because another CPU can free compound page.
6109 * This check already skips compound tails of THP
6110 * because their page->_count is zero at all time.
6112 if (!atomic_read(&page
->_count
)) {
6113 if (PageBuddy(page
))
6114 iter
+= (1 << page_order(page
)) - 1;
6119 * The HWPoisoned page may be not in buddy system, and
6120 * page_count() is not 0.
6122 if (skip_hwpoisoned_pages
&& PageHWPoison(page
))
6128 * If there are RECLAIMABLE pages, we need to check it.
6129 * But now, memory offline itself doesn't call shrink_slab()
6130 * and it still to be fixed.
6133 * If the page is not RAM, page_count()should be 0.
6134 * we don't need more check. This is an _used_ not-movable page.
6136 * The problematic thing here is PG_reserved pages. PG_reserved
6137 * is set to both of a memory hole page and a _used_ kernel
6146 bool is_pageblock_removable_nolock(struct page
*page
)
6152 * We have to be careful here because we are iterating over memory
6153 * sections which are not zone aware so we might end up outside of
6154 * the zone but still within the section.
6155 * We have to take care about the node as well. If the node is offline
6156 * its NODE_DATA will be NULL - see page_zone.
6158 if (!node_online(page_to_nid(page
)))
6161 zone
= page_zone(page
);
6162 pfn
= page_to_pfn(page
);
6163 if (!zone_spans_pfn(zone
, pfn
))
6166 return !has_unmovable_pages(zone
, page
, 0, true);
6171 static unsigned long pfn_max_align_down(unsigned long pfn
)
6173 return pfn
& ~(max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6174 pageblock_nr_pages
) - 1);
6177 static unsigned long pfn_max_align_up(unsigned long pfn
)
6179 return ALIGN(pfn
, max_t(unsigned long, MAX_ORDER_NR_PAGES
,
6180 pageblock_nr_pages
));
6183 /* [start, end) must belong to a single zone. */
6184 static int __alloc_contig_migrate_range(struct compact_control
*cc
,
6185 unsigned long start
, unsigned long end
)
6187 /* This function is based on compact_zone() from compaction.c. */
6188 unsigned long nr_reclaimed
;
6189 unsigned long pfn
= start
;
6190 unsigned int tries
= 0;
6195 while (pfn
< end
|| !list_empty(&cc
->migratepages
)) {
6196 if (fatal_signal_pending(current
)) {
6201 if (list_empty(&cc
->migratepages
)) {
6202 cc
->nr_migratepages
= 0;
6203 pfn
= isolate_migratepages_range(cc
->zone
, cc
,
6210 } else if (++tries
== 5) {
6211 ret
= ret
< 0 ? ret
: -EBUSY
;
6215 nr_reclaimed
= reclaim_clean_pages_from_list(cc
->zone
,
6217 cc
->nr_migratepages
-= nr_reclaimed
;
6219 ret
= migrate_pages(&cc
->migratepages
, alloc_migrate_target
,
6220 0, MIGRATE_SYNC
, MR_CMA
);
6223 putback_movable_pages(&cc
->migratepages
);
6230 * alloc_contig_range() -- tries to allocate given range of pages
6231 * @start: start PFN to allocate
6232 * @end: one-past-the-last PFN to allocate
6233 * @migratetype: migratetype of the underlaying pageblocks (either
6234 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6235 * in range must have the same migratetype and it must
6236 * be either of the two.
6238 * The PFN range does not have to be pageblock or MAX_ORDER_NR_PAGES
6239 * aligned, however it's the caller's responsibility to guarantee that
6240 * we are the only thread that changes migrate type of pageblocks the
6243 * The PFN range must belong to a single zone.
6245 * Returns zero on success or negative error code. On success all
6246 * pages which PFN is in [start, end) are allocated for the caller and
6247 * need to be freed with free_contig_range().
6249 int alloc_contig_range(unsigned long start
, unsigned long end
,
6250 unsigned migratetype
)
6252 unsigned long outer_start
, outer_end
;
6255 struct compact_control cc
= {
6256 .nr_migratepages
= 0,
6258 .zone
= page_zone(pfn_to_page(start
)),
6260 .ignore_skip_hint
= true,
6262 INIT_LIST_HEAD(&cc
.migratepages
);
6265 * What we do here is we mark all pageblocks in range as
6266 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6267 * have different sizes, and due to the way page allocator
6268 * work, we align the range to biggest of the two pages so
6269 * that page allocator won't try to merge buddies from
6270 * different pageblocks and change MIGRATE_ISOLATE to some
6271 * other migration type.
6273 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6274 * migrate the pages from an unaligned range (ie. pages that
6275 * we are interested in). This will put all the pages in
6276 * range back to page allocator as MIGRATE_ISOLATE.
