Loading...
1/*
2 * linux/mm/page_alloc.c
3 *
4 * Manages the free list, the system allocates free pages here.
5 * Note that kmalloc() lives in slab.c
6 *
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)
15 */
16
17#include <linux/stddef.h>
18#include <linux/mm.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/memory.h>
55#include <linux/compaction.h>
56#include <trace/events/kmem.h>
57#include <linux/ftrace_event.h>
58#include <linux/memcontrol.h>
59#include <linux/prefetch.h>
60
61#include <asm/tlbflush.h>
62#include <asm/div64.h>
63#include "internal.h"
64
65#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
66DEFINE_PER_CPU(int, numa_node);
67EXPORT_PER_CPU_SYMBOL(numa_node);
68#endif
69
70#ifdef CONFIG_HAVE_MEMORYLESS_NODES
71/*
72 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
73 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
74 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
75 * defined in <linux/topology.h>.
76 */
77DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
78EXPORT_PER_CPU_SYMBOL(_numa_mem_);
79#endif
80
81/*
82 * Array of node states.
83 */
84nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
85 [N_POSSIBLE] = NODE_MASK_ALL,
86 [N_ONLINE] = { { [0] = 1UL } },
87#ifndef CONFIG_NUMA
88 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
89#ifdef CONFIG_HIGHMEM
90 [N_HIGH_MEMORY] = { { [0] = 1UL } },
91#endif
92 [N_CPU] = { { [0] = 1UL } },
93#endif /* NUMA */
94};
95EXPORT_SYMBOL(node_states);
96
97unsigned long totalram_pages __read_mostly;
98unsigned long totalreserve_pages __read_mostly;
99int percpu_pagelist_fraction;
100gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
101
102#ifdef CONFIG_PM_SLEEP
103/*
104 * The following functions are used by the suspend/hibernate code to temporarily
105 * change gfp_allowed_mask in order to avoid using I/O during memory allocations
106 * while devices are suspended. To avoid races with the suspend/hibernate code,
107 * they should always be called with pm_mutex held (gfp_allowed_mask also should
108 * only be modified with pm_mutex held, unless the suspend/hibernate code is
109 * guaranteed not to run in parallel with that modification).
110 */
111
112static gfp_t saved_gfp_mask;
113
114void pm_restore_gfp_mask(void)
115{
116 WARN_ON(!mutex_is_locked(&pm_mutex));
117 if (saved_gfp_mask) {
118 gfp_allowed_mask = saved_gfp_mask;
119 saved_gfp_mask = 0;
120 }
121}
122
123void pm_restrict_gfp_mask(void)
124{
125 WARN_ON(!mutex_is_locked(&pm_mutex));
126 WARN_ON(saved_gfp_mask);
127 saved_gfp_mask = gfp_allowed_mask;
128 gfp_allowed_mask &= ~GFP_IOFS;
129}
130#endif /* CONFIG_PM_SLEEP */
131
132#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
133int pageblock_order __read_mostly;
134#endif
135
136static void __free_pages_ok(struct page *page, unsigned int order);
137
138/*
139 * results with 256, 32 in the lowmem_reserve sysctl:
140 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
141 * 1G machine -> (16M dma, 784M normal, 224M high)
142 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
143 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
144 * HIGHMEM allocation will (224M+784M)/256 of ram reserved in ZONE_DMA
145 *
146 * TBD: should special case ZONE_DMA32 machines here - in those we normally
147 * don't need any ZONE_NORMAL reservation
148 */
149int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES-1] = {
150#ifdef CONFIG_ZONE_DMA
151 256,
152#endif
153#ifdef CONFIG_ZONE_DMA32
154 256,
155#endif
156#ifdef CONFIG_HIGHMEM
157 32,
158#endif
159 32,
160};
161
162EXPORT_SYMBOL(totalram_pages);
163
164static char * const zone_names[MAX_NR_ZONES] = {
165#ifdef CONFIG_ZONE_DMA
166 "DMA",
167#endif
168#ifdef CONFIG_ZONE_DMA32
169 "DMA32",
170#endif
171 "Normal",
172#ifdef CONFIG_HIGHMEM
173 "HighMem",
174#endif
175 "Movable",
176};
177
178int min_free_kbytes = 1024;
179
180static unsigned long __meminitdata nr_kernel_pages;
181static unsigned long __meminitdata nr_all_pages;
182static unsigned long __meminitdata dma_reserve;
183
184#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
185 /*
186 * MAX_ACTIVE_REGIONS determines the maximum number of distinct
187 * ranges of memory (RAM) that may be registered with add_active_range().
188 * Ranges passed to add_active_range() will be merged if possible
189 * so the number of times add_active_range() can be called is
190 * related to the number of nodes and the number of holes
191 */
192 #ifdef CONFIG_MAX_ACTIVE_REGIONS
193 /* Allow an architecture to set MAX_ACTIVE_REGIONS to save memory */
194 #define MAX_ACTIVE_REGIONS CONFIG_MAX_ACTIVE_REGIONS
195 #else
196 #if MAX_NUMNODES >= 32
197 /* If there can be many nodes, allow up to 50 holes per node */
198 #define MAX_ACTIVE_REGIONS (MAX_NUMNODES*50)
199 #else
200 /* By default, allow up to 256 distinct regions */
201 #define MAX_ACTIVE_REGIONS 256
202 #endif
203 #endif
204
205 static struct node_active_region __meminitdata early_node_map[MAX_ACTIVE_REGIONS];
206 static int __meminitdata nr_nodemap_entries;
207 static unsigned long __meminitdata arch_zone_lowest_possible_pfn[MAX_NR_ZONES];
208 static unsigned long __meminitdata arch_zone_highest_possible_pfn[MAX_NR_ZONES];
209 static unsigned long __initdata required_kernelcore;
210 static unsigned long __initdata required_movablecore;
211 static unsigned long __meminitdata zone_movable_pfn[MAX_NUMNODES];
212
213 /* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
214 int movable_zone;
215 EXPORT_SYMBOL(movable_zone);
216#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
217
218#if MAX_NUMNODES > 1
219int nr_node_ids __read_mostly = MAX_NUMNODES;
220int nr_online_nodes __read_mostly = 1;
221EXPORT_SYMBOL(nr_node_ids);
222EXPORT_SYMBOL(nr_online_nodes);
223#endif
224
225int page_group_by_mobility_disabled __read_mostly;
226
227static void set_pageblock_migratetype(struct page *page, int migratetype)
228{
229
230 if (unlikely(page_group_by_mobility_disabled))
231 migratetype = MIGRATE_UNMOVABLE;
232
233 set_pageblock_flags_group(page, (unsigned long)migratetype,
234 PB_migrate, PB_migrate_end);
235}
236
237bool oom_killer_disabled __read_mostly;
238
239#ifdef CONFIG_DEBUG_VM
240static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
241{
242 int ret = 0;
243 unsigned seq;
244 unsigned long pfn = page_to_pfn(page);
245
246 do {
247 seq = zone_span_seqbegin(zone);
248 if (pfn >= zone->zone_start_pfn + zone->spanned_pages)
249 ret = 1;
250 else if (pfn < zone->zone_start_pfn)
251 ret = 1;
252 } while (zone_span_seqretry(zone, seq));
253
254 return ret;
255}
256
257static int page_is_consistent(struct zone *zone, struct page *page)
258{
259 if (!pfn_valid_within(page_to_pfn(page)))
260 return 0;
261 if (zone != page_zone(page))
262 return 0;
263
264 return 1;
265}
266/*
267 * Temporary debugging check for pages not lying within a given zone.
268 */
269static int bad_range(struct zone *zone, struct page *page)
270{
271 if (page_outside_zone_boundaries(zone, page))
272 return 1;
273 if (!page_is_consistent(zone, page))
274 return 1;
275
276 return 0;
277}
278#else
279static inline int bad_range(struct zone *zone, struct page *page)
280{
281 return 0;
282}
283#endif
284
285static void bad_page(struct page *page)
286{
287 static unsigned long resume;
288 static unsigned long nr_shown;
289 static unsigned long nr_unshown;
290
291 /* Don't complain about poisoned pages */
292 if (PageHWPoison(page)) {
293 reset_page_mapcount(page); /* remove PageBuddy */
294 return;
295 }
296
297 /*
298 * Allow a burst of 60 reports, then keep quiet for that minute;
299 * or allow a steady drip of one report per second.
300 */
301 if (nr_shown == 60) {
302 if (time_before(jiffies, resume)) {
303 nr_unshown++;
304 goto out;
305 }
306 if (nr_unshown) {
307 printk(KERN_ALERT
308 "BUG: Bad page state: %lu messages suppressed\n",
309 nr_unshown);
310 nr_unshown = 0;
311 }
312 nr_shown = 0;
313 }
314 if (nr_shown++ == 0)
315 resume = jiffies + 60 * HZ;
316
317 printk(KERN_ALERT "BUG: Bad page state in process %s pfn:%05lx\n",
318 current->comm, page_to_pfn(page));
319 dump_page(page);
320
321 dump_stack();
322out:
323 /* Leave bad fields for debug, except PageBuddy could make trouble */
324 reset_page_mapcount(page); /* remove PageBuddy */
325 add_taint(TAINT_BAD_PAGE);
326}
327
328/*
329 * Higher-order pages are called "compound pages". They are structured thusly:
330 *
331 * The first PAGE_SIZE page is called the "head page".
332 *
333 * The remaining PAGE_SIZE pages are called "tail pages".
334 *
335 * All pages have PG_compound set. All pages have their ->private pointing at
336 * the head page (even the head page has this).
337 *
338 * The first tail page's ->lru.next holds the address of the compound page's
339 * put_page() function. Its ->lru.prev holds the order of allocation.
340 * This usage means that zero-order pages may not be compound.
341 */
342
343static void free_compound_page(struct page *page)
344{
345 __free_pages_ok(page, compound_order(page));
346}
347
348void prep_compound_page(struct page *page, unsigned long order)
349{
350 int i;
351 int nr_pages = 1 << order;
352
353 set_compound_page_dtor(page, free_compound_page);
354 set_compound_order(page, order);
355 __SetPageHead(page);
356 for (i = 1; i < nr_pages; i++) {
357 struct page *p = page + i;
358
359 __SetPageTail(p);
360 p->first_page = page;
361 }
362}
363
364/* update __split_huge_page_refcount if you change this function */
365static int destroy_compound_page(struct page *page, unsigned long order)
366{
367 int i;
368 int nr_pages = 1 << order;
369 int bad = 0;
370
371 if (unlikely(compound_order(page) != order) ||
372 unlikely(!PageHead(page))) {
373 bad_page(page);
374 bad++;
375 }
376
377 __ClearPageHead(page);
378
379 for (i = 1; i < nr_pages; i++) {
380 struct page *p = page + i;
381
382 if (unlikely(!PageTail(p) || (p->first_page != page))) {
383 bad_page(page);
384 bad++;
385 }
386 __ClearPageTail(p);
387 }
388
389 return bad;
390}
391
392static inline void prep_zero_page(struct page *page, int order, gfp_t gfp_flags)
393{
394 int i;
395
396 /*
397 * clear_highpage() will use KM_USER0, so it's a bug to use __GFP_ZERO
398 * and __GFP_HIGHMEM from hard or soft interrupt context.
399 */
400 VM_BUG_ON((gfp_flags & __GFP_HIGHMEM) && in_interrupt());
401 for (i = 0; i < (1 << order); i++)
402 clear_highpage(page + i);
403}
404
405static inline void set_page_order(struct page *page, int order)
406{
407 set_page_private(page, order);
408 __SetPageBuddy(page);
409}
410
411static inline void rmv_page_order(struct page *page)
412{
413 __ClearPageBuddy(page);
414 set_page_private(page, 0);
415}
416
417/*
418 * Locate the struct page for both the matching buddy in our
419 * pair (buddy1) and the combined O(n+1) page they form (page).
420 *
421 * 1) Any buddy B1 will have an order O twin B2 which satisfies
422 * the following equation:
423 * B2 = B1 ^ (1 << O)
424 * For example, if the starting buddy (buddy2) is #8 its order
425 * 1 buddy is #10:
426 * B2 = 8 ^ (1 << 1) = 8 ^ 2 = 10
427 *
428 * 2) Any buddy B will have an order O+1 parent P which
429 * satisfies the following equation:
430 * P = B & ~(1 << O)
431 *
432 * Assumption: *_mem_map is contiguous at least up to MAX_ORDER
433 */
434static inline unsigned long
435__find_buddy_index(unsigned long page_idx, unsigned int order)
436{
437 return page_idx ^ (1 << order);
438}
439
440/*
441 * This function checks whether a page is free && is the buddy
442 * we can do coalesce a page and its buddy if
443 * (a) the buddy is not in a hole &&
444 * (b) the buddy is in the buddy system &&
445 * (c) a page and its buddy have the same order &&
446 * (d) a page and its buddy are in the same zone.
447 *
448 * For recording whether a page is in the buddy system, we set ->_mapcount -2.
449 * Setting, clearing, and testing _mapcount -2 is serialized by zone->lock.
450 *
451 * For recording page's order, we use page_private(page).
452 */
453static inline int page_is_buddy(struct page *page, struct page *buddy,
454 int order)
455{
456 if (!pfn_valid_within(page_to_pfn(buddy)))
457 return 0;
458
459 if (page_zone_id(page) != page_zone_id(buddy))
460 return 0;
461
462 if (PageBuddy(buddy) && page_order(buddy) == order) {
463 VM_BUG_ON(page_count(buddy) != 0);
464 return 1;
465 }
466 return 0;
467}
468
469/*
470 * Freeing function for a buddy system allocator.
471 *
472 * The concept of a buddy system is to maintain direct-mapped table
473 * (containing bit values) for memory blocks of various "orders".
474 * The bottom level table contains the map for the smallest allocatable
475 * units of memory (here, pages), and each level above it describes
476 * pairs of units from the levels below, hence, "buddies".
477 * At a high level, all that happens here is marking the table entry
478 * at the bottom level available, and propagating the changes upward
479 * as necessary, plus some accounting needed to play nicely with other
480 * parts of the VM system.
481 * At each level, we keep a list of pages, which are heads of continuous
482 * free pages of length of (1 << order) and marked with _mapcount -2. Page's
483 * order is recorded in page_private(page) field.
484 * So when we are allocating or freeing one, we can derive the state of the
485 * other. That is, if we allocate a small block, and both were
486 * free, the remainder of the region must be split into blocks.
487 * If a block is freed, and its buddy is also free, then this
488 * triggers coalescing into a block of larger size.
489 *
490 * -- wli
491 */
492
493static inline void __free_one_page(struct page *page,
494 struct zone *zone, unsigned int order,
495 int migratetype)
496{
497 unsigned long page_idx;
498 unsigned long combined_idx;
499 unsigned long uninitialized_var(buddy_idx);
500 struct page *buddy;
501
502 if (unlikely(PageCompound(page)))
503 if (unlikely(destroy_compound_page(page, order)))
504 return;
505
506 VM_BUG_ON(migratetype == -1);
507
508 page_idx = page_to_pfn(page) & ((1 << MAX_ORDER) - 1);
509
510 VM_BUG_ON(page_idx & ((1 << order) - 1));
511 VM_BUG_ON(bad_range(zone, page));
512
513 while (order < MAX_ORDER-1) {
514 buddy_idx = __find_buddy_index(page_idx, order);
515 buddy = page + (buddy_idx - page_idx);
516 if (!page_is_buddy(page, buddy, order))
517 break;
518
519 /* Our buddy is free, merge with it and move up one order. */
520 list_del(&buddy->lru);
521 zone->free_area[order].nr_free--;
522 rmv_page_order(buddy);
523 combined_idx = buddy_idx & page_idx;
524 page = page + (combined_idx - page_idx);
525 page_idx = combined_idx;
526 order++;
527 }
528 set_page_order(page, order);
529
530 /*
531 * If this is not the largest possible page, check if the buddy
532 * of the next-highest order is free. If it is, it's possible
533 * that pages are being freed that will coalesce soon. In case,
534 * that is happening, add the free page to the tail of the list
535 * so it's less likely to be used soon and more likely to be merged
536 * as a higher order page
537 */
538 if ((order < MAX_ORDER-2) && pfn_valid_within(page_to_pfn(buddy))) {
539 struct page *higher_page, *higher_buddy;
540 combined_idx = buddy_idx & page_idx;
541 higher_page = page + (combined_idx - page_idx);
542 buddy_idx = __find_buddy_index(combined_idx, order + 1);
543 higher_buddy = page + (buddy_idx - combined_idx);
544 if (page_is_buddy(higher_page, higher_buddy, order + 1)) {
545 list_add_tail(&page->lru,
546 &zone->free_area[order].free_list[migratetype]);
547 goto out;
548 }
549 }
550
551 list_add(&page->lru, &zone->free_area[order].free_list[migratetype]);
552out:
553 zone->free_area[order].nr_free++;
554}
555
556/*
557 * free_page_mlock() -- clean up attempts to free and mlocked() page.
558 * Page should not be on lru, so no need to fix that up.
559 * free_pages_check() will verify...
560 */
561static inline void free_page_mlock(struct page *page)
562{
563 __dec_zone_page_state(page, NR_MLOCK);
564 __count_vm_event(UNEVICTABLE_MLOCKFREED);
565}
566
567static inline int free_pages_check(struct page *page)
568{
569 if (unlikely(page_mapcount(page) |
570 (page->mapping != NULL) |
571 (atomic_read(&page->_count) != 0) |
572 (page->flags & PAGE_FLAGS_CHECK_AT_FREE) |
573 (mem_cgroup_bad_page_check(page)))) {
574 bad_page(page);
575 return 1;
576 }
577 if (page->flags & PAGE_FLAGS_CHECK_AT_PREP)
578 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
579 return 0;
580}
581
582/*
583 * Frees a number of pages from the PCP lists
584 * Assumes all pages on list are in same zone, and of same order.
585 * count is the number of pages to free.
586 *
587 * If the zone was previously in an "all pages pinned" state then look to
588 * see if this freeing clears that state.
589 *
590 * And clear the zone's pages_scanned counter, to hold off the "all pages are
591 * pinned" detection logic.
592 */
593static void free_pcppages_bulk(struct zone *zone, int count,
594 struct per_cpu_pages *pcp)
595{
596 int migratetype = 0;
597 int batch_free = 0;
598 int to_free = count;
599
600 spin_lock(&zone->lock);
601 zone->all_unreclaimable = 0;
602 zone->pages_scanned = 0;
603
604 while (to_free) {
605 struct page *page;
606 struct list_head *list;
607
608 /*
609 * Remove pages from lists in a round-robin fashion. A
610 * batch_free count is maintained that is incremented when an
611 * empty list is encountered. This is so more pages are freed
612 * off fuller lists instead of spinning excessively around empty
613 * lists
614 */
615 do {
616 batch_free++;
617 if (++migratetype == MIGRATE_PCPTYPES)
618 migratetype = 0;
619 list = &pcp->lists[migratetype];
620 } while (list_empty(list));
621
622 /* This is the only non-empty list. Free them all. */
623 if (batch_free == MIGRATE_PCPTYPES)
624 batch_free = to_free;
625
626 do {
627 page = list_entry(list->prev, struct page, lru);
628 /* must delete as __free_one_page list manipulates */
629 list_del(&page->lru);
630 /* MIGRATE_MOVABLE list may include MIGRATE_RESERVEs */
631 __free_one_page(page, zone, 0, page_private(page));
632 trace_mm_page_pcpu_drain(page, 0, page_private(page));
633 } while (--to_free && --batch_free && !list_empty(list));
634 }
635 __mod_zone_page_state(zone, NR_FREE_PAGES, count);
636 spin_unlock(&zone->lock);
637}
638
639static void free_one_page(struct zone *zone, struct page *page, int order,
640 int migratetype)
641{
642 spin_lock(&zone->lock);
643 zone->all_unreclaimable = 0;
644 zone->pages_scanned = 0;
645
646 __free_one_page(page, zone, order, migratetype);
647 __mod_zone_page_state(zone, NR_FREE_PAGES, 1 << order);
648 spin_unlock(&zone->lock);
649}
650
651static bool free_pages_prepare(struct page *page, unsigned int order)
652{
653 int i;
654 int bad = 0;
655
656 trace_mm_page_free_direct(page, order);
657 kmemcheck_free_shadow(page, order);
658
659 if (PageAnon(page))
660 page->mapping = NULL;
661 for (i = 0; i < (1 << order); i++)
662 bad += free_pages_check(page + i);
663 if (bad)
664 return false;
665
666 if (!PageHighMem(page)) {
667 debug_check_no_locks_freed(page_address(page),PAGE_SIZE<<order);
668 debug_check_no_obj_freed(page_address(page),
669 PAGE_SIZE << order);
670 }
671 arch_free_page(page, order);
672 kernel_map_pages(page, 1 << order, 0);
673
674 return true;
675}
676
677static void __free_pages_ok(struct page *page, unsigned int order)
678{
679 unsigned long flags;
680 int wasMlocked = __TestClearPageMlocked(page);
681
682 if (!free_pages_prepare(page, order))
683 return;
684
685 local_irq_save(flags);
686 if (unlikely(wasMlocked))
687 free_page_mlock(page);
688 __count_vm_events(PGFREE, 1 << order);
689 free_one_page(page_zone(page), page, order,
690 get_pageblock_migratetype(page));
691 local_irq_restore(flags);
692}
693
694/*
695 * permit the bootmem allocator to evade page validation on high-order frees
696 */
697void __meminit __free_pages_bootmem(struct page *page, unsigned int order)
698{
699 if (order == 0) {
700 __ClearPageReserved(page);
701 set_page_count(page, 0);
702 set_page_refcounted(page);
703 __free_page(page);
704 } else {
705 int loop;
706
707 prefetchw(page);
708 for (loop = 0; loop < BITS_PER_LONG; loop++) {
709 struct page *p = &page[loop];
710
711 if (loop + 1 < BITS_PER_LONG)
712 prefetchw(p + 1);
713 __ClearPageReserved(p);
714 set_page_count(p, 0);
715 }
716
717 set_page_refcounted(page);
718 __free_pages(page, order);
719 }
720}
721
722
723/*
724 * The order of subdivision here is critical for the IO subsystem.
725 * Please do not alter this order without good reasons and regression
726 * testing. Specifically, as large blocks of memory are subdivided,
727 * the order in which smaller blocks are delivered depends on the order
728 * they're subdivided in this function. This is the primary factor
729 * influencing the order in which pages are delivered to the IO
730 * subsystem according to empirical testing, and this is also justified
731 * by considering the behavior of a buddy system containing a single
732 * large block of memory acted on by a series of small allocations.
733 * This behavior is a critical factor in sglist merging's success.
734 *
735 * -- wli
736 */
737static inline void expand(struct zone *zone, struct page *page,
738 int low, int high, struct free_area *area,
739 int migratetype)
740{
741 unsigned long size = 1 << high;
742
743 while (high > low) {
744 area--;
745 high--;
746 size >>= 1;
747 VM_BUG_ON(bad_range(zone, &page[size]));
748 list_add(&page[size].lru, &area->free_list[migratetype]);
749 area->nr_free++;
750 set_page_order(&page[size], high);
751 }
752}
753
754/*
755 * This page is about to be returned from the page allocator
756 */
757static inline int check_new_page(struct page *page)
758{
759 if (unlikely(page_mapcount(page) |
760 (page->mapping != NULL) |
761 (atomic_read(&page->_count) != 0) |
762 (page->flags & PAGE_FLAGS_CHECK_AT_PREP) |
763 (mem_cgroup_bad_page_check(page)))) {
764 bad_page(page);
765 return 1;
766 }
767 return 0;
768}
769
770static int prep_new_page(struct page *page, int order, gfp_t gfp_flags)
771{
772 int i;
773
774 for (i = 0; i < (1 << order); i++) {
775 struct page *p = page + i;
776 if (unlikely(check_new_page(p)))
777 return 1;
778 }
779
780 set_page_private(page, 0);
781 set_page_refcounted(page);
782
783 arch_alloc_page(page, order);
784 kernel_map_pages(page, 1 << order, 1);
785
786 if (gfp_flags & __GFP_ZERO)
787 prep_zero_page(page, order, gfp_flags);
788
789 if (order && (gfp_flags & __GFP_COMP))
790 prep_compound_page(page, order);
791
792 return 0;
793}
794
795/*
796 * Go through the free lists for the given migratetype and remove
797 * the smallest available page from the freelists
798 */
799static inline
800struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
801 int migratetype)
802{
803 unsigned int current_order;
804 struct free_area * area;
805 struct page *page;
806
807 /* Find a page of the appropriate size in the preferred list */
808 for (current_order = order; current_order < MAX_ORDER; ++current_order) {
809 area = &(zone->free_area[current_order]);
810 if (list_empty(&area->free_list[migratetype]))
811 continue;
812
813 page = list_entry(area->free_list[migratetype].next,
814 struct page, lru);
815 list_del(&page->lru);
816 rmv_page_order(page);
817 area->nr_free--;
818 expand(zone, page, order, current_order, area, migratetype);
819 return page;
820 }
821
822 return NULL;
823}
824
825
826/*
827 * This array describes the order lists are fallen back to when
828 * the free lists for the desirable migrate type are depleted
829 */
830static int fallbacks[MIGRATE_TYPES][MIGRATE_TYPES-1] = {
831 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
832 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE, MIGRATE_RESERVE },
833 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE, MIGRATE_RESERVE },
834 [MIGRATE_RESERVE] = { MIGRATE_RESERVE, MIGRATE_RESERVE, MIGRATE_RESERVE }, /* Never used */
835};
836
837/*
838 * Move the free pages in a range to the free lists of the requested type.
839 * Note that start_page and end_pages are not aligned on a pageblock
840 * boundary. If alignment is required, use move_freepages_block()
841 */
842static int move_freepages(struct zone *zone,
843 struct page *start_page, struct page *end_page,
844 int migratetype)
845{
846 struct page *page;
847 unsigned long order;
848 int pages_moved = 0;
849
850#ifndef CONFIG_HOLES_IN_ZONE
851 /*
852 * page_zone is not safe to call in this context when
853 * CONFIG_HOLES_IN_ZONE is set. This bug check is probably redundant
854 * anyway as we check zone boundaries in move_freepages_block().
855 * Remove at a later date when no bug reports exist related to
856 * grouping pages by mobility
857 */
858 BUG_ON(page_zone(start_page) != page_zone(end_page));
859#endif
860
861 for (page = start_page; page <= end_page;) {
862 /* Make sure we are not inadvertently changing nodes */
863 VM_BUG_ON(page_to_nid(page) != zone_to_nid(zone));
864
865 if (!pfn_valid_within(page_to_pfn(page))) {
866 page++;
867 continue;
868 }
869
870 if (!PageBuddy(page)) {
871 page++;
872 continue;
873 }
874
875 order = page_order(page);
876 list_move(&page->lru,
877 &zone->free_area[order].free_list[migratetype]);
878 page += 1 << order;
879 pages_moved += 1 << order;
880 }
881
882 return pages_moved;
883}
884
885static int move_freepages_block(struct zone *zone, struct page *page,
886 int migratetype)
887{
888 unsigned long start_pfn, end_pfn;
889 struct page *start_page, *end_page;
890
891 start_pfn = page_to_pfn(page);
892 start_pfn = start_pfn & ~(pageblock_nr_pages-1);
893 start_page = pfn_to_page(start_pfn);
894 end_page = start_page + pageblock_nr_pages - 1;
895 end_pfn = start_pfn + pageblock_nr_pages - 1;
896
897 /* Do not cross zone boundaries */
898 if (start_pfn < zone->zone_start_pfn)
899 start_page = page;
900 if (end_pfn >= zone->zone_start_pfn + zone->spanned_pages)
901 return 0;
902
903 return move_freepages(zone, start_page, end_page, migratetype);
904}
905
906static void change_pageblock_range(struct page *pageblock_page,
907 int start_order, int migratetype)
908{
909 int nr_pageblocks = 1 << (start_order - pageblock_order);
910
911 while (nr_pageblocks--) {
912 set_pageblock_migratetype(pageblock_page, migratetype);
913 pageblock_page += pageblock_nr_pages;
914 }
915}
916
917/* Remove an element from the buddy allocator from the fallback list */
918static inline struct page *
919__rmqueue_fallback(struct zone *zone, int order, int start_migratetype)
920{
921 struct free_area * area;
922 int current_order;
923 struct page *page;
924 int migratetype, i;
925
926 /* Find the largest possible block of pages in the other list */
927 for (current_order = MAX_ORDER-1; current_order >= order;
928 --current_order) {
929 for (i = 0; i < MIGRATE_TYPES - 1; i++) {
930 migratetype = fallbacks[start_migratetype][i];
931
932 /* MIGRATE_RESERVE handled later if necessary */
933 if (migratetype == MIGRATE_RESERVE)
934 continue;
935
936 area = &(zone->free_area[current_order]);
937 if (list_empty(&area->free_list[migratetype]))
938 continue;
939
940 page = list_entry(area->free_list[migratetype].next,
941 struct page, lru);
942 area->nr_free--;
943
944 /*
945 * If breaking a large block of pages, move all free
946 * pages to the preferred allocation list. If falling
947 * back for a reclaimable kernel allocation, be more
948 * aggressive about taking ownership of free pages
949 */
950 if (unlikely(current_order >= (pageblock_order >> 1)) ||
951 start_migratetype == MIGRATE_RECLAIMABLE ||
952 page_group_by_mobility_disabled) {
953 unsigned long pages;
954 pages = move_freepages_block(zone, page,
955 start_migratetype);
956
957 /* Claim the whole block if over half of it is free */
958 if (pages >= (1 << (pageblock_order-1)) ||
959 page_group_by_mobility_disabled)
960 set_pageblock_migratetype(page,
961 start_migratetype);
962
963 migratetype = start_migratetype;
964 }
965
966 /* Remove the page from the freelists */
967 list_del(&page->lru);
968 rmv_page_order(page);
969
970 /* Take ownership for orders >= pageblock_order */
971 if (current_order >= pageblock_order)
972 change_pageblock_range(page, current_order,
973 start_migratetype);
974
975 expand(zone, page, order, current_order, area, migratetype);
976
977 trace_mm_page_alloc_extfrag(page, order, current_order,
978 start_migratetype, migratetype);
979
980 return page;
981 }
982 }
983
984 return NULL;
985}
986
987/*
988 * Do the hard work of removing an element from the buddy allocator.
989 * Call me with the zone->lock already held.
990 */
991static struct page *__rmqueue(struct zone *zone, unsigned int order,
992 int migratetype)
993{
994 struct page *page;
995
996retry_reserve:
997 page = __rmqueue_smallest(zone, order, migratetype);
998
999 if (unlikely(!page) && migratetype != MIGRATE_RESERVE) {
1000 page = __rmqueue_fallback(zone, order, migratetype);
1001
1002 /*
1003 * Use MIGRATE_RESERVE rather than fail an allocation. goto
1004 * is used because __rmqueue_smallest is an inline function
1005 * and we want just one call site
1006 */
1007 if (!page) {
1008 migratetype = MIGRATE_RESERVE;
1009 goto retry_reserve;
1010 }
1011 }
1012
1013 trace_mm_page_alloc_zone_locked(page, order, migratetype);
1014 return page;
1015}
1016
1017/*
1018 * Obtain a specified number of elements from the buddy allocator, all under
1019 * a single hold of the lock, for efficiency. Add them to the supplied list.
1020 * Returns the number of new pages which were placed at *list.
1021 */
1022static int rmqueue_bulk(struct zone *zone, unsigned int order,
1023 unsigned long count, struct list_head *list,
1024 int migratetype, int cold)
1025{
1026 int i;
1027
1028 spin_lock(&zone->lock);
1029 for (i = 0; i < count; ++i) {
1030 struct page *page = __rmqueue(zone, order, migratetype);
1031 if (unlikely(page == NULL))
1032 break;
1033
1034 /*
1035 * Split buddy pages returned by expand() are received here
1036 * in physical page order. The page is added to the callers and
1037 * list and the list head then moves forward. From the callers
1038 * perspective, the linked list is ordered by page number in
1039 * some conditions. This is useful for IO devices that can
1040 * merge IO requests if the physical pages are ordered
1041 * properly.
1042 */
1043 if (likely(cold == 0))
1044 list_add(&page->lru, list);
1045 else
1046 list_add_tail(&page->lru, list);
1047 set_page_private(page, migratetype);
1048 list = &page->lru;
1049 }
1050 __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
1051 spin_unlock(&zone->lock);
1052 return i;
1053}
1054
1055#ifdef CONFIG_NUMA
1056/*
1057 * Called from the vmstat counter updater to drain pagesets of this
1058 * currently executing processor on remote nodes after they have
1059 * expired.
1060 *
1061 * Note that this function must be called with the thread pinned to
1062 * a single processor.
1063 */
1064void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
1065{
1066 unsigned long flags;
1067 int to_drain;
1068
1069 local_irq_save(flags);
1070 if (pcp->count >= pcp->batch)
1071 to_drain = pcp->batch;
1072 else
1073 to_drain = pcp->count;
1074 free_pcppages_bulk(zone, to_drain, pcp);
1075 pcp->count -= to_drain;
1076 local_irq_restore(flags);
1077}
1078#endif
1079
1080/*
1081 * Drain pages of the indicated processor.
1082 *
1083 * The processor must either be the current processor and the
1084 * thread pinned to the current processor or a processor that
1085 * is not online.
1086 */
1087static void drain_pages(unsigned int cpu)
1088{
1089 unsigned long flags;
1090 struct zone *zone;
1091
1092 for_each_populated_zone(zone) {
1093 struct per_cpu_pageset *pset;
1094 struct per_cpu_pages *pcp;
1095
1096 local_irq_save(flags);
1097 pset = per_cpu_ptr(zone->pageset, cpu);
1098
1099 pcp = &pset->pcp;
1100 if (pcp->count) {
1101 free_pcppages_bulk(zone, pcp->count, pcp);
1102 pcp->count = 0;
1103 }
1104 local_irq_restore(flags);
1105 }
1106}
1107
1108/*
1109 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
1110 */
1111void drain_local_pages(void *arg)
1112{
1113 drain_pages(smp_processor_id());
1114}
1115
1116/*
1117 * Spill all the per-cpu pages from all CPUs back into the buddy allocator
1118 */
1119void drain_all_pages(void)
1120{
1121 on_each_cpu(drain_local_pages, NULL, 1);
1122}
1123
1124#ifdef CONFIG_HIBERNATION
1125
1126void mark_free_pages(struct zone *zone)
1127{
1128 unsigned long pfn, max_zone_pfn;
1129 unsigned long flags;
1130 int order, t;
1131 struct list_head *curr;
1132
1133 if (!zone->spanned_pages)
1134 return;
1135
1136 spin_lock_irqsave(&zone->lock, flags);
1137
1138 max_zone_pfn = zone->zone_start_pfn + zone->spanned_pages;
1139 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++)
1140 if (pfn_valid(pfn)) {
1141 struct page *page = pfn_to_page(pfn);
1142
1143 if (!swsusp_page_is_forbidden(page))
1144 swsusp_unset_page_free(page);
1145 }
1146
1147 for_each_migratetype_order(order, t) {
1148 list_for_each(curr, &zone->free_area[order].free_list[t]) {
1149 unsigned long i;
1150
1151 pfn = page_to_pfn(list_entry(curr, struct page, lru));
1152 for (i = 0; i < (1UL << order); i++)
1153 swsusp_set_page_free(pfn_to_page(pfn + i));
1154 }
1155 }
1156 spin_unlock_irqrestore(&zone->lock, flags);
1157}
1158#endif /* CONFIG_PM */
1159
1160/*
1161 * Free a 0-order page
1162 * cold == 1 ? free a cold page : free a hot page
1163 */
1164void free_hot_cold_page(struct page *page, int cold)
1165{
1166 struct zone *zone = page_zone(page);
1167 struct per_cpu_pages *pcp;
1168 unsigned long flags;
1169 int migratetype;
1170 int wasMlocked = __TestClearPageMlocked(page);
1171
1172 if (!free_pages_prepare(page, 0))
1173 return;
1174
1175 migratetype = get_pageblock_migratetype(page);
1176 set_page_private(page, migratetype);
1177 local_irq_save(flags);
1178 if (unlikely(wasMlocked))
1179 free_page_mlock(page);
1180 __count_vm_event(PGFREE);
1181
1182 /*
1183 * We only track unmovable, reclaimable and movable on pcp lists.
1184 * Free ISOLATE pages back to the allocator because they are being
1185 * offlined but treat RESERVE as movable pages so we can get those
1186 * areas back if necessary. Otherwise, we may have to free
1187 * excessively into the page allocator
1188 */
1189 if (migratetype >= MIGRATE_PCPTYPES) {
1190 if (unlikely(migratetype == MIGRATE_ISOLATE)) {
1191 free_one_page(zone, page, 0, migratetype);
1192 goto out;
1193 }
1194 migratetype = MIGRATE_MOVABLE;
1195 }
1196
1197 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1198 if (cold)
1199 list_add_tail(&page->lru, &pcp->lists[migratetype]);
1200 else
1201 list_add(&page->lru, &pcp->lists[migratetype]);
1202 pcp->count++;
1203 if (pcp->count >= pcp->high) {
1204 free_pcppages_bulk(zone, pcp->batch, pcp);
1205 pcp->count -= pcp->batch;
1206 }
1207
1208out:
1209 local_irq_restore(flags);
1210}
1211
1212/*
1213 * split_page takes a non-compound higher-order page, and splits it into
1214 * n (1<<order) sub-pages: page[0..n]
1215 * Each sub-page must be freed individually.
1216 *
1217 * Note: this is probably too low level an operation for use in drivers.
1218 * Please consult with lkml before using this in your driver.
1219 */
1220void split_page(struct page *page, unsigned int order)
1221{
1222 int i;
1223
1224 VM_BUG_ON(PageCompound(page));
1225 VM_BUG_ON(!page_count(page));
1226
1227#ifdef CONFIG_KMEMCHECK
1228 /*
1229 * Split shadow pages too, because free(page[0]) would
1230 * otherwise free the whole shadow.
1231 */
1232 if (kmemcheck_page_is_tracked(page))
1233 split_page(virt_to_page(page[0].shadow), order);
1234#endif
1235
1236 for (i = 1; i < (1 << order); i++)
1237 set_page_refcounted(page + i);
1238}
1239
1240/*
1241 * Similar to split_page except the page is already free. As this is only
1242 * being used for migration, the migratetype of the block also changes.
1243 * As this is called with interrupts disabled, the caller is responsible
1244 * for calling arch_alloc_page() and kernel_map_page() after interrupts
1245 * are enabled.
1246 *
1247 * Note: this is probably too low level an operation for use in drivers.
1248 * Please consult with lkml before using this in your driver.
1249 */
1250int split_free_page(struct page *page)
1251{
1252 unsigned int order;
1253 unsigned long watermark;
1254 struct zone *zone;
1255
1256 BUG_ON(!PageBuddy(page));
1257
1258 zone = page_zone(page);
1259 order = page_order(page);
1260
1261 /* Obey watermarks as if the page was being allocated */
1262 watermark = low_wmark_pages(zone) + (1 << order);
1263 if (!zone_watermark_ok(zone, 0, watermark, 0, 0))
1264 return 0;
1265
1266 /* Remove page from free list */
1267 list_del(&page->lru);
1268 zone->free_area[order].nr_free--;
1269 rmv_page_order(page);
1270 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1UL << order));
1271
1272 /* Split into individual pages */
1273 set_page_refcounted(page);
1274 split_page(page, order);
1275
1276 if (order >= pageblock_order - 1) {
1277 struct page *endpage = page + (1 << order) - 1;
1278 for (; page < endpage; page += pageblock_nr_pages)
1279 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
1280 }
1281
1282 return 1 << order;
1283}
1284
1285/*
1286 * Really, prep_compound_page() should be called from __rmqueue_bulk(). But
1287 * we cheat by calling it from here, in the order > 0 path. Saves a branch
1288 * or two.
1289 */
1290static inline
1291struct page *buffered_rmqueue(struct zone *preferred_zone,
1292 struct zone *zone, int order, gfp_t gfp_flags,
1293 int migratetype)
1294{
1295 unsigned long flags;
1296 struct page *page;
1297 int cold = !!(gfp_flags & __GFP_COLD);
1298
1299again:
1300 if (likely(order == 0)) {
1301 struct per_cpu_pages *pcp;
1302 struct list_head *list;
1303
1304 local_irq_save(flags);
1305 pcp = &this_cpu_ptr(zone->pageset)->pcp;
1306 list = &pcp->lists[migratetype];
1307 if (list_empty(list)) {
1308 pcp->count += rmqueue_bulk(zone, 0,
1309 pcp->batch, list,
1310 migratetype, cold);
1311 if (unlikely(list_empty(list)))
1312 goto failed;
1313 }
1314
1315 if (cold)
1316 page = list_entry(list->prev, struct page, lru);
1317 else
1318 page = list_entry(list->next, struct page, lru);
1319
1320 list_del(&page->lru);
1321 pcp->count--;
1322 } else {
1323 if (unlikely(gfp_flags & __GFP_NOFAIL)) {
1324 /*
1325 * __GFP_NOFAIL is not to be used in new code.
1326 *
1327 * All __GFP_NOFAIL callers should be fixed so that they
1328 * properly detect and handle allocation failures.
1329 *
1330 * We most definitely don't want callers attempting to
1331 * allocate greater than order-1 page units with
1332 * __GFP_NOFAIL.
1333 */
1334 WARN_ON_ONCE(order > 1);
1335 }
1336 spin_lock_irqsave(&zone->lock, flags);
1337 page = __rmqueue(zone, order, migratetype);
1338 spin_unlock(&zone->lock);
1339 if (!page)
1340 goto failed;
1341 __mod_zone_page_state(zone, NR_FREE_PAGES, -(1 << order));
1342 }
1343
1344 __count_zone_vm_events(PGALLOC, zone, 1 << order);
1345 zone_statistics(preferred_zone, zone, gfp_flags);
1346 local_irq_restore(flags);
1347
1348 VM_BUG_ON(bad_range(zone, page));
1349 if (prep_new_page(page, order, gfp_flags))
1350 goto again;
1351 return page;
1352
1353failed:
1354 local_irq_restore(flags);
1355 return NULL;
1356}
1357
1358/* The ALLOC_WMARK bits are used as an index to zone->watermark */
1359#define ALLOC_WMARK_MIN WMARK_MIN
1360#define ALLOC_WMARK_LOW WMARK_LOW
1361#define ALLOC_WMARK_HIGH WMARK_HIGH
1362#define ALLOC_NO_WATERMARKS 0x04 /* don't check watermarks at all */
1363
1364/* Mask to get the watermark bits */
1365#define ALLOC_WMARK_MASK (ALLOC_NO_WATERMARKS-1)
1366
1367#define ALLOC_HARDER 0x10 /* try to alloc harder */
1368#define ALLOC_HIGH 0x20 /* __GFP_HIGH set */
1369#define ALLOC_CPUSET 0x40 /* check for correct cpuset */
1370
1371#ifdef CONFIG_FAIL_PAGE_ALLOC
1372
1373static struct {
1374 struct fault_attr attr;
1375
1376 u32 ignore_gfp_highmem;
1377 u32 ignore_gfp_wait;
1378 u32 min_order;
1379} fail_page_alloc = {
1380 .attr = FAULT_ATTR_INITIALIZER,
1381 .ignore_gfp_wait = 1,
1382 .ignore_gfp_highmem = 1,
1383 .min_order = 1,
1384};
1385
1386static int __init setup_fail_page_alloc(char *str)
1387{
1388 return setup_fault_attr(&fail_page_alloc.attr, str);
1389}
1390__setup("fail_page_alloc=", setup_fail_page_alloc);
1391
1392static int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1393{
1394 if (order < fail_page_alloc.min_order)
1395 return 0;
1396 if (gfp_mask & __GFP_NOFAIL)
1397 return 0;
1398 if (fail_page_alloc.ignore_gfp_highmem && (gfp_mask & __GFP_HIGHMEM))
1399 return 0;
1400 if (fail_page_alloc.ignore_gfp_wait && (gfp_mask & __GFP_WAIT))
1401 return 0;
1402
1403 return should_fail(&fail_page_alloc.attr, 1 << order);
1404}
1405
1406#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
1407
1408static int __init fail_page_alloc_debugfs(void)
1409{
1410 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
1411 struct dentry *dir;
1412
1413 dir = fault_create_debugfs_attr("fail_page_alloc", NULL,
1414 &fail_page_alloc.attr);
1415 if (IS_ERR(dir))
1416 return PTR_ERR(dir);
1417
1418 if (!debugfs_create_bool("ignore-gfp-wait", mode, dir,
1419 &fail_page_alloc.ignore_gfp_wait))
1420 goto fail;
1421 if (!debugfs_create_bool("ignore-gfp-highmem", mode, dir,
1422 &fail_page_alloc.ignore_gfp_highmem))
1423 goto fail;
1424 if (!debugfs_create_u32("min-order", mode, dir,
1425 &fail_page_alloc.min_order))
1426 goto fail;
1427
1428 return 0;
1429fail:
1430 debugfs_remove_recursive(dir);
1431
1432 return -ENOMEM;
1433}
1434
1435late_initcall(fail_page_alloc_debugfs);
1436
1437#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
1438
1439#else /* CONFIG_FAIL_PAGE_ALLOC */
1440
1441static inline int should_fail_alloc_page(gfp_t gfp_mask, unsigned int order)
1442{
1443 return 0;
1444}
1445
1446#endif /* CONFIG_FAIL_PAGE_ALLOC */
1447
1448/*
1449 * Return true if free pages are above 'mark'. This takes into account the order
1450 * of the allocation.
1451 */
1452static bool __zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1453 int classzone_idx, int alloc_flags, long free_pages)
1454{
1455 /* free_pages my go negative - that's OK */
1456 long min = mark;
1457 int o;
1458
1459 free_pages -= (1 << order) + 1;
1460 if (alloc_flags & ALLOC_HIGH)
1461 min -= min / 2;
1462 if (alloc_flags & ALLOC_HARDER)
1463 min -= min / 4;
1464
1465 if (free_pages <= min + z->lowmem_reserve[classzone_idx])
1466 return false;
1467 for (o = 0; o < order; o++) {
1468 /* At the next order, this order's pages become unavailable */
1469 free_pages -= z->free_area[o].nr_free << o;
1470
1471 /* Require fewer higher order pages to be free */
1472 min >>= 1;
1473
1474 if (free_pages <= min)
1475 return false;
1476 }
1477 return true;
1478}
1479
1480bool zone_watermark_ok(struct zone *z, int order, unsigned long mark,
1481 int classzone_idx, int alloc_flags)
1482{
1483 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1484 zone_page_state(z, NR_FREE_PAGES));
1485}
1486
1487bool zone_watermark_ok_safe(struct zone *z, int order, unsigned long mark,
1488 int classzone_idx, int alloc_flags)
1489{
1490 long free_pages = zone_page_state(z, NR_FREE_PAGES);
1491
1492 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
1493 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
1494
1495 return __zone_watermark_ok(z, order, mark, classzone_idx, alloc_flags,
1496 free_pages);
1497}
1498
1499#ifdef CONFIG_NUMA
1500/*
1501 * zlc_setup - Setup for "zonelist cache". Uses cached zone data to
1502 * skip over zones that are not allowed by the cpuset, or that have
1503 * been recently (in last second) found to be nearly full. See further
1504 * comments in mmzone.h. Reduces cache footprint of zonelist scans
1505 * that have to skip over a lot of full or unallowed zones.
1506 *
1507 * If the zonelist cache is present in the passed in zonelist, then
1508 * returns a pointer to the allowed node mask (either the current
1509 * tasks mems_allowed, or node_states[N_HIGH_MEMORY].)
1510 *
1511 * If the zonelist cache is not available for this zonelist, does
1512 * nothing and returns NULL.
1513 *
1514 * If the fullzones BITMAP in the zonelist cache is stale (more than
1515 * a second since last zap'd) then we zap it out (clear its bits.)
1516 *
1517 * We hold off even calling zlc_setup, until after we've checked the
1518 * first zone in the zonelist, on the theory that most allocations will
1519 * be satisfied from that first zone, so best to examine that zone as
1520 * quickly as we can.
1521 */
1522static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1523{
1524 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1525 nodemask_t *allowednodes; /* zonelist_cache approximation */
1526
1527 zlc = zonelist->zlcache_ptr;
1528 if (!zlc)
1529 return NULL;
1530
1531 if (time_after(jiffies, zlc->last_full_zap + HZ)) {
1532 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1533 zlc->last_full_zap = jiffies;
1534 }
1535
1536 allowednodes = !in_interrupt() && (alloc_flags & ALLOC_CPUSET) ?
1537 &cpuset_current_mems_allowed :
1538 &node_states[N_HIGH_MEMORY];
1539 return allowednodes;
1540}
1541
1542/*
1543 * Given 'z' scanning a zonelist, run a couple of quick checks to see
1544 * if it is worth looking at further for free memory:
1545 * 1) Check that the zone isn't thought to be full (doesn't have its
1546 * bit set in the zonelist_cache fullzones BITMAP).
1547 * 2) Check that the zones node (obtained from the zonelist_cache
1548 * z_to_n[] mapping) is allowed in the passed in allowednodes mask.
1549 * Return true (non-zero) if zone is worth looking at further, or
1550 * else return false (zero) if it is not.
1551 *
1552 * This check -ignores- the distinction between various watermarks,
1553 * such as GFP_HIGH, GFP_ATOMIC, PF_MEMALLOC, ... If a zone is
1554 * found to be full for any variation of these watermarks, it will
1555 * be considered full for up to one second by all requests, unless
1556 * we are so low on memory on all allowed nodes that we are forced
1557 * into the second scan of the zonelist.
1558 *
1559 * In the second scan we ignore this zonelist cache and exactly
1560 * apply the watermarks to all zones, even it is slower to do so.
1561 * We are low on memory in the second scan, and should leave no stone
1562 * unturned looking for a free page.
1563 */
1564static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1565 nodemask_t *allowednodes)
1566{
1567 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1568 int i; /* index of *z in zonelist zones */
1569 int n; /* node that zone *z is on */
1570
1571 zlc = zonelist->zlcache_ptr;
1572 if (!zlc)
1573 return 1;
1574
1575 i = z - zonelist->_zonerefs;
1576 n = zlc->z_to_n[i];
1577
1578 /* This zone is worth trying if it is allowed but not full */
1579 return node_isset(n, *allowednodes) && !test_bit(i, zlc->fullzones);
1580}
1581
1582/*
1583 * Given 'z' scanning a zonelist, set the corresponding bit in
1584 * zlc->fullzones, so that subsequent attempts to allocate a page
1585 * from that zone don't waste time re-examining it.
1586 */
1587static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1588{
1589 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1590 int i; /* index of *z in zonelist zones */
1591
1592 zlc = zonelist->zlcache_ptr;
1593 if (!zlc)
1594 return;
1595
1596 i = z - zonelist->_zonerefs;
1597
1598 set_bit(i, zlc->fullzones);
1599}
1600
1601/*
1602 * clear all zones full, called after direct reclaim makes progress so that
1603 * a zone that was recently full is not skipped over for up to a second
1604 */
1605static void zlc_clear_zones_full(struct zonelist *zonelist)
1606{
1607 struct zonelist_cache *zlc; /* cached zonelist speedup info */
1608
1609 zlc = zonelist->zlcache_ptr;
1610 if (!zlc)
1611 return;
1612
1613 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
1614}
1615
1616#else /* CONFIG_NUMA */
1617
1618static nodemask_t *zlc_setup(struct zonelist *zonelist, int alloc_flags)
1619{
1620 return NULL;
1621}
1622
1623static int zlc_zone_worth_trying(struct zonelist *zonelist, struct zoneref *z,
1624 nodemask_t *allowednodes)
1625{
1626 return 1;
1627}
1628
1629static void zlc_mark_zone_full(struct zonelist *zonelist, struct zoneref *z)
1630{
1631}
1632
1633static void zlc_clear_zones_full(struct zonelist *zonelist)
1634{
1635}
1636#endif /* CONFIG_NUMA */
1637
1638/*
1639 * get_page_from_freelist goes through the zonelist trying to allocate
1640 * a page.
1641 */
1642static struct page *
1643get_page_from_freelist(gfp_t gfp_mask, nodemask_t *nodemask, unsigned int order,
1644 struct zonelist *zonelist, int high_zoneidx, int alloc_flags,
1645 struct zone *preferred_zone, int migratetype)
1646{
1647 struct zoneref *z;
1648 struct page *page = NULL;
1649 int classzone_idx;
1650 struct zone *zone;
1651 nodemask_t *allowednodes = NULL;/* zonelist_cache approximation */
1652 int zlc_active = 0; /* set if using zonelist_cache */
1653 int did_zlc_setup = 0; /* just call zlc_setup() one time */
1654
1655 classzone_idx = zone_idx(preferred_zone);
1656zonelist_scan:
1657 /*
1658 * Scan zonelist, looking for a zone with enough free.
1659 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
1660 */
1661 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1662 high_zoneidx, nodemask) {
1663 if (NUMA_BUILD && zlc_active &&
1664 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1665 continue;
1666 if ((alloc_flags & ALLOC_CPUSET) &&
1667 !cpuset_zone_allowed_softwall(zone, gfp_mask))
1668 continue;
1669
1670 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
1671 if (!(alloc_flags & ALLOC_NO_WATERMARKS)) {
1672 unsigned long mark;
1673 int ret;
1674
1675 mark = zone->watermark[alloc_flags & ALLOC_WMARK_MASK];
1676 if (zone_watermark_ok(zone, order, mark,
1677 classzone_idx, alloc_flags))
1678 goto try_this_zone;
1679
1680 if (NUMA_BUILD && !did_zlc_setup && nr_online_nodes > 1) {
1681 /*
1682 * we do zlc_setup if there are multiple nodes
1683 * and before considering the first zone allowed
1684 * by the cpuset.
1685 */
1686 allowednodes = zlc_setup(zonelist, alloc_flags);
1687 zlc_active = 1;
1688 did_zlc_setup = 1;
1689 }
1690
1691 if (zone_reclaim_mode == 0)
1692 goto this_zone_full;
1693
1694 /*
1695 * As we may have just activated ZLC, check if the first
1696 * eligible zone has failed zone_reclaim recently.
1697 */
1698 if (NUMA_BUILD && zlc_active &&
1699 !zlc_zone_worth_trying(zonelist, z, allowednodes))
1700 continue;
1701
1702 ret = zone_reclaim(zone, gfp_mask, order);
1703 switch (ret) {
1704 case ZONE_RECLAIM_NOSCAN:
1705 /* did not scan */
1706 continue;
1707 case ZONE_RECLAIM_FULL:
1708 /* scanned but unreclaimable */
1709 continue;
1710 default:
1711 /* did we reclaim enough */
1712 if (!zone_watermark_ok(zone, order, mark,
1713 classzone_idx, alloc_flags))
1714 goto this_zone_full;
1715 }
1716 }
1717
1718try_this_zone:
1719 page = buffered_rmqueue(preferred_zone, zone, order,
1720 gfp_mask, migratetype);
1721 if (page)
1722 break;
1723this_zone_full:
1724 if (NUMA_BUILD)
1725 zlc_mark_zone_full(zonelist, z);
1726 }
1727
1728 if (unlikely(NUMA_BUILD && page == NULL && zlc_active)) {
1729 /* Disable zlc cache for second zonelist scan */
1730 zlc_active = 0;
1731 goto zonelist_scan;
1732 }
1733 return page;
1734}
1735
1736/*
1737 * Large machines with many possible nodes should not always dump per-node
1738 * meminfo in irq context.
1739 */
1740static inline bool should_suppress_show_mem(void)
1741{
1742 bool ret = false;
1743
1744#if NODES_SHIFT > 8
1745 ret = in_interrupt();
1746#endif
1747 return ret;
1748}
1749
1750static DEFINE_RATELIMIT_STATE(nopage_rs,
1751 DEFAULT_RATELIMIT_INTERVAL,
1752 DEFAULT_RATELIMIT_BURST);
1753
1754void warn_alloc_failed(gfp_t gfp_mask, int order, const char *fmt, ...)
1755{
1756 va_list args;
1757 unsigned int filter = SHOW_MEM_FILTER_NODES;
1758
1759 if ((gfp_mask & __GFP_NOWARN) || !__ratelimit(&nopage_rs))
1760 return;
1761
1762 /*
1763 * This documents exceptions given to allocations in certain
1764 * contexts that are allowed to allocate outside current's set
1765 * of allowed nodes.
1766 */
1767 if (!(gfp_mask & __GFP_NOMEMALLOC))
1768 if (test_thread_flag(TIF_MEMDIE) ||
1769 (current->flags & (PF_MEMALLOC | PF_EXITING)))
1770 filter &= ~SHOW_MEM_FILTER_NODES;
1771 if (in_interrupt() || !(gfp_mask & __GFP_WAIT))
1772 filter &= ~SHOW_MEM_FILTER_NODES;
1773
1774 if (fmt) {
1775 printk(KERN_WARNING);
1776 va_start(args, fmt);
1777 vprintk(fmt, args);
1778 va_end(args);
1779 }
1780
1781 pr_warning("%s: page allocation failure: order:%d, mode:0x%x\n",
1782 current->comm, order, gfp_mask);
1783
1784 dump_stack();
1785 if (!should_suppress_show_mem())
1786 show_mem(filter);
1787}
1788
1789static inline int
1790should_alloc_retry(gfp_t gfp_mask, unsigned int order,
1791 unsigned long pages_reclaimed)
1792{
1793 /* Do not loop if specifically requested */
1794 if (gfp_mask & __GFP_NORETRY)
1795 return 0;
1796
1797 /*
1798 * In this implementation, order <= PAGE_ALLOC_COSTLY_ORDER
1799 * means __GFP_NOFAIL, but that may not be true in other
1800 * implementations.
1801 */
1802 if (order <= PAGE_ALLOC_COSTLY_ORDER)
1803 return 1;
1804
1805 /*
1806 * For order > PAGE_ALLOC_COSTLY_ORDER, if __GFP_REPEAT is
1807 * specified, then we retry until we no longer reclaim any pages
1808 * (above), or we've reclaimed an order of pages at least as
1809 * large as the allocation's order. In both cases, if the
1810 * allocation still fails, we stop retrying.
1811 */
1812 if (gfp_mask & __GFP_REPEAT && pages_reclaimed < (1 << order))
1813 return 1;
1814
1815 /*
1816 * Don't let big-order allocations loop unless the caller
1817 * explicitly requests that.
1818 */
1819 if (gfp_mask & __GFP_NOFAIL)
1820 return 1;
1821
1822 return 0;
1823}
1824
1825static inline struct page *
1826__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
1827 struct zonelist *zonelist, enum zone_type high_zoneidx,
1828 nodemask_t *nodemask, struct zone *preferred_zone,
1829 int migratetype)
1830{
1831 struct page *page;
1832
1833 /* Acquire the OOM killer lock for the zones in zonelist */
1834 if (!try_set_zonelist_oom(zonelist, gfp_mask)) {
1835 schedule_timeout_uninterruptible(1);
1836 return NULL;
1837 }
1838
1839 /*
1840 * Go through the zonelist yet one more time, keep very high watermark
1841 * here, this is only to catch a parallel oom killing, we must fail if
1842 * we're still under heavy pressure.
1843 */
1844 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask,
1845 order, zonelist, high_zoneidx,
1846 ALLOC_WMARK_HIGH|ALLOC_CPUSET,
1847 preferred_zone, migratetype);
1848 if (page)
1849 goto out;
1850
1851 if (!(gfp_mask & __GFP_NOFAIL)) {
1852 /* The OOM killer will not help higher order allocs */
1853 if (order > PAGE_ALLOC_COSTLY_ORDER)
1854 goto out;
1855 /* The OOM killer does not needlessly kill tasks for lowmem */
1856 if (high_zoneidx < ZONE_NORMAL)
1857 goto out;
1858 /*
1859 * GFP_THISNODE contains __GFP_NORETRY and we never hit this.
1860 * Sanity check for bare calls of __GFP_THISNODE, not real OOM.
1861 * The caller should handle page allocation failure by itself if
1862 * it specifies __GFP_THISNODE.
1863 * Note: Hugepage uses it but will hit PAGE_ALLOC_COSTLY_ORDER.
1864 */
1865 if (gfp_mask & __GFP_THISNODE)
1866 goto out;
1867 }
1868 /* Exhausted what can be done so it's blamo time */
1869 out_of_memory(zonelist, gfp_mask, order, nodemask);
1870
1871out:
1872 clear_zonelist_oom(zonelist, gfp_mask);
1873 return page;
1874}
1875
1876#ifdef CONFIG_COMPACTION
1877/* Try memory compaction for high-order allocations before reclaim */
1878static struct page *
1879__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1880 struct zonelist *zonelist, enum zone_type high_zoneidx,
1881 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1882 int migratetype, unsigned long *did_some_progress,
1883 bool sync_migration)
1884{
1885 struct page *page;
1886
1887 if (!order || compaction_deferred(preferred_zone))
1888 return NULL;
1889
1890 current->flags |= PF_MEMALLOC;
1891 *did_some_progress = try_to_compact_pages(zonelist, order, gfp_mask,
1892 nodemask, sync_migration);
1893 current->flags &= ~PF_MEMALLOC;
1894 if (*did_some_progress != COMPACT_SKIPPED) {
1895
1896 /* Page migration frees to the PCP lists but we want merging */
1897 drain_pages(get_cpu());
1898 put_cpu();
1899
1900 page = get_page_from_freelist(gfp_mask, nodemask,
1901 order, zonelist, high_zoneidx,
1902 alloc_flags, preferred_zone,
1903 migratetype);
1904 if (page) {
1905 preferred_zone->compact_considered = 0;
1906 preferred_zone->compact_defer_shift = 0;
1907 count_vm_event(COMPACTSUCCESS);
1908 return page;
1909 }
1910
1911 /*
1912 * It's bad if compaction run occurs and fails.
1913 * The most likely reason is that pages exist,
1914 * but not enough to satisfy watermarks.
1915 */
1916 count_vm_event(COMPACTFAIL);
1917 defer_compaction(preferred_zone);
1918
1919 cond_resched();
1920 }
1921
1922 return NULL;
1923}
1924#else
1925static inline struct page *
1926__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
1927 struct zonelist *zonelist, enum zone_type high_zoneidx,
1928 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1929 int migratetype, unsigned long *did_some_progress,
1930 bool sync_migration)
1931{
1932 return NULL;
1933}
1934#endif /* CONFIG_COMPACTION */
1935
1936/* The really slow allocator path where we enter direct reclaim */
1937static inline struct page *
1938__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
1939 struct zonelist *zonelist, enum zone_type high_zoneidx,
1940 nodemask_t *nodemask, int alloc_flags, struct zone *preferred_zone,
1941 int migratetype, unsigned long *did_some_progress)
1942{
1943 struct page *page = NULL;
1944 struct reclaim_state reclaim_state;
1945 bool drained = false;
1946
1947 cond_resched();
1948
1949 /* We now go into synchronous reclaim */
1950 cpuset_memory_pressure_bump();
1951 current->flags |= PF_MEMALLOC;
1952 lockdep_set_current_reclaim_state(gfp_mask);
1953 reclaim_state.reclaimed_slab = 0;
1954 current->reclaim_state = &reclaim_state;
1955
1956 *did_some_progress = try_to_free_pages(zonelist, order, gfp_mask, nodemask);
1957
1958 current->reclaim_state = NULL;
1959 lockdep_clear_current_reclaim_state();
1960 current->flags &= ~PF_MEMALLOC;
1961
1962 cond_resched();
1963
1964 if (unlikely(!(*did_some_progress)))
1965 return NULL;
1966
1967 /* After successful reclaim, reconsider all zones for allocation */
1968 if (NUMA_BUILD)
1969 zlc_clear_zones_full(zonelist);
1970
1971retry:
1972 page = get_page_from_freelist(gfp_mask, nodemask, order,
1973 zonelist, high_zoneidx,
1974 alloc_flags, preferred_zone,
1975 migratetype);
1976
1977 /*
1978 * If an allocation failed after direct reclaim, it could be because
1979 * pages are pinned on the per-cpu lists. Drain them and try again
1980 */
1981 if (!page && !drained) {
1982 drain_all_pages();
1983 drained = true;
1984 goto retry;
1985 }
1986
1987 return page;
1988}
1989
1990/*
1991 * This is called in the allocator slow-path if the allocation request is of
1992 * sufficient urgency to ignore watermarks and take other desperate measures
1993 */
1994static inline struct page *
1995__alloc_pages_high_priority(gfp_t gfp_mask, unsigned int order,
1996 struct zonelist *zonelist, enum zone_type high_zoneidx,
1997 nodemask_t *nodemask, struct zone *preferred_zone,
1998 int migratetype)
1999{
2000 struct page *page;
2001
2002 do {
2003 page = get_page_from_freelist(gfp_mask, nodemask, order,
2004 zonelist, high_zoneidx, ALLOC_NO_WATERMARKS,
2005 preferred_zone, migratetype);
2006
2007 if (!page && gfp_mask & __GFP_NOFAIL)
2008 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2009 } while (!page && (gfp_mask & __GFP_NOFAIL));
2010
2011 return page;
2012}
2013
2014static inline
2015void wake_all_kswapd(unsigned int order, struct zonelist *zonelist,
2016 enum zone_type high_zoneidx,
2017 enum zone_type classzone_idx)
2018{
2019 struct zoneref *z;
2020 struct zone *zone;
2021
2022 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx)
2023 wakeup_kswapd(zone, order, classzone_idx);
2024}
2025
2026static inline int
2027gfp_to_alloc_flags(gfp_t gfp_mask)
2028{
2029 int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
2030 const gfp_t wait = gfp_mask & __GFP_WAIT;
2031
2032 /* __GFP_HIGH is assumed to be the same as ALLOC_HIGH to save a branch. */
2033 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_HIGH);
2034
2035 /*
2036 * The caller may dip into page reserves a bit more if the caller
2037 * cannot run direct reclaim, or if the caller has realtime scheduling
2038 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
2039 * set both ALLOC_HARDER (!wait) and ALLOC_HIGH (__GFP_HIGH).
2040 */
2041 alloc_flags |= (__force int) (gfp_mask & __GFP_HIGH);
2042
2043 if (!wait) {
2044 /*
2045 * Not worth trying to allocate harder for
2046 * __GFP_NOMEMALLOC even if it can't schedule.
2047 */
2048 if (!(gfp_mask & __GFP_NOMEMALLOC))
2049 alloc_flags |= ALLOC_HARDER;
2050 /*
2051 * Ignore cpuset if GFP_ATOMIC (!wait) rather than fail alloc.
2052 * See also cpuset_zone_allowed() comment in kernel/cpuset.c.
2053 */
2054 alloc_flags &= ~ALLOC_CPUSET;
2055 } else if (unlikely(rt_task(current)) && !in_interrupt())
2056 alloc_flags |= ALLOC_HARDER;
2057
2058 if (likely(!(gfp_mask & __GFP_NOMEMALLOC))) {
2059 if (!in_interrupt() &&
2060 ((current->flags & PF_MEMALLOC) ||
2061 unlikely(test_thread_flag(TIF_MEMDIE))))
2062 alloc_flags |= ALLOC_NO_WATERMARKS;
2063 }
2064
2065 return alloc_flags;
2066}
2067
2068static inline struct page *
2069__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
2070 struct zonelist *zonelist, enum zone_type high_zoneidx,
2071 nodemask_t *nodemask, struct zone *preferred_zone,
2072 int migratetype)
2073{
2074 const gfp_t wait = gfp_mask & __GFP_WAIT;
2075 struct page *page = NULL;
2076 int alloc_flags;
2077 unsigned long pages_reclaimed = 0;
2078 unsigned long did_some_progress;
2079 bool sync_migration = false;
2080
2081 /*
2082 * In the slowpath, we sanity check order to avoid ever trying to
2083 * reclaim >= MAX_ORDER areas which will never succeed. Callers may
2084 * be using allocators in order of preference for an area that is
2085 * too large.
2086 */
2087 if (order >= MAX_ORDER) {
2088 WARN_ON_ONCE(!(gfp_mask & __GFP_NOWARN));
2089 return NULL;
2090 }
2091
2092 /*
2093 * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
2094 * __GFP_NOWARN set) should not cause reclaim since the subsystem
2095 * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
2096 * using a larger set of nodes after it has established that the
2097 * allowed per node queues are empty and that nodes are
2098 * over allocated.
2099 */
2100 if (NUMA_BUILD && (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
2101 goto nopage;
2102
2103restart:
2104 if (!(gfp_mask & __GFP_NO_KSWAPD))
2105 wake_all_kswapd(order, zonelist, high_zoneidx,
2106 zone_idx(preferred_zone));
2107
2108 /*
2109 * OK, we're below the kswapd watermark and have kicked background
2110 * reclaim. Now things get more complex, so set up alloc_flags according
2111 * to how we want to proceed.
2112 */
2113 alloc_flags = gfp_to_alloc_flags(gfp_mask);
2114
2115 /*
2116 * Find the true preferred zone if the allocation is unconstrained by
2117 * cpusets.
2118 */
2119 if (!(alloc_flags & ALLOC_CPUSET) && !nodemask)
2120 first_zones_zonelist(zonelist, high_zoneidx, NULL,
2121 &preferred_zone);
2122
2123rebalance:
2124 /* This is the last chance, in general, before the goto nopage. */
2125 page = get_page_from_freelist(gfp_mask, nodemask, order, zonelist,
2126 high_zoneidx, alloc_flags & ~ALLOC_NO_WATERMARKS,
2127 preferred_zone, migratetype);
2128 if (page)
2129 goto got_pg;
2130
2131 /* Allocate without watermarks if the context allows */
2132 if (alloc_flags & ALLOC_NO_WATERMARKS) {
2133 page = __alloc_pages_high_priority(gfp_mask, order,
2134 zonelist, high_zoneidx, nodemask,
2135 preferred_zone, migratetype);
2136 if (page)
2137 goto got_pg;
2138 }
2139
2140 /* Atomic allocations - we can't balance anything */
2141 if (!wait)
2142 goto nopage;
2143
2144 /* Avoid recursion of direct reclaim */
2145 if (current->flags & PF_MEMALLOC)
2146 goto nopage;
2147
2148 /* Avoid allocations with no watermarks from looping endlessly */
2149 if (test_thread_flag(TIF_MEMDIE) && !(gfp_mask & __GFP_NOFAIL))
2150 goto nopage;
2151
2152 /*
2153 * Try direct compaction. The first pass is asynchronous. Subsequent
2154 * attempts after direct reclaim are synchronous
2155 */
2156 page = __alloc_pages_direct_compact(gfp_mask, order,
2157 zonelist, high_zoneidx,
2158 nodemask,
2159 alloc_flags, preferred_zone,
2160 migratetype, &did_some_progress,
2161 sync_migration);
2162 if (page)
2163 goto got_pg;
2164 sync_migration = true;
2165
2166 /* Try direct reclaim and then allocating */
2167 page = __alloc_pages_direct_reclaim(gfp_mask, order,
2168 zonelist, high_zoneidx,
2169 nodemask,
2170 alloc_flags, preferred_zone,
2171 migratetype, &did_some_progress);
2172 if (page)
2173 goto got_pg;
2174
2175 /*
2176 * If we failed to make any progress reclaiming, then we are
2177 * running out of options and have to consider going OOM
2178 */
2179 if (!did_some_progress) {
2180 if ((gfp_mask & __GFP_FS) && !(gfp_mask & __GFP_NORETRY)) {
2181 if (oom_killer_disabled)
2182 goto nopage;
2183 page = __alloc_pages_may_oom(gfp_mask, order,
2184 zonelist, high_zoneidx,
2185 nodemask, preferred_zone,
2186 migratetype);
2187 if (page)
2188 goto got_pg;
2189
2190 if (!(gfp_mask & __GFP_NOFAIL)) {
2191 /*
2192 * The oom killer is not called for high-order
2193 * allocations that may fail, so if no progress
2194 * is being made, there are no other options and
2195 * retrying is unlikely to help.
2196 */
2197 if (order > PAGE_ALLOC_COSTLY_ORDER)
2198 goto nopage;
2199 /*
2200 * The oom killer is not called for lowmem
2201 * allocations to prevent needlessly killing
2202 * innocent tasks.
2203 */
2204 if (high_zoneidx < ZONE_NORMAL)
2205 goto nopage;
2206 }
2207
2208 goto restart;
2209 }
2210 }
2211
2212 /* Check if we should retry the allocation */
2213 pages_reclaimed += did_some_progress;
2214 if (should_alloc_retry(gfp_mask, order, pages_reclaimed)) {
2215 /* Wait for some write requests to complete then retry */
2216 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/50);
2217 goto rebalance;
2218 } else {
2219 /*
2220 * High-order allocations do not necessarily loop after
2221 * direct reclaim and reclaim/compaction depends on compaction
2222 * being called after reclaim so call directly if necessary
2223 */
2224 page = __alloc_pages_direct_compact(gfp_mask, order,
2225 zonelist, high_zoneidx,
2226 nodemask,
2227 alloc_flags, preferred_zone,
2228 migratetype, &did_some_progress,
2229 sync_migration);
2230 if (page)
2231 goto got_pg;
2232 }
2233
2234nopage:
2235 warn_alloc_failed(gfp_mask, order, NULL);
2236 return page;
2237got_pg:
2238 if (kmemcheck_enabled)
2239 kmemcheck_pagealloc_alloc(page, order, gfp_mask);
2240 return page;
2241
2242}
2243
2244/*
2245 * This is the 'heart' of the zoned buddy allocator.
2246 */
2247struct page *
2248__alloc_pages_nodemask(gfp_t gfp_mask, unsigned int order,
2249 struct zonelist *zonelist, nodemask_t *nodemask)
2250{
2251 enum zone_type high_zoneidx = gfp_zone(gfp_mask);
2252 struct zone *preferred_zone;
2253 struct page *page;
2254 int migratetype = allocflags_to_migratetype(gfp_mask);
2255
2256 gfp_mask &= gfp_allowed_mask;
2257
2258 lockdep_trace_alloc(gfp_mask);
2259
2260 might_sleep_if(gfp_mask & __GFP_WAIT);
2261
2262 if (should_fail_alloc_page(gfp_mask, order))
2263 return NULL;
2264
2265 /*
2266 * Check the zones suitable for the gfp_mask contain at least one
2267 * valid zone. It's possible to have an empty zonelist as a result
2268 * of GFP_THISNODE and a memoryless node
2269 */
2270 if (unlikely(!zonelist->_zonerefs->zone))
2271 return NULL;
2272
2273 get_mems_allowed();
2274 /* The preferred zone is used for statistics later */
2275 first_zones_zonelist(zonelist, high_zoneidx,
2276 nodemask ? : &cpuset_current_mems_allowed,
2277 &preferred_zone);
2278 if (!preferred_zone) {
2279 put_mems_allowed();
2280 return NULL;
2281 }
2282
2283 /* First allocation attempt */
2284 page = get_page_from_freelist(gfp_mask|__GFP_HARDWALL, nodemask, order,
2285 zonelist, high_zoneidx, ALLOC_WMARK_LOW|ALLOC_CPUSET,
2286 preferred_zone, migratetype);
2287 if (unlikely(!page))
2288 page = __alloc_pages_slowpath(gfp_mask, order,
2289 zonelist, high_zoneidx, nodemask,
2290 preferred_zone, migratetype);
2291 put_mems_allowed();
2292
2293 trace_mm_page_alloc(page, order, gfp_mask, migratetype);
2294 return page;
2295}
2296EXPORT_SYMBOL(__alloc_pages_nodemask);
2297
2298/*
2299 * Common helper functions.
2300 */
2301unsigned long __get_free_pages(gfp_t gfp_mask, unsigned int order)
2302{
2303 struct page *page;
2304
2305 /*
2306 * __get_free_pages() returns a 32-bit address, which cannot represent
2307 * a highmem page
2308 */
2309 VM_BUG_ON((gfp_mask & __GFP_HIGHMEM) != 0);
2310
2311 page = alloc_pages(gfp_mask, order);
2312 if (!page)
2313 return 0;
2314 return (unsigned long) page_address(page);
2315}
2316EXPORT_SYMBOL(__get_free_pages);
2317
2318unsigned long get_zeroed_page(gfp_t gfp_mask)
2319{
2320 return __get_free_pages(gfp_mask | __GFP_ZERO, 0);
2321}
2322EXPORT_SYMBOL(get_zeroed_page);
2323
2324void __pagevec_free(struct pagevec *pvec)
2325{
2326 int i = pagevec_count(pvec);
2327
2328 while (--i >= 0) {
2329 trace_mm_pagevec_free(pvec->pages[i], pvec->cold);
2330 free_hot_cold_page(pvec->pages[i], pvec->cold);
2331 }
2332}
2333
2334void __free_pages(struct page *page, unsigned int order)
2335{
2336 if (put_page_testzero(page)) {
2337 if (order == 0)
2338 free_hot_cold_page(page, 0);
2339 else
2340 __free_pages_ok(page, order);
2341 }
2342}
2343
2344EXPORT_SYMBOL(__free_pages);
2345
2346void free_pages(unsigned long addr, unsigned int order)
2347{
2348 if (addr != 0) {
2349 VM_BUG_ON(!virt_addr_valid((void *)addr));
2350 __free_pages(virt_to_page((void *)addr), order);
2351 }
2352}
2353
2354EXPORT_SYMBOL(free_pages);
2355
2356static void *make_alloc_exact(unsigned long addr, unsigned order, size_t size)
2357{
2358 if (addr) {
2359 unsigned long alloc_end = addr + (PAGE_SIZE << order);
2360 unsigned long used = addr + PAGE_ALIGN(size);
2361
2362 split_page(virt_to_page((void *)addr), order);
2363 while (used < alloc_end) {
2364 free_page(used);
2365 used += PAGE_SIZE;
2366 }
2367 }
2368 return (void *)addr;
2369}
2370
2371/**
2372 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
2373 * @size: the number of bytes to allocate
2374 * @gfp_mask: GFP flags for the allocation
2375 *
2376 * This function is similar to alloc_pages(), except that it allocates the
2377 * minimum number of pages to satisfy the request. alloc_pages() can only
2378 * allocate memory in power-of-two pages.
2379 *
2380 * This function is also limited by MAX_ORDER.
2381 *
2382 * Memory allocated by this function must be released by free_pages_exact().
2383 */
2384void *alloc_pages_exact(size_t size, gfp_t gfp_mask)
2385{
2386 unsigned int order = get_order(size);
2387 unsigned long addr;
2388
2389 addr = __get_free_pages(gfp_mask, order);
2390 return make_alloc_exact(addr, order, size);
2391}
2392EXPORT_SYMBOL(alloc_pages_exact);
2393
2394/**
2395 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
2396 * pages on a node.
2397 * @nid: the preferred node ID where memory should be allocated
2398 * @size: the number of bytes to allocate
2399 * @gfp_mask: GFP flags for the allocation
2400 *
2401 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
2402 * back.
2403 * Note this is not alloc_pages_exact_node() which allocates on a specific node,
2404 * but is not exact.
2405 */
2406void *alloc_pages_exact_nid(int nid, size_t size, gfp_t gfp_mask)
2407{
2408 unsigned order = get_order(size);
2409 struct page *p = alloc_pages_node(nid, gfp_mask, order);
2410 if (!p)
2411 return NULL;
2412 return make_alloc_exact((unsigned long)page_address(p), order, size);
2413}
2414EXPORT_SYMBOL(alloc_pages_exact_nid);
2415
2416/**
2417 * free_pages_exact - release memory allocated via alloc_pages_exact()
2418 * @virt: the value returned by alloc_pages_exact.
2419 * @size: size of allocation, same value as passed to alloc_pages_exact().
2420 *
2421 * Release the memory allocated by a previous call to alloc_pages_exact.
2422 */
2423void free_pages_exact(void *virt, size_t size)
2424{
2425 unsigned long addr = (unsigned long)virt;
2426 unsigned long end = addr + PAGE_ALIGN(size);
2427
2428 while (addr < end) {
2429 free_page(addr);
2430 addr += PAGE_SIZE;
2431 }
2432}
2433EXPORT_SYMBOL(free_pages_exact);
2434
2435static unsigned int nr_free_zone_pages(int offset)
2436{
2437 struct zoneref *z;
2438 struct zone *zone;
2439
2440 /* Just pick one node, since fallback list is circular */
2441 unsigned int sum = 0;
2442
2443 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
2444
2445 for_each_zone_zonelist(zone, z, zonelist, offset) {
2446 unsigned long size = zone->present_pages;
2447 unsigned long high = high_wmark_pages(zone);
2448 if (size > high)
2449 sum += size - high;
2450 }
2451
2452 return sum;
2453}
2454
2455/*
2456 * Amount of free RAM allocatable within ZONE_DMA and ZONE_NORMAL
2457 */
2458unsigned int nr_free_buffer_pages(void)
2459{
2460 return nr_free_zone_pages(gfp_zone(GFP_USER));
2461}
2462EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
2463
2464/*
2465 * Amount of free RAM allocatable within all zones
2466 */
2467unsigned int nr_free_pagecache_pages(void)
2468{
2469 return nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
2470}
2471
2472static inline void show_node(struct zone *zone)
2473{
2474 if (NUMA_BUILD)
2475 printk("Node %d ", zone_to_nid(zone));
2476}
2477
2478void si_meminfo(struct sysinfo *val)
2479{
2480 val->totalram = totalram_pages;
2481 val->sharedram = 0;
2482 val->freeram = global_page_state(NR_FREE_PAGES);
2483 val->bufferram = nr_blockdev_pages();
2484 val->totalhigh = totalhigh_pages;
2485 val->freehigh = nr_free_highpages();
2486 val->mem_unit = PAGE_SIZE;
2487}
2488
2489EXPORT_SYMBOL(si_meminfo);
2490
2491#ifdef CONFIG_NUMA
2492void si_meminfo_node(struct sysinfo *val, int nid)
2493{
2494 pg_data_t *pgdat = NODE_DATA(nid);
2495
2496 val->totalram = pgdat->node_present_pages;
2497 val->freeram = node_page_state(nid, NR_FREE_PAGES);
2498#ifdef CONFIG_HIGHMEM
2499 val->totalhigh = pgdat->node_zones[ZONE_HIGHMEM].present_pages;
2500 val->freehigh = zone_page_state(&pgdat->node_zones[ZONE_HIGHMEM],
2501 NR_FREE_PAGES);
2502#else
2503 val->totalhigh = 0;
2504 val->freehigh = 0;
2505#endif
2506 val->mem_unit = PAGE_SIZE;
2507}
2508#endif
2509
2510/*
2511 * Determine whether the node should be displayed or not, depending on whether
2512 * SHOW_MEM_FILTER_NODES was passed to show_free_areas().
2513 */
2514bool skip_free_areas_node(unsigned int flags, int nid)
2515{
2516 bool ret = false;
2517
2518 if (!(flags & SHOW_MEM_FILTER_NODES))
2519 goto out;
2520
2521 get_mems_allowed();
2522 ret = !node_isset(nid, cpuset_current_mems_allowed);
2523 put_mems_allowed();
2524out:
2525 return ret;
2526}
2527
2528#define K(x) ((x) << (PAGE_SHIFT-10))
2529
2530/*
2531 * Show free area list (used inside shift_scroll-lock stuff)
2532 * We also calculate the percentage fragmentation. We do this by counting the
2533 * memory on each free list with the exception of the first item on the list.
2534 * Suppresses nodes that are not allowed by current's cpuset if
2535 * SHOW_MEM_FILTER_NODES is passed.
2536 */
2537void show_free_areas(unsigned int filter)
2538{
2539 int cpu;
2540 struct zone *zone;
2541
2542 for_each_populated_zone(zone) {
2543 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2544 continue;
2545 show_node(zone);
2546 printk("%s per-cpu:\n", zone->name);
2547
2548 for_each_online_cpu(cpu) {
2549 struct per_cpu_pageset *pageset;
2550
2551 pageset = per_cpu_ptr(zone->pageset, cpu);
2552
2553 printk("CPU %4d: hi:%5d, btch:%4d usd:%4d\n",
2554 cpu, pageset->pcp.high,
2555 pageset->pcp.batch, pageset->pcp.count);
2556 }
2557 }
2558
2559 printk("active_anon:%lu inactive_anon:%lu isolated_anon:%lu\n"
2560 " active_file:%lu inactive_file:%lu isolated_file:%lu\n"
2561 " unevictable:%lu"
2562 " dirty:%lu writeback:%lu unstable:%lu\n"
2563 " free:%lu slab_reclaimable:%lu slab_unreclaimable:%lu\n"
2564 " mapped:%lu shmem:%lu pagetables:%lu bounce:%lu\n",
2565 global_page_state(NR_ACTIVE_ANON),
2566 global_page_state(NR_INACTIVE_ANON),
2567 global_page_state(NR_ISOLATED_ANON),
2568 global_page_state(NR_ACTIVE_FILE),
2569 global_page_state(NR_INACTIVE_FILE),
2570 global_page_state(NR_ISOLATED_FILE),
2571 global_page_state(NR_UNEVICTABLE),
2572 global_page_state(NR_FILE_DIRTY),
2573 global_page_state(NR_WRITEBACK),
2574 global_page_state(NR_UNSTABLE_NFS),
2575 global_page_state(NR_FREE_PAGES),
2576 global_page_state(NR_SLAB_RECLAIMABLE),
2577 global_page_state(NR_SLAB_UNRECLAIMABLE),
2578 global_page_state(NR_FILE_MAPPED),
2579 global_page_state(NR_SHMEM),
2580 global_page_state(NR_PAGETABLE),
2581 global_page_state(NR_BOUNCE));
2582
2583 for_each_populated_zone(zone) {
2584 int i;
2585
2586 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2587 continue;
2588 show_node(zone);
2589 printk("%s"
2590 " free:%lukB"
2591 " min:%lukB"
2592 " low:%lukB"
2593 " high:%lukB"
2594 " active_anon:%lukB"
2595 " inactive_anon:%lukB"
2596 " active_file:%lukB"
2597 " inactive_file:%lukB"
2598 " unevictable:%lukB"
2599 " isolated(anon):%lukB"
2600 " isolated(file):%lukB"
2601 " present:%lukB"
2602 " mlocked:%lukB"
2603 " dirty:%lukB"
2604 " writeback:%lukB"
2605 " mapped:%lukB"
2606 " shmem:%lukB"
2607 " slab_reclaimable:%lukB"
2608 " slab_unreclaimable:%lukB"
2609 " kernel_stack:%lukB"
2610 " pagetables:%lukB"
2611 " unstable:%lukB"
2612 " bounce:%lukB"
2613 " writeback_tmp:%lukB"
2614 " pages_scanned:%lu"
2615 " all_unreclaimable? %s"
2616 "\n",
2617 zone->name,
2618 K(zone_page_state(zone, NR_FREE_PAGES)),
2619 K(min_wmark_pages(zone)),
2620 K(low_wmark_pages(zone)),
2621 K(high_wmark_pages(zone)),
2622 K(zone_page_state(zone, NR_ACTIVE_ANON)),
2623 K(zone_page_state(zone, NR_INACTIVE_ANON)),
2624 K(zone_page_state(zone, NR_ACTIVE_FILE)),
2625 K(zone_page_state(zone, NR_INACTIVE_FILE)),
2626 K(zone_page_state(zone, NR_UNEVICTABLE)),
2627 K(zone_page_state(zone, NR_ISOLATED_ANON)),
2628 K(zone_page_state(zone, NR_ISOLATED_FILE)),
2629 K(zone->present_pages),
2630 K(zone_page_state(zone, NR_MLOCK)),
2631 K(zone_page_state(zone, NR_FILE_DIRTY)),
2632 K(zone_page_state(zone, NR_WRITEBACK)),
2633 K(zone_page_state(zone, NR_FILE_MAPPED)),
2634 K(zone_page_state(zone, NR_SHMEM)),
2635 K(zone_page_state(zone, NR_SLAB_RECLAIMABLE)),
2636 K(zone_page_state(zone, NR_SLAB_UNRECLAIMABLE)),
2637 zone_page_state(zone, NR_KERNEL_STACK) *
2638 THREAD_SIZE / 1024,
2639 K(zone_page_state(zone, NR_PAGETABLE)),
2640 K(zone_page_state(zone, NR_UNSTABLE_NFS)),
2641 K(zone_page_state(zone, NR_BOUNCE)),
2642 K(zone_page_state(zone, NR_WRITEBACK_TEMP)),
2643 zone->pages_scanned,
2644 (zone->all_unreclaimable ? "yes" : "no")
2645 );
2646 printk("lowmem_reserve[]:");
2647 for (i = 0; i < MAX_NR_ZONES; i++)
2648 printk(" %lu", zone->lowmem_reserve[i]);
2649 printk("\n");
2650 }
2651
2652 for_each_populated_zone(zone) {
2653 unsigned long nr[MAX_ORDER], flags, order, total = 0;
2654
2655 if (skip_free_areas_node(filter, zone_to_nid(zone)))
2656 continue;
2657 show_node(zone);
2658 printk("%s: ", zone->name);
2659
2660 spin_lock_irqsave(&zone->lock, flags);
2661 for (order = 0; order < MAX_ORDER; order++) {
2662 nr[order] = zone->free_area[order].nr_free;
2663 total += nr[order] << order;
2664 }
2665 spin_unlock_irqrestore(&zone->lock, flags);
2666 for (order = 0; order < MAX_ORDER; order++)
2667 printk("%lu*%lukB ", nr[order], K(1UL) << order);
2668 printk("= %lukB\n", K(total));
2669 }
2670
2671 printk("%ld total pagecache pages\n", global_page_state(NR_FILE_PAGES));
2672
2673 show_swap_cache_info();
2674}
2675
2676static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
2677{
2678 zoneref->zone = zone;
2679 zoneref->zone_idx = zone_idx(zone);
2680}
2681
2682/*
2683 * Builds allocation fallback zone lists.
2684 *
2685 * Add all populated zones of a node to the zonelist.
2686 */
2687static int build_zonelists_node(pg_data_t *pgdat, struct zonelist *zonelist,
2688 int nr_zones, enum zone_type zone_type)
2689{
2690 struct zone *zone;
2691
2692 BUG_ON(zone_type >= MAX_NR_ZONES);
2693 zone_type++;
2694
2695 do {
2696 zone_type--;
2697 zone = pgdat->node_zones + zone_type;
2698 if (populated_zone(zone)) {
2699 zoneref_set_zone(zone,
2700 &zonelist->_zonerefs[nr_zones++]);
2701 check_highest_zone(zone_type);
2702 }
2703
2704 } while (zone_type);
2705 return nr_zones;
2706}
2707
2708
2709/*
2710 * zonelist_order:
2711 * 0 = automatic detection of better ordering.
2712 * 1 = order by ([node] distance, -zonetype)
2713 * 2 = order by (-zonetype, [node] distance)
2714 *
2715 * If not NUMA, ZONELIST_ORDER_ZONE and ZONELIST_ORDER_NODE will create
2716 * the same zonelist. So only NUMA can configure this param.
2717 */
2718#define ZONELIST_ORDER_DEFAULT 0
2719#define ZONELIST_ORDER_NODE 1
2720#define ZONELIST_ORDER_ZONE 2
2721
2722/* zonelist order in the kernel.
2723 * set_zonelist_order() will set this to NODE or ZONE.
2724 */
2725static int current_zonelist_order = ZONELIST_ORDER_DEFAULT;
2726static char zonelist_order_name[3][8] = {"Default", "Node", "Zone"};
2727
2728
2729#ifdef CONFIG_NUMA
2730/* The value user specified ....changed by config */
2731static int user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2732/* string for sysctl */
2733#define NUMA_ZONELIST_ORDER_LEN 16
2734char numa_zonelist_order[16] = "default";
2735
2736/*
2737 * interface for configure zonelist ordering.
2738 * command line option "numa_zonelist_order"
2739 * = "[dD]efault - default, automatic configuration.
2740 * = "[nN]ode - order by node locality, then by zone within node
2741 * = "[zZ]one - order by zone, then by locality within zone
2742 */
2743
2744static int __parse_numa_zonelist_order(char *s)
2745{
2746 if (*s == 'd' || *s == 'D') {
2747 user_zonelist_order = ZONELIST_ORDER_DEFAULT;
2748 } else if (*s == 'n' || *s == 'N') {
2749 user_zonelist_order = ZONELIST_ORDER_NODE;
2750 } else if (*s == 'z' || *s == 'Z') {
2751 user_zonelist_order = ZONELIST_ORDER_ZONE;
2752 } else {
2753 printk(KERN_WARNING
2754 "Ignoring invalid numa_zonelist_order value: "
2755 "%s\n", s);
2756 return -EINVAL;
2757 }
2758 return 0;
2759}
2760
2761static __init int setup_numa_zonelist_order(char *s)
2762{
2763 int ret;
2764
2765 if (!s)
2766 return 0;
2767
2768 ret = __parse_numa_zonelist_order(s);
2769 if (ret == 0)
2770 strlcpy(numa_zonelist_order, s, NUMA_ZONELIST_ORDER_LEN);
2771
2772 return ret;
2773}
2774early_param("numa_zonelist_order", setup_numa_zonelist_order);
2775
2776/*
2777 * sysctl handler for numa_zonelist_order
2778 */
2779int numa_zonelist_order_handler(ctl_table *table, int write,
2780 void __user *buffer, size_t *length,
2781 loff_t *ppos)
2782{
2783 char saved_string[NUMA_ZONELIST_ORDER_LEN];
2784 int ret;
2785 static DEFINE_MUTEX(zl_order_mutex);
2786
2787 mutex_lock(&zl_order_mutex);
2788 if (write)
2789 strcpy(saved_string, (char*)table->data);
2790 ret = proc_dostring(table, write, buffer, length, ppos);
2791 if (ret)
2792 goto out;
2793 if (write) {
2794 int oldval = user_zonelist_order;
2795 if (__parse_numa_zonelist_order((char*)table->data)) {
2796 /*
2797 * bogus value. restore saved string
2798 */
2799 strncpy((char*)table->data, saved_string,
2800 NUMA_ZONELIST_ORDER_LEN);
2801 user_zonelist_order = oldval;
2802 } else if (oldval != user_zonelist_order) {
2803 mutex_lock(&zonelists_mutex);
2804 build_all_zonelists(NULL);
2805 mutex_unlock(&zonelists_mutex);
2806 }
2807 }
2808out:
2809 mutex_unlock(&zl_order_mutex);
2810 return ret;
2811}
2812
2813
2814#define MAX_NODE_LOAD (nr_online_nodes)
2815static int node_load[MAX_NUMNODES];
2816
2817/**
2818 * find_next_best_node - find the next node that should appear in a given node's fallback list
2819 * @node: node whose fallback list we're appending
2820 * @used_node_mask: nodemask_t of already used nodes
2821 *
2822 * We use a number of factors to determine which is the next node that should
2823 * appear on a given node's fallback list. The node should not have appeared
2824 * already in @node's fallback list, and it should be the next closest node
2825 * according to the distance array (which contains arbitrary distance values
2826 * from each node to each node in the system), and should also prefer nodes
2827 * with no CPUs, since presumably they'll have very little allocation pressure
2828 * on them otherwise.
2829 * It returns -1 if no node is found.
2830 */
2831static int find_next_best_node(int node, nodemask_t *used_node_mask)
2832{
2833 int n, val;
2834 int min_val = INT_MAX;
2835 int best_node = -1;
2836 const struct cpumask *tmp = cpumask_of_node(0);
2837
2838 /* Use the local node if we haven't already */
2839 if (!node_isset(node, *used_node_mask)) {
2840 node_set(node, *used_node_mask);
2841 return node;
2842 }
2843
2844 for_each_node_state(n, N_HIGH_MEMORY) {
2845
2846 /* Don't want a node to appear more than once */
2847 if (node_isset(n, *used_node_mask))
2848 continue;
2849
2850 /* Use the distance array to find the distance */
2851 val = node_distance(node, n);
2852
2853 /* Penalize nodes under us ("prefer the next node") */
2854 val += (n < node);
2855
2856 /* Give preference to headless and unused nodes */
2857 tmp = cpumask_of_node(n);
2858 if (!cpumask_empty(tmp))
2859 val += PENALTY_FOR_NODE_WITH_CPUS;
2860
2861 /* Slight preference for less loaded node */
2862 val *= (MAX_NODE_LOAD*MAX_NUMNODES);
2863 val += node_load[n];
2864
2865 if (val < min_val) {
2866 min_val = val;
2867 best_node = n;
2868 }
2869 }
2870
2871 if (best_node >= 0)
2872 node_set(best_node, *used_node_mask);
2873
2874 return best_node;
2875}
2876
2877
2878/*
2879 * Build zonelists ordered by node and zones within node.
2880 * This results in maximum locality--normal zone overflows into local
2881 * DMA zone, if any--but risks exhausting DMA zone.
2882 */
2883static void build_zonelists_in_node_order(pg_data_t *pgdat, int node)
2884{
2885 int j;
2886 struct zonelist *zonelist;
2887
2888 zonelist = &pgdat->node_zonelists[0];
2889 for (j = 0; zonelist->_zonerefs[j].zone != NULL; j++)
2890 ;
2891 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
2892 MAX_NR_ZONES - 1);
2893 zonelist->_zonerefs[j].zone = NULL;
2894 zonelist->_zonerefs[j].zone_idx = 0;
2895}
2896
2897/*
2898 * Build gfp_thisnode zonelists
2899 */
2900static void build_thisnode_zonelists(pg_data_t *pgdat)
2901{
2902 int j;
2903 struct zonelist *zonelist;
2904
2905 zonelist = &pgdat->node_zonelists[1];
2906 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
2907 zonelist->_zonerefs[j].zone = NULL;
2908 zonelist->_zonerefs[j].zone_idx = 0;
2909}
2910
2911/*
2912 * Build zonelists ordered by zone and nodes within zones.
2913 * This results in conserving DMA zone[s] until all Normal memory is
2914 * exhausted, but results in overflowing to remote node while memory
2915 * may still exist in local DMA zone.
2916 */
2917static int node_order[MAX_NUMNODES];
2918
2919static void build_zonelists_in_zone_order(pg_data_t *pgdat, int nr_nodes)
2920{
2921 int pos, j, node;
2922 int zone_type; /* needs to be signed */
2923 struct zone *z;
2924 struct zonelist *zonelist;
2925
2926 zonelist = &pgdat->node_zonelists[0];
2927 pos = 0;
2928 for (zone_type = MAX_NR_ZONES - 1; zone_type >= 0; zone_type--) {
2929 for (j = 0; j < nr_nodes; j++) {
2930 node = node_order[j];
2931 z = &NODE_DATA(node)->node_zones[zone_type];
2932 if (populated_zone(z)) {
2933 zoneref_set_zone(z,
2934 &zonelist->_zonerefs[pos++]);
2935 check_highest_zone(zone_type);
2936 }
2937 }
2938 }
2939 zonelist->_zonerefs[pos].zone = NULL;
2940 zonelist->_zonerefs[pos].zone_idx = 0;
2941}
2942
2943static int default_zonelist_order(void)
2944{
2945 int nid, zone_type;
2946 unsigned long low_kmem_size,total_size;
2947 struct zone *z;
2948 int average_size;
2949 /*
2950 * ZONE_DMA and ZONE_DMA32 can be very small area in the system.
2951 * If they are really small and used heavily, the system can fall
2952 * into OOM very easily.
2953 * This function detect ZONE_DMA/DMA32 size and configures zone order.
2954 */
2955 /* Is there ZONE_NORMAL ? (ex. ppc has only DMA zone..) */
2956 low_kmem_size = 0;
2957 total_size = 0;
2958 for_each_online_node(nid) {
2959 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2960 z = &NODE_DATA(nid)->node_zones[zone_type];
2961 if (populated_zone(z)) {
2962 if (zone_type < ZONE_NORMAL)
2963 low_kmem_size += z->present_pages;
2964 total_size += z->present_pages;
2965 } else if (zone_type == ZONE_NORMAL) {
2966 /*
2967 * If any node has only lowmem, then node order
2968 * is preferred to allow kernel allocations
2969 * locally; otherwise, they can easily infringe
2970 * on other nodes when there is an abundance of
2971 * lowmem available to allocate from.
2972 */
2973 return ZONELIST_ORDER_NODE;
2974 }
2975 }
2976 }
2977 if (!low_kmem_size || /* there are no DMA area. */
2978 low_kmem_size > total_size/2) /* DMA/DMA32 is big. */
2979 return ZONELIST_ORDER_NODE;
2980 /*
2981 * look into each node's config.
2982 * If there is a node whose DMA/DMA32 memory is very big area on
2983 * local memory, NODE_ORDER may be suitable.
2984 */
2985 average_size = total_size /
2986 (nodes_weight(node_states[N_HIGH_MEMORY]) + 1);
2987 for_each_online_node(nid) {
2988 low_kmem_size = 0;
2989 total_size = 0;
2990 for (zone_type = 0; zone_type < MAX_NR_ZONES; zone_type++) {
2991 z = &NODE_DATA(nid)->node_zones[zone_type];
2992 if (populated_zone(z)) {
2993 if (zone_type < ZONE_NORMAL)
2994 low_kmem_size += z->present_pages;
2995 total_size += z->present_pages;
2996 }
2997 }
2998 if (low_kmem_size &&
2999 total_size > average_size && /* ignore small node */
3000 low_kmem_size > total_size * 70/100)
3001 return ZONELIST_ORDER_NODE;
3002 }
3003 return ZONELIST_ORDER_ZONE;
3004}
3005
3006static void set_zonelist_order(void)
3007{
3008 if (user_zonelist_order == ZONELIST_ORDER_DEFAULT)
3009 current_zonelist_order = default_zonelist_order();
3010 else
3011 current_zonelist_order = user_zonelist_order;
3012}
3013
3014static void build_zonelists(pg_data_t *pgdat)
3015{
3016 int j, node, load;
3017 enum zone_type i;
3018 nodemask_t used_mask;
3019 int local_node, prev_node;
3020 struct zonelist *zonelist;
3021 int order = current_zonelist_order;
3022
3023 /* initialize zonelists */
3024 for (i = 0; i < MAX_ZONELISTS; i++) {
3025 zonelist = pgdat->node_zonelists + i;
3026 zonelist->_zonerefs[0].zone = NULL;
3027 zonelist->_zonerefs[0].zone_idx = 0;
3028 }
3029
3030 /* NUMA-aware ordering of nodes */
3031 local_node = pgdat->node_id;
3032 load = nr_online_nodes;
3033 prev_node = local_node;
3034 nodes_clear(used_mask);
3035
3036 memset(node_order, 0, sizeof(node_order));
3037 j = 0;
3038
3039 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
3040 int distance = node_distance(local_node, node);
3041
3042 /*
3043 * If another node is sufficiently far away then it is better
3044 * to reclaim pages in a zone before going off node.
3045 */
3046 if (distance > RECLAIM_DISTANCE)
3047 zone_reclaim_mode = 1;
3048
3049 /*
3050 * We don't want to pressure a particular node.
3051 * So adding penalty to the first node in same
3052 * distance group to make it round-robin.
3053 */
3054 if (distance != node_distance(local_node, prev_node))
3055 node_load[node] = load;
3056
3057 prev_node = node;
3058 load--;
3059 if (order == ZONELIST_ORDER_NODE)
3060 build_zonelists_in_node_order(pgdat, node);
3061 else
3062 node_order[j++] = node; /* remember order */
3063 }
3064
3065 if (order == ZONELIST_ORDER_ZONE) {
3066 /* calculate node order -- i.e., DMA last! */
3067 build_zonelists_in_zone_order(pgdat, j);
3068 }
3069
3070 build_thisnode_zonelists(pgdat);
3071}
3072
3073/* Construct the zonelist performance cache - see further mmzone.h */
3074static void build_zonelist_cache(pg_data_t *pgdat)
3075{
3076 struct zonelist *zonelist;
3077 struct zonelist_cache *zlc;
3078 struct zoneref *z;
3079
3080 zonelist = &pgdat->node_zonelists[0];
3081 zonelist->zlcache_ptr = zlc = &zonelist->zlcache;
3082 bitmap_zero(zlc->fullzones, MAX_ZONES_PER_ZONELIST);
3083 for (z = zonelist->_zonerefs; z->zone; z++)
3084 zlc->z_to_n[z - zonelist->_zonerefs] = zonelist_node_idx(z);
3085}
3086
3087#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3088/*
3089 * Return node id of node used for "local" allocations.
3090 * I.e., first node id of first zone in arg node's generic zonelist.
3091 * Used for initializing percpu 'numa_mem', which is used primarily
3092 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
3093 */
3094int local_memory_node(int node)
3095{
3096 struct zone *zone;
3097
3098 (void)first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
3099 gfp_zone(GFP_KERNEL),
3100 NULL,
3101 &zone);
3102 return zone->node;
3103}
3104#endif
3105
3106#else /* CONFIG_NUMA */
3107
3108static void set_zonelist_order(void)
3109{
3110 current_zonelist_order = ZONELIST_ORDER_ZONE;
3111}
3112
3113static void build_zonelists(pg_data_t *pgdat)
3114{
3115 int node, local_node;
3116 enum zone_type j;
3117 struct zonelist *zonelist;
3118
3119 local_node = pgdat->node_id;
3120
3121 zonelist = &pgdat->node_zonelists[0];
3122 j = build_zonelists_node(pgdat, zonelist, 0, MAX_NR_ZONES - 1);
3123
3124 /*
3125 * Now we build the zonelist so that it contains the zones
3126 * of all the other nodes.
3127 * We don't want to pressure a particular node, so when
3128 * building the zones for node N, we make sure that the
3129 * zones coming right after the local ones are those from
3130 * node N+1 (modulo N)
3131 */
3132 for (node = local_node + 1; node < MAX_NUMNODES; node++) {
3133 if (!node_online(node))
3134 continue;
3135 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3136 MAX_NR_ZONES - 1);
3137 }
3138 for (node = 0; node < local_node; node++) {
3139 if (!node_online(node))
3140 continue;
3141 j = build_zonelists_node(NODE_DATA(node), zonelist, j,
3142 MAX_NR_ZONES - 1);
3143 }
3144
3145 zonelist->_zonerefs[j].zone = NULL;
3146 zonelist->_zonerefs[j].zone_idx = 0;
3147}
3148
3149/* non-NUMA variant of zonelist performance cache - just NULL zlcache_ptr */
3150static void build_zonelist_cache(pg_data_t *pgdat)
3151{
3152 pgdat->node_zonelists[0].zlcache_ptr = NULL;
3153}
3154
3155#endif /* CONFIG_NUMA */
3156
3157/*
3158 * Boot pageset table. One per cpu which is going to be used for all
3159 * zones and all nodes. The parameters will be set in such a way
3160 * that an item put on a list will immediately be handed over to
3161 * the buddy list. This is safe since pageset manipulation is done
3162 * with interrupts disabled.
3163 *
3164 * The boot_pagesets must be kept even after bootup is complete for
3165 * unused processors and/or zones. They do play a role for bootstrapping
3166 * hotplugged processors.
3167 *
3168 * zoneinfo_show() and maybe other functions do
3169 * not check if the processor is online before following the pageset pointer.
3170 * Other parts of the kernel may not check if the zone is available.
3171 */
3172static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch);
3173static DEFINE_PER_CPU(struct per_cpu_pageset, boot_pageset);
3174static void setup_zone_pageset(struct zone *zone);
3175
3176/*
3177 * Global mutex to protect against size modification of zonelists
3178 * as well as to serialize pageset setup for the new populated zone.
3179 */
3180DEFINE_MUTEX(zonelists_mutex);
3181
3182/* return values int ....just for stop_machine() */
3183static __init_refok int __build_all_zonelists(void *data)
3184{
3185 int nid;
3186 int cpu;
3187
3188#ifdef CONFIG_NUMA
3189 memset(node_load, 0, sizeof(node_load));
3190#endif
3191 for_each_online_node(nid) {
3192 pg_data_t *pgdat = NODE_DATA(nid);
3193
3194 build_zonelists(pgdat);
3195 build_zonelist_cache(pgdat);
3196 }
3197
3198 /*
3199 * Initialize the boot_pagesets that are going to be used
3200 * for bootstrapping processors. The real pagesets for
3201 * each zone will be allocated later when the per cpu
3202 * allocator is available.
3203 *
3204 * boot_pagesets are used also for bootstrapping offline
3205 * cpus if the system is already booted because the pagesets
3206 * are needed to initialize allocators on a specific cpu too.
3207 * F.e. the percpu allocator needs the page allocator which
3208 * needs the percpu allocator in order to allocate its pagesets
3209 * (a chicken-egg dilemma).
3210 */
3211 for_each_possible_cpu(cpu) {
3212 setup_pageset(&per_cpu(boot_pageset, cpu), 0);
3213
3214#ifdef CONFIG_HAVE_MEMORYLESS_NODES
3215 /*
3216 * We now know the "local memory node" for each node--
3217 * i.e., the node of the first zone in the generic zonelist.
3218 * Set up numa_mem percpu variable for on-line cpus. During
3219 * boot, only the boot cpu should be on-line; we'll init the
3220 * secondary cpus' numa_mem as they come on-line. During
3221 * node/memory hotplug, we'll fixup all on-line cpus.
3222 */
3223 if (cpu_online(cpu))
3224 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
3225#endif
3226 }
3227
3228 return 0;
3229}
3230
3231/*
3232 * Called with zonelists_mutex held always
3233 * unless system_state == SYSTEM_BOOTING.
3234 */
3235void __ref build_all_zonelists(void *data)
3236{
3237 set_zonelist_order();
3238
3239 if (system_state == SYSTEM_BOOTING) {
3240 __build_all_zonelists(NULL);
3241 mminit_verify_zonelist();
3242 cpuset_init_current_mems_allowed();
3243 } else {
3244 /* we have to stop all cpus to guarantee there is no user
3245 of zonelist */
3246#ifdef CONFIG_MEMORY_HOTPLUG
3247 if (data)
3248 setup_zone_pageset((struct zone *)data);
3249#endif
3250 stop_machine(__build_all_zonelists, NULL, NULL);
3251 /* cpuset refresh routine should be here */
3252 }
3253 vm_total_pages = nr_free_pagecache_pages();
3254 /*
3255 * Disable grouping by mobility if the number of pages in the
3256 * system is too low to allow the mechanism to work. It would be
3257 * more accurate, but expensive to check per-zone. This check is
3258 * made on memory-hotadd so a system can start with mobility
3259 * disabled and enable it later
3260 */
3261 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
3262 page_group_by_mobility_disabled = 1;
3263 else
3264 page_group_by_mobility_disabled = 0;
3265
3266 printk("Built %i zonelists in %s order, mobility grouping %s. "
3267 "Total pages: %ld\n",
3268 nr_online_nodes,
3269 zonelist_order_name[current_zonelist_order],
3270 page_group_by_mobility_disabled ? "off" : "on",
3271 vm_total_pages);
3272#ifdef CONFIG_NUMA
3273 printk("Policy zone: %s\n", zone_names[policy_zone]);
3274#endif
3275}
3276
3277/*
3278 * Helper functions to size the waitqueue hash table.
3279 * Essentially these want to choose hash table sizes sufficiently
3280 * large so that collisions trying to wait on pages are rare.
3281 * But in fact, the number of active page waitqueues on typical
3282 * systems is ridiculously low, less than 200. So this is even
3283 * conservative, even though it seems large.
3284 *
3285 * The constant PAGES_PER_WAITQUEUE specifies the ratio of pages to
3286 * waitqueues, i.e. the size of the waitq table given the number of pages.
3287 */
3288#define PAGES_PER_WAITQUEUE 256
3289
3290#ifndef CONFIG_MEMORY_HOTPLUG
3291static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3292{
3293 unsigned long size = 1;
3294
3295 pages /= PAGES_PER_WAITQUEUE;
3296
3297 while (size < pages)
3298 size <<= 1;
3299
3300 /*
3301 * Once we have dozens or even hundreds of threads sleeping
3302 * on IO we've got bigger problems than wait queue collision.
3303 * Limit the size of the wait table to a reasonable size.
3304 */
3305 size = min(size, 4096UL);
3306
3307 return max(size, 4UL);
3308}
3309#else
3310/*
3311 * A zone's size might be changed by hot-add, so it is not possible to determine
3312 * a suitable size for its wait_table. So we use the maximum size now.
3313 *
3314 * The max wait table size = 4096 x sizeof(wait_queue_head_t). ie:
3315 *
3316 * i386 (preemption config) : 4096 x 16 = 64Kbyte.
3317 * ia64, x86-64 (no preemption): 4096 x 20 = 80Kbyte.
3318 * ia64, x86-64 (preemption) : 4096 x 24 = 96Kbyte.
3319 *
3320 * The maximum entries are prepared when a zone's memory is (512K + 256) pages
3321 * or more by the traditional way. (See above). It equals:
3322 *
3323 * i386, x86-64, powerpc(4K page size) : = ( 2G + 1M)byte.
3324 * ia64(16K page size) : = ( 8G + 4M)byte.
3325 * powerpc (64K page size) : = (32G +16M)byte.
3326 */
3327static inline unsigned long wait_table_hash_nr_entries(unsigned long pages)
3328{
3329 return 4096UL;
3330}
3331#endif
3332
3333/*
3334 * This is an integer logarithm so that shifts can be used later
3335 * to extract the more random high bits from the multiplicative
3336 * hash function before the remainder is taken.
3337 */
3338static inline unsigned long wait_table_bits(unsigned long size)
3339{
3340 return ffz(~size);
3341}
3342
3343#define LONG_ALIGN(x) (((x)+(sizeof(long))-1)&~((sizeof(long))-1))
3344
3345/*
3346 * Check if a pageblock contains reserved pages
3347 */
3348static int pageblock_is_reserved(unsigned long start_pfn, unsigned long end_pfn)
3349{
3350 unsigned long pfn;
3351
3352 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3353 if (!pfn_valid_within(pfn) || PageReserved(pfn_to_page(pfn)))
3354 return 1;
3355 }
3356 return 0;
3357}
3358
3359/*
3360 * Mark a number of pageblocks as MIGRATE_RESERVE. The number
3361 * of blocks reserved is based on min_wmark_pages(zone). The memory within
3362 * the reserve will tend to store contiguous free pages. Setting min_free_kbytes
3363 * higher will lead to a bigger reserve which will get freed as contiguous
3364 * blocks as reclaim kicks in
3365 */
3366static void setup_zone_migrate_reserve(struct zone *zone)
3367{
3368 unsigned long start_pfn, pfn, end_pfn, block_end_pfn;
3369 struct page *page;
3370 unsigned long block_migratetype;
3371 int reserve;
3372
3373 /* Get the start pfn, end pfn and the number of blocks to reserve */
3374 start_pfn = zone->zone_start_pfn;
3375 end_pfn = start_pfn + zone->spanned_pages;
3376 reserve = roundup(min_wmark_pages(zone), pageblock_nr_pages) >>
3377 pageblock_order;
3378
3379 /*
3380 * Reserve blocks are generally in place to help high-order atomic
3381 * allocations that are short-lived. A min_free_kbytes value that
3382 * would result in more than 2 reserve blocks for atomic allocations
3383 * is assumed to be in place to help anti-fragmentation for the
3384 * future allocation of hugepages at runtime.
3385 */
3386 reserve = min(2, reserve);
3387
3388 for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
3389 if (!pfn_valid(pfn))
3390 continue;
3391 page = pfn_to_page(pfn);
3392
3393 /* Watch out for overlapping nodes */
3394 if (page_to_nid(page) != zone_to_nid(zone))
3395 continue;
3396
3397 /* Blocks with reserved pages will never free, skip them. */
3398 block_end_pfn = min(pfn + pageblock_nr_pages, end_pfn);
3399 if (pageblock_is_reserved(pfn, block_end_pfn))
3400 continue;
3401
3402 block_migratetype = get_pageblock_migratetype(page);
3403
3404 /* If this block is reserved, account for it */
3405 if (reserve > 0 && block_migratetype == MIGRATE_RESERVE) {
3406 reserve--;
3407 continue;
3408 }
3409
3410 /* Suitable for reserving if this block is movable */
3411 if (reserve > 0 && block_migratetype == MIGRATE_MOVABLE) {
3412 set_pageblock_migratetype(page, MIGRATE_RESERVE);
3413 move_freepages_block(zone, page, MIGRATE_RESERVE);
3414 reserve--;
3415 continue;
3416 }
3417
3418 /*
3419 * If the reserve is met and this is a previous reserved block,
3420 * take it back
3421 */
3422 if (block_migratetype == MIGRATE_RESERVE) {
3423 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3424 move_freepages_block(zone, page, MIGRATE_MOVABLE);
3425 }
3426 }
3427}
3428
3429/*
3430 * Initially all pages are reserved - free ones are freed
3431 * up by free_all_bootmem() once the early boot process is
3432 * done. Non-atomic initialization, single-pass.
3433 */
3434void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
3435 unsigned long start_pfn, enum memmap_context context)
3436{
3437 struct page *page;
3438 unsigned long end_pfn = start_pfn + size;
3439 unsigned long pfn;
3440 struct zone *z;
3441
3442 if (highest_memmap_pfn < end_pfn - 1)
3443 highest_memmap_pfn = end_pfn - 1;
3444
3445 z = &NODE_DATA(nid)->node_zones[zone];
3446 for (pfn = start_pfn; pfn < end_pfn; pfn++) {
3447 /*
3448 * There can be holes in boot-time mem_map[]s
3449 * handed to this function. They do not
3450 * exist on hotplugged memory.
3451 */
3452 if (context == MEMMAP_EARLY) {
3453 if (!early_pfn_valid(pfn))
3454 continue;
3455 if (!early_pfn_in_nid(pfn, nid))
3456 continue;
3457 }
3458 page = pfn_to_page(pfn);
3459 set_page_links(page, zone, nid, pfn);
3460 mminit_verify_page_links(page, zone, nid, pfn);
3461 init_page_count(page);
3462 reset_page_mapcount(page);
3463 SetPageReserved(page);
3464 /*
3465 * Mark the block movable so that blocks are reserved for
3466 * movable at startup. This will force kernel allocations
3467 * to reserve their blocks rather than leaking throughout
3468 * the address space during boot when many long-lived
3469 * kernel allocations are made. Later some blocks near
3470 * the start are marked MIGRATE_RESERVE by
3471 * setup_zone_migrate_reserve()
3472 *
3473 * bitmap is created for zone's valid pfn range. but memmap
3474 * can be created for invalid pages (for alignment)
3475 * check here not to call set_pageblock_migratetype() against
3476 * pfn out of zone.
3477 */
3478 if ((z->zone_start_pfn <= pfn)
3479 && (pfn < z->zone_start_pfn + z->spanned_pages)
3480 && !(pfn & (pageblock_nr_pages - 1)))
3481 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
3482
3483 INIT_LIST_HEAD(&page->lru);
3484#ifdef WANT_PAGE_VIRTUAL
3485 /* The shift won't overflow because ZONE_NORMAL is below 4G. */
3486 if (!is_highmem_idx(zone))
3487 set_page_address(page, __va(pfn << PAGE_SHIFT));
3488#endif
3489 }
3490}
3491
3492static void __meminit zone_init_free_lists(struct zone *zone)
3493{
3494 int order, t;
3495 for_each_migratetype_order(order, t) {
3496 INIT_LIST_HEAD(&zone->free_area[order].free_list[t]);
3497 zone->free_area[order].nr_free = 0;
3498 }
3499}
3500
3501#ifndef __HAVE_ARCH_MEMMAP_INIT
3502#define memmap_init(size, nid, zone, start_pfn) \
3503 memmap_init_zone((size), (nid), (zone), (start_pfn), MEMMAP_EARLY)
3504#endif
3505
3506static int zone_batchsize(struct zone *zone)
3507{
3508#ifdef CONFIG_MMU
3509 int batch;
3510
3511 /*
3512 * The per-cpu-pages pools are set to around 1000th of the
3513 * size of the zone. But no more than 1/2 of a meg.
3514 *
3515 * OK, so we don't know how big the cache is. So guess.
3516 */
3517 batch = zone->present_pages / 1024;
3518 if (batch * PAGE_SIZE > 512 * 1024)
3519 batch = (512 * 1024) / PAGE_SIZE;
3520 batch /= 4; /* We effectively *= 4 below */
3521 if (batch < 1)
3522 batch = 1;
3523
3524 /*
3525 * Clamp the batch to a 2^n - 1 value. Having a power
3526 * of 2 value was found to be more likely to have
3527 * suboptimal cache aliasing properties in some cases.
3528 *
3529 * For example if 2 tasks are alternately allocating
3530 * batches of pages, one task can end up with a lot
3531 * of pages of one half of the possible page colors
3532 * and the other with pages of the other colors.
3533 */
3534 batch = rounddown_pow_of_two(batch + batch/2) - 1;
3535
3536 return batch;
3537
3538#else
3539 /* The deferral and batching of frees should be suppressed under NOMMU
3540 * conditions.
3541 *
3542 * The problem is that NOMMU needs to be able to allocate large chunks
3543 * of contiguous memory as there's no hardware page translation to
3544 * assemble apparent contiguous memory from discontiguous pages.
3545 *
3546 * Queueing large contiguous runs of pages for batching, however,
3547 * causes the pages to actually be freed in smaller chunks. As there
3548 * can be a significant delay between the individual batches being
3549 * recycled, this leads to the once large chunks of space being
3550 * fragmented and becoming unavailable for high-order allocations.
3551 */
3552 return 0;
3553#endif
3554}
3555
3556static void setup_pageset(struct per_cpu_pageset *p, unsigned long batch)
3557{
3558 struct per_cpu_pages *pcp;
3559 int migratetype;
3560
3561 memset(p, 0, sizeof(*p));
3562
3563 pcp = &p->pcp;
3564 pcp->count = 0;
3565 pcp->high = 6 * batch;
3566 pcp->batch = max(1UL, 1 * batch);
3567 for (migratetype = 0; migratetype < MIGRATE_PCPTYPES; migratetype++)
3568 INIT_LIST_HEAD(&pcp->lists[migratetype]);
3569}
3570
3571/*
3572 * setup_pagelist_highmark() sets the high water mark for hot per_cpu_pagelist
3573 * to the value high for the pageset p.
3574 */
3575
3576static void setup_pagelist_highmark(struct per_cpu_pageset *p,
3577 unsigned long high)
3578{
3579 struct per_cpu_pages *pcp;
3580
3581 pcp = &p->pcp;
3582 pcp->high = high;
3583 pcp->batch = max(1UL, high/4);
3584 if ((high/4) > (PAGE_SHIFT * 8))
3585 pcp->batch = PAGE_SHIFT * 8;
3586}
3587
3588static void setup_zone_pageset(struct zone *zone)
3589{
3590 int cpu;
3591
3592 zone->pageset = alloc_percpu(struct per_cpu_pageset);
3593
3594 for_each_possible_cpu(cpu) {
3595 struct per_cpu_pageset *pcp = per_cpu_ptr(zone->pageset, cpu);
3596
3597 setup_pageset(pcp, zone_batchsize(zone));
3598
3599 if (percpu_pagelist_fraction)
3600 setup_pagelist_highmark(pcp,
3601 (zone->present_pages /
3602 percpu_pagelist_fraction));
3603 }
3604}
3605
3606/*
3607 * Allocate per cpu pagesets and initialize them.
3608 * Before this call only boot pagesets were available.
3609 */
3610void __init setup_per_cpu_pageset(void)
3611{
3612 struct zone *zone;
3613
3614 for_each_populated_zone(zone)
3615 setup_zone_pageset(zone);
3616}
3617
3618static noinline __init_refok
3619int zone_wait_table_init(struct zone *zone, unsigned long zone_size_pages)
3620{
3621 int i;
3622 struct pglist_data *pgdat = zone->zone_pgdat;
3623 size_t alloc_size;
3624
3625 /*
3626 * The per-page waitqueue mechanism uses hashed waitqueues
3627 * per zone.
3628 */
3629 zone->wait_table_hash_nr_entries =
3630 wait_table_hash_nr_entries(zone_size_pages);
3631 zone->wait_table_bits =
3632 wait_table_bits(zone->wait_table_hash_nr_entries);
3633 alloc_size = zone->wait_table_hash_nr_entries
3634 * sizeof(wait_queue_head_t);
3635
3636 if (!slab_is_available()) {
3637 zone->wait_table = (wait_queue_head_t *)
3638 alloc_bootmem_node_nopanic(pgdat, alloc_size);
3639 } else {
3640 /*
3641 * This case means that a zone whose size was 0 gets new memory
3642 * via memory hot-add.
3643 * But it may be the case that a new node was hot-added. In
3644 * this case vmalloc() will not be able to use this new node's
3645 * memory - this wait_table must be initialized to use this new
3646 * node itself as well.
3647 * To use this new node's memory, further consideration will be
3648 * necessary.
3649 */
3650 zone->wait_table = vmalloc(alloc_size);
3651 }
3652 if (!zone->wait_table)
3653 return -ENOMEM;
3654
3655 for(i = 0; i < zone->wait_table_hash_nr_entries; ++i)
3656 init_waitqueue_head(zone->wait_table + i);
3657
3658 return 0;
3659}
3660
3661static int __zone_pcp_update(void *data)
3662{
3663 struct zone *zone = data;
3664 int cpu;
3665 unsigned long batch = zone_batchsize(zone), flags;
3666
3667 for_each_possible_cpu(cpu) {
3668 struct per_cpu_pageset *pset;
3669 struct per_cpu_pages *pcp;
3670
3671 pset = per_cpu_ptr(zone->pageset, cpu);
3672 pcp = &pset->pcp;
3673
3674 local_irq_save(flags);
3675 free_pcppages_bulk(zone, pcp->count, pcp);
3676 setup_pageset(pset, batch);
3677 local_irq_restore(flags);
3678 }
3679 return 0;
3680}
3681
3682void zone_pcp_update(struct zone *zone)
3683{
3684 stop_machine(__zone_pcp_update, zone, NULL);
3685}
3686
3687static __meminit void zone_pcp_init(struct zone *zone)
3688{
3689 /*
3690 * per cpu subsystem is not up at this point. The following code
3691 * relies on the ability of the linker to provide the
3692 * offset of a (static) per cpu variable into the per cpu area.
3693 */
3694 zone->pageset = &boot_pageset;
3695
3696 if (zone->present_pages)
3697 printk(KERN_DEBUG " %s zone: %lu pages, LIFO batch:%u\n",
3698 zone->name, zone->present_pages,
3699 zone_batchsize(zone));
3700}
3701
3702__meminit int init_currently_empty_zone(struct zone *zone,
3703 unsigned long zone_start_pfn,
3704 unsigned long size,
3705 enum memmap_context context)
3706{
3707 struct pglist_data *pgdat = zone->zone_pgdat;
3708 int ret;
3709 ret = zone_wait_table_init(zone, size);
3710 if (ret)
3711 return ret;
3712 pgdat->nr_zones = zone_idx(zone) + 1;
3713
3714 zone->zone_start_pfn = zone_start_pfn;
3715
3716 mminit_dprintk(MMINIT_TRACE, "memmap_init",
3717 "Initialising map node %d zone %lu pfns %lu -> %lu\n",
3718 pgdat->node_id,
3719 (unsigned long)zone_idx(zone),
3720 zone_start_pfn, (zone_start_pfn + size));
3721
3722 zone_init_free_lists(zone);
3723
3724 return 0;
3725}
3726
3727#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
3728/*
3729 * Basic iterator support. Return the first range of PFNs for a node
3730 * Note: nid == MAX_NUMNODES returns first region regardless of node
3731 */
3732static int __meminit first_active_region_index_in_nid(int nid)
3733{
3734 int i;
3735
3736 for (i = 0; i < nr_nodemap_entries; i++)
3737 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3738 return i;
3739
3740 return -1;
3741}
3742
3743/*
3744 * Basic iterator support. Return the next active range of PFNs for a node
3745 * Note: nid == MAX_NUMNODES returns next region regardless of node
3746 */
3747static int __meminit next_active_region_index_in_nid(int index, int nid)
3748{
3749 for (index = index + 1; index < nr_nodemap_entries; index++)
3750 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3751 return index;
3752
3753 return -1;
3754}
3755
3756#ifndef CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID
3757/*
3758 * Required by SPARSEMEM. Given a PFN, return what node the PFN is on.
3759 * Architectures may implement their own version but if add_active_range()
3760 * was used and there are no special requirements, this is a convenient
3761 * alternative
3762 */
3763int __meminit __early_pfn_to_nid(unsigned long pfn)
3764{
3765 int i;
3766
3767 for (i = 0; i < nr_nodemap_entries; i++) {
3768 unsigned long start_pfn = early_node_map[i].start_pfn;
3769 unsigned long end_pfn = early_node_map[i].end_pfn;
3770
3771 if (start_pfn <= pfn && pfn < end_pfn)
3772 return early_node_map[i].nid;
3773 }
3774 /* This is a memory hole */
3775 return -1;
3776}
3777#endif /* CONFIG_HAVE_ARCH_EARLY_PFN_TO_NID */
3778
3779int __meminit early_pfn_to_nid(unsigned long pfn)
3780{
3781 int nid;
3782
3783 nid = __early_pfn_to_nid(pfn);
3784 if (nid >= 0)
3785 return nid;
3786 /* just returns 0 */
3787 return 0;
3788}
3789
3790#ifdef CONFIG_NODES_SPAN_OTHER_NODES
3791bool __meminit early_pfn_in_nid(unsigned long pfn, int node)
3792{
3793 int nid;
3794
3795 nid = __early_pfn_to_nid(pfn);
3796 if (nid >= 0 && nid != node)
3797 return false;
3798 return true;
3799}
3800#endif
3801
3802/* Basic iterator support to walk early_node_map[] */
3803#define for_each_active_range_index_in_nid(i, nid) \
3804 for (i = first_active_region_index_in_nid(nid); i != -1; \
3805 i = next_active_region_index_in_nid(i, nid))
3806
3807/**
3808 * free_bootmem_with_active_regions - Call free_bootmem_node for each active range
3809 * @nid: The node to free memory on. If MAX_NUMNODES, all nodes are freed.
3810 * @max_low_pfn: The highest PFN that will be passed to free_bootmem_node
3811 *
3812 * If an architecture guarantees that all ranges registered with
3813 * add_active_ranges() contain no holes and may be freed, this
3814 * this function may be used instead of calling free_bootmem() manually.
3815 */
3816void __init free_bootmem_with_active_regions(int nid,
3817 unsigned long max_low_pfn)
3818{
3819 int i;
3820
3821 for_each_active_range_index_in_nid(i, nid) {
3822 unsigned long size_pages = 0;
3823 unsigned long end_pfn = early_node_map[i].end_pfn;
3824
3825 if (early_node_map[i].start_pfn >= max_low_pfn)
3826 continue;
3827
3828 if (end_pfn > max_low_pfn)
3829 end_pfn = max_low_pfn;
3830
3831 size_pages = end_pfn - early_node_map[i].start_pfn;
3832 free_bootmem_node(NODE_DATA(early_node_map[i].nid),
3833 PFN_PHYS(early_node_map[i].start_pfn),
3834 size_pages << PAGE_SHIFT);
3835 }
3836}
3837
3838#ifdef CONFIG_HAVE_MEMBLOCK
3839/*
3840 * Basic iterator support. Return the last range of PFNs for a node
3841 * Note: nid == MAX_NUMNODES returns last region regardless of node
3842 */
3843static int __meminit last_active_region_index_in_nid(int nid)
3844{
3845 int i;
3846
3847 for (i = nr_nodemap_entries - 1; i >= 0; i--)
3848 if (nid == MAX_NUMNODES || early_node_map[i].nid == nid)
3849 return i;
3850
3851 return -1;
3852}
3853
3854/*
3855 * Basic iterator support. Return the previous active range of PFNs for a node
3856 * Note: nid == MAX_NUMNODES returns next region regardless of node
3857 */
3858static int __meminit previous_active_region_index_in_nid(int index, int nid)
3859{
3860 for (index = index - 1; index >= 0; index--)
3861 if (nid == MAX_NUMNODES || early_node_map[index].nid == nid)
3862 return index;
3863
3864 return -1;
3865}
3866
3867#define for_each_active_range_index_in_nid_reverse(i, nid) \
3868 for (i = last_active_region_index_in_nid(nid); i != -1; \
3869 i = previous_active_region_index_in_nid(i, nid))
3870
3871u64 __init find_memory_core_early(int nid, u64 size, u64 align,
3872 u64 goal, u64 limit)
3873{
3874 int i;
3875
3876 /* Need to go over early_node_map to find out good range for node */
3877 for_each_active_range_index_in_nid_reverse(i, nid) {
3878 u64 addr;
3879 u64 ei_start, ei_last;
3880 u64 final_start, final_end;
3881
3882 ei_last = early_node_map[i].end_pfn;
3883 ei_last <<= PAGE_SHIFT;
3884 ei_start = early_node_map[i].start_pfn;
3885 ei_start <<= PAGE_SHIFT;
3886
3887 final_start = max(ei_start, goal);
3888 final_end = min(ei_last, limit);
3889
3890 if (final_start >= final_end)
3891 continue;
3892
3893 addr = memblock_find_in_range(final_start, final_end, size, align);
3894
3895 if (addr == MEMBLOCK_ERROR)
3896 continue;
3897
3898 return addr;
3899 }
3900
3901 return MEMBLOCK_ERROR;
3902}
3903#endif
3904
3905int __init add_from_early_node_map(struct range *range, int az,
3906 int nr_range, int nid)
3907{
3908 int i;
3909 u64 start, end;
3910
3911 /* need to go over early_node_map to find out good range for node */
3912 for_each_active_range_index_in_nid(i, nid) {
3913 start = early_node_map[i].start_pfn;
3914 end = early_node_map[i].end_pfn;
3915 nr_range = add_range(range, az, nr_range, start, end);
3916 }
3917 return nr_range;
3918}
3919
3920void __init work_with_active_regions(int nid, work_fn_t work_fn, void *data)
3921{
3922 int i;
3923 int ret;
3924
3925 for_each_active_range_index_in_nid(i, nid) {
3926 ret = work_fn(early_node_map[i].start_pfn,
3927 early_node_map[i].end_pfn, data);
3928 if (ret)
3929 break;
3930 }
3931}
3932/**
3933 * sparse_memory_present_with_active_regions - Call memory_present for each active range
3934 * @nid: The node to call memory_present for. If MAX_NUMNODES, all nodes will be used.
3935 *
3936 * If an architecture guarantees that all ranges registered with
3937 * add_active_ranges() contain no holes and may be freed, this
3938 * function may be used instead of calling memory_present() manually.
3939 */
3940void __init sparse_memory_present_with_active_regions(int nid)
3941{
3942 int i;
3943
3944 for_each_active_range_index_in_nid(i, nid)
3945 memory_present(early_node_map[i].nid,
3946 early_node_map[i].start_pfn,
3947 early_node_map[i].end_pfn);
3948}
3949
3950/**
3951 * get_pfn_range_for_nid - Return the start and end page frames for a node
3952 * @nid: The nid to return the range for. If MAX_NUMNODES, the min and max PFN are returned.
3953 * @start_pfn: Passed by reference. On return, it will have the node start_pfn.
3954 * @end_pfn: Passed by reference. On return, it will have the node end_pfn.
3955 *
3956 * It returns the start and end page frame of a node based on information
3957 * provided by an arch calling add_active_range(). If called for a node
3958 * with no available memory, a warning is printed and the start and end
3959 * PFNs will be 0.
3960 */
3961void __meminit get_pfn_range_for_nid(unsigned int nid,
3962 unsigned long *start_pfn, unsigned long *end_pfn)
3963{
3964 int i;
3965 *start_pfn = -1UL;
3966 *end_pfn = 0;
3967
3968 for_each_active_range_index_in_nid(i, nid) {
3969 *start_pfn = min(*start_pfn, early_node_map[i].start_pfn);
3970 *end_pfn = max(*end_pfn, early_node_map[i].end_pfn);
3971 }
3972
3973 if (*start_pfn == -1UL)
3974 *start_pfn = 0;
3975}
3976
3977/*
3978 * This finds a zone that can be used for ZONE_MOVABLE pages. The
3979 * assumption is made that zones within a node are ordered in monotonic
3980 * increasing memory addresses so that the "highest" populated zone is used
3981 */
3982static void __init find_usable_zone_for_movable(void)
3983{
3984 int zone_index;
3985 for (zone_index = MAX_NR_ZONES - 1; zone_index >= 0; zone_index--) {
3986 if (zone_index == ZONE_MOVABLE)
3987 continue;
3988
3989 if (arch_zone_highest_possible_pfn[zone_index] >
3990 arch_zone_lowest_possible_pfn[zone_index])
3991 break;
3992 }
3993
3994 VM_BUG_ON(zone_index == -1);
3995 movable_zone = zone_index;
3996}
3997
3998/*
3999 * The zone ranges provided by the architecture do not include ZONE_MOVABLE
4000 * because it is sized independent of architecture. Unlike the other zones,
4001 * the starting point for ZONE_MOVABLE is not fixed. It may be different
4002 * in each node depending on the size of each node and how evenly kernelcore
4003 * is distributed. This helper function adjusts the zone ranges
4004 * provided by the architecture for a given node by using the end of the
4005 * highest usable zone for ZONE_MOVABLE. This preserves the assumption that
4006 * zones within a node are in order of monotonic increases memory addresses
4007 */
4008static void __meminit adjust_zone_range_for_zone_movable(int nid,
4009 unsigned long zone_type,
4010 unsigned long node_start_pfn,
4011 unsigned long node_end_pfn,
4012 unsigned long *zone_start_pfn,
4013 unsigned long *zone_end_pfn)
4014{
4015 /* Only adjust if ZONE_MOVABLE is on this node */
4016 if (zone_movable_pfn[nid]) {
4017 /* Size ZONE_MOVABLE */
4018 if (zone_type == ZONE_MOVABLE) {
4019 *zone_start_pfn = zone_movable_pfn[nid];
4020 *zone_end_pfn = min(node_end_pfn,
4021 arch_zone_highest_possible_pfn[movable_zone]);
4022
4023 /* Adjust for ZONE_MOVABLE starting within this range */
4024 } else if (*zone_start_pfn < zone_movable_pfn[nid] &&
4025 *zone_end_pfn > zone_movable_pfn[nid]) {
4026 *zone_end_pfn = zone_movable_pfn[nid];
4027
4028 /* Check if this whole range is within ZONE_MOVABLE */
4029 } else if (*zone_start_pfn >= zone_movable_pfn[nid])
4030 *zone_start_pfn = *zone_end_pfn;
4031 }
4032}
4033
4034/*
4035 * Return the number of pages a zone spans in a node, including holes
4036 * present_pages = zone_spanned_pages_in_node() - zone_absent_pages_in_node()
4037 */
4038static unsigned long __meminit zone_spanned_pages_in_node(int nid,
4039 unsigned long zone_type,
4040 unsigned long *ignored)
4041{
4042 unsigned long node_start_pfn, node_end_pfn;
4043 unsigned long zone_start_pfn, zone_end_pfn;
4044
4045 /* Get the start and end of the node and zone */
4046 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4047 zone_start_pfn = arch_zone_lowest_possible_pfn[zone_type];
4048 zone_end_pfn = arch_zone_highest_possible_pfn[zone_type];
4049 adjust_zone_range_for_zone_movable(nid, zone_type,
4050 node_start_pfn, node_end_pfn,
4051 &zone_start_pfn, &zone_end_pfn);
4052
4053 /* Check that this node has pages within the zone's required range */
4054 if (zone_end_pfn < node_start_pfn || zone_start_pfn > node_end_pfn)
4055 return 0;
4056
4057 /* Move the zone boundaries inside the node if necessary */
4058 zone_end_pfn = min(zone_end_pfn, node_end_pfn);
4059 zone_start_pfn = max(zone_start_pfn, node_start_pfn);
4060
4061 /* Return the spanned pages */
4062 return zone_end_pfn - zone_start_pfn;
4063}
4064
4065/*
4066 * Return the number of holes in a range on a node. If nid is MAX_NUMNODES,
4067 * then all holes in the requested range will be accounted for.
4068 */
4069unsigned long __meminit __absent_pages_in_range(int nid,
4070 unsigned long range_start_pfn,
4071 unsigned long range_end_pfn)
4072{
4073 int i = 0;
4074 unsigned long prev_end_pfn = 0, hole_pages = 0;
4075 unsigned long start_pfn;
4076
4077 /* Find the end_pfn of the first active range of pfns in the node */
4078 i = first_active_region_index_in_nid(nid);
4079 if (i == -1)
4080 return 0;
4081
4082 prev_end_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4083
4084 /* Account for ranges before physical memory on this node */
4085 if (early_node_map[i].start_pfn > range_start_pfn)
4086 hole_pages = prev_end_pfn - range_start_pfn;
4087
4088 /* Find all holes for the zone within the node */
4089 for (; i != -1; i = next_active_region_index_in_nid(i, nid)) {
4090
4091 /* No need to continue if prev_end_pfn is outside the zone */
4092 if (prev_end_pfn >= range_end_pfn)
4093 break;
4094
4095 /* Make sure the end of the zone is not within the hole */
4096 start_pfn = min(early_node_map[i].start_pfn, range_end_pfn);
4097 prev_end_pfn = max(prev_end_pfn, range_start_pfn);
4098
4099 /* Update the hole size cound and move on */
4100 if (start_pfn > range_start_pfn) {
4101 BUG_ON(prev_end_pfn > start_pfn);
4102 hole_pages += start_pfn - prev_end_pfn;
4103 }
4104 prev_end_pfn = early_node_map[i].end_pfn;
4105 }
4106
4107 /* Account for ranges past physical memory on this node */
4108 if (range_end_pfn > prev_end_pfn)
4109 hole_pages += range_end_pfn -
4110 max(range_start_pfn, prev_end_pfn);
4111
4112 return hole_pages;
4113}
4114
4115/**
4116 * absent_pages_in_range - Return number of page frames in holes within a range
4117 * @start_pfn: The start PFN to start searching for holes
4118 * @end_pfn: The end PFN to stop searching for holes
4119 *
4120 * It returns the number of pages frames in memory holes within a range.
4121 */
4122unsigned long __init absent_pages_in_range(unsigned long start_pfn,
4123 unsigned long end_pfn)
4124{
4125 return __absent_pages_in_range(MAX_NUMNODES, start_pfn, end_pfn);
4126}
4127
4128/* Return the number of page frames in holes in a zone on a node */
4129static unsigned long __meminit zone_absent_pages_in_node(int nid,
4130 unsigned long zone_type,
4131 unsigned long *ignored)
4132{
4133 unsigned long node_start_pfn, node_end_pfn;
4134 unsigned long zone_start_pfn, zone_end_pfn;
4135
4136 get_pfn_range_for_nid(nid, &node_start_pfn, &node_end_pfn);
4137 zone_start_pfn = max(arch_zone_lowest_possible_pfn[zone_type],
4138 node_start_pfn);
4139 zone_end_pfn = min(arch_zone_highest_possible_pfn[zone_type],
4140 node_end_pfn);
4141
4142 adjust_zone_range_for_zone_movable(nid, zone_type,
4143 node_start_pfn, node_end_pfn,
4144 &zone_start_pfn, &zone_end_pfn);
4145 return __absent_pages_in_range(nid, zone_start_pfn, zone_end_pfn);
4146}
4147
4148#else
4149static inline unsigned long __meminit zone_spanned_pages_in_node(int nid,
4150 unsigned long zone_type,
4151 unsigned long *zones_size)
4152{
4153 return zones_size[zone_type];
4154}
4155
4156static inline unsigned long __meminit zone_absent_pages_in_node(int nid,
4157 unsigned long zone_type,
4158 unsigned long *zholes_size)
4159{
4160 if (!zholes_size)
4161 return 0;
4162
4163 return zholes_size[zone_type];
4164}
4165
4166#endif
4167
4168static void __meminit calculate_node_totalpages(struct pglist_data *pgdat,
4169 unsigned long *zones_size, unsigned long *zholes_size)
4170{
4171 unsigned long realtotalpages, totalpages = 0;
4172 enum zone_type i;
4173
4174 for (i = 0; i < MAX_NR_ZONES; i++)
4175 totalpages += zone_spanned_pages_in_node(pgdat->node_id, i,
4176 zones_size);
4177 pgdat->node_spanned_pages = totalpages;
4178
4179 realtotalpages = totalpages;
4180 for (i = 0; i < MAX_NR_ZONES; i++)
4181 realtotalpages -=
4182 zone_absent_pages_in_node(pgdat->node_id, i,
4183 zholes_size);
4184 pgdat->node_present_pages = realtotalpages;
4185 printk(KERN_DEBUG "On node %d totalpages: %lu\n", pgdat->node_id,
4186 realtotalpages);
4187}
4188
4189#ifndef CONFIG_SPARSEMEM
4190/*
4191 * Calculate the size of the zone->blockflags rounded to an unsigned long
4192 * Start by making sure zonesize is a multiple of pageblock_order by rounding
4193 * up. Then use 1 NR_PAGEBLOCK_BITS worth of bits per pageblock, finally
4194 * round what is now in bits to nearest long in bits, then return it in
4195 * bytes.
4196 */
4197static unsigned long __init usemap_size(unsigned long zonesize)
4198{
4199 unsigned long usemapsize;
4200
4201 usemapsize = roundup(zonesize, pageblock_nr_pages);
4202 usemapsize = usemapsize >> pageblock_order;
4203 usemapsize *= NR_PAGEBLOCK_BITS;
4204 usemapsize = roundup(usemapsize, 8 * sizeof(unsigned long));
4205
4206 return usemapsize / 8;
4207}
4208
4209static void __init setup_usemap(struct pglist_data *pgdat,
4210 struct zone *zone, unsigned long zonesize)
4211{
4212 unsigned long usemapsize = usemap_size(zonesize);
4213 zone->pageblock_flags = NULL;
4214 if (usemapsize)
4215 zone->pageblock_flags = alloc_bootmem_node_nopanic(pgdat,
4216 usemapsize);
4217}
4218#else
4219static inline void setup_usemap(struct pglist_data *pgdat,
4220 struct zone *zone, unsigned long zonesize) {}
4221#endif /* CONFIG_SPARSEMEM */
4222
4223#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
4224
4225/* Return a sensible default order for the pageblock size. */
4226static inline int pageblock_default_order(void)
4227{
4228 if (HPAGE_SHIFT > PAGE_SHIFT)
4229 return HUGETLB_PAGE_ORDER;
4230
4231 return MAX_ORDER-1;
4232}
4233
4234/* Initialise the number of pages represented by NR_PAGEBLOCK_BITS */
4235static inline void __init set_pageblock_order(unsigned int order)
4236{
4237 /* Check that pageblock_nr_pages has not already been setup */
4238 if (pageblock_order)
4239 return;
4240
4241 /*
4242 * Assume the largest contiguous order of interest is a huge page.
4243 * This value may be variable depending on boot parameters on IA64
4244 */
4245 pageblock_order = order;
4246}
4247#else /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4248
4249/*
4250 * When CONFIG_HUGETLB_PAGE_SIZE_VARIABLE is not set, set_pageblock_order()
4251 * and pageblock_default_order() are unused as pageblock_order is set
4252 * at compile-time. See include/linux/pageblock-flags.h for the values of
4253 * pageblock_order based on the kernel config
4254 */
4255static inline int pageblock_default_order(unsigned int order)
4256{
4257 return MAX_ORDER-1;
4258}
4259#define set_pageblock_order(x) do {} while (0)
4260
4261#endif /* CONFIG_HUGETLB_PAGE_SIZE_VARIABLE */
4262
4263/*
4264 * Set up the zone data structures:
4265 * - mark all pages reserved
4266 * - mark all memory queues empty
4267 * - clear the memory bitmaps
4268 */
4269static void __paginginit free_area_init_core(struct pglist_data *pgdat,
4270 unsigned long *zones_size, unsigned long *zholes_size)
4271{
4272 enum zone_type j;
4273 int nid = pgdat->node_id;
4274 unsigned long zone_start_pfn = pgdat->node_start_pfn;
4275 int ret;
4276
4277 pgdat_resize_init(pgdat);
4278 pgdat->nr_zones = 0;
4279 init_waitqueue_head(&pgdat->kswapd_wait);
4280 pgdat->kswapd_max_order = 0;
4281 pgdat_page_cgroup_init(pgdat);
4282
4283 for (j = 0; j < MAX_NR_ZONES; j++) {
4284 struct zone *zone = pgdat->node_zones + j;
4285 unsigned long size, realsize, memmap_pages;
4286 enum lru_list l;
4287
4288 size = zone_spanned_pages_in_node(nid, j, zones_size);
4289 realsize = size - zone_absent_pages_in_node(nid, j,
4290 zholes_size);
4291
4292 /*
4293 * Adjust realsize so that it accounts for how much memory
4294 * is used by this zone for memmap. This affects the watermark
4295 * and per-cpu initialisations
4296 */
4297 memmap_pages =
4298 PAGE_ALIGN(size * sizeof(struct page)) >> PAGE_SHIFT;
4299 if (realsize >= memmap_pages) {
4300 realsize -= memmap_pages;
4301 if (memmap_pages)
4302 printk(KERN_DEBUG
4303 " %s zone: %lu pages used for memmap\n",
4304 zone_names[j], memmap_pages);
4305 } else
4306 printk(KERN_WARNING
4307 " %s zone: %lu pages exceeds realsize %lu\n",
4308 zone_names[j], memmap_pages, realsize);
4309
4310 /* Account for reserved pages */
4311 if (j == 0 && realsize > dma_reserve) {
4312 realsize -= dma_reserve;
4313 printk(KERN_DEBUG " %s zone: %lu pages reserved\n",
4314 zone_names[0], dma_reserve);
4315 }
4316
4317 if (!is_highmem_idx(j))
4318 nr_kernel_pages += realsize;
4319 nr_all_pages += realsize;
4320
4321 zone->spanned_pages = size;
4322 zone->present_pages = realsize;
4323#ifdef CONFIG_NUMA
4324 zone->node = nid;
4325 zone->min_unmapped_pages = (realsize*sysctl_min_unmapped_ratio)
4326 / 100;
4327 zone->min_slab_pages = (realsize * sysctl_min_slab_ratio) / 100;
4328#endif
4329 zone->name = zone_names[j];
4330 spin_lock_init(&zone->lock);
4331 spin_lock_init(&zone->lru_lock);
4332 zone_seqlock_init(zone);
4333 zone->zone_pgdat = pgdat;
4334
4335 zone_pcp_init(zone);
4336 for_each_lru(l)
4337 INIT_LIST_HEAD(&zone->lru[l].list);
4338 zone->reclaim_stat.recent_rotated[0] = 0;
4339 zone->reclaim_stat.recent_rotated[1] = 0;
4340 zone->reclaim_stat.recent_scanned[0] = 0;
4341 zone->reclaim_stat.recent_scanned[1] = 0;
4342 zap_zone_vm_stats(zone);
4343 zone->flags = 0;
4344 if (!size)
4345 continue;
4346
4347 set_pageblock_order(pageblock_default_order());
4348 setup_usemap(pgdat, zone, size);
4349 ret = init_currently_empty_zone(zone, zone_start_pfn,
4350 size, MEMMAP_EARLY);
4351 BUG_ON(ret);
4352 memmap_init(size, nid, j, zone_start_pfn);
4353 zone_start_pfn += size;
4354 }
4355}
4356
4357static void __init_refok alloc_node_mem_map(struct pglist_data *pgdat)
4358{
4359 /* Skip empty nodes */
4360 if (!pgdat->node_spanned_pages)
4361 return;
4362
4363#ifdef CONFIG_FLAT_NODE_MEM_MAP
4364 /* ia64 gets its own node_mem_map, before this, without bootmem */
4365 if (!pgdat->node_mem_map) {
4366 unsigned long size, start, end;
4367 struct page *map;
4368
4369 /*
4370 * The zone's endpoints aren't required to be MAX_ORDER
4371 * aligned but the node_mem_map endpoints must be in order
4372 * for the buddy allocator to function correctly.
4373 */
4374 start = pgdat->node_start_pfn & ~(MAX_ORDER_NR_PAGES - 1);
4375 end = pgdat->node_start_pfn + pgdat->node_spanned_pages;
4376 end = ALIGN(end, MAX_ORDER_NR_PAGES);
4377 size = (end - start) * sizeof(struct page);
4378 map = alloc_remap(pgdat->node_id, size);
4379 if (!map)
4380 map = alloc_bootmem_node_nopanic(pgdat, size);
4381 pgdat->node_mem_map = map + (pgdat->node_start_pfn - start);
4382 }
4383#ifndef CONFIG_NEED_MULTIPLE_NODES
4384 /*
4385 * With no DISCONTIG, the global mem_map is just set as node 0's
4386 */
4387 if (pgdat == NODE_DATA(0)) {
4388 mem_map = NODE_DATA(0)->node_mem_map;
4389#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4390 if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
4391 mem_map -= (pgdat->node_start_pfn - ARCH_PFN_OFFSET);
4392#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4393 }
4394#endif
4395#endif /* CONFIG_FLAT_NODE_MEM_MAP */
4396}
4397
4398void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
4399 unsigned long node_start_pfn, unsigned long *zholes_size)
4400{
4401 pg_data_t *pgdat = NODE_DATA(nid);
4402
4403 pgdat->node_id = nid;
4404 pgdat->node_start_pfn = node_start_pfn;
4405 calculate_node_totalpages(pgdat, zones_size, zholes_size);
4406
4407 alloc_node_mem_map(pgdat);
4408#ifdef CONFIG_FLAT_NODE_MEM_MAP
4409 printk(KERN_DEBUG "free_area_init_node: node %d, pgdat %08lx, node_mem_map %08lx\n",
4410 nid, (unsigned long)pgdat,
4411 (unsigned long)pgdat->node_mem_map);
4412#endif
4413
4414 free_area_init_core(pgdat, zones_size, zholes_size);
4415}
4416
4417#ifdef CONFIG_ARCH_POPULATES_NODE_MAP
4418
4419#if MAX_NUMNODES > 1
4420/*
4421 * Figure out the number of possible node ids.
4422 */
4423static void __init setup_nr_node_ids(void)
4424{
4425 unsigned int node;
4426 unsigned int highest = 0;
4427
4428 for_each_node_mask(node, node_possible_map)
4429 highest = node;
4430 nr_node_ids = highest + 1;
4431}
4432#else
4433static inline void setup_nr_node_ids(void)
4434{
4435}
4436#endif
4437
4438/**
4439 * add_active_range - Register a range of PFNs backed by physical memory
4440 * @nid: The node ID the range resides on
4441 * @start_pfn: The start PFN of the available physical memory
4442 * @end_pfn: The end PFN of the available physical memory
4443 *
4444 * These ranges are stored in an early_node_map[] and later used by
4445 * free_area_init_nodes() to calculate zone sizes and holes. If the
4446 * range spans a memory hole, it is up to the architecture to ensure
4447 * the memory is not freed by the bootmem allocator. If possible
4448 * the range being registered will be merged with existing ranges.
4449 */
4450void __init add_active_range(unsigned int nid, unsigned long start_pfn,
4451 unsigned long end_pfn)
4452{
4453 int i;
4454
4455 mminit_dprintk(MMINIT_TRACE, "memory_register",
4456 "Entering add_active_range(%d, %#lx, %#lx) "
4457 "%d entries of %d used\n",
4458 nid, start_pfn, end_pfn,
4459 nr_nodemap_entries, MAX_ACTIVE_REGIONS);
4460
4461 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
4462
4463 /* Merge with existing active regions if possible */
4464 for (i = 0; i < nr_nodemap_entries; i++) {
4465 if (early_node_map[i].nid != nid)
4466 continue;
4467
4468 /* Skip if an existing region covers this new one */
4469 if (start_pfn >= early_node_map[i].start_pfn &&
4470 end_pfn <= early_node_map[i].end_pfn)
4471 return;
4472
4473 /* Merge forward if suitable */
4474 if (start_pfn <= early_node_map[i].end_pfn &&
4475 end_pfn > early_node_map[i].end_pfn) {
4476 early_node_map[i].end_pfn = end_pfn;
4477 return;
4478 }
4479
4480 /* Merge backward if suitable */
4481 if (start_pfn < early_node_map[i].start_pfn &&
4482 end_pfn >= early_node_map[i].start_pfn) {
4483 early_node_map[i].start_pfn = start_pfn;
4484 return;
4485 }
4486 }
4487
4488 /* Check that early_node_map is large enough */
4489 if (i >= MAX_ACTIVE_REGIONS) {
4490 printk(KERN_CRIT "More than %d memory regions, truncating\n",
4491 MAX_ACTIVE_REGIONS);
4492 return;
4493 }
4494
4495 early_node_map[i].nid = nid;
4496 early_node_map[i].start_pfn = start_pfn;
4497 early_node_map[i].end_pfn = end_pfn;
4498 nr_nodemap_entries = i + 1;
4499}
4500
4501/**
4502 * remove_active_range - Shrink an existing registered range of PFNs
4503 * @nid: The node id the range is on that should be shrunk
4504 * @start_pfn: The new PFN of the range
4505 * @end_pfn: The new PFN of the range
4506 *
4507 * i386 with NUMA use alloc_remap() to store a node_mem_map on a local node.
4508 * The map is kept near the end physical page range that has already been
4509 * registered. This function allows an arch to shrink an existing registered
4510 * range.
4511 */
4512void __init remove_active_range(unsigned int nid, unsigned long start_pfn,
4513 unsigned long end_pfn)
4514{
4515 int i, j;
4516 int removed = 0;
4517
4518 printk(KERN_DEBUG "remove_active_range (%d, %lu, %lu)\n",
4519 nid, start_pfn, end_pfn);
4520
4521 /* Find the old active region end and shrink */
4522 for_each_active_range_index_in_nid(i, nid) {
4523 if (early_node_map[i].start_pfn >= start_pfn &&
4524 early_node_map[i].end_pfn <= end_pfn) {
4525 /* clear it */
4526 early_node_map[i].start_pfn = 0;
4527 early_node_map[i].end_pfn = 0;
4528 removed = 1;
4529 continue;
4530 }
4531 if (early_node_map[i].start_pfn < start_pfn &&
4532 early_node_map[i].end_pfn > start_pfn) {
4533 unsigned long temp_end_pfn = early_node_map[i].end_pfn;
4534 early_node_map[i].end_pfn = start_pfn;
4535 if (temp_end_pfn > end_pfn)
4536 add_active_range(nid, end_pfn, temp_end_pfn);
4537 continue;
4538 }
4539 if (early_node_map[i].start_pfn >= start_pfn &&
4540 early_node_map[i].end_pfn > end_pfn &&
4541 early_node_map[i].start_pfn < end_pfn) {
4542 early_node_map[i].start_pfn = end_pfn;
4543 continue;
4544 }
4545 }
4546
4547 if (!removed)
4548 return;
4549
4550 /* remove the blank ones */
4551 for (i = nr_nodemap_entries - 1; i > 0; i--) {
4552 if (early_node_map[i].nid != nid)
4553 continue;
4554 if (early_node_map[i].end_pfn)
4555 continue;
4556 /* we found it, get rid of it */
4557 for (j = i; j < nr_nodemap_entries - 1; j++)
4558 memcpy(&early_node_map[j], &early_node_map[j+1],
4559 sizeof(early_node_map[j]));
4560 j = nr_nodemap_entries - 1;
4561 memset(&early_node_map[j], 0, sizeof(early_node_map[j]));
4562 nr_nodemap_entries--;
4563 }
4564}
4565
4566/**
4567 * remove_all_active_ranges - Remove all currently registered regions
4568 *
4569 * During discovery, it may be found that a table like SRAT is invalid
4570 * and an alternative discovery method must be used. This function removes
4571 * all currently registered regions.
4572 */
4573void __init remove_all_active_ranges(void)
4574{
4575 memset(early_node_map, 0, sizeof(early_node_map));
4576 nr_nodemap_entries = 0;
4577}
4578
4579/* Compare two active node_active_regions */
4580static int __init cmp_node_active_region(const void *a, const void *b)
4581{
4582 struct node_active_region *arange = (struct node_active_region *)a;
4583 struct node_active_region *brange = (struct node_active_region *)b;
4584
4585 /* Done this way to avoid overflows */
4586 if (arange->start_pfn > brange->start_pfn)
4587 return 1;
4588 if (arange->start_pfn < brange->start_pfn)
4589 return -1;
4590
4591 return 0;
4592}
4593
4594/* sort the node_map by start_pfn */
4595void __init sort_node_map(void)
4596{
4597 sort(early_node_map, (size_t)nr_nodemap_entries,
4598 sizeof(struct node_active_region),
4599 cmp_node_active_region, NULL);
4600}
4601
4602/**
4603 * node_map_pfn_alignment - determine the maximum internode alignment
4604 *
4605 * This function should be called after node map is populated and sorted.
4606 * It calculates the maximum power of two alignment which can distinguish
4607 * all the nodes.
4608 *
4609 * For example, if all nodes are 1GiB and aligned to 1GiB, the return value
4610 * would indicate 1GiB alignment with (1 << (30 - PAGE_SHIFT)). If the
4611 * nodes are shifted by 256MiB, 256MiB. Note that if only the last node is
4612 * shifted, 1GiB is enough and this function will indicate so.
4613 *
4614 * This is used to test whether pfn -> nid mapping of the chosen memory
4615 * model has fine enough granularity to avoid incorrect mapping for the
4616 * populated node map.
4617 *
4618 * Returns the determined alignment in pfn's. 0 if there is no alignment
4619 * requirement (single node).
4620 */
4621unsigned long __init node_map_pfn_alignment(void)
4622{
4623 unsigned long accl_mask = 0, last_end = 0;
4624 int last_nid = -1;
4625 int i;
4626
4627 for_each_active_range_index_in_nid(i, MAX_NUMNODES) {
4628 int nid = early_node_map[i].nid;
4629 unsigned long start = early_node_map[i].start_pfn;
4630 unsigned long end = early_node_map[i].end_pfn;
4631 unsigned long mask;
4632
4633 if (!start || last_nid < 0 || last_nid == nid) {
4634 last_nid = nid;
4635 last_end = end;
4636 continue;
4637 }
4638
4639 /*
4640 * Start with a mask granular enough to pin-point to the
4641 * start pfn and tick off bits one-by-one until it becomes
4642 * too coarse to separate the current node from the last.
4643 */
4644 mask = ~((1 << __ffs(start)) - 1);
4645 while (mask && last_end <= (start & (mask << 1)))
4646 mask <<= 1;
4647
4648 /* accumulate all internode masks */
4649 accl_mask |= mask;
4650 }
4651
4652 /* convert mask to number of pages */
4653 return ~accl_mask + 1;
4654}
4655
4656/* Find the lowest pfn for a node */
4657static unsigned long __init find_min_pfn_for_node(int nid)
4658{
4659 int i;
4660 unsigned long min_pfn = ULONG_MAX;
4661
4662 /* Assuming a sorted map, the first range found has the starting pfn */
4663 for_each_active_range_index_in_nid(i, nid)
4664 min_pfn = min(min_pfn, early_node_map[i].start_pfn);
4665
4666 if (min_pfn == ULONG_MAX) {
4667 printk(KERN_WARNING
4668 "Could not find start_pfn for node %d\n", nid);
4669 return 0;
4670 }
4671
4672 return min_pfn;
4673}
4674
4675/**
4676 * find_min_pfn_with_active_regions - Find the minimum PFN registered
4677 *
4678 * It returns the minimum PFN based on information provided via
4679 * add_active_range().
4680 */
4681unsigned long __init find_min_pfn_with_active_regions(void)
4682{
4683 return find_min_pfn_for_node(MAX_NUMNODES);
4684}
4685
4686/*
4687 * early_calculate_totalpages()
4688 * Sum pages in active regions for movable zone.
4689 * Populate N_HIGH_MEMORY for calculating usable_nodes.
4690 */
4691static unsigned long __init early_calculate_totalpages(void)
4692{
4693 int i;
4694 unsigned long totalpages = 0;
4695
4696 for (i = 0; i < nr_nodemap_entries; i++) {
4697 unsigned long pages = early_node_map[i].end_pfn -
4698 early_node_map[i].start_pfn;
4699 totalpages += pages;
4700 if (pages)
4701 node_set_state(early_node_map[i].nid, N_HIGH_MEMORY);
4702 }
4703 return totalpages;
4704}
4705
4706/*
4707 * Find the PFN the Movable zone begins in each node. Kernel memory
4708 * is spread evenly between nodes as long as the nodes have enough
4709 * memory. When they don't, some nodes will have more kernelcore than
4710 * others
4711 */
4712static void __init find_zone_movable_pfns_for_nodes(unsigned long *movable_pfn)
4713{
4714 int i, nid;
4715 unsigned long usable_startpfn;
4716 unsigned long kernelcore_node, kernelcore_remaining;
4717 /* save the state before borrow the nodemask */
4718 nodemask_t saved_node_state = node_states[N_HIGH_MEMORY];
4719 unsigned long totalpages = early_calculate_totalpages();
4720 int usable_nodes = nodes_weight(node_states[N_HIGH_MEMORY]);
4721
4722 /*
4723 * If movablecore was specified, calculate what size of
4724 * kernelcore that corresponds so that memory usable for
4725 * any allocation type is evenly spread. If both kernelcore
4726 * and movablecore are specified, then the value of kernelcore
4727 * will be used for required_kernelcore if it's greater than
4728 * what movablecore would have allowed.
4729 */
4730 if (required_movablecore) {
4731 unsigned long corepages;
4732
4733 /*
4734 * Round-up so that ZONE_MOVABLE is at least as large as what
4735 * was requested by the user
4736 */
4737 required_movablecore =
4738 roundup(required_movablecore, MAX_ORDER_NR_PAGES);
4739 corepages = totalpages - required_movablecore;
4740
4741 required_kernelcore = max(required_kernelcore, corepages);
4742 }
4743
4744 /* If kernelcore was not specified, there is no ZONE_MOVABLE */
4745 if (!required_kernelcore)
4746 goto out;
4747
4748 /* usable_startpfn is the lowest possible pfn ZONE_MOVABLE can be at */
4749 find_usable_zone_for_movable();
4750 usable_startpfn = arch_zone_lowest_possible_pfn[movable_zone];
4751
4752restart:
4753 /* Spread kernelcore memory as evenly as possible throughout nodes */
4754 kernelcore_node = required_kernelcore / usable_nodes;
4755 for_each_node_state(nid, N_HIGH_MEMORY) {
4756 /*
4757 * Recalculate kernelcore_node if the division per node
4758 * now exceeds what is necessary to satisfy the requested
4759 * amount of memory for the kernel
4760 */
4761 if (required_kernelcore < kernelcore_node)
4762 kernelcore_node = required_kernelcore / usable_nodes;
4763
4764 /*
4765 * As the map is walked, we track how much memory is usable
4766 * by the kernel using kernelcore_remaining. When it is
4767 * 0, the rest of the node is usable by ZONE_MOVABLE
4768 */
4769 kernelcore_remaining = kernelcore_node;
4770
4771 /* Go through each range of PFNs within this node */
4772 for_each_active_range_index_in_nid(i, nid) {
4773 unsigned long start_pfn, end_pfn;
4774 unsigned long size_pages;
4775
4776 start_pfn = max(early_node_map[i].start_pfn,
4777 zone_movable_pfn[nid]);
4778 end_pfn = early_node_map[i].end_pfn;
4779 if (start_pfn >= end_pfn)
4780 continue;
4781
4782 /* Account for what is only usable for kernelcore */
4783 if (start_pfn < usable_startpfn) {
4784 unsigned long kernel_pages;
4785 kernel_pages = min(end_pfn, usable_startpfn)
4786 - start_pfn;
4787
4788 kernelcore_remaining -= min(kernel_pages,
4789 kernelcore_remaining);
4790 required_kernelcore -= min(kernel_pages,
4791 required_kernelcore);
4792
4793 /* Continue if range is now fully accounted */
4794 if (end_pfn <= usable_startpfn) {
4795
4796 /*
4797 * Push zone_movable_pfn to the end so
4798 * that if we have to rebalance
4799 * kernelcore across nodes, we will
4800 * not double account here
4801 */
4802 zone_movable_pfn[nid] = end_pfn;
4803 continue;
4804 }
4805 start_pfn = usable_startpfn;
4806 }
4807
4808 /*
4809 * The usable PFN range for ZONE_MOVABLE is from
4810 * start_pfn->end_pfn. Calculate size_pages as the
4811 * number of pages used as kernelcore
4812 */
4813 size_pages = end_pfn - start_pfn;
4814 if (size_pages > kernelcore_remaining)
4815 size_pages = kernelcore_remaining;
4816 zone_movable_pfn[nid] = start_pfn + size_pages;
4817
4818 /*
4819 * Some kernelcore has been met, update counts and
4820 * break if the kernelcore for this node has been
4821 * satisified
4822 */
4823 required_kernelcore -= min(required_kernelcore,
4824 size_pages);
4825 kernelcore_remaining -= size_pages;
4826 if (!kernelcore_remaining)
4827 break;
4828 }
4829 }
4830
4831 /*
4832 * If there is still required_kernelcore, we do another pass with one
4833 * less node in the count. This will push zone_movable_pfn[nid] further
4834 * along on the nodes that still have memory until kernelcore is
4835 * satisified
4836 */
4837 usable_nodes--;
4838 if (usable_nodes && required_kernelcore > usable_nodes)
4839 goto restart;
4840
4841 /* Align start of ZONE_MOVABLE on all nids to MAX_ORDER_NR_PAGES */
4842 for (nid = 0; nid < MAX_NUMNODES; nid++)
4843 zone_movable_pfn[nid] =
4844 roundup(zone_movable_pfn[nid], MAX_ORDER_NR_PAGES);
4845
4846out:
4847 /* restore the node_state */
4848 node_states[N_HIGH_MEMORY] = saved_node_state;
4849}
4850
4851/* Any regular memory on that node ? */
4852static void check_for_regular_memory(pg_data_t *pgdat)
4853{
4854#ifdef CONFIG_HIGHMEM
4855 enum zone_type zone_type;
4856
4857 for (zone_type = 0; zone_type <= ZONE_NORMAL; zone_type++) {
4858 struct zone *zone = &pgdat->node_zones[zone_type];
4859 if (zone->present_pages)
4860 node_set_state(zone_to_nid(zone), N_NORMAL_MEMORY);
4861 }
4862#endif
4863}
4864
4865/**
4866 * free_area_init_nodes - Initialise all pg_data_t and zone data
4867 * @max_zone_pfn: an array of max PFNs for each zone
4868 *
4869 * This will call free_area_init_node() for each active node in the system.
4870 * Using the page ranges provided by add_active_range(), the size of each
4871 * zone in each node and their holes is calculated. If the maximum PFN
4872 * between two adjacent zones match, it is assumed that the zone is empty.
4873 * For example, if arch_max_dma_pfn == arch_max_dma32_pfn, it is assumed
4874 * that arch_max_dma32_pfn has no pages. It is also assumed that a zone
4875 * starts where the previous one ended. For example, ZONE_DMA32 starts
4876 * at arch_max_dma_pfn.
4877 */
4878void __init free_area_init_nodes(unsigned long *max_zone_pfn)
4879{
4880 unsigned long nid;
4881 int i;
4882
4883 /* Sort early_node_map as initialisation assumes it is sorted */
4884 sort_node_map();
4885
4886 /* Record where the zone boundaries are */
4887 memset(arch_zone_lowest_possible_pfn, 0,
4888 sizeof(arch_zone_lowest_possible_pfn));
4889 memset(arch_zone_highest_possible_pfn, 0,
4890 sizeof(arch_zone_highest_possible_pfn));
4891 arch_zone_lowest_possible_pfn[0] = find_min_pfn_with_active_regions();
4892 arch_zone_highest_possible_pfn[0] = max_zone_pfn[0];
4893 for (i = 1; i < MAX_NR_ZONES; i++) {
4894 if (i == ZONE_MOVABLE)
4895 continue;
4896 arch_zone_lowest_possible_pfn[i] =
4897 arch_zone_highest_possible_pfn[i-1];
4898 arch_zone_highest_possible_pfn[i] =
4899 max(max_zone_pfn[i], arch_zone_lowest_possible_pfn[i]);
4900 }
4901 arch_zone_lowest_possible_pfn[ZONE_MOVABLE] = 0;
4902 arch_zone_highest_possible_pfn[ZONE_MOVABLE] = 0;
4903
4904 /* Find the PFNs that ZONE_MOVABLE begins at in each node */
4905 memset(zone_movable_pfn, 0, sizeof(zone_movable_pfn));
4906 find_zone_movable_pfns_for_nodes(zone_movable_pfn);
4907
4908 /* Print out the zone ranges */
4909 printk("Zone PFN ranges:\n");
4910 for (i = 0; i < MAX_NR_ZONES; i++) {
4911 if (i == ZONE_MOVABLE)
4912 continue;
4913 printk(" %-8s ", zone_names[i]);
4914 if (arch_zone_lowest_possible_pfn[i] ==
4915 arch_zone_highest_possible_pfn[i])
4916 printk("empty\n");
4917 else
4918 printk("%0#10lx -> %0#10lx\n",
4919 arch_zone_lowest_possible_pfn[i],
4920 arch_zone_highest_possible_pfn[i]);
4921 }
4922
4923 /* Print out the PFNs ZONE_MOVABLE begins at in each node */
4924 printk("Movable zone start PFN for each node\n");
4925 for (i = 0; i < MAX_NUMNODES; i++) {
4926 if (zone_movable_pfn[i])
4927 printk(" Node %d: %lu\n", i, zone_movable_pfn[i]);
4928 }
4929
4930 /* Print out the early_node_map[] */
4931 printk("early_node_map[%d] active PFN ranges\n", nr_nodemap_entries);
4932 for (i = 0; i < nr_nodemap_entries; i++)
4933 printk(" %3d: %0#10lx -> %0#10lx\n", early_node_map[i].nid,
4934 early_node_map[i].start_pfn,
4935 early_node_map[i].end_pfn);
4936
4937 /* Initialise every node */
4938 mminit_verify_pageflags_layout();
4939 setup_nr_node_ids();
4940 for_each_online_node(nid) {
4941 pg_data_t *pgdat = NODE_DATA(nid);
4942 free_area_init_node(nid, NULL,
4943 find_min_pfn_for_node(nid), NULL);
4944
4945 /* Any memory on that node */
4946 if (pgdat->node_present_pages)
4947 node_set_state(nid, N_HIGH_MEMORY);
4948 check_for_regular_memory(pgdat);
4949 }
4950}
4951
4952static int __init cmdline_parse_core(char *p, unsigned long *core)
4953{
4954 unsigned long long coremem;
4955 if (!p)
4956 return -EINVAL;
4957
4958 coremem = memparse(p, &p);
4959 *core = coremem >> PAGE_SHIFT;
4960
4961 /* Paranoid check that UL is enough for the coremem value */
4962 WARN_ON((coremem >> PAGE_SHIFT) > ULONG_MAX);
4963
4964 return 0;
4965}
4966
4967/*
4968 * kernelcore=size sets the amount of memory for use for allocations that
4969 * cannot be reclaimed or migrated.
4970 */
4971static int __init cmdline_parse_kernelcore(char *p)
4972{
4973 return cmdline_parse_core(p, &required_kernelcore);
4974}
4975
4976/*
4977 * movablecore=size sets the amount of memory for use for allocations that
4978 * can be reclaimed or migrated.
4979 */
4980static int __init cmdline_parse_movablecore(char *p)
4981{
4982 return cmdline_parse_core(p, &required_movablecore);
4983}
4984
4985early_param("kernelcore", cmdline_parse_kernelcore);
4986early_param("movablecore", cmdline_parse_movablecore);
4987
4988#endif /* CONFIG_ARCH_POPULATES_NODE_MAP */
4989
4990/**
4991 * set_dma_reserve - set the specified number of pages reserved in the first zone
4992 * @new_dma_reserve: The number of pages to mark reserved
4993 *
4994 * The per-cpu batchsize and zone watermarks are determined by present_pages.
4995 * In the DMA zone, a significant percentage may be consumed by kernel image
4996 * and other unfreeable allocations which can skew the watermarks badly. This
4997 * function may optionally be used to account for unfreeable pages in the
4998 * first zone (e.g., ZONE_DMA). The effect will be lower watermarks and
4999 * smaller per-cpu batchsize.
5000 */
5001void __init set_dma_reserve(unsigned long new_dma_reserve)
5002{
5003 dma_reserve = new_dma_reserve;
5004}
5005
5006void __init free_area_init(unsigned long *zones_size)
5007{
5008 free_area_init_node(0, zones_size,
5009 __pa(PAGE_OFFSET) >> PAGE_SHIFT, NULL);
5010}
5011
5012static int page_alloc_cpu_notify(struct notifier_block *self,
5013 unsigned long action, void *hcpu)
5014{
5015 int cpu = (unsigned long)hcpu;
5016
5017 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN) {
5018 drain_pages(cpu);
5019
5020 /*
5021 * Spill the event counters of the dead processor
5022 * into the current processors event counters.
5023 * This artificially elevates the count of the current
5024 * processor.
5025 */
5026 vm_events_fold_cpu(cpu);
5027
5028 /*
5029 * Zero the differential counters of the dead processor
5030 * so that the vm statistics are consistent.
5031 *
5032 * This is only okay since the processor is dead and cannot
5033 * race with what we are doing.
5034 */
5035 refresh_cpu_vm_stats(cpu);
5036 }
5037 return NOTIFY_OK;
5038}
5039
5040void __init page_alloc_init(void)
5041{
5042 hotcpu_notifier(page_alloc_cpu_notify, 0);
5043}
5044
5045/*
5046 * calculate_totalreserve_pages - called when sysctl_lower_zone_reserve_ratio
5047 * or min_free_kbytes changes.
5048 */
5049static void calculate_totalreserve_pages(void)
5050{
5051 struct pglist_data *pgdat;
5052 unsigned long reserve_pages = 0;
5053 enum zone_type i, j;
5054
5055 for_each_online_pgdat(pgdat) {
5056 for (i = 0; i < MAX_NR_ZONES; i++) {
5057 struct zone *zone = pgdat->node_zones + i;
5058 unsigned long max = 0;
5059
5060 /* Find valid and maximum lowmem_reserve in the zone */
5061 for (j = i; j < MAX_NR_ZONES; j++) {
5062 if (zone->lowmem_reserve[j] > max)
5063 max = zone->lowmem_reserve[j];
5064 }
5065
5066 /* we treat the high watermark as reserved pages. */
5067 max += high_wmark_pages(zone);
5068
5069 if (max > zone->present_pages)
5070 max = zone->present_pages;
5071 reserve_pages += max;
5072 }
5073 }
5074 totalreserve_pages = reserve_pages;
5075}
5076
5077/*
5078 * setup_per_zone_lowmem_reserve - called whenever
5079 * sysctl_lower_zone_reserve_ratio changes. Ensures that each zone
5080 * has a correct pages reserved value, so an adequate number of
5081 * pages are left in the zone after a successful __alloc_pages().
5082 */
5083static void setup_per_zone_lowmem_reserve(void)
5084{
5085 struct pglist_data *pgdat;
5086 enum zone_type j, idx;
5087
5088 for_each_online_pgdat(pgdat) {
5089 for (j = 0; j < MAX_NR_ZONES; j++) {
5090 struct zone *zone = pgdat->node_zones + j;
5091 unsigned long present_pages = zone->present_pages;
5092
5093 zone->lowmem_reserve[j] = 0;
5094
5095 idx = j;
5096 while (idx) {
5097 struct zone *lower_zone;
5098
5099 idx--;
5100
5101 if (sysctl_lowmem_reserve_ratio[idx] < 1)
5102 sysctl_lowmem_reserve_ratio[idx] = 1;
5103
5104 lower_zone = pgdat->node_zones + idx;
5105 lower_zone->lowmem_reserve[j] = present_pages /
5106 sysctl_lowmem_reserve_ratio[idx];
5107 present_pages += lower_zone->present_pages;
5108 }
5109 }
5110 }
5111
5112 /* update totalreserve_pages */
5113 calculate_totalreserve_pages();
5114}
5115
5116/**
5117 * setup_per_zone_wmarks - called when min_free_kbytes changes
5118 * or when memory is hot-{added|removed}
5119 *
5120 * Ensures that the watermark[min,low,high] values for each zone are set
5121 * correctly with respect to min_free_kbytes.
5122 */
5123void setup_per_zone_wmarks(void)
5124{
5125 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5126 unsigned long lowmem_pages = 0;
5127 struct zone *zone;
5128 unsigned long flags;
5129
5130 /* Calculate total number of !ZONE_HIGHMEM pages */
5131 for_each_zone(zone) {
5132 if (!is_highmem(zone))
5133 lowmem_pages += zone->present_pages;
5134 }
5135
5136 for_each_zone(zone) {
5137 u64 tmp;
5138
5139 spin_lock_irqsave(&zone->lock, flags);
5140 tmp = (u64)pages_min * zone->present_pages;
5141 do_div(tmp, lowmem_pages);
5142 if (is_highmem(zone)) {
5143 /*
5144 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5145 * need highmem pages, so cap pages_min to a small
5146 * value here.
5147 *
5148 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5149 * deltas controls asynch page reclaim, and so should
5150 * not be capped for highmem.
5151 */
5152 int min_pages;
5153
5154 min_pages = zone->present_pages / 1024;
5155 if (min_pages < SWAP_CLUSTER_MAX)
5156 min_pages = SWAP_CLUSTER_MAX;
5157 if (min_pages > 128)
5158 min_pages = 128;
5159 zone->watermark[WMARK_MIN] = min_pages;
5160 } else {
5161 /*
5162 * If it's a lowmem zone, reserve a number of pages
5163 * proportionate to the zone's size.
5164 */
5165 zone->watermark[WMARK_MIN] = tmp;
5166 }
5167
5168 zone->watermark[WMARK_LOW] = min_wmark_pages(zone) + (tmp >> 2);
5169 zone->watermark[WMARK_HIGH] = min_wmark_pages(zone) + (tmp >> 1);
5170 setup_zone_migrate_reserve(zone);
5171 spin_unlock_irqrestore(&zone->lock, flags);
5172 }
5173
5174 /* update totalreserve_pages */
5175 calculate_totalreserve_pages();
5176}
5177
5178/*
5179 * The inactive anon list should be small enough that the VM never has to
5180 * do too much work, but large enough that each inactive page has a chance
5181 * to be referenced again before it is swapped out.
5182 *
5183 * The inactive_anon ratio is the target ratio of ACTIVE_ANON to
5184 * INACTIVE_ANON pages on this zone's LRU, maintained by the
5185 * pageout code. A zone->inactive_ratio of 3 means 3:1 or 25% of
5186 * the anonymous pages are kept on the inactive list.
5187 *
5188 * total target max
5189 * memory ratio inactive anon
5190 * -------------------------------------
5191 * 10MB 1 5MB
5192 * 100MB 1 50MB
5193 * 1GB 3 250MB
5194 * 10GB 10 0.9GB
5195 * 100GB 31 3GB
5196 * 1TB 101 10GB
5197 * 10TB 320 32GB
5198 */
5199static void __meminit calculate_zone_inactive_ratio(struct zone *zone)
5200{
5201 unsigned int gb, ratio;
5202
5203 /* Zone size in gigabytes */
5204 gb = zone->present_pages >> (30 - PAGE_SHIFT);
5205 if (gb)
5206 ratio = int_sqrt(10 * gb);
5207 else
5208 ratio = 1;
5209
5210 zone->inactive_ratio = ratio;
5211}
5212
5213static void __meminit setup_per_zone_inactive_ratio(void)
5214{
5215 struct zone *zone;
5216
5217 for_each_zone(zone)
5218 calculate_zone_inactive_ratio(zone);
5219}
5220
5221/*
5222 * Initialise min_free_kbytes.
5223 *
5224 * For small machines we want it small (128k min). For large machines
5225 * we want it large (64MB max). But it is not linear, because network
5226 * bandwidth does not increase linearly with machine size. We use
5227 *
5228 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5229 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5230 *
5231 * which yields
5232 *
5233 * 16MB: 512k
5234 * 32MB: 724k
5235 * 64MB: 1024k
5236 * 128MB: 1448k
5237 * 256MB: 2048k
5238 * 512MB: 2896k
5239 * 1024MB: 4096k
5240 * 2048MB: 5792k
5241 * 4096MB: 8192k
5242 * 8192MB: 11584k
5243 * 16384MB: 16384k
5244 */
5245int __meminit init_per_zone_wmark_min(void)
5246{
5247 unsigned long lowmem_kbytes;
5248
5249 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5250
5251 min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5252 if (min_free_kbytes < 128)
5253 min_free_kbytes = 128;
5254 if (min_free_kbytes > 65536)
5255 min_free_kbytes = 65536;
5256 setup_per_zone_wmarks();
5257 refresh_zone_stat_thresholds();
5258 setup_per_zone_lowmem_reserve();
5259 setup_per_zone_inactive_ratio();
5260 return 0;
5261}
5262module_init(init_per_zone_wmark_min)
5263
5264/*
5265 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
5266 * that we can call two helper functions whenever min_free_kbytes
5267 * changes.
5268 */
5269int min_free_kbytes_sysctl_handler(ctl_table *table, int write,
5270 void __user *buffer, size_t *length, loff_t *ppos)
5271{
5272 proc_dointvec(table, write, buffer, length, ppos);
5273 if (write)
5274 setup_per_zone_wmarks();
5275 return 0;
5276}
5277
5278#ifdef CONFIG_NUMA
5279int sysctl_min_unmapped_ratio_sysctl_handler(ctl_table *table, int write,
5280 void __user *buffer, size_t *length, loff_t *ppos)
5281{
5282 struct zone *zone;
5283 int rc;
5284
5285 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5286 if (rc)
5287 return rc;
5288
5289 for_each_zone(zone)
5290 zone->min_unmapped_pages = (zone->present_pages *
5291 sysctl_min_unmapped_ratio) / 100;
5292 return 0;
5293}
5294
5295int sysctl_min_slab_ratio_sysctl_handler(ctl_table *table, int write,
5296 void __user *buffer, size_t *length, loff_t *ppos)
5297{
5298 struct zone *zone;
5299 int rc;
5300
5301 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
5302 if (rc)
5303 return rc;
5304
5305 for_each_zone(zone)
5306 zone->min_slab_pages = (zone->present_pages *
5307 sysctl_min_slab_ratio) / 100;
5308 return 0;
5309}
5310#endif
5311
5312/*
5313 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
5314 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
5315 * whenever sysctl_lowmem_reserve_ratio changes.
5316 *
5317 * The reserve ratio obviously has absolutely no relation with the
5318 * minimum watermarks. The lowmem reserve ratio can only make sense
5319 * if in function of the boot time zone sizes.
5320 */
5321int lowmem_reserve_ratio_sysctl_handler(ctl_table *table, int write,
5322 void __user *buffer, size_t *length, loff_t *ppos)
5323{
5324 proc_dointvec_minmax(table, write, buffer, length, ppos);
5325 setup_per_zone_lowmem_reserve();
5326 return 0;
5327}
5328
5329/*
5330 * percpu_pagelist_fraction - changes the pcp->high for each zone on each
5331 * cpu. It is the fraction of total pages in each zone that a hot per cpu pagelist
5332 * can have before it gets flushed back to buddy allocator.
5333 */
5334
5335int percpu_pagelist_fraction_sysctl_handler(ctl_table *table, int write,
5336 void __user *buffer, size_t *length, loff_t *ppos)
5337{
5338 struct zone *zone;
5339 unsigned int cpu;
5340 int ret;
5341
5342 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
5343 if (!write || (ret == -EINVAL))
5344 return ret;
5345 for_each_populated_zone(zone) {
5346 for_each_possible_cpu(cpu) {
5347 unsigned long high;
5348 high = zone->present_pages / percpu_pagelist_fraction;
5349 setup_pagelist_highmark(
5350 per_cpu_ptr(zone->pageset, cpu), high);
5351 }
5352 }
5353 return 0;
5354}
5355
5356int hashdist = HASHDIST_DEFAULT;
5357
5358#ifdef CONFIG_NUMA
5359static int __init set_hashdist(char *str)
5360{
5361 if (!str)
5362 return 0;
5363 hashdist = simple_strtoul(str, &str, 0);
5364 return 1;
5365}
5366__setup("hashdist=", set_hashdist);
5367#endif
5368
5369/*
5370 * allocate a large system hash table from bootmem
5371 * - it is assumed that the hash table must contain an exact power-of-2
5372 * quantity of entries
5373 * - limit is the number of hash buckets, not the total allocation size
5374 */
5375void *__init alloc_large_system_hash(const char *tablename,
5376 unsigned long bucketsize,
5377 unsigned long numentries,
5378 int scale,
5379 int flags,
5380 unsigned int *_hash_shift,
5381 unsigned int *_hash_mask,
5382 unsigned long limit)
5383{
5384 unsigned long long max = limit;
5385 unsigned long log2qty, size;
5386 void *table = NULL;
5387
5388 /* allow the kernel cmdline to have a say */
5389 if (!numentries) {
5390 /* round applicable memory size up to nearest megabyte */
5391 numentries = nr_kernel_pages;
5392 numentries += (1UL << (20 - PAGE_SHIFT)) - 1;
5393 numentries >>= 20 - PAGE_SHIFT;
5394 numentries <<= 20 - PAGE_SHIFT;
5395
5396 /* limit to 1 bucket per 2^scale bytes of low memory */
5397 if (scale > PAGE_SHIFT)
5398 numentries >>= (scale - PAGE_SHIFT);
5399 else
5400 numentries <<= (PAGE_SHIFT - scale);
5401
5402 /* Make sure we've got at least a 0-order allocation.. */
5403 if (unlikely(flags & HASH_SMALL)) {
5404 /* Makes no sense without HASH_EARLY */
5405 WARN_ON(!(flags & HASH_EARLY));
5406 if (!(numentries >> *_hash_shift)) {
5407 numentries = 1UL << *_hash_shift;
5408 BUG_ON(!numentries);
5409 }
5410 } else if (unlikely((numentries * bucketsize) < PAGE_SIZE))
5411 numentries = PAGE_SIZE / bucketsize;
5412 }
5413 numentries = roundup_pow_of_two(numentries);
5414
5415 /* limit allocation size to 1/16 total memory by default */
5416 if (max == 0) {
5417 max = ((unsigned long long)nr_all_pages << PAGE_SHIFT) >> 4;
5418 do_div(max, bucketsize);
5419 }
5420
5421 if (numentries > max)
5422 numentries = max;
5423
5424 log2qty = ilog2(numentries);
5425
5426 do {
5427 size = bucketsize << log2qty;
5428 if (flags & HASH_EARLY)
5429 table = alloc_bootmem_nopanic(size);
5430 else if (hashdist)
5431 table = __vmalloc(size, GFP_ATOMIC, PAGE_KERNEL);
5432 else {
5433 /*
5434 * If bucketsize is not a power-of-two, we may free
5435 * some pages at the end of hash table which
5436 * alloc_pages_exact() automatically does
5437 */
5438 if (get_order(size) < MAX_ORDER) {
5439 table = alloc_pages_exact(size, GFP_ATOMIC);
5440 kmemleak_alloc(table, size, 1, GFP_ATOMIC);
5441 }
5442 }
5443 } while (!table && size > PAGE_SIZE && --log2qty);
5444
5445 if (!table)
5446 panic("Failed to allocate %s hash table\n", tablename);
5447
5448 printk(KERN_INFO "%s hash table entries: %ld (order: %d, %lu bytes)\n",
5449 tablename,
5450 (1UL << log2qty),
5451 ilog2(size) - PAGE_SHIFT,
5452 size);
5453
5454 if (_hash_shift)
5455 *_hash_shift = log2qty;
5456 if (_hash_mask)
5457 *_hash_mask = (1 << log2qty) - 1;
5458
5459 return table;
5460}
5461
5462/* Return a pointer to the bitmap storing bits affecting a block of pages */
5463static inline unsigned long *get_pageblock_bitmap(struct zone *zone,
5464 unsigned long pfn)
5465{
5466#ifdef CONFIG_SPARSEMEM
5467 return __pfn_to_section(pfn)->pageblock_flags;
5468#else
5469 return zone->pageblock_flags;
5470#endif /* CONFIG_SPARSEMEM */
5471}
5472
5473static inline int pfn_to_bitidx(struct zone *zone, unsigned long pfn)
5474{
5475#ifdef CONFIG_SPARSEMEM
5476 pfn &= (PAGES_PER_SECTION-1);
5477 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5478#else
5479 pfn = pfn - zone->zone_start_pfn;
5480 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
5481#endif /* CONFIG_SPARSEMEM */
5482}
5483
5484/**
5485 * get_pageblock_flags_group - Return the requested group of flags for the pageblock_nr_pages block of pages
5486 * @page: The page within the block of interest
5487 * @start_bitidx: The first bit of interest to retrieve
5488 * @end_bitidx: The last bit of interest
5489 * returns pageblock_bits flags
5490 */
5491unsigned long get_pageblock_flags_group(struct page *page,
5492 int start_bitidx, int end_bitidx)
5493{
5494 struct zone *zone;
5495 unsigned long *bitmap;
5496 unsigned long pfn, bitidx;
5497 unsigned long flags = 0;
5498 unsigned long value = 1;
5499
5500 zone = page_zone(page);
5501 pfn = page_to_pfn(page);
5502 bitmap = get_pageblock_bitmap(zone, pfn);
5503 bitidx = pfn_to_bitidx(zone, pfn);
5504
5505 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5506 if (test_bit(bitidx + start_bitidx, bitmap))
5507 flags |= value;
5508
5509 return flags;
5510}
5511
5512/**
5513 * set_pageblock_flags_group - Set the requested group of flags for a pageblock_nr_pages block of pages
5514 * @page: The page within the block of interest
5515 * @start_bitidx: The first bit of interest
5516 * @end_bitidx: The last bit of interest
5517 * @flags: The flags to set
5518 */
5519void set_pageblock_flags_group(struct page *page, unsigned long flags,
5520 int start_bitidx, int end_bitidx)
5521{
5522 struct zone *zone;
5523 unsigned long *bitmap;
5524 unsigned long pfn, bitidx;
5525 unsigned long value = 1;
5526
5527 zone = page_zone(page);
5528 pfn = page_to_pfn(page);
5529 bitmap = get_pageblock_bitmap(zone, pfn);
5530 bitidx = pfn_to_bitidx(zone, pfn);
5531 VM_BUG_ON(pfn < zone->zone_start_pfn);
5532 VM_BUG_ON(pfn >= zone->zone_start_pfn + zone->spanned_pages);
5533
5534 for (; start_bitidx <= end_bitidx; start_bitidx++, value <<= 1)
5535 if (flags & value)
5536 __set_bit(bitidx + start_bitidx, bitmap);
5537 else
5538 __clear_bit(bitidx + start_bitidx, bitmap);
5539}
5540
5541/*
5542 * This is designed as sub function...plz see page_isolation.c also.
5543 * set/clear page block's type to be ISOLATE.
5544 * page allocater never alloc memory from ISOLATE block.
5545 */
5546
5547static int
5548__count_immobile_pages(struct zone *zone, struct page *page, int count)
5549{
5550 unsigned long pfn, iter, found;
5551 /*
5552 * For avoiding noise data, lru_add_drain_all() should be called
5553 * If ZONE_MOVABLE, the zone never contains immobile pages
5554 */
5555 if (zone_idx(zone) == ZONE_MOVABLE)
5556 return true;
5557
5558 if (get_pageblock_migratetype(page) == MIGRATE_MOVABLE)
5559 return true;
5560
5561 pfn = page_to_pfn(page);
5562 for (found = 0, iter = 0; iter < pageblock_nr_pages; iter++) {
5563 unsigned long check = pfn + iter;
5564
5565 if (!pfn_valid_within(check))
5566 continue;
5567
5568 page = pfn_to_page(check);
5569 if (!page_count(page)) {
5570 if (PageBuddy(page))
5571 iter += (1 << page_order(page)) - 1;
5572 continue;
5573 }
5574 if (!PageLRU(page))
5575 found++;
5576 /*
5577 * If there are RECLAIMABLE pages, we need to check it.
5578 * But now, memory offline itself doesn't call shrink_slab()
5579 * and it still to be fixed.
5580 */
5581 /*
5582 * If the page is not RAM, page_count()should be 0.
5583 * we don't need more check. This is an _used_ not-movable page.
5584 *
5585 * The problematic thing here is PG_reserved pages. PG_reserved
5586 * is set to both of a memory hole page and a _used_ kernel
5587 * page at boot.
5588 */
5589 if (found > count)
5590 return false;
5591 }
5592 return true;
5593}
5594
5595bool is_pageblock_removable_nolock(struct page *page)
5596{
5597 struct zone *zone = page_zone(page);
5598 return __count_immobile_pages(zone, page, 0);
5599}
5600
5601int set_migratetype_isolate(struct page *page)
5602{
5603 struct zone *zone;
5604 unsigned long flags, pfn;
5605 struct memory_isolate_notify arg;
5606 int notifier_ret;
5607 int ret = -EBUSY;
5608
5609 zone = page_zone(page);
5610
5611 spin_lock_irqsave(&zone->lock, flags);
5612
5613 pfn = page_to_pfn(page);
5614 arg.start_pfn = pfn;
5615 arg.nr_pages = pageblock_nr_pages;
5616 arg.pages_found = 0;
5617
5618 /*
5619 * It may be possible to isolate a pageblock even if the
5620 * migratetype is not MIGRATE_MOVABLE. The memory isolation
5621 * notifier chain is used by balloon drivers to return the
5622 * number of pages in a range that are held by the balloon
5623 * driver to shrink memory. If all the pages are accounted for
5624 * by balloons, are free, or on the LRU, isolation can continue.
5625 * Later, for example, when memory hotplug notifier runs, these
5626 * pages reported as "can be isolated" should be isolated(freed)
5627 * by the balloon driver through the memory notifier chain.
5628 */
5629 notifier_ret = memory_isolate_notify(MEM_ISOLATE_COUNT, &arg);
5630 notifier_ret = notifier_to_errno(notifier_ret);
5631 if (notifier_ret)
5632 goto out;
5633 /*
5634 * FIXME: Now, memory hotplug doesn't call shrink_slab() by itself.
5635 * We just check MOVABLE pages.
5636 */
5637 if (__count_immobile_pages(zone, page, arg.pages_found))
5638 ret = 0;
5639
5640 /*
5641 * immobile means "not-on-lru" paes. If immobile is larger than
5642 * removable-by-driver pages reported by notifier, we'll fail.
5643 */
5644
5645out:
5646 if (!ret) {
5647 set_pageblock_migratetype(page, MIGRATE_ISOLATE);
5648 move_freepages_block(zone, page, MIGRATE_ISOLATE);
5649 }
5650
5651 spin_unlock_irqrestore(&zone->lock, flags);
5652 if (!ret)
5653 drain_all_pages();
5654 return ret;
5655}
5656
5657void unset_migratetype_isolate(struct page *page)
5658{
5659 struct zone *zone;
5660 unsigned long flags;
5661 zone = page_zone(page);
5662 spin_lock_irqsave(&zone->lock, flags);
5663 if (get_pageblock_migratetype(page) != MIGRATE_ISOLATE)
5664 goto out;
5665 set_pageblock_migratetype(page, MIGRATE_MOVABLE);
5666 move_freepages_block(zone, page, MIGRATE_MOVABLE);
5667out:
5668 spin_unlock_irqrestore(&zone->lock, flags);
5669}
5670
5671#ifdef CONFIG_MEMORY_HOTREMOVE
5672/*
5673 * All pages in the range must be isolated before calling this.
5674 */
5675void
5676__offline_isolated_pages(unsigned long start_pfn, unsigned long end_pfn)
5677{
5678 struct page *page;
5679 struct zone *zone;
5680 int order, i;
5681 unsigned long pfn;
5682 unsigned long flags;
5683 /* find the first valid pfn */
5684 for (pfn = start_pfn; pfn < end_pfn; pfn++)
5685 if (pfn_valid(pfn))
5686 break;
5687 if (pfn == end_pfn)
5688 return;
5689 zone = page_zone(pfn_to_page(pfn));
5690 spin_lock_irqsave(&zone->lock, flags);
5691 pfn = start_pfn;
5692 while (pfn < end_pfn) {
5693 if (!pfn_valid(pfn)) {
5694 pfn++;
5695 continue;
5696 }
5697 page = pfn_to_page(pfn);
5698 BUG_ON(page_count(page));
5699 BUG_ON(!PageBuddy(page));
5700 order = page_order(page);
5701#ifdef CONFIG_DEBUG_VM
5702 printk(KERN_INFO "remove from free list %lx %d %lx\n",
5703 pfn, 1 << order, end_pfn);
5704#endif
5705 list_del(&page->lru);
5706 rmv_page_order(page);
5707 zone->free_area[order].nr_free--;
5708 __mod_zone_page_state(zone, NR_FREE_PAGES,
5709 - (1UL << order));
5710 for (i = 0; i < (1 << order); i++)
5711 SetPageReserved((page+i));
5712 pfn += (1 << order);
5713 }
5714 spin_unlock_irqrestore(&zone->lock, flags);
5715}
5716#endif
5717
5718#ifdef CONFIG_MEMORY_FAILURE
5719bool is_free_buddy_page(struct page *page)
5720{
5721 struct zone *zone = page_zone(page);
5722 unsigned long pfn = page_to_pfn(page);
5723 unsigned long flags;
5724 int order;
5725
5726 spin_lock_irqsave(&zone->lock, flags);
5727 for (order = 0; order < MAX_ORDER; order++) {
5728 struct page *page_head = page - (pfn & ((1 << order) - 1));
5729
5730 if (PageBuddy(page_head) && page_order(page_head) >= order)
5731 break;
5732 }
5733 spin_unlock_irqrestore(&zone->lock, flags);
5734
5735 return order < MAX_ORDER;
5736}
5737#endif
5738
5739static struct trace_print_flags pageflag_names[] = {
5740 {1UL << PG_locked, "locked" },
5741 {1UL << PG_error, "error" },
5742 {1UL << PG_referenced, "referenced" },
5743 {1UL << PG_uptodate, "uptodate" },
5744 {1UL << PG_dirty, "dirty" },
5745 {1UL << PG_lru, "lru" },
5746 {1UL << PG_active, "active" },
5747 {1UL << PG_slab, "slab" },
5748 {1UL << PG_owner_priv_1, "owner_priv_1" },
5749 {1UL << PG_arch_1, "arch_1" },
5750 {1UL << PG_reserved, "reserved" },
5751 {1UL << PG_private, "private" },
5752 {1UL << PG_private_2, "private_2" },
5753 {1UL << PG_writeback, "writeback" },
5754#ifdef CONFIG_PAGEFLAGS_EXTENDED
5755 {1UL << PG_head, "head" },
5756 {1UL << PG_tail, "tail" },
5757#else
5758 {1UL << PG_compound, "compound" },
5759#endif
5760 {1UL << PG_swapcache, "swapcache" },
5761 {1UL << PG_mappedtodisk, "mappedtodisk" },
5762 {1UL << PG_reclaim, "reclaim" },
5763 {1UL << PG_swapbacked, "swapbacked" },
5764 {1UL << PG_unevictable, "unevictable" },
5765#ifdef CONFIG_MMU
5766 {1UL << PG_mlocked, "mlocked" },
5767#endif
5768#ifdef CONFIG_ARCH_USES_PG_UNCACHED
5769 {1UL << PG_uncached, "uncached" },
5770#endif
5771#ifdef CONFIG_MEMORY_FAILURE
5772 {1UL << PG_hwpoison, "hwpoison" },
5773#endif
5774 {-1UL, NULL },
5775};
5776
5777static void dump_page_flags(unsigned long flags)
5778{
5779 const char *delim = "";
5780 unsigned long mask;
5781 int i;
5782
5783 printk(KERN_ALERT "page flags: %#lx(", flags);
5784
5785 /* remove zone id */
5786 flags &= (1UL << NR_PAGEFLAGS) - 1;
5787
5788 for (i = 0; pageflag_names[i].name && flags; i++) {
5789
5790 mask = pageflag_names[i].mask;
5791 if ((flags & mask) != mask)
5792 continue;
5793
5794 flags &= ~mask;
5795 printk("%s%s", delim, pageflag_names[i].name);
5796 delim = "|";
5797 }
5798
5799 /* check for left over flags */
5800 if (flags)
5801 printk("%s%#lx", delim, flags);
5802
5803 printk(")\n");
5804}
5805
5806void dump_page(struct page *page)
5807{
5808 printk(KERN_ALERT
5809 "page:%p count:%d mapcount:%d mapping:%p index:%#lx\n",
5810 page, atomic_read(&page->_count), page_mapcount(page),
5811 page->mapping, page->index);
5812 dump_page_flags(page->flags);
5813 mem_cgroup_print_bad_page(page);
5814}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/page_alloc.c
4 *
5 * Manages the free list, the system allocates free pages here.
6 * Note that kmalloc() lives in slab.c
7 *
8 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
9 * Swap reorganised 29.12.95, Stephen Tweedie
10 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
11 * Reshaped it to be a zoned allocator, Ingo Molnar, Red Hat, 1999
12 * Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
13 * Zone balancing, Kanoj Sarcar, SGI, Jan 2000
14 * Per cpu hot/cold page lists, bulk allocation, Martin J. Bligh, Sept 2002
15 * (lots of bits borrowed from Ingo Molnar & Andrew Morton)
16 */
17
18#include <linux/stddef.h>
19#include <linux/mm.h>
20#include <linux/highmem.h>
21#include <linux/interrupt.h>
22#include <linux/jiffies.h>
23#include <linux/compiler.h>
24#include <linux/kernel.h>
25#include <linux/kasan.h>
26#include <linux/kmsan.h>
27#include <linux/module.h>
28#include <linux/suspend.h>
29#include <linux/ratelimit.h>
30#include <linux/oom.h>
31#include <linux/topology.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/pagevec.h>
36#include <linux/memory_hotplug.h>
37#include <linux/nodemask.h>
38#include <linux/vmstat.h>
39#include <linux/fault-inject.h>
40#include <linux/compaction.h>
41#include <trace/events/kmem.h>
42#include <trace/events/oom.h>
43#include <linux/prefetch.h>
44#include <linux/mm_inline.h>
45#include <linux/mmu_notifier.h>
46#include <linux/migrate.h>
47#include <linux/sched/mm.h>
48#include <linux/page_owner.h>
49#include <linux/page_table_check.h>
50#include <linux/memcontrol.h>
51#include <linux/ftrace.h>
52#include <linux/lockdep.h>
53#include <linux/psi.h>
54#include <linux/khugepaged.h>
55#include <linux/delayacct.h>
56#include <linux/cacheinfo.h>
57#include <linux/pgalloc_tag.h>
58#include <asm/div64.h>
59#include "internal.h"
60#include "shuffle.h"
61#include "page_reporting.h"
62
63/* Free Page Internal flags: for internal, non-pcp variants of free_pages(). */
64typedef int __bitwise fpi_t;
65
66/* No special request */
67#define FPI_NONE ((__force fpi_t)0)
68
69/*
70 * Skip free page reporting notification for the (possibly merged) page.
71 * This does not hinder free page reporting from grabbing the page,
72 * reporting it and marking it "reported" - it only skips notifying
73 * the free page reporting infrastructure about a newly freed page. For
74 * example, used when temporarily pulling a page from a freelist and
75 * putting it back unmodified.
76 */
77#define FPI_SKIP_REPORT_NOTIFY ((__force fpi_t)BIT(0))
78
79/*
80 * Place the (possibly merged) page to the tail of the freelist. Will ignore
81 * page shuffling (relevant code - e.g., memory onlining - is expected to
82 * shuffle the whole zone).
83 *
84 * Note: No code should rely on this flag for correctness - it's purely
85 * to allow for optimizations when handing back either fresh pages
86 * (memory onlining) or untouched pages (page isolation, free page
87 * reporting).
88 */
89#define FPI_TO_TAIL ((__force fpi_t)BIT(1))
90
91/* prevent >1 _updater_ of zone percpu pageset ->high and ->batch fields */
92static DEFINE_MUTEX(pcp_batch_high_lock);
93#define MIN_PERCPU_PAGELIST_HIGH_FRACTION (8)
94
95#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT_RT)
96/*
97 * On SMP, spin_trylock is sufficient protection.
98 * On PREEMPT_RT, spin_trylock is equivalent on both SMP and UP.
99 */
100#define pcp_trylock_prepare(flags) do { } while (0)
101#define pcp_trylock_finish(flag) do { } while (0)
102#else
103
104/* UP spin_trylock always succeeds so disable IRQs to prevent re-entrancy. */
105#define pcp_trylock_prepare(flags) local_irq_save(flags)
106#define pcp_trylock_finish(flags) local_irq_restore(flags)
107#endif
108
109/*
110 * Locking a pcp requires a PCP lookup followed by a spinlock. To avoid
111 * a migration causing the wrong PCP to be locked and remote memory being
112 * potentially allocated, pin the task to the CPU for the lookup+lock.
113 * preempt_disable is used on !RT because it is faster than migrate_disable.
114 * migrate_disable is used on RT because otherwise RT spinlock usage is
115 * interfered with and a high priority task cannot preempt the allocator.
116 */
117#ifndef CONFIG_PREEMPT_RT
118#define pcpu_task_pin() preempt_disable()
119#define pcpu_task_unpin() preempt_enable()
120#else
121#define pcpu_task_pin() migrate_disable()
122#define pcpu_task_unpin() migrate_enable()
123#endif
124
125/*
126 * Generic helper to lookup and a per-cpu variable with an embedded spinlock.
127 * Return value should be used with equivalent unlock helper.
128 */
129#define pcpu_spin_lock(type, member, ptr) \
130({ \
131 type *_ret; \
132 pcpu_task_pin(); \
133 _ret = this_cpu_ptr(ptr); \
134 spin_lock(&_ret->member); \
135 _ret; \
136})
137
138#define pcpu_spin_trylock(type, member, ptr) \
139({ \
140 type *_ret; \
141 pcpu_task_pin(); \
142 _ret = this_cpu_ptr(ptr); \
143 if (!spin_trylock(&_ret->member)) { \
144 pcpu_task_unpin(); \
145 _ret = NULL; \
146 } \
147 _ret; \
148})
149
150#define pcpu_spin_unlock(member, ptr) \
151({ \
152 spin_unlock(&ptr->member); \
153 pcpu_task_unpin(); \
154})
155
156/* struct per_cpu_pages specific helpers. */
157#define pcp_spin_lock(ptr) \
158 pcpu_spin_lock(struct per_cpu_pages, lock, ptr)
159
160#define pcp_spin_trylock(ptr) \
161 pcpu_spin_trylock(struct per_cpu_pages, lock, ptr)
162
163#define pcp_spin_unlock(ptr) \
164 pcpu_spin_unlock(lock, ptr)
165
166#ifdef CONFIG_USE_PERCPU_NUMA_NODE_ID
167DEFINE_PER_CPU(int, numa_node);
168EXPORT_PER_CPU_SYMBOL(numa_node);
169#endif
170
171DEFINE_STATIC_KEY_TRUE(vm_numa_stat_key);
172
173#ifdef CONFIG_HAVE_MEMORYLESS_NODES
174/*
175 * N.B., Do NOT reference the '_numa_mem_' per cpu variable directly.
176 * It will not be defined when CONFIG_HAVE_MEMORYLESS_NODES is not defined.
177 * Use the accessor functions set_numa_mem(), numa_mem_id() and cpu_to_mem()
178 * defined in <linux/topology.h>.
179 */
180DEFINE_PER_CPU(int, _numa_mem_); /* Kernel "local memory" node */
181EXPORT_PER_CPU_SYMBOL(_numa_mem_);
182#endif
183
184static DEFINE_MUTEX(pcpu_drain_mutex);
185
186#ifdef CONFIG_GCC_PLUGIN_LATENT_ENTROPY
187volatile unsigned long latent_entropy __latent_entropy;
188EXPORT_SYMBOL(latent_entropy);
189#endif
190
191/*
192 * Array of node states.
193 */
194nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
195 [N_POSSIBLE] = NODE_MASK_ALL,
196 [N_ONLINE] = { { [0] = 1UL } },
197#ifndef CONFIG_NUMA
198 [N_NORMAL_MEMORY] = { { [0] = 1UL } },
199#ifdef CONFIG_HIGHMEM
200 [N_HIGH_MEMORY] = { { [0] = 1UL } },
201#endif
202 [N_MEMORY] = { { [0] = 1UL } },
203 [N_CPU] = { { [0] = 1UL } },
204#endif /* NUMA */
205};
206EXPORT_SYMBOL(node_states);
207
208gfp_t gfp_allowed_mask __read_mostly = GFP_BOOT_MASK;
209
210#ifdef CONFIG_HUGETLB_PAGE_SIZE_VARIABLE
211unsigned int pageblock_order __read_mostly;
212#endif
213
214static void __free_pages_ok(struct page *page, unsigned int order,
215 fpi_t fpi_flags);
216
217/*
218 * results with 256, 32 in the lowmem_reserve sysctl:
219 * 1G machine -> (16M dma, 800M-16M normal, 1G-800M high)
220 * 1G machine -> (16M dma, 784M normal, 224M high)
221 * NORMAL allocation will leave 784M/256 of ram reserved in the ZONE_DMA
222 * HIGHMEM allocation will leave 224M/32 of ram reserved in ZONE_NORMAL
223 * HIGHMEM allocation will leave (224M+784M)/256 of ram reserved in ZONE_DMA
224 *
225 * TBD: should special case ZONE_DMA32 machines here - in those we normally
226 * don't need any ZONE_NORMAL reservation
227 */
228static int sysctl_lowmem_reserve_ratio[MAX_NR_ZONES] = {
229#ifdef CONFIG_ZONE_DMA
230 [ZONE_DMA] = 256,
231#endif
232#ifdef CONFIG_ZONE_DMA32
233 [ZONE_DMA32] = 256,
234#endif
235 [ZONE_NORMAL] = 32,
236#ifdef CONFIG_HIGHMEM
237 [ZONE_HIGHMEM] = 0,
238#endif
239 [ZONE_MOVABLE] = 0,
240};
241
242char * const zone_names[MAX_NR_ZONES] = {
243#ifdef CONFIG_ZONE_DMA
244 "DMA",
245#endif
246#ifdef CONFIG_ZONE_DMA32
247 "DMA32",
248#endif
249 "Normal",
250#ifdef CONFIG_HIGHMEM
251 "HighMem",
252#endif
253 "Movable",
254#ifdef CONFIG_ZONE_DEVICE
255 "Device",
256#endif
257};
258
259const char * const migratetype_names[MIGRATE_TYPES] = {
260 "Unmovable",
261 "Movable",
262 "Reclaimable",
263 "HighAtomic",
264#ifdef CONFIG_CMA
265 "CMA",
266#endif
267#ifdef CONFIG_MEMORY_ISOLATION
268 "Isolate",
269#endif
270};
271
272int min_free_kbytes = 1024;
273int user_min_free_kbytes = -1;
274static int watermark_boost_factor __read_mostly = 15000;
275static int watermark_scale_factor = 10;
276
277/* movable_zone is the "real" zone pages in ZONE_MOVABLE are taken from */
278int movable_zone;
279EXPORT_SYMBOL(movable_zone);
280
281#if MAX_NUMNODES > 1
282unsigned int nr_node_ids __read_mostly = MAX_NUMNODES;
283unsigned int nr_online_nodes __read_mostly = 1;
284EXPORT_SYMBOL(nr_node_ids);
285EXPORT_SYMBOL(nr_online_nodes);
286#endif
287
288static bool page_contains_unaccepted(struct page *page, unsigned int order);
289static bool cond_accept_memory(struct zone *zone, unsigned int order);
290static bool __free_unaccepted(struct page *page);
291
292int page_group_by_mobility_disabled __read_mostly;
293
294#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
295/*
296 * During boot we initialize deferred pages on-demand, as needed, but once
297 * page_alloc_init_late() has finished, the deferred pages are all initialized,
298 * and we can permanently disable that path.
299 */
300DEFINE_STATIC_KEY_TRUE(deferred_pages);
301
302static inline bool deferred_pages_enabled(void)
303{
304 return static_branch_unlikely(&deferred_pages);
305}
306
307/*
308 * deferred_grow_zone() is __init, but it is called from
309 * get_page_from_freelist() during early boot until deferred_pages permanently
310 * disables this call. This is why we have refdata wrapper to avoid warning,
311 * and to ensure that the function body gets unloaded.
312 */
313static bool __ref
314_deferred_grow_zone(struct zone *zone, unsigned int order)
315{
316 return deferred_grow_zone(zone, order);
317}
318#else
319static inline bool deferred_pages_enabled(void)
320{
321 return false;
322}
323
324static inline bool _deferred_grow_zone(struct zone *zone, unsigned int order)
325{
326 return false;
327}
328#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
329
330/* Return a pointer to the bitmap storing bits affecting a block of pages */
331static inline unsigned long *get_pageblock_bitmap(const struct page *page,
332 unsigned long pfn)
333{
334#ifdef CONFIG_SPARSEMEM
335 return section_to_usemap(__pfn_to_section(pfn));
336#else
337 return page_zone(page)->pageblock_flags;
338#endif /* CONFIG_SPARSEMEM */
339}
340
341static inline int pfn_to_bitidx(const struct page *page, unsigned long pfn)
342{
343#ifdef CONFIG_SPARSEMEM
344 pfn &= (PAGES_PER_SECTION-1);
345#else
346 pfn = pfn - pageblock_start_pfn(page_zone(page)->zone_start_pfn);
347#endif /* CONFIG_SPARSEMEM */
348 return (pfn >> pageblock_order) * NR_PAGEBLOCK_BITS;
349}
350
351/**
352 * get_pfnblock_flags_mask - Return the requested group of flags for the pageblock_nr_pages block of pages
353 * @page: The page within the block of interest
354 * @pfn: The target page frame number
355 * @mask: mask of bits that the caller is interested in
356 *
357 * Return: pageblock_bits flags
358 */
359unsigned long get_pfnblock_flags_mask(const struct page *page,
360 unsigned long pfn, unsigned long mask)
361{
362 unsigned long *bitmap;
363 unsigned long bitidx, word_bitidx;
364 unsigned long word;
365
366 bitmap = get_pageblock_bitmap(page, pfn);
367 bitidx = pfn_to_bitidx(page, pfn);
368 word_bitidx = bitidx / BITS_PER_LONG;
369 bitidx &= (BITS_PER_LONG-1);
370 /*
371 * This races, without locks, with set_pfnblock_flags_mask(). Ensure
372 * a consistent read of the memory array, so that results, even though
373 * racy, are not corrupted.
374 */
375 word = READ_ONCE(bitmap[word_bitidx]);
376 return (word >> bitidx) & mask;
377}
378
379static __always_inline int get_pfnblock_migratetype(const struct page *page,
380 unsigned long pfn)
381{
382 return get_pfnblock_flags_mask(page, pfn, MIGRATETYPE_MASK);
383}
384
385/**
386 * set_pfnblock_flags_mask - Set the requested group of flags for a pageblock_nr_pages block of pages
387 * @page: The page within the block of interest
388 * @flags: The flags to set
389 * @pfn: The target page frame number
390 * @mask: mask of bits that the caller is interested in
391 */
392void set_pfnblock_flags_mask(struct page *page, unsigned long flags,
393 unsigned long pfn,
394 unsigned long mask)
395{
396 unsigned long *bitmap;
397 unsigned long bitidx, word_bitidx;
398 unsigned long word;
399
400 BUILD_BUG_ON(NR_PAGEBLOCK_BITS != 4);
401 BUILD_BUG_ON(MIGRATE_TYPES > (1 << PB_migratetype_bits));
402
403 bitmap = get_pageblock_bitmap(page, pfn);
404 bitidx = pfn_to_bitidx(page, pfn);
405 word_bitidx = bitidx / BITS_PER_LONG;
406 bitidx &= (BITS_PER_LONG-1);
407
408 VM_BUG_ON_PAGE(!zone_spans_pfn(page_zone(page), pfn), page);
409
410 mask <<= bitidx;
411 flags <<= bitidx;
412
413 word = READ_ONCE(bitmap[word_bitidx]);
414 do {
415 } while (!try_cmpxchg(&bitmap[word_bitidx], &word, (word & ~mask) | flags));
416}
417
418void set_pageblock_migratetype(struct page *page, int migratetype)
419{
420 if (unlikely(page_group_by_mobility_disabled &&
421 migratetype < MIGRATE_PCPTYPES))
422 migratetype = MIGRATE_UNMOVABLE;
423
424 set_pfnblock_flags_mask(page, (unsigned long)migratetype,
425 page_to_pfn(page), MIGRATETYPE_MASK);
426}
427
428#ifdef CONFIG_DEBUG_VM
429static int page_outside_zone_boundaries(struct zone *zone, struct page *page)
430{
431 int ret;
432 unsigned seq;
433 unsigned long pfn = page_to_pfn(page);
434 unsigned long sp, start_pfn;
435
436 do {
437 seq = zone_span_seqbegin(zone);
438 start_pfn = zone->zone_start_pfn;
439 sp = zone->spanned_pages;
440 ret = !zone_spans_pfn(zone, pfn);
441 } while (zone_span_seqretry(zone, seq));
442
443 if (ret)
444 pr_err("page 0x%lx outside node %d zone %s [ 0x%lx - 0x%lx ]\n",
445 pfn, zone_to_nid(zone), zone->name,
446 start_pfn, start_pfn + sp);
447
448 return ret;
449}
450
451/*
452 * Temporary debugging check for pages not lying within a given zone.
453 */
454static bool __maybe_unused bad_range(struct zone *zone, struct page *page)
455{
456 if (page_outside_zone_boundaries(zone, page))
457 return true;
458 if (zone != page_zone(page))
459 return true;
460
461 return false;
462}
463#else
464static inline bool __maybe_unused bad_range(struct zone *zone, struct page *page)
465{
466 return false;
467}
468#endif
469
470static void bad_page(struct page *page, const char *reason)
471{
472 static unsigned long resume;
473 static unsigned long nr_shown;
474 static unsigned long nr_unshown;
475
476 /*
477 * Allow a burst of 60 reports, then keep quiet for that minute;
478 * or allow a steady drip of one report per second.
479 */
480 if (nr_shown == 60) {
481 if (time_before(jiffies, resume)) {
482 nr_unshown++;
483 goto out;
484 }
485 if (nr_unshown) {
486 pr_alert(
487 "BUG: Bad page state: %lu messages suppressed\n",
488 nr_unshown);
489 nr_unshown = 0;
490 }
491 nr_shown = 0;
492 }
493 if (nr_shown++ == 0)
494 resume = jiffies + 60 * HZ;
495
496 pr_alert("BUG: Bad page state in process %s pfn:%05lx\n",
497 current->comm, page_to_pfn(page));
498 dump_page(page, reason);
499
500 print_modules();
501 dump_stack();
502out:
503 /* Leave bad fields for debug, except PageBuddy could make trouble */
504 if (PageBuddy(page))
505 __ClearPageBuddy(page);
506 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
507}
508
509static inline unsigned int order_to_pindex(int migratetype, int order)
510{
511 bool __maybe_unused movable;
512
513#ifdef CONFIG_TRANSPARENT_HUGEPAGE
514 if (order > PAGE_ALLOC_COSTLY_ORDER) {
515 VM_BUG_ON(order != HPAGE_PMD_ORDER);
516
517 movable = migratetype == MIGRATE_MOVABLE;
518
519 return NR_LOWORDER_PCP_LISTS + movable;
520 }
521#else
522 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
523#endif
524
525 return (MIGRATE_PCPTYPES * order) + migratetype;
526}
527
528static inline int pindex_to_order(unsigned int pindex)
529{
530 int order = pindex / MIGRATE_PCPTYPES;
531
532#ifdef CONFIG_TRANSPARENT_HUGEPAGE
533 if (pindex >= NR_LOWORDER_PCP_LISTS)
534 order = HPAGE_PMD_ORDER;
535#else
536 VM_BUG_ON(order > PAGE_ALLOC_COSTLY_ORDER);
537#endif
538
539 return order;
540}
541
542static inline bool pcp_allowed_order(unsigned int order)
543{
544 if (order <= PAGE_ALLOC_COSTLY_ORDER)
545 return true;
546#ifdef CONFIG_TRANSPARENT_HUGEPAGE
547 if (order == HPAGE_PMD_ORDER)
548 return true;
549#endif
550 return false;
551}
552
553/*
554 * Higher-order pages are called "compound pages". They are structured thusly:
555 *
556 * The first PAGE_SIZE page is called the "head page" and have PG_head set.
557 *
558 * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
559 * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
560 *
561 * The first tail page's ->compound_order holds the order of allocation.
562 * This usage means that zero-order pages may not be compound.
563 */
564
565void prep_compound_page(struct page *page, unsigned int order)
566{
567 int i;
568 int nr_pages = 1 << order;
569
570 __SetPageHead(page);
571 for (i = 1; i < nr_pages; i++)
572 prep_compound_tail(page, i);
573
574 prep_compound_head(page, order);
575}
576
577static inline void set_buddy_order(struct page *page, unsigned int order)
578{
579 set_page_private(page, order);
580 __SetPageBuddy(page);
581}
582
583#ifdef CONFIG_COMPACTION
584static inline struct capture_control *task_capc(struct zone *zone)
585{
586 struct capture_control *capc = current->capture_control;
587
588 return unlikely(capc) &&
589 !(current->flags & PF_KTHREAD) &&
590 !capc->page &&
591 capc->cc->zone == zone ? capc : NULL;
592}
593
594static inline bool
595compaction_capture(struct capture_control *capc, struct page *page,
596 int order, int migratetype)
597{
598 if (!capc || order != capc->cc->order)
599 return false;
600
601 /* Do not accidentally pollute CMA or isolated regions*/
602 if (is_migrate_cma(migratetype) ||
603 is_migrate_isolate(migratetype))
604 return false;
605
606 /*
607 * Do not let lower order allocations pollute a movable pageblock
608 * unless compaction is also requesting movable pages.
609 * This might let an unmovable request use a reclaimable pageblock
610 * and vice-versa but no more than normal fallback logic which can
611 * have trouble finding a high-order free page.
612 */
613 if (order < pageblock_order && migratetype == MIGRATE_MOVABLE &&
614 capc->cc->migratetype != MIGRATE_MOVABLE)
615 return false;
616
617 capc->page = page;
618 return true;
619}
620
621#else
622static inline struct capture_control *task_capc(struct zone *zone)
623{
624 return NULL;
625}
626
627static inline bool
628compaction_capture(struct capture_control *capc, struct page *page,
629 int order, int migratetype)
630{
631 return false;
632}
633#endif /* CONFIG_COMPACTION */
634
635static inline void account_freepages(struct zone *zone, int nr_pages,
636 int migratetype)
637{
638 lockdep_assert_held(&zone->lock);
639
640 if (is_migrate_isolate(migratetype))
641 return;
642
643 __mod_zone_page_state(zone, NR_FREE_PAGES, nr_pages);
644
645 if (is_migrate_cma(migratetype))
646 __mod_zone_page_state(zone, NR_FREE_CMA_PAGES, nr_pages);
647 else if (is_migrate_highatomic(migratetype))
648 WRITE_ONCE(zone->nr_free_highatomic,
649 zone->nr_free_highatomic + nr_pages);
650}
651
652/* Used for pages not on another list */
653static inline void __add_to_free_list(struct page *page, struct zone *zone,
654 unsigned int order, int migratetype,
655 bool tail)
656{
657 struct free_area *area = &zone->free_area[order];
658
659 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
660 "page type is %lu, passed migratetype is %d (nr=%d)\n",
661 get_pageblock_migratetype(page), migratetype, 1 << order);
662
663 if (tail)
664 list_add_tail(&page->buddy_list, &area->free_list[migratetype]);
665 else
666 list_add(&page->buddy_list, &area->free_list[migratetype]);
667 area->nr_free++;
668}
669
670/*
671 * Used for pages which are on another list. Move the pages to the tail
672 * of the list - so the moved pages won't immediately be considered for
673 * allocation again (e.g., optimization for memory onlining).
674 */
675static inline void move_to_free_list(struct page *page, struct zone *zone,
676 unsigned int order, int old_mt, int new_mt)
677{
678 struct free_area *area = &zone->free_area[order];
679
680 /* Free page moving can fail, so it happens before the type update */
681 VM_WARN_ONCE(get_pageblock_migratetype(page) != old_mt,
682 "page type is %lu, passed migratetype is %d (nr=%d)\n",
683 get_pageblock_migratetype(page), old_mt, 1 << order);
684
685 list_move_tail(&page->buddy_list, &area->free_list[new_mt]);
686
687 account_freepages(zone, -(1 << order), old_mt);
688 account_freepages(zone, 1 << order, new_mt);
689}
690
691static inline void __del_page_from_free_list(struct page *page, struct zone *zone,
692 unsigned int order, int migratetype)
693{
694 VM_WARN_ONCE(get_pageblock_migratetype(page) != migratetype,
695 "page type is %lu, passed migratetype is %d (nr=%d)\n",
696 get_pageblock_migratetype(page), migratetype, 1 << order);
697
698 /* clear reported state and update reported page count */
699 if (page_reported(page))
700 __ClearPageReported(page);
701
702 list_del(&page->buddy_list);
703 __ClearPageBuddy(page);
704 set_page_private(page, 0);
705 zone->free_area[order].nr_free--;
706}
707
708static inline void del_page_from_free_list(struct page *page, struct zone *zone,
709 unsigned int order, int migratetype)
710{
711 __del_page_from_free_list(page, zone, order, migratetype);
712 account_freepages(zone, -(1 << order), migratetype);
713}
714
715static inline struct page *get_page_from_free_area(struct free_area *area,
716 int migratetype)
717{
718 return list_first_entry_or_null(&area->free_list[migratetype],
719 struct page, buddy_list);
720}
721
722/*
723 * If this is less than the 2nd largest possible page, check if the buddy
724 * of the next-higher order is free. If it is, it's possible
725 * that pages are being freed that will coalesce soon. In case,
726 * that is happening, add the free page to the tail of the list
727 * so it's less likely to be used soon and more likely to be merged
728 * as a 2-level higher order page
729 */
730static inline bool
731buddy_merge_likely(unsigned long pfn, unsigned long buddy_pfn,
732 struct page *page, unsigned int order)
733{
734 unsigned long higher_page_pfn;
735 struct page *higher_page;
736
737 if (order >= MAX_PAGE_ORDER - 1)
738 return false;
739
740 higher_page_pfn = buddy_pfn & pfn;
741 higher_page = page + (higher_page_pfn - pfn);
742
743 return find_buddy_page_pfn(higher_page, higher_page_pfn, order + 1,
744 NULL) != NULL;
745}
746
747/*
748 * Freeing function for a buddy system allocator.
749 *
750 * The concept of a buddy system is to maintain direct-mapped table
751 * (containing bit values) for memory blocks of various "orders".
752 * The bottom level table contains the map for the smallest allocatable
753 * units of memory (here, pages), and each level above it describes
754 * pairs of units from the levels below, hence, "buddies".
755 * At a high level, all that happens here is marking the table entry
756 * at the bottom level available, and propagating the changes upward
757 * as necessary, plus some accounting needed to play nicely with other
758 * parts of the VM system.
759 * At each level, we keep a list of pages, which are heads of continuous
760 * free pages of length of (1 << order) and marked with PageBuddy.
761 * Page's order is recorded in page_private(page) field.
762 * So when we are allocating or freeing one, we can derive the state of the
763 * other. That is, if we allocate a small block, and both were
764 * free, the remainder of the region must be split into blocks.
765 * If a block is freed, and its buddy is also free, then this
766 * triggers coalescing into a block of larger size.
767 *
768 * -- nyc
769 */
770
771static inline void __free_one_page(struct page *page,
772 unsigned long pfn,
773 struct zone *zone, unsigned int order,
774 int migratetype, fpi_t fpi_flags)
775{
776 struct capture_control *capc = task_capc(zone);
777 unsigned long buddy_pfn = 0;
778 unsigned long combined_pfn;
779 struct page *buddy;
780 bool to_tail;
781
782 VM_BUG_ON(!zone_is_initialized(zone));
783 VM_BUG_ON_PAGE(page->flags & PAGE_FLAGS_CHECK_AT_PREP, page);
784
785 VM_BUG_ON(migratetype == -1);
786 VM_BUG_ON_PAGE(pfn & ((1 << order) - 1), page);
787 VM_BUG_ON_PAGE(bad_range(zone, page), page);
788
789 account_freepages(zone, 1 << order, migratetype);
790
791 while (order < MAX_PAGE_ORDER) {
792 int buddy_mt = migratetype;
793
794 if (compaction_capture(capc, page, order, migratetype)) {
795 account_freepages(zone, -(1 << order), migratetype);
796 return;
797 }
798
799 buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
800 if (!buddy)
801 goto done_merging;
802
803 if (unlikely(order >= pageblock_order)) {
804 /*
805 * We want to prevent merge between freepages on pageblock
806 * without fallbacks and normal pageblock. Without this,
807 * pageblock isolation could cause incorrect freepage or CMA
808 * accounting or HIGHATOMIC accounting.
809 */
810 buddy_mt = get_pfnblock_migratetype(buddy, buddy_pfn);
811
812 if (migratetype != buddy_mt &&
813 (!migratetype_is_mergeable(migratetype) ||
814 !migratetype_is_mergeable(buddy_mt)))
815 goto done_merging;
816 }
817
818 /*
819 * Our buddy is free or it is CONFIG_DEBUG_PAGEALLOC guard page,
820 * merge with it and move up one order.
821 */
822 if (page_is_guard(buddy))
823 clear_page_guard(zone, buddy, order);
824 else
825 __del_page_from_free_list(buddy, zone, order, buddy_mt);
826
827 if (unlikely(buddy_mt != migratetype)) {
828 /*
829 * Match buddy type. This ensures that an
830 * expand() down the line puts the sub-blocks
831 * on the right freelists.
832 */
833 set_pageblock_migratetype(buddy, migratetype);
834 }
835
836 combined_pfn = buddy_pfn & pfn;
837 page = page + (combined_pfn - pfn);
838 pfn = combined_pfn;
839 order++;
840 }
841
842done_merging:
843 set_buddy_order(page, order);
844
845 if (fpi_flags & FPI_TO_TAIL)
846 to_tail = true;
847 else if (is_shuffle_order(order))
848 to_tail = shuffle_pick_tail();
849 else
850 to_tail = buddy_merge_likely(pfn, buddy_pfn, page, order);
851
852 __add_to_free_list(page, zone, order, migratetype, to_tail);
853
854 /* Notify page reporting subsystem of freed page */
855 if (!(fpi_flags & FPI_SKIP_REPORT_NOTIFY))
856 page_reporting_notify_free(order);
857}
858
859/*
860 * A bad page could be due to a number of fields. Instead of multiple branches,
861 * try and check multiple fields with one check. The caller must do a detailed
862 * check if necessary.
863 */
864static inline bool page_expected_state(struct page *page,
865 unsigned long check_flags)
866{
867 if (unlikely(atomic_read(&page->_mapcount) != -1))
868 return false;
869
870 if (unlikely((unsigned long)page->mapping |
871 page_ref_count(page) |
872#ifdef CONFIG_MEMCG
873 page->memcg_data |
874#endif
875#ifdef CONFIG_PAGE_POOL
876 ((page->pp_magic & ~0x3UL) == PP_SIGNATURE) |
877#endif
878 (page->flags & check_flags)))
879 return false;
880
881 return true;
882}
883
884static const char *page_bad_reason(struct page *page, unsigned long flags)
885{
886 const char *bad_reason = NULL;
887
888 if (unlikely(atomic_read(&page->_mapcount) != -1))
889 bad_reason = "nonzero mapcount";
890 if (unlikely(page->mapping != NULL))
891 bad_reason = "non-NULL mapping";
892 if (unlikely(page_ref_count(page) != 0))
893 bad_reason = "nonzero _refcount";
894 if (unlikely(page->flags & flags)) {
895 if (flags == PAGE_FLAGS_CHECK_AT_PREP)
896 bad_reason = "PAGE_FLAGS_CHECK_AT_PREP flag(s) set";
897 else
898 bad_reason = "PAGE_FLAGS_CHECK_AT_FREE flag(s) set";
899 }
900#ifdef CONFIG_MEMCG
901 if (unlikely(page->memcg_data))
902 bad_reason = "page still charged to cgroup";
903#endif
904#ifdef CONFIG_PAGE_POOL
905 if (unlikely((page->pp_magic & ~0x3UL) == PP_SIGNATURE))
906 bad_reason = "page_pool leak";
907#endif
908 return bad_reason;
909}
910
911static void free_page_is_bad_report(struct page *page)
912{
913 bad_page(page,
914 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_FREE));
915}
916
917static inline bool free_page_is_bad(struct page *page)
918{
919 if (likely(page_expected_state(page, PAGE_FLAGS_CHECK_AT_FREE)))
920 return false;
921
922 /* Something has gone sideways, find it */
923 free_page_is_bad_report(page);
924 return true;
925}
926
927static inline bool is_check_pages_enabled(void)
928{
929 return static_branch_unlikely(&check_pages_enabled);
930}
931
932static int free_tail_page_prepare(struct page *head_page, struct page *page)
933{
934 struct folio *folio = (struct folio *)head_page;
935 int ret = 1;
936
937 /*
938 * We rely page->lru.next never has bit 0 set, unless the page
939 * is PageTail(). Let's make sure that's true even for poisoned ->lru.
940 */
941 BUILD_BUG_ON((unsigned long)LIST_POISON1 & 1);
942
943 if (!is_check_pages_enabled()) {
944 ret = 0;
945 goto out;
946 }
947 switch (page - head_page) {
948 case 1:
949 /* the first tail page: these may be in place of ->mapping */
950 if (unlikely(folio_entire_mapcount(folio))) {
951 bad_page(page, "nonzero entire_mapcount");
952 goto out;
953 }
954 if (unlikely(folio_large_mapcount(folio))) {
955 bad_page(page, "nonzero large_mapcount");
956 goto out;
957 }
958 if (unlikely(atomic_read(&folio->_nr_pages_mapped))) {
959 bad_page(page, "nonzero nr_pages_mapped");
960 goto out;
961 }
962 if (unlikely(atomic_read(&folio->_pincount))) {
963 bad_page(page, "nonzero pincount");
964 goto out;
965 }
966 break;
967 case 2:
968 /* the second tail page: deferred_list overlaps ->mapping */
969 if (unlikely(!list_empty(&folio->_deferred_list))) {
970 bad_page(page, "on deferred list");
971 goto out;
972 }
973 break;
974 default:
975 if (page->mapping != TAIL_MAPPING) {
976 bad_page(page, "corrupted mapping in tail page");
977 goto out;
978 }
979 break;
980 }
981 if (unlikely(!PageTail(page))) {
982 bad_page(page, "PageTail not set");
983 goto out;
984 }
985 if (unlikely(compound_head(page) != head_page)) {
986 bad_page(page, "compound_head not consistent");
987 goto out;
988 }
989 ret = 0;
990out:
991 page->mapping = NULL;
992 clear_compound_head(page);
993 return ret;
994}
995
996/*
997 * Skip KASAN memory poisoning when either:
998 *
999 * 1. For generic KASAN: deferred memory initialization has not yet completed.
1000 * Tag-based KASAN modes skip pages freed via deferred memory initialization
1001 * using page tags instead (see below).
1002 * 2. For tag-based KASAN modes: the page has a match-all KASAN tag, indicating
1003 * that error detection is disabled for accesses via the page address.
1004 *
1005 * Pages will have match-all tags in the following circumstances:
1006 *
1007 * 1. Pages are being initialized for the first time, including during deferred
1008 * memory init; see the call to page_kasan_tag_reset in __init_single_page.
1009 * 2. The allocation was not unpoisoned due to __GFP_SKIP_KASAN, with the
1010 * exception of pages unpoisoned by kasan_unpoison_vmalloc.
1011 * 3. The allocation was excluded from being checked due to sampling,
1012 * see the call to kasan_unpoison_pages.
1013 *
1014 * Poisoning pages during deferred memory init will greatly lengthen the
1015 * process and cause problem in large memory systems as the deferred pages
1016 * initialization is done with interrupt disabled.
1017 *
1018 * Assuming that there will be no reference to those newly initialized
1019 * pages before they are ever allocated, this should have no effect on
1020 * KASAN memory tracking as the poison will be properly inserted at page
1021 * allocation time. The only corner case is when pages are allocated by
1022 * on-demand allocation and then freed again before the deferred pages
1023 * initialization is done, but this is not likely to happen.
1024 */
1025static inline bool should_skip_kasan_poison(struct page *page)
1026{
1027 if (IS_ENABLED(CONFIG_KASAN_GENERIC))
1028 return deferred_pages_enabled();
1029
1030 return page_kasan_tag(page) == KASAN_TAG_KERNEL;
1031}
1032
1033static void kernel_init_pages(struct page *page, int numpages)
1034{
1035 int i;
1036
1037 /* s390's use of memset() could override KASAN redzones. */
1038 kasan_disable_current();
1039 for (i = 0; i < numpages; i++)
1040 clear_highpage_kasan_tagged(page + i);
1041 kasan_enable_current();
1042}
1043
1044__always_inline bool free_pages_prepare(struct page *page,
1045 unsigned int order)
1046{
1047 int bad = 0;
1048 bool skip_kasan_poison = should_skip_kasan_poison(page);
1049 bool init = want_init_on_free();
1050 bool compound = PageCompound(page);
1051 struct folio *folio = page_folio(page);
1052
1053 VM_BUG_ON_PAGE(PageTail(page), page);
1054
1055 trace_mm_page_free(page, order);
1056 kmsan_free_page(page, order);
1057
1058 if (memcg_kmem_online() && PageMemcgKmem(page))
1059 __memcg_kmem_uncharge_page(page, order);
1060
1061 /*
1062 * In rare cases, when truncation or holepunching raced with
1063 * munlock after VM_LOCKED was cleared, Mlocked may still be
1064 * found set here. This does not indicate a problem, unless
1065 * "unevictable_pgs_cleared" appears worryingly large.
1066 */
1067 if (unlikely(folio_test_mlocked(folio))) {
1068 long nr_pages = folio_nr_pages(folio);
1069
1070 __folio_clear_mlocked(folio);
1071 zone_stat_mod_folio(folio, NR_MLOCK, -nr_pages);
1072 count_vm_events(UNEVICTABLE_PGCLEARED, nr_pages);
1073 }
1074
1075 if (unlikely(PageHWPoison(page)) && !order) {
1076 /* Do not let hwpoison pages hit pcplists/buddy */
1077 reset_page_owner(page, order);
1078 page_table_check_free(page, order);
1079 pgalloc_tag_sub(page, 1 << order);
1080
1081 /*
1082 * The page is isolated and accounted for.
1083 * Mark the codetag as empty to avoid accounting error
1084 * when the page is freed by unpoison_memory().
1085 */
1086 clear_page_tag_ref(page);
1087 return false;
1088 }
1089
1090 VM_BUG_ON_PAGE(compound && compound_order(page) != order, page);
1091
1092 /*
1093 * Check tail pages before head page information is cleared to
1094 * avoid checking PageCompound for order-0 pages.
1095 */
1096 if (unlikely(order)) {
1097 int i;
1098
1099 if (compound)
1100 page[1].flags &= ~PAGE_FLAGS_SECOND;
1101 for (i = 1; i < (1 << order); i++) {
1102 if (compound)
1103 bad += free_tail_page_prepare(page, page + i);
1104 if (is_check_pages_enabled()) {
1105 if (free_page_is_bad(page + i)) {
1106 bad++;
1107 continue;
1108 }
1109 }
1110 (page + i)->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1111 }
1112 }
1113 if (PageMappingFlags(page)) {
1114 if (PageAnon(page))
1115 mod_mthp_stat(order, MTHP_STAT_NR_ANON, -1);
1116 page->mapping = NULL;
1117 }
1118 if (is_check_pages_enabled()) {
1119 if (free_page_is_bad(page))
1120 bad++;
1121 if (bad)
1122 return false;
1123 }
1124
1125 page_cpupid_reset_last(page);
1126 page->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
1127 reset_page_owner(page, order);
1128 page_table_check_free(page, order);
1129 pgalloc_tag_sub(page, 1 << order);
1130
1131 if (!PageHighMem(page)) {
1132 debug_check_no_locks_freed(page_address(page),
1133 PAGE_SIZE << order);
1134 debug_check_no_obj_freed(page_address(page),
1135 PAGE_SIZE << order);
1136 }
1137
1138 kernel_poison_pages(page, 1 << order);
1139
1140 /*
1141 * As memory initialization might be integrated into KASAN,
1142 * KASAN poisoning and memory initialization code must be
1143 * kept together to avoid discrepancies in behavior.
1144 *
1145 * With hardware tag-based KASAN, memory tags must be set before the
1146 * page becomes unavailable via debug_pagealloc or arch_free_page.
1147 */
1148 if (!skip_kasan_poison) {
1149 kasan_poison_pages(page, order, init);
1150
1151 /* Memory is already initialized if KASAN did it internally. */
1152 if (kasan_has_integrated_init())
1153 init = false;
1154 }
1155 if (init)
1156 kernel_init_pages(page, 1 << order);
1157
1158 /*
1159 * arch_free_page() can make the page's contents inaccessible. s390
1160 * does this. So nothing which can access the page's contents should
1161 * happen after this.
1162 */
1163 arch_free_page(page, order);
1164
1165 debug_pagealloc_unmap_pages(page, 1 << order);
1166
1167 return true;
1168}
1169
1170/*
1171 * Frees a number of pages from the PCP lists
1172 * Assumes all pages on list are in same zone.
1173 * count is the number of pages to free.
1174 */
1175static void free_pcppages_bulk(struct zone *zone, int count,
1176 struct per_cpu_pages *pcp,
1177 int pindex)
1178{
1179 unsigned long flags;
1180 unsigned int order;
1181 struct page *page;
1182
1183 /*
1184 * Ensure proper count is passed which otherwise would stuck in the
1185 * below while (list_empty(list)) loop.
1186 */
1187 count = min(pcp->count, count);
1188
1189 /* Ensure requested pindex is drained first. */
1190 pindex = pindex - 1;
1191
1192 spin_lock_irqsave(&zone->lock, flags);
1193
1194 while (count > 0) {
1195 struct list_head *list;
1196 int nr_pages;
1197
1198 /* Remove pages from lists in a round-robin fashion. */
1199 do {
1200 if (++pindex > NR_PCP_LISTS - 1)
1201 pindex = 0;
1202 list = &pcp->lists[pindex];
1203 } while (list_empty(list));
1204
1205 order = pindex_to_order(pindex);
1206 nr_pages = 1 << order;
1207 do {
1208 unsigned long pfn;
1209 int mt;
1210
1211 page = list_last_entry(list, struct page, pcp_list);
1212 pfn = page_to_pfn(page);
1213 mt = get_pfnblock_migratetype(page, pfn);
1214
1215 /* must delete to avoid corrupting pcp list */
1216 list_del(&page->pcp_list);
1217 count -= nr_pages;
1218 pcp->count -= nr_pages;
1219
1220 __free_one_page(page, pfn, zone, order, mt, FPI_NONE);
1221 trace_mm_page_pcpu_drain(page, order, mt);
1222 } while (count > 0 && !list_empty(list));
1223 }
1224
1225 spin_unlock_irqrestore(&zone->lock, flags);
1226}
1227
1228/* Split a multi-block free page into its individual pageblocks. */
1229static void split_large_buddy(struct zone *zone, struct page *page,
1230 unsigned long pfn, int order, fpi_t fpi)
1231{
1232 unsigned long end = pfn + (1 << order);
1233
1234 VM_WARN_ON_ONCE(!IS_ALIGNED(pfn, 1 << order));
1235 /* Caller removed page from freelist, buddy info cleared! */
1236 VM_WARN_ON_ONCE(PageBuddy(page));
1237
1238 if (order > pageblock_order)
1239 order = pageblock_order;
1240
1241 do {
1242 int mt = get_pfnblock_migratetype(page, pfn);
1243
1244 __free_one_page(page, pfn, zone, order, mt, fpi);
1245 pfn += 1 << order;
1246 if (pfn == end)
1247 break;
1248 page = pfn_to_page(pfn);
1249 } while (1);
1250}
1251
1252static void free_one_page(struct zone *zone, struct page *page,
1253 unsigned long pfn, unsigned int order,
1254 fpi_t fpi_flags)
1255{
1256 unsigned long flags;
1257
1258 spin_lock_irqsave(&zone->lock, flags);
1259 split_large_buddy(zone, page, pfn, order, fpi_flags);
1260 spin_unlock_irqrestore(&zone->lock, flags);
1261
1262 __count_vm_events(PGFREE, 1 << order);
1263}
1264
1265static void __free_pages_ok(struct page *page, unsigned int order,
1266 fpi_t fpi_flags)
1267{
1268 unsigned long pfn = page_to_pfn(page);
1269 struct zone *zone = page_zone(page);
1270
1271 if (free_pages_prepare(page, order))
1272 free_one_page(zone, page, pfn, order, fpi_flags);
1273}
1274
1275void __meminit __free_pages_core(struct page *page, unsigned int order,
1276 enum meminit_context context)
1277{
1278 unsigned int nr_pages = 1 << order;
1279 struct page *p = page;
1280 unsigned int loop;
1281
1282 /*
1283 * When initializing the memmap, __init_single_page() sets the refcount
1284 * of all pages to 1 ("allocated"/"not free"). We have to set the
1285 * refcount of all involved pages to 0.
1286 *
1287 * Note that hotplugged memory pages are initialized to PageOffline().
1288 * Pages freed from memblock might be marked as reserved.
1289 */
1290 if (IS_ENABLED(CONFIG_MEMORY_HOTPLUG) &&
1291 unlikely(context == MEMINIT_HOTPLUG)) {
1292 for (loop = 0; loop < nr_pages; loop++, p++) {
1293 VM_WARN_ON_ONCE(PageReserved(p));
1294 __ClearPageOffline(p);
1295 set_page_count(p, 0);
1296 }
1297
1298 /*
1299 * Freeing the page with debug_pagealloc enabled will try to
1300 * unmap it; some archs don't like double-unmappings, so
1301 * map it first.
1302 */
1303 debug_pagealloc_map_pages(page, nr_pages);
1304 adjust_managed_page_count(page, nr_pages);
1305 } else {
1306 for (loop = 0; loop < nr_pages; loop++, p++) {
1307 __ClearPageReserved(p);
1308 set_page_count(p, 0);
1309 }
1310
1311 /* memblock adjusts totalram_pages() manually. */
1312 atomic_long_add(nr_pages, &page_zone(page)->managed_pages);
1313 }
1314
1315 if (page_contains_unaccepted(page, order)) {
1316 if (order == MAX_PAGE_ORDER && __free_unaccepted(page))
1317 return;
1318
1319 accept_memory(page_to_phys(page), PAGE_SIZE << order);
1320 }
1321
1322 /*
1323 * Bypass PCP and place fresh pages right to the tail, primarily
1324 * relevant for memory onlining.
1325 */
1326 __free_pages_ok(page, order, FPI_TO_TAIL);
1327}
1328
1329/*
1330 * Check that the whole (or subset of) a pageblock given by the interval of
1331 * [start_pfn, end_pfn) is valid and within the same zone, before scanning it
1332 * with the migration of free compaction scanner.
1333 *
1334 * Return struct page pointer of start_pfn, or NULL if checks were not passed.
1335 *
1336 * It's possible on some configurations to have a setup like node0 node1 node0
1337 * i.e. it's possible that all pages within a zones range of pages do not
1338 * belong to a single zone. We assume that a border between node0 and node1
1339 * can occur within a single pageblock, but not a node0 node1 node0
1340 * interleaving within a single pageblock. It is therefore sufficient to check
1341 * the first and last page of a pageblock and avoid checking each individual
1342 * page in a pageblock.
1343 *
1344 * Note: the function may return non-NULL struct page even for a page block
1345 * which contains a memory hole (i.e. there is no physical memory for a subset
1346 * of the pfn range). For example, if the pageblock order is MAX_PAGE_ORDER, which
1347 * will fall into 2 sub-sections, and the end pfn of the pageblock may be hole
1348 * even though the start pfn is online and valid. This should be safe most of
1349 * the time because struct pages are still initialized via init_unavailable_range()
1350 * and pfn walkers shouldn't touch any physical memory range for which they do
1351 * not recognize any specific metadata in struct pages.
1352 */
1353struct page *__pageblock_pfn_to_page(unsigned long start_pfn,
1354 unsigned long end_pfn, struct zone *zone)
1355{
1356 struct page *start_page;
1357 struct page *end_page;
1358
1359 /* end_pfn is one past the range we are checking */
1360 end_pfn--;
1361
1362 if (!pfn_valid(end_pfn))
1363 return NULL;
1364
1365 start_page = pfn_to_online_page(start_pfn);
1366 if (!start_page)
1367 return NULL;
1368
1369 if (page_zone(start_page) != zone)
1370 return NULL;
1371
1372 end_page = pfn_to_page(end_pfn);
1373
1374 /* This gives a shorter code than deriving page_zone(end_page) */
1375 if (page_zone_id(start_page) != page_zone_id(end_page))
1376 return NULL;
1377
1378 return start_page;
1379}
1380
1381/*
1382 * The order of subdivision here is critical for the IO subsystem.
1383 * Please do not alter this order without good reasons and regression
1384 * testing. Specifically, as large blocks of memory are subdivided,
1385 * the order in which smaller blocks are delivered depends on the order
1386 * they're subdivided in this function. This is the primary factor
1387 * influencing the order in which pages are delivered to the IO
1388 * subsystem according to empirical testing, and this is also justified
1389 * by considering the behavior of a buddy system containing a single
1390 * large block of memory acted on by a series of small allocations.
1391 * This behavior is a critical factor in sglist merging's success.
1392 *
1393 * -- nyc
1394 */
1395static inline unsigned int expand(struct zone *zone, struct page *page, int low,
1396 int high, int migratetype)
1397{
1398 unsigned int size = 1 << high;
1399 unsigned int nr_added = 0;
1400
1401 while (high > low) {
1402 high--;
1403 size >>= 1;
1404 VM_BUG_ON_PAGE(bad_range(zone, &page[size]), &page[size]);
1405
1406 /*
1407 * Mark as guard pages (or page), that will allow to
1408 * merge back to allocator when buddy will be freed.
1409 * Corresponding page table entries will not be touched,
1410 * pages will stay not present in virtual address space
1411 */
1412 if (set_page_guard(zone, &page[size], high))
1413 continue;
1414
1415 __add_to_free_list(&page[size], zone, high, migratetype, false);
1416 set_buddy_order(&page[size], high);
1417 nr_added += size;
1418 }
1419
1420 return nr_added;
1421}
1422
1423static __always_inline void page_del_and_expand(struct zone *zone,
1424 struct page *page, int low,
1425 int high, int migratetype)
1426{
1427 int nr_pages = 1 << high;
1428
1429 __del_page_from_free_list(page, zone, high, migratetype);
1430 nr_pages -= expand(zone, page, low, high, migratetype);
1431 account_freepages(zone, -nr_pages, migratetype);
1432}
1433
1434static void check_new_page_bad(struct page *page)
1435{
1436 if (unlikely(page->flags & __PG_HWPOISON)) {
1437 /* Don't complain about hwpoisoned pages */
1438 if (PageBuddy(page))
1439 __ClearPageBuddy(page);
1440 return;
1441 }
1442
1443 bad_page(page,
1444 page_bad_reason(page, PAGE_FLAGS_CHECK_AT_PREP));
1445}
1446
1447/*
1448 * This page is about to be returned from the page allocator
1449 */
1450static bool check_new_page(struct page *page)
1451{
1452 if (likely(page_expected_state(page,
1453 PAGE_FLAGS_CHECK_AT_PREP|__PG_HWPOISON)))
1454 return false;
1455
1456 check_new_page_bad(page);
1457 return true;
1458}
1459
1460static inline bool check_new_pages(struct page *page, unsigned int order)
1461{
1462 if (is_check_pages_enabled()) {
1463 for (int i = 0; i < (1 << order); i++) {
1464 struct page *p = page + i;
1465
1466 if (check_new_page(p))
1467 return true;
1468 }
1469 }
1470
1471 return false;
1472}
1473
1474static inline bool should_skip_kasan_unpoison(gfp_t flags)
1475{
1476 /* Don't skip if a software KASAN mode is enabled. */
1477 if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
1478 IS_ENABLED(CONFIG_KASAN_SW_TAGS))
1479 return false;
1480
1481 /* Skip, if hardware tag-based KASAN is not enabled. */
1482 if (!kasan_hw_tags_enabled())
1483 return true;
1484
1485 /*
1486 * With hardware tag-based KASAN enabled, skip if this has been
1487 * requested via __GFP_SKIP_KASAN.
1488 */
1489 return flags & __GFP_SKIP_KASAN;
1490}
1491
1492static inline bool should_skip_init(gfp_t flags)
1493{
1494 /* Don't skip, if hardware tag-based KASAN is not enabled. */
1495 if (!kasan_hw_tags_enabled())
1496 return false;
1497
1498 /* For hardware tag-based KASAN, skip if requested. */
1499 return (flags & __GFP_SKIP_ZERO);
1500}
1501
1502inline void post_alloc_hook(struct page *page, unsigned int order,
1503 gfp_t gfp_flags)
1504{
1505 bool init = !want_init_on_free() && want_init_on_alloc(gfp_flags) &&
1506 !should_skip_init(gfp_flags);
1507 bool zero_tags = init && (gfp_flags & __GFP_ZEROTAGS);
1508 int i;
1509
1510 set_page_private(page, 0);
1511 set_page_refcounted(page);
1512
1513 arch_alloc_page(page, order);
1514 debug_pagealloc_map_pages(page, 1 << order);
1515
1516 /*
1517 * Page unpoisoning must happen before memory initialization.
1518 * Otherwise, the poison pattern will be overwritten for __GFP_ZERO
1519 * allocations and the page unpoisoning code will complain.
1520 */
1521 kernel_unpoison_pages(page, 1 << order);
1522
1523 /*
1524 * As memory initialization might be integrated into KASAN,
1525 * KASAN unpoisoning and memory initializion code must be
1526 * kept together to avoid discrepancies in behavior.
1527 */
1528
1529 /*
1530 * If memory tags should be zeroed
1531 * (which happens only when memory should be initialized as well).
1532 */
1533 if (zero_tags) {
1534 /* Initialize both memory and memory tags. */
1535 for (i = 0; i != 1 << order; ++i)
1536 tag_clear_highpage(page + i);
1537
1538 /* Take note that memory was initialized by the loop above. */
1539 init = false;
1540 }
1541 if (!should_skip_kasan_unpoison(gfp_flags) &&
1542 kasan_unpoison_pages(page, order, init)) {
1543 /* Take note that memory was initialized by KASAN. */
1544 if (kasan_has_integrated_init())
1545 init = false;
1546 } else {
1547 /*
1548 * If memory tags have not been set by KASAN, reset the page
1549 * tags to ensure page_address() dereferencing does not fault.
1550 */
1551 for (i = 0; i != 1 << order; ++i)
1552 page_kasan_tag_reset(page + i);
1553 }
1554 /* If memory is still not initialized, initialize it now. */
1555 if (init)
1556 kernel_init_pages(page, 1 << order);
1557
1558 set_page_owner(page, order, gfp_flags);
1559 page_table_check_alloc(page, order);
1560 pgalloc_tag_add(page, current, 1 << order);
1561}
1562
1563static void prep_new_page(struct page *page, unsigned int order, gfp_t gfp_flags,
1564 unsigned int alloc_flags)
1565{
1566 post_alloc_hook(page, order, gfp_flags);
1567
1568 if (order && (gfp_flags & __GFP_COMP))
1569 prep_compound_page(page, order);
1570
1571 /*
1572 * page is set pfmemalloc when ALLOC_NO_WATERMARKS was necessary to
1573 * allocate the page. The expectation is that the caller is taking
1574 * steps that will free more memory. The caller should avoid the page
1575 * being used for !PFMEMALLOC purposes.
1576 */
1577 if (alloc_flags & ALLOC_NO_WATERMARKS)
1578 set_page_pfmemalloc(page);
1579 else
1580 clear_page_pfmemalloc(page);
1581}
1582
1583/*
1584 * Go through the free lists for the given migratetype and remove
1585 * the smallest available page from the freelists
1586 */
1587static __always_inline
1588struct page *__rmqueue_smallest(struct zone *zone, unsigned int order,
1589 int migratetype)
1590{
1591 unsigned int current_order;
1592 struct free_area *area;
1593 struct page *page;
1594
1595 /* Find a page of the appropriate size in the preferred list */
1596 for (current_order = order; current_order < NR_PAGE_ORDERS; ++current_order) {
1597 area = &(zone->free_area[current_order]);
1598 page = get_page_from_free_area(area, migratetype);
1599 if (!page)
1600 continue;
1601
1602 page_del_and_expand(zone, page, order, current_order,
1603 migratetype);
1604 trace_mm_page_alloc_zone_locked(page, order, migratetype,
1605 pcp_allowed_order(order) &&
1606 migratetype < MIGRATE_PCPTYPES);
1607 return page;
1608 }
1609
1610 return NULL;
1611}
1612
1613
1614/*
1615 * This array describes the order lists are fallen back to when
1616 * the free lists for the desirable migrate type are depleted
1617 *
1618 * The other migratetypes do not have fallbacks.
1619 */
1620static int fallbacks[MIGRATE_PCPTYPES][MIGRATE_PCPTYPES - 1] = {
1621 [MIGRATE_UNMOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_MOVABLE },
1622 [MIGRATE_MOVABLE] = { MIGRATE_RECLAIMABLE, MIGRATE_UNMOVABLE },
1623 [MIGRATE_RECLAIMABLE] = { MIGRATE_UNMOVABLE, MIGRATE_MOVABLE },
1624};
1625
1626#ifdef CONFIG_CMA
1627static __always_inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1628 unsigned int order)
1629{
1630 return __rmqueue_smallest(zone, order, MIGRATE_CMA);
1631}
1632#else
1633static inline struct page *__rmqueue_cma_fallback(struct zone *zone,
1634 unsigned int order) { return NULL; }
1635#endif
1636
1637/*
1638 * Change the type of a block and move all its free pages to that
1639 * type's freelist.
1640 */
1641static int __move_freepages_block(struct zone *zone, unsigned long start_pfn,
1642 int old_mt, int new_mt)
1643{
1644 struct page *page;
1645 unsigned long pfn, end_pfn;
1646 unsigned int order;
1647 int pages_moved = 0;
1648
1649 VM_WARN_ON(start_pfn & (pageblock_nr_pages - 1));
1650 end_pfn = pageblock_end_pfn(start_pfn);
1651
1652 for (pfn = start_pfn; pfn < end_pfn;) {
1653 page = pfn_to_page(pfn);
1654 if (!PageBuddy(page)) {
1655 pfn++;
1656 continue;
1657 }
1658
1659 /* Make sure we are not inadvertently changing nodes */
1660 VM_BUG_ON_PAGE(page_to_nid(page) != zone_to_nid(zone), page);
1661 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1662
1663 order = buddy_order(page);
1664
1665 move_to_free_list(page, zone, order, old_mt, new_mt);
1666
1667 pfn += 1 << order;
1668 pages_moved += 1 << order;
1669 }
1670
1671 set_pageblock_migratetype(pfn_to_page(start_pfn), new_mt);
1672
1673 return pages_moved;
1674}
1675
1676static bool prep_move_freepages_block(struct zone *zone, struct page *page,
1677 unsigned long *start_pfn,
1678 int *num_free, int *num_movable)
1679{
1680 unsigned long pfn, start, end;
1681
1682 pfn = page_to_pfn(page);
1683 start = pageblock_start_pfn(pfn);
1684 end = pageblock_end_pfn(pfn);
1685
1686 /*
1687 * The caller only has the lock for @zone, don't touch ranges
1688 * that straddle into other zones. While we could move part of
1689 * the range that's inside the zone, this call is usually
1690 * accompanied by other operations such as migratetype updates
1691 * which also should be locked.
1692 */
1693 if (!zone_spans_pfn(zone, start))
1694 return false;
1695 if (!zone_spans_pfn(zone, end - 1))
1696 return false;
1697
1698 *start_pfn = start;
1699
1700 if (num_free) {
1701 *num_free = 0;
1702 *num_movable = 0;
1703 for (pfn = start; pfn < end;) {
1704 page = pfn_to_page(pfn);
1705 if (PageBuddy(page)) {
1706 int nr = 1 << buddy_order(page);
1707
1708 *num_free += nr;
1709 pfn += nr;
1710 continue;
1711 }
1712 /*
1713 * We assume that pages that could be isolated for
1714 * migration are movable. But we don't actually try
1715 * isolating, as that would be expensive.
1716 */
1717 if (PageLRU(page) || __PageMovable(page))
1718 (*num_movable)++;
1719 pfn++;
1720 }
1721 }
1722
1723 return true;
1724}
1725
1726static int move_freepages_block(struct zone *zone, struct page *page,
1727 int old_mt, int new_mt)
1728{
1729 unsigned long start_pfn;
1730
1731 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
1732 return -1;
1733
1734 return __move_freepages_block(zone, start_pfn, old_mt, new_mt);
1735}
1736
1737#ifdef CONFIG_MEMORY_ISOLATION
1738/* Look for a buddy that straddles start_pfn */
1739static unsigned long find_large_buddy(unsigned long start_pfn)
1740{
1741 int order = 0;
1742 struct page *page;
1743 unsigned long pfn = start_pfn;
1744
1745 while (!PageBuddy(page = pfn_to_page(pfn))) {
1746 /* Nothing found */
1747 if (++order > MAX_PAGE_ORDER)
1748 return start_pfn;
1749 pfn &= ~0UL << order;
1750 }
1751
1752 /*
1753 * Found a preceding buddy, but does it straddle?
1754 */
1755 if (pfn + (1 << buddy_order(page)) > start_pfn)
1756 return pfn;
1757
1758 /* Nothing found */
1759 return start_pfn;
1760}
1761
1762/**
1763 * move_freepages_block_isolate - move free pages in block for page isolation
1764 * @zone: the zone
1765 * @page: the pageblock page
1766 * @migratetype: migratetype to set on the pageblock
1767 *
1768 * This is similar to move_freepages_block(), but handles the special
1769 * case encountered in page isolation, where the block of interest
1770 * might be part of a larger buddy spanning multiple pageblocks.
1771 *
1772 * Unlike the regular page allocator path, which moves pages while
1773 * stealing buddies off the freelist, page isolation is interested in
1774 * arbitrary pfn ranges that may have overlapping buddies on both ends.
1775 *
1776 * This function handles that. Straddling buddies are split into
1777 * individual pageblocks. Only the block of interest is moved.
1778 *
1779 * Returns %true if pages could be moved, %false otherwise.
1780 */
1781bool move_freepages_block_isolate(struct zone *zone, struct page *page,
1782 int migratetype)
1783{
1784 unsigned long start_pfn, pfn;
1785
1786 if (!prep_move_freepages_block(zone, page, &start_pfn, NULL, NULL))
1787 return false;
1788
1789 /* No splits needed if buddies can't span multiple blocks */
1790 if (pageblock_order == MAX_PAGE_ORDER)
1791 goto move;
1792
1793 /* We're a tail block in a larger buddy */
1794 pfn = find_large_buddy(start_pfn);
1795 if (pfn != start_pfn) {
1796 struct page *buddy = pfn_to_page(pfn);
1797 int order = buddy_order(buddy);
1798
1799 del_page_from_free_list(buddy, zone, order,
1800 get_pfnblock_migratetype(buddy, pfn));
1801 set_pageblock_migratetype(page, migratetype);
1802 split_large_buddy(zone, buddy, pfn, order, FPI_NONE);
1803 return true;
1804 }
1805
1806 /* We're the starting block of a larger buddy */
1807 if (PageBuddy(page) && buddy_order(page) > pageblock_order) {
1808 int order = buddy_order(page);
1809
1810 del_page_from_free_list(page, zone, order,
1811 get_pfnblock_migratetype(page, pfn));
1812 set_pageblock_migratetype(page, migratetype);
1813 split_large_buddy(zone, page, pfn, order, FPI_NONE);
1814 return true;
1815 }
1816move:
1817 __move_freepages_block(zone, start_pfn,
1818 get_pfnblock_migratetype(page, start_pfn),
1819 migratetype);
1820 return true;
1821}
1822#endif /* CONFIG_MEMORY_ISOLATION */
1823
1824static void change_pageblock_range(struct page *pageblock_page,
1825 int start_order, int migratetype)
1826{
1827 int nr_pageblocks = 1 << (start_order - pageblock_order);
1828
1829 while (nr_pageblocks--) {
1830 set_pageblock_migratetype(pageblock_page, migratetype);
1831 pageblock_page += pageblock_nr_pages;
1832 }
1833}
1834
1835/*
1836 * When we are falling back to another migratetype during allocation, try to
1837 * steal extra free pages from the same pageblocks to satisfy further
1838 * allocations, instead of polluting multiple pageblocks.
1839 *
1840 * If we are stealing a relatively large buddy page, it is likely there will
1841 * be more free pages in the pageblock, so try to steal them all. For
1842 * reclaimable and unmovable allocations, we steal regardless of page size,
1843 * as fragmentation caused by those allocations polluting movable pageblocks
1844 * is worse than movable allocations stealing from unmovable and reclaimable
1845 * pageblocks.
1846 */
1847static bool can_steal_fallback(unsigned int order, int start_mt)
1848{
1849 /*
1850 * Leaving this order check is intended, although there is
1851 * relaxed order check in next check. The reason is that
1852 * we can actually steal whole pageblock if this condition met,
1853 * but, below check doesn't guarantee it and that is just heuristic
1854 * so could be changed anytime.
1855 */
1856 if (order >= pageblock_order)
1857 return true;
1858
1859 if (order >= pageblock_order / 2 ||
1860 start_mt == MIGRATE_RECLAIMABLE ||
1861 start_mt == MIGRATE_UNMOVABLE ||
1862 page_group_by_mobility_disabled)
1863 return true;
1864
1865 return false;
1866}
1867
1868static inline bool boost_watermark(struct zone *zone)
1869{
1870 unsigned long max_boost;
1871
1872 if (!watermark_boost_factor)
1873 return false;
1874 /*
1875 * Don't bother in zones that are unlikely to produce results.
1876 * On small machines, including kdump capture kernels running
1877 * in a small area, boosting the watermark can cause an out of
1878 * memory situation immediately.
1879 */
1880 if ((pageblock_nr_pages * 4) > zone_managed_pages(zone))
1881 return false;
1882
1883 max_boost = mult_frac(zone->_watermark[WMARK_HIGH],
1884 watermark_boost_factor, 10000);
1885
1886 /*
1887 * high watermark may be uninitialised if fragmentation occurs
1888 * very early in boot so do not boost. We do not fall
1889 * through and boost by pageblock_nr_pages as failing
1890 * allocations that early means that reclaim is not going
1891 * to help and it may even be impossible to reclaim the
1892 * boosted watermark resulting in a hang.
1893 */
1894 if (!max_boost)
1895 return false;
1896
1897 max_boost = max(pageblock_nr_pages, max_boost);
1898
1899 zone->watermark_boost = min(zone->watermark_boost + pageblock_nr_pages,
1900 max_boost);
1901
1902 return true;
1903}
1904
1905/*
1906 * This function implements actual steal behaviour. If order is large enough, we
1907 * can claim the whole pageblock for the requested migratetype. If not, we check
1908 * the pageblock for constituent pages; if at least half of the pages are free
1909 * or compatible, we can still claim the whole block, so pages freed in the
1910 * future will be put on the correct free list. Otherwise, we isolate exactly
1911 * the order we need from the fallback block and leave its migratetype alone.
1912 */
1913static struct page *
1914steal_suitable_fallback(struct zone *zone, struct page *page,
1915 int current_order, int order, int start_type,
1916 unsigned int alloc_flags, bool whole_block)
1917{
1918 int free_pages, movable_pages, alike_pages;
1919 unsigned long start_pfn;
1920 int block_type;
1921
1922 block_type = get_pageblock_migratetype(page);
1923
1924 /*
1925 * This can happen due to races and we want to prevent broken
1926 * highatomic accounting.
1927 */
1928 if (is_migrate_highatomic(block_type))
1929 goto single_page;
1930
1931 /* Take ownership for orders >= pageblock_order */
1932 if (current_order >= pageblock_order) {
1933 unsigned int nr_added;
1934
1935 del_page_from_free_list(page, zone, current_order, block_type);
1936 change_pageblock_range(page, current_order, start_type);
1937 nr_added = expand(zone, page, order, current_order, start_type);
1938 account_freepages(zone, nr_added, start_type);
1939 return page;
1940 }
1941
1942 /*
1943 * Boost watermarks to increase reclaim pressure to reduce the
1944 * likelihood of future fallbacks. Wake kswapd now as the node
1945 * may be balanced overall and kswapd will not wake naturally.
1946 */
1947 if (boost_watermark(zone) && (alloc_flags & ALLOC_KSWAPD))
1948 set_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
1949
1950 /* We are not allowed to try stealing from the whole block */
1951 if (!whole_block)
1952 goto single_page;
1953
1954 /* moving whole block can fail due to zone boundary conditions */
1955 if (!prep_move_freepages_block(zone, page, &start_pfn, &free_pages,
1956 &movable_pages))
1957 goto single_page;
1958
1959 /*
1960 * Determine how many pages are compatible with our allocation.
1961 * For movable allocation, it's the number of movable pages which
1962 * we just obtained. For other types it's a bit more tricky.
1963 */
1964 if (start_type == MIGRATE_MOVABLE) {
1965 alike_pages = movable_pages;
1966 } else {
1967 /*
1968 * If we are falling back a RECLAIMABLE or UNMOVABLE allocation
1969 * to MOVABLE pageblock, consider all non-movable pages as
1970 * compatible. If it's UNMOVABLE falling back to RECLAIMABLE or
1971 * vice versa, be conservative since we can't distinguish the
1972 * exact migratetype of non-movable pages.
1973 */
1974 if (block_type == MIGRATE_MOVABLE)
1975 alike_pages = pageblock_nr_pages
1976 - (free_pages + movable_pages);
1977 else
1978 alike_pages = 0;
1979 }
1980 /*
1981 * If a sufficient number of pages in the block are either free or of
1982 * compatible migratability as our allocation, claim the whole block.
1983 */
1984 if (free_pages + alike_pages >= (1 << (pageblock_order-1)) ||
1985 page_group_by_mobility_disabled) {
1986 __move_freepages_block(zone, start_pfn, block_type, start_type);
1987 return __rmqueue_smallest(zone, order, start_type);
1988 }
1989
1990single_page:
1991 page_del_and_expand(zone, page, order, current_order, block_type);
1992 return page;
1993}
1994
1995/*
1996 * Check whether there is a suitable fallback freepage with requested order.
1997 * If only_stealable is true, this function returns fallback_mt only if
1998 * we can steal other freepages all together. This would help to reduce
1999 * fragmentation due to mixed migratetype pages in one pageblock.
2000 */
2001int find_suitable_fallback(struct free_area *area, unsigned int order,
2002 int migratetype, bool only_stealable, bool *can_steal)
2003{
2004 int i;
2005 int fallback_mt;
2006
2007 if (area->nr_free == 0)
2008 return -1;
2009
2010 *can_steal = false;
2011 for (i = 0; i < MIGRATE_PCPTYPES - 1 ; i++) {
2012 fallback_mt = fallbacks[migratetype][i];
2013 if (free_area_empty(area, fallback_mt))
2014 continue;
2015
2016 if (can_steal_fallback(order, migratetype))
2017 *can_steal = true;
2018
2019 if (!only_stealable)
2020 return fallback_mt;
2021
2022 if (*can_steal)
2023 return fallback_mt;
2024 }
2025
2026 return -1;
2027}
2028
2029/*
2030 * Reserve the pageblock(s) surrounding an allocation request for
2031 * exclusive use of high-order atomic allocations if there are no
2032 * empty page blocks that contain a page with a suitable order
2033 */
2034static void reserve_highatomic_pageblock(struct page *page, int order,
2035 struct zone *zone)
2036{
2037 int mt;
2038 unsigned long max_managed, flags;
2039
2040 /*
2041 * The number reserved as: minimum is 1 pageblock, maximum is
2042 * roughly 1% of a zone. But if 1% of a zone falls below a
2043 * pageblock size, then don't reserve any pageblocks.
2044 * Check is race-prone but harmless.
2045 */
2046 if ((zone_managed_pages(zone) / 100) < pageblock_nr_pages)
2047 return;
2048 max_managed = ALIGN((zone_managed_pages(zone) / 100), pageblock_nr_pages);
2049 if (zone->nr_reserved_highatomic >= max_managed)
2050 return;
2051
2052 spin_lock_irqsave(&zone->lock, flags);
2053
2054 /* Recheck the nr_reserved_highatomic limit under the lock */
2055 if (zone->nr_reserved_highatomic >= max_managed)
2056 goto out_unlock;
2057
2058 /* Yoink! */
2059 mt = get_pageblock_migratetype(page);
2060 /* Only reserve normal pageblocks (i.e., they can merge with others) */
2061 if (!migratetype_is_mergeable(mt))
2062 goto out_unlock;
2063
2064 if (order < pageblock_order) {
2065 if (move_freepages_block(zone, page, mt, MIGRATE_HIGHATOMIC) == -1)
2066 goto out_unlock;
2067 zone->nr_reserved_highatomic += pageblock_nr_pages;
2068 } else {
2069 change_pageblock_range(page, order, MIGRATE_HIGHATOMIC);
2070 zone->nr_reserved_highatomic += 1 << order;
2071 }
2072
2073out_unlock:
2074 spin_unlock_irqrestore(&zone->lock, flags);
2075}
2076
2077/*
2078 * Used when an allocation is about to fail under memory pressure. This
2079 * potentially hurts the reliability of high-order allocations when under
2080 * intense memory pressure but failed atomic allocations should be easier
2081 * to recover from than an OOM.
2082 *
2083 * If @force is true, try to unreserve pageblocks even though highatomic
2084 * pageblock is exhausted.
2085 */
2086static bool unreserve_highatomic_pageblock(const struct alloc_context *ac,
2087 bool force)
2088{
2089 struct zonelist *zonelist = ac->zonelist;
2090 unsigned long flags;
2091 struct zoneref *z;
2092 struct zone *zone;
2093 struct page *page;
2094 int order;
2095 int ret;
2096
2097 for_each_zone_zonelist_nodemask(zone, z, zonelist, ac->highest_zoneidx,
2098 ac->nodemask) {
2099 /*
2100 * Preserve at least one pageblock unless memory pressure
2101 * is really high.
2102 */
2103 if (!force && zone->nr_reserved_highatomic <=
2104 pageblock_nr_pages)
2105 continue;
2106
2107 spin_lock_irqsave(&zone->lock, flags);
2108 for (order = 0; order < NR_PAGE_ORDERS; order++) {
2109 struct free_area *area = &(zone->free_area[order]);
2110 int mt;
2111
2112 page = get_page_from_free_area(area, MIGRATE_HIGHATOMIC);
2113 if (!page)
2114 continue;
2115
2116 mt = get_pageblock_migratetype(page);
2117 /*
2118 * In page freeing path, migratetype change is racy so
2119 * we can counter several free pages in a pageblock
2120 * in this loop although we changed the pageblock type
2121 * from highatomic to ac->migratetype. So we should
2122 * adjust the count once.
2123 */
2124 if (is_migrate_highatomic(mt)) {
2125 unsigned long size;
2126 /*
2127 * It should never happen but changes to
2128 * locking could inadvertently allow a per-cpu
2129 * drain to add pages to MIGRATE_HIGHATOMIC
2130 * while unreserving so be safe and watch for
2131 * underflows.
2132 */
2133 size = max(pageblock_nr_pages, 1UL << order);
2134 size = min(size, zone->nr_reserved_highatomic);
2135 zone->nr_reserved_highatomic -= size;
2136 }
2137
2138 /*
2139 * Convert to ac->migratetype and avoid the normal
2140 * pageblock stealing heuristics. Minimally, the caller
2141 * is doing the work and needs the pages. More
2142 * importantly, if the block was always converted to
2143 * MIGRATE_UNMOVABLE or another type then the number
2144 * of pageblocks that cannot be completely freed
2145 * may increase.
2146 */
2147 if (order < pageblock_order)
2148 ret = move_freepages_block(zone, page, mt,
2149 ac->migratetype);
2150 else {
2151 move_to_free_list(page, zone, order, mt,
2152 ac->migratetype);
2153 change_pageblock_range(page, order,
2154 ac->migratetype);
2155 ret = 1;
2156 }
2157 /*
2158 * Reserving the block(s) already succeeded,
2159 * so this should not fail on zone boundaries.
2160 */
2161 WARN_ON_ONCE(ret == -1);
2162 if (ret > 0) {
2163 spin_unlock_irqrestore(&zone->lock, flags);
2164 return ret;
2165 }
2166 }
2167 spin_unlock_irqrestore(&zone->lock, flags);
2168 }
2169
2170 return false;
2171}
2172
2173/*
2174 * Try finding a free buddy page on the fallback list and put it on the free
2175 * list of requested migratetype, possibly along with other pages from the same
2176 * block, depending on fragmentation avoidance heuristics. Returns true if
2177 * fallback was found so that __rmqueue_smallest() can grab it.
2178 *
2179 * The use of signed ints for order and current_order is a deliberate
2180 * deviation from the rest of this file, to make the for loop
2181 * condition simpler.
2182 */
2183static __always_inline struct page *
2184__rmqueue_fallback(struct zone *zone, int order, int start_migratetype,
2185 unsigned int alloc_flags)
2186{
2187 struct free_area *area;
2188 int current_order;
2189 int min_order = order;
2190 struct page *page;
2191 int fallback_mt;
2192 bool can_steal;
2193
2194 /*
2195 * Do not steal pages from freelists belonging to other pageblocks
2196 * i.e. orders < pageblock_order. If there are no local zones free,
2197 * the zonelists will be reiterated without ALLOC_NOFRAGMENT.
2198 */
2199 if (order < pageblock_order && alloc_flags & ALLOC_NOFRAGMENT)
2200 min_order = pageblock_order;
2201
2202 /*
2203 * Find the largest available free page in the other list. This roughly
2204 * approximates finding the pageblock with the most free pages, which
2205 * would be too costly to do exactly.
2206 */
2207 for (current_order = MAX_PAGE_ORDER; current_order >= min_order;
2208 --current_order) {
2209 area = &(zone->free_area[current_order]);
2210 fallback_mt = find_suitable_fallback(area, current_order,
2211 start_migratetype, false, &can_steal);
2212 if (fallback_mt == -1)
2213 continue;
2214
2215 /*
2216 * We cannot steal all free pages from the pageblock and the
2217 * requested migratetype is movable. In that case it's better to
2218 * steal and split the smallest available page instead of the
2219 * largest available page, because even if the next movable
2220 * allocation falls back into a different pageblock than this
2221 * one, it won't cause permanent fragmentation.
2222 */
2223 if (!can_steal && start_migratetype == MIGRATE_MOVABLE
2224 && current_order > order)
2225 goto find_smallest;
2226
2227 goto do_steal;
2228 }
2229
2230 return NULL;
2231
2232find_smallest:
2233 for (current_order = order; current_order < NR_PAGE_ORDERS; current_order++) {
2234 area = &(zone->free_area[current_order]);
2235 fallback_mt = find_suitable_fallback(area, current_order,
2236 start_migratetype, false, &can_steal);
2237 if (fallback_mt != -1)
2238 break;
2239 }
2240
2241 /*
2242 * This should not happen - we already found a suitable fallback
2243 * when looking for the largest page.
2244 */
2245 VM_BUG_ON(current_order > MAX_PAGE_ORDER);
2246
2247do_steal:
2248 page = get_page_from_free_area(area, fallback_mt);
2249
2250 /* take off list, maybe claim block, expand remainder */
2251 page = steal_suitable_fallback(zone, page, current_order, order,
2252 start_migratetype, alloc_flags, can_steal);
2253
2254 trace_mm_page_alloc_extfrag(page, order, current_order,
2255 start_migratetype, fallback_mt);
2256
2257 return page;
2258}
2259
2260/*
2261 * Do the hard work of removing an element from the buddy allocator.
2262 * Call me with the zone->lock already held.
2263 */
2264static __always_inline struct page *
2265__rmqueue(struct zone *zone, unsigned int order, int migratetype,
2266 unsigned int alloc_flags)
2267{
2268 struct page *page;
2269
2270 if (IS_ENABLED(CONFIG_CMA)) {
2271 /*
2272 * Balance movable allocations between regular and CMA areas by
2273 * allocating from CMA when over half of the zone's free memory
2274 * is in the CMA area.
2275 */
2276 if (alloc_flags & ALLOC_CMA &&
2277 zone_page_state(zone, NR_FREE_CMA_PAGES) >
2278 zone_page_state(zone, NR_FREE_PAGES) / 2) {
2279 page = __rmqueue_cma_fallback(zone, order);
2280 if (page)
2281 return page;
2282 }
2283 }
2284
2285 page = __rmqueue_smallest(zone, order, migratetype);
2286 if (unlikely(!page)) {
2287 if (alloc_flags & ALLOC_CMA)
2288 page = __rmqueue_cma_fallback(zone, order);
2289
2290 if (!page)
2291 page = __rmqueue_fallback(zone, order, migratetype,
2292 alloc_flags);
2293 }
2294 return page;
2295}
2296
2297/*
2298 * Obtain a specified number of elements from the buddy allocator, all under
2299 * a single hold of the lock, for efficiency. Add them to the supplied list.
2300 * Returns the number of new pages which were placed at *list.
2301 */
2302static int rmqueue_bulk(struct zone *zone, unsigned int order,
2303 unsigned long count, struct list_head *list,
2304 int migratetype, unsigned int alloc_flags)
2305{
2306 unsigned long flags;
2307 int i;
2308
2309 spin_lock_irqsave(&zone->lock, flags);
2310 for (i = 0; i < count; ++i) {
2311 struct page *page = __rmqueue(zone, order, migratetype,
2312 alloc_flags);
2313 if (unlikely(page == NULL))
2314 break;
2315
2316 /*
2317 * Split buddy pages returned by expand() are received here in
2318 * physical page order. The page is added to the tail of
2319 * caller's list. From the callers perspective, the linked list
2320 * is ordered by page number under some conditions. This is
2321 * useful for IO devices that can forward direction from the
2322 * head, thus also in the physical page order. This is useful
2323 * for IO devices that can merge IO requests if the physical
2324 * pages are ordered properly.
2325 */
2326 list_add_tail(&page->pcp_list, list);
2327 }
2328 spin_unlock_irqrestore(&zone->lock, flags);
2329
2330 return i;
2331}
2332
2333/*
2334 * Called from the vmstat counter updater to decay the PCP high.
2335 * Return whether there are addition works to do.
2336 */
2337int decay_pcp_high(struct zone *zone, struct per_cpu_pages *pcp)
2338{
2339 int high_min, to_drain, batch;
2340 int todo = 0;
2341
2342 high_min = READ_ONCE(pcp->high_min);
2343 batch = READ_ONCE(pcp->batch);
2344 /*
2345 * Decrease pcp->high periodically to try to free possible
2346 * idle PCP pages. And, avoid to free too many pages to
2347 * control latency. This caps pcp->high decrement too.
2348 */
2349 if (pcp->high > high_min) {
2350 pcp->high = max3(pcp->count - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2351 pcp->high - (pcp->high >> 3), high_min);
2352 if (pcp->high > high_min)
2353 todo++;
2354 }
2355
2356 to_drain = pcp->count - pcp->high;
2357 if (to_drain > 0) {
2358 spin_lock(&pcp->lock);
2359 free_pcppages_bulk(zone, to_drain, pcp, 0);
2360 spin_unlock(&pcp->lock);
2361 todo++;
2362 }
2363
2364 return todo;
2365}
2366
2367#ifdef CONFIG_NUMA
2368/*
2369 * Called from the vmstat counter updater to drain pagesets of this
2370 * currently executing processor on remote nodes after they have
2371 * expired.
2372 */
2373void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
2374{
2375 int to_drain, batch;
2376
2377 batch = READ_ONCE(pcp->batch);
2378 to_drain = min(pcp->count, batch);
2379 if (to_drain > 0) {
2380 spin_lock(&pcp->lock);
2381 free_pcppages_bulk(zone, to_drain, pcp, 0);
2382 spin_unlock(&pcp->lock);
2383 }
2384}
2385#endif
2386
2387/*
2388 * Drain pcplists of the indicated processor and zone.
2389 */
2390static void drain_pages_zone(unsigned int cpu, struct zone *zone)
2391{
2392 struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2393 int count;
2394
2395 do {
2396 spin_lock(&pcp->lock);
2397 count = pcp->count;
2398 if (count) {
2399 int to_drain = min(count,
2400 pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
2401
2402 free_pcppages_bulk(zone, to_drain, pcp, 0);
2403 count -= to_drain;
2404 }
2405 spin_unlock(&pcp->lock);
2406 } while (count);
2407}
2408
2409/*
2410 * Drain pcplists of all zones on the indicated processor.
2411 */
2412static void drain_pages(unsigned int cpu)
2413{
2414 struct zone *zone;
2415
2416 for_each_populated_zone(zone) {
2417 drain_pages_zone(cpu, zone);
2418 }
2419}
2420
2421/*
2422 * Spill all of this CPU's per-cpu pages back into the buddy allocator.
2423 */
2424void drain_local_pages(struct zone *zone)
2425{
2426 int cpu = smp_processor_id();
2427
2428 if (zone)
2429 drain_pages_zone(cpu, zone);
2430 else
2431 drain_pages(cpu);
2432}
2433
2434/*
2435 * The implementation of drain_all_pages(), exposing an extra parameter to
2436 * drain on all cpus.
2437 *
2438 * drain_all_pages() is optimized to only execute on cpus where pcplists are
2439 * not empty. The check for non-emptiness can however race with a free to
2440 * pcplist that has not yet increased the pcp->count from 0 to 1. Callers
2441 * that need the guarantee that every CPU has drained can disable the
2442 * optimizing racy check.
2443 */
2444static void __drain_all_pages(struct zone *zone, bool force_all_cpus)
2445{
2446 int cpu;
2447
2448 /*
2449 * Allocate in the BSS so we won't require allocation in
2450 * direct reclaim path for CONFIG_CPUMASK_OFFSTACK=y
2451 */
2452 static cpumask_t cpus_with_pcps;
2453
2454 /*
2455 * Do not drain if one is already in progress unless it's specific to
2456 * a zone. Such callers are primarily CMA and memory hotplug and need
2457 * the drain to be complete when the call returns.
2458 */
2459 if (unlikely(!mutex_trylock(&pcpu_drain_mutex))) {
2460 if (!zone)
2461 return;
2462 mutex_lock(&pcpu_drain_mutex);
2463 }
2464
2465 /*
2466 * We don't care about racing with CPU hotplug event
2467 * as offline notification will cause the notified
2468 * cpu to drain that CPU pcps and on_each_cpu_mask
2469 * disables preemption as part of its processing
2470 */
2471 for_each_online_cpu(cpu) {
2472 struct per_cpu_pages *pcp;
2473 struct zone *z;
2474 bool has_pcps = false;
2475
2476 if (force_all_cpus) {
2477 /*
2478 * The pcp.count check is racy, some callers need a
2479 * guarantee that no cpu is missed.
2480 */
2481 has_pcps = true;
2482 } else if (zone) {
2483 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
2484 if (pcp->count)
2485 has_pcps = true;
2486 } else {
2487 for_each_populated_zone(z) {
2488 pcp = per_cpu_ptr(z->per_cpu_pageset, cpu);
2489 if (pcp->count) {
2490 has_pcps = true;
2491 break;
2492 }
2493 }
2494 }
2495
2496 if (has_pcps)
2497 cpumask_set_cpu(cpu, &cpus_with_pcps);
2498 else
2499 cpumask_clear_cpu(cpu, &cpus_with_pcps);
2500 }
2501
2502 for_each_cpu(cpu, &cpus_with_pcps) {
2503 if (zone)
2504 drain_pages_zone(cpu, zone);
2505 else
2506 drain_pages(cpu);
2507 }
2508
2509 mutex_unlock(&pcpu_drain_mutex);
2510}
2511
2512/*
2513 * Spill all the per-cpu pages from all CPUs back into the buddy allocator.
2514 *
2515 * When zone parameter is non-NULL, spill just the single zone's pages.
2516 */
2517void drain_all_pages(struct zone *zone)
2518{
2519 __drain_all_pages(zone, false);
2520}
2521
2522static int nr_pcp_free(struct per_cpu_pages *pcp, int batch, int high, bool free_high)
2523{
2524 int min_nr_free, max_nr_free;
2525
2526 /* Free as much as possible if batch freeing high-order pages. */
2527 if (unlikely(free_high))
2528 return min(pcp->count, batch << CONFIG_PCP_BATCH_SCALE_MAX);
2529
2530 /* Check for PCP disabled or boot pageset */
2531 if (unlikely(high < batch))
2532 return 1;
2533
2534 /* Leave at least pcp->batch pages on the list */
2535 min_nr_free = batch;
2536 max_nr_free = high - batch;
2537
2538 /*
2539 * Increase the batch number to the number of the consecutive
2540 * freed pages to reduce zone lock contention.
2541 */
2542 batch = clamp_t(int, pcp->free_count, min_nr_free, max_nr_free);
2543
2544 return batch;
2545}
2546
2547static int nr_pcp_high(struct per_cpu_pages *pcp, struct zone *zone,
2548 int batch, bool free_high)
2549{
2550 int high, high_min, high_max;
2551
2552 high_min = READ_ONCE(pcp->high_min);
2553 high_max = READ_ONCE(pcp->high_max);
2554 high = pcp->high = clamp(pcp->high, high_min, high_max);
2555
2556 if (unlikely(!high))
2557 return 0;
2558
2559 if (unlikely(free_high)) {
2560 pcp->high = max(high - (batch << CONFIG_PCP_BATCH_SCALE_MAX),
2561 high_min);
2562 return 0;
2563 }
2564
2565 /*
2566 * If reclaim is active, limit the number of pages that can be
2567 * stored on pcp lists
2568 */
2569 if (test_bit(ZONE_RECLAIM_ACTIVE, &zone->flags)) {
2570 int free_count = max_t(int, pcp->free_count, batch);
2571
2572 pcp->high = max(high - free_count, high_min);
2573 return min(batch << 2, pcp->high);
2574 }
2575
2576 if (high_min == high_max)
2577 return high;
2578
2579 if (test_bit(ZONE_BELOW_HIGH, &zone->flags)) {
2580 int free_count = max_t(int, pcp->free_count, batch);
2581
2582 pcp->high = max(high - free_count, high_min);
2583 high = max(pcp->count, high_min);
2584 } else if (pcp->count >= high) {
2585 int need_high = pcp->free_count + batch;
2586
2587 /* pcp->high should be large enough to hold batch freed pages */
2588 if (pcp->high < need_high)
2589 pcp->high = clamp(need_high, high_min, high_max);
2590 }
2591
2592 return high;
2593}
2594
2595static void free_unref_page_commit(struct zone *zone, struct per_cpu_pages *pcp,
2596 struct page *page, int migratetype,
2597 unsigned int order)
2598{
2599 int high, batch;
2600 int pindex;
2601 bool free_high = false;
2602
2603 /*
2604 * On freeing, reduce the number of pages that are batch allocated.
2605 * See nr_pcp_alloc() where alloc_factor is increased for subsequent
2606 * allocations.
2607 */
2608 pcp->alloc_factor >>= 1;
2609 __count_vm_events(PGFREE, 1 << order);
2610 pindex = order_to_pindex(migratetype, order);
2611 list_add(&page->pcp_list, &pcp->lists[pindex]);
2612 pcp->count += 1 << order;
2613
2614 batch = READ_ONCE(pcp->batch);
2615 /*
2616 * As high-order pages other than THP's stored on PCP can contribute
2617 * to fragmentation, limit the number stored when PCP is heavily
2618 * freeing without allocation. The remainder after bulk freeing
2619 * stops will be drained from vmstat refresh context.
2620 */
2621 if (order && order <= PAGE_ALLOC_COSTLY_ORDER) {
2622 free_high = (pcp->free_count >= batch &&
2623 (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) &&
2624 (!(pcp->flags & PCPF_FREE_HIGH_BATCH) ||
2625 pcp->count >= READ_ONCE(batch)));
2626 pcp->flags |= PCPF_PREV_FREE_HIGH_ORDER;
2627 } else if (pcp->flags & PCPF_PREV_FREE_HIGH_ORDER) {
2628 pcp->flags &= ~PCPF_PREV_FREE_HIGH_ORDER;
2629 }
2630 if (pcp->free_count < (batch << CONFIG_PCP_BATCH_SCALE_MAX))
2631 pcp->free_count += (1 << order);
2632 high = nr_pcp_high(pcp, zone, batch, free_high);
2633 if (pcp->count >= high) {
2634 free_pcppages_bulk(zone, nr_pcp_free(pcp, batch, high, free_high),
2635 pcp, pindex);
2636 if (test_bit(ZONE_BELOW_HIGH, &zone->flags) &&
2637 zone_watermark_ok(zone, 0, high_wmark_pages(zone),
2638 ZONE_MOVABLE, 0))
2639 clear_bit(ZONE_BELOW_HIGH, &zone->flags);
2640 }
2641}
2642
2643/*
2644 * Free a pcp page
2645 */
2646void free_unref_page(struct page *page, unsigned int order)
2647{
2648 unsigned long __maybe_unused UP_flags;
2649 struct per_cpu_pages *pcp;
2650 struct zone *zone;
2651 unsigned long pfn = page_to_pfn(page);
2652 int migratetype;
2653
2654 if (!pcp_allowed_order(order)) {
2655 __free_pages_ok(page, order, FPI_NONE);
2656 return;
2657 }
2658
2659 if (!free_pages_prepare(page, order))
2660 return;
2661
2662 /*
2663 * We only track unmovable, reclaimable and movable on pcp lists.
2664 * Place ISOLATE pages on the isolated list because they are being
2665 * offlined but treat HIGHATOMIC and CMA as movable pages so we can
2666 * get those areas back if necessary. Otherwise, we may have to free
2667 * excessively into the page allocator
2668 */
2669 migratetype = get_pfnblock_migratetype(page, pfn);
2670 if (unlikely(migratetype >= MIGRATE_PCPTYPES)) {
2671 if (unlikely(is_migrate_isolate(migratetype))) {
2672 free_one_page(page_zone(page), page, pfn, order, FPI_NONE);
2673 return;
2674 }
2675 migratetype = MIGRATE_MOVABLE;
2676 }
2677
2678 zone = page_zone(page);
2679 pcp_trylock_prepare(UP_flags);
2680 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2681 if (pcp) {
2682 free_unref_page_commit(zone, pcp, page, migratetype, order);
2683 pcp_spin_unlock(pcp);
2684 } else {
2685 free_one_page(zone, page, pfn, order, FPI_NONE);
2686 }
2687 pcp_trylock_finish(UP_flags);
2688}
2689
2690/*
2691 * Free a batch of folios
2692 */
2693void free_unref_folios(struct folio_batch *folios)
2694{
2695 unsigned long __maybe_unused UP_flags;
2696 struct per_cpu_pages *pcp = NULL;
2697 struct zone *locked_zone = NULL;
2698 int i, j;
2699
2700 /* Prepare folios for freeing */
2701 for (i = 0, j = 0; i < folios->nr; i++) {
2702 struct folio *folio = folios->folios[i];
2703 unsigned long pfn = folio_pfn(folio);
2704 unsigned int order = folio_order(folio);
2705
2706 if (!free_pages_prepare(&folio->page, order))
2707 continue;
2708 /*
2709 * Free orders not handled on the PCP directly to the
2710 * allocator.
2711 */
2712 if (!pcp_allowed_order(order)) {
2713 free_one_page(folio_zone(folio), &folio->page,
2714 pfn, order, FPI_NONE);
2715 continue;
2716 }
2717 folio->private = (void *)(unsigned long)order;
2718 if (j != i)
2719 folios->folios[j] = folio;
2720 j++;
2721 }
2722 folios->nr = j;
2723
2724 for (i = 0; i < folios->nr; i++) {
2725 struct folio *folio = folios->folios[i];
2726 struct zone *zone = folio_zone(folio);
2727 unsigned long pfn = folio_pfn(folio);
2728 unsigned int order = (unsigned long)folio->private;
2729 int migratetype;
2730
2731 folio->private = NULL;
2732 migratetype = get_pfnblock_migratetype(&folio->page, pfn);
2733
2734 /* Different zone requires a different pcp lock */
2735 if (zone != locked_zone ||
2736 is_migrate_isolate(migratetype)) {
2737 if (pcp) {
2738 pcp_spin_unlock(pcp);
2739 pcp_trylock_finish(UP_flags);
2740 locked_zone = NULL;
2741 pcp = NULL;
2742 }
2743
2744 /*
2745 * Free isolated pages directly to the
2746 * allocator, see comment in free_unref_page.
2747 */
2748 if (is_migrate_isolate(migratetype)) {
2749 free_one_page(zone, &folio->page, pfn,
2750 order, FPI_NONE);
2751 continue;
2752 }
2753
2754 /*
2755 * trylock is necessary as folios may be getting freed
2756 * from IRQ or SoftIRQ context after an IO completion.
2757 */
2758 pcp_trylock_prepare(UP_flags);
2759 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
2760 if (unlikely(!pcp)) {
2761 pcp_trylock_finish(UP_flags);
2762 free_one_page(zone, &folio->page, pfn,
2763 order, FPI_NONE);
2764 continue;
2765 }
2766 locked_zone = zone;
2767 }
2768
2769 /*
2770 * Non-isolated types over MIGRATE_PCPTYPES get added
2771 * to the MIGRATE_MOVABLE pcp list.
2772 */
2773 if (unlikely(migratetype >= MIGRATE_PCPTYPES))
2774 migratetype = MIGRATE_MOVABLE;
2775
2776 trace_mm_page_free_batched(&folio->page);
2777 free_unref_page_commit(zone, pcp, &folio->page, migratetype,
2778 order);
2779 }
2780
2781 if (pcp) {
2782 pcp_spin_unlock(pcp);
2783 pcp_trylock