6278 * When this is done, we take the pages in range from page
6279 * allocator removing them from the buddy system. This way
6280 * page allocator will never consider using them.
6282 * This lets us mark the pageblocks back as
6283 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6284 * aligned range but not in the unaligned, original range are
6285 * put back to page allocator so that buddy can use them.
6288 ret
= start_isolate_page_range(pfn_max_align_down(start
),
6289 pfn_max_align_up(end
), migratetype
,
6294 ret
= __alloc_contig_migrate_range(&cc
, start
, end
);
6299 * Pages from [start, end) are within a MAX_ORDER_NR_PAGES
6300 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6301 * more, all pages in [start, end) are free in page allocator.
6302 * What we are going to do is to allocate all pages from
6303 * [start, end) (that is remove them from page allocator).
6305 * The only problem is that pages at the beginning and at the
6306 * end of interesting range may be not aligned with pages that
6307 * page allocator holds, ie. they can be part of higher order
6308 * pages. Because of this, we reserve the bigger range and
6309 * once this is done free the pages we are not interested in.
6311 * We don't have to hold zone->lock here because the pages are
6312 * isolated thus they won't get removed from buddy.
6315 lru_add_drain_all();
6319 outer_start
= start
;
6320 while (!PageBuddy(pfn_to_page(outer_start
))) {
6321 if (++order
>= MAX_ORDER
) {
6325 outer_start
&= ~0UL << order
;
6328 /* Make sure the range is really isolated. */
6329 if (test_pages_isolated(outer_start
, end
, false)) {
6330 pr_warn("alloc_contig_range test_pages_isolated(%lx, %lx) failed\n",
6337 /* Grab isolated pages from freelists. */
6338 outer_end
= isolate_freepages_range(&cc
, outer_start
, end
);
6344 /* Free head and tail (if any) */
6345 if (start
!= outer_start
)
6346 free_contig_range(outer_start
, start
- outer_start
);
6347 if (end
!= outer_end
)
6348 free_contig_range(end
, outer_end
- end
);
6351 undo_isolate_page_range(pfn_max_align_down(start
),
6352 pfn_max_align_up(end
), migratetype
);
6356 void free_contig_range(unsigned long pfn
, unsigned nr_pages
)
6358 unsigned int count
= 0;
6360 for (; nr_pages
--; pfn
++) {
6361 struct page
*page
= pfn_to_page(pfn
);
6363 count
+= page_count(page
) != 1;
6366 WARN(count
!= 0, "%d pages are still in use!\n", count
);
6370 #ifdef CONFIG_MEMORY_HOTPLUG
6372 * The zone indicated has a new number of managed_pages; batch sizes and percpu
6373 * page high values need to be recalulated.
6375 void __meminit
zone_pcp_update(struct zone
*zone
)
6378 mutex_lock(&pcp_batch_high_lock
);
6379 for_each_possible_cpu(cpu
)
6380 pageset_set_high_and_batch(zone
,
6381 per_cpu_ptr(zone
->pageset
, cpu
));
6382 mutex_unlock(&pcp_batch_high_lock
);
6386 void zone_pcp_reset(struct zone
*zone
)
6388 unsigned long flags
;
6390 struct per_cpu_pageset
*pset
;
6392 /* avoid races with drain_pages() */
6393 local_irq_save(flags
);
6394 if (zone
->pageset
!= &boot_pageset
) {
6395 for_each_online_cpu(cpu
) {
6396 pset
= per_cpu_ptr(zone
->pageset
, cpu
);
6397 drain_zonestat(zone
, pset
);
6399 free_percpu(zone
->pageset
);
6400 zone
->pageset
= &boot_pageset
;
6402 local_irq_restore(flags
);
6405 #ifdef CONFIG_MEMORY_HOTREMOVE
6407 * All pages in the range must be isolated before calling this.
6410 __offline_isolated_pages(unsigned long start_pfn
, unsigned long end_pfn
)
6416 unsigned long flags
;
6417 /* find the first valid pfn */
6418 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
++)
6423 zone
= page_zone(pfn_to_page(pfn
));
6424 spin_lock_irqsave(&zone
->lock
, flags
);
6426 while (pfn
< end_pfn
) {
6427 if (!pfn_valid(pfn
)) {
6431 page
= pfn_to_page(pfn
);
6433 * The HWPoisoned page may be not in buddy system, and
6434 * page_count() is not 0.
6436 if (unlikely(!PageBuddy(page
) && PageHWPoison(page
))) {
6438 SetPageReserved(page
);
6442 BUG_ON(page_count(page
));
6443 BUG_ON(!PageBuddy(page
));
6444 order
= page_order(page
);
6445 #ifdef CONFIG_DEBUG_VM
6446 printk(KERN_INFO
"remove from free list %lx %d %lx\n",
6447 pfn
, 1 << order
, end_pfn
);
6449 list_del(&page
->lru
);
6450 rmv_page_order(page
);
6451 zone
->free_area
[order
].nr_free
--;
6452 for (i
= 0; i
< (1 << order
); i
++)
6453 SetPageReserved((page
+i
));
6454 pfn
+= (1 << order
);
6456 spin_unlock_irqrestore(&zone
->lock
, flags
);
6460 #ifdef CONFIG_MEMORY_FAILURE
6461 bool is_free_buddy_page(struct page
*page
)
6463 struct zone
*zone
= page_zone(page
);
6464 unsigned long pfn
= page_to_pfn(page
);
6465 unsigned long flags
;
6468 spin_lock_irqsave(&zone
->lock
, flags
);
6469 for (order
= 0; order
< MAX_ORDER
; order
++) {
6470 struct page
*page_head
= page
- (pfn
& ((1 << order
) - 1));
6472 if (PageBuddy(page_head
) && page_order(page_head
) >= order
)
6475 spin_unlock_irqrestore(&zone
->lock
, flags
);
6477 return order
< MAX_ORDER
;
6481 static const struct trace_print_flags pageflag_names
[] = {
6482 {1UL << PG_locked
, "locked" },
6483 {1UL << PG_error
, "error" },
6484 {1UL << PG_referenced
, "referenced" },
6485 {1UL << PG_uptodate
, "uptodate" },
6486 {1UL << PG_dirty
, "dirty" },
6487 {1UL << PG_lru
, "lru" },
6488 {1UL << PG_active
, "active" },
6489 {1UL << PG_slab
, "slab" },
6490 {1UL << PG_owner_priv_1
, "owner_priv_1" },
6491 {1UL << PG_arch_1
, "arch_1" },
6492 {1UL << PG_reserved
, "reserved" },
6493 {1UL << PG_private
, "private" },
6494 {1UL << PG_private_2
, "private_2" },
6495 {1UL << PG_writeback
, "writeback" },
6496 #ifdef CONFIG_PAGEFLAGS_EXTENDED
6497 {1UL << PG_head
, "head" },
6498 {1UL << PG_tail
, "tail" },
6500 {1UL << PG_compound
, "compound" },
6502 {1UL << PG_swapcache
, "swapcache" },
6503 {1UL << PG_mappedtodisk
, "mappedtodisk" },
6504 {1UL << PG_reclaim
, "reclaim" },
6505 {1UL << PG_swapbacked
, "swapbacked" },
6506 {1UL << PG_unevictable
, "unevictable" },
6508 {1UL << PG_mlocked
, "mlocked" },
6510 #ifdef CONFIG_ARCH_USES_PG_UNCACHED
6511 {1UL << PG_uncached
, "uncached" },
6513 #ifdef CONFIG_MEMORY_FAILURE
6514 {1UL << PG_hwpoison
, "hwpoison" },
6516 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6517 {1UL << PG_compound_lock
, "compound_lock" },
6521 static void dump_page_flags(unsigned long flags
)
6523 const char *delim
= "";
6527 BUILD_BUG_ON(ARRAY_SIZE(pageflag_names
) != __NR_PAGEFLAGS
);
6529 printk(KERN_ALERT
"page flags: %#lx(", flags
);
6531 /* remove zone id */
6532 flags
&= (1UL << NR_PAGEFLAGS
) - 1;
6534 for (i
= 0; i
< ARRAY_SIZE(pageflag_names
) && flags
; i
++) {
6536 mask
= pageflag_names
[i
].mask
;
6537 if ((flags
& mask
) != mask
)
6541 printk("%s%s", delim
, pageflag_names
[i
].name
);
6545 /* check for left over flags */
6547 printk("%s%#lx", delim
, flags
);
6552 void dump_page_badflags(struct page
*page
, const char *reason
,
6553 unsigned long badflags
)
6556 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
6557 page
, atomic_read(&page
->_count
), page_mapcount(page
),
6558 page
->mapping
, page
->index
);
6559 dump_page_flags(page
->flags
);
6561 pr_alert("page dumped because: %s\n", reason
);
6562 if (page
->flags
& badflags
) {
6563 pr_alert("bad because of flags:\n");
6564 dump_page_flags(page
->flags
& badflags
);
6566 mem_cgroup_print_bad_page(page
);
6569 void dump_page(struct page
*page
, const char *reason
)
6571 dump_page_badflags(page
, reason
, 0);
6573 EXPORT_SYMBOL(dump_page
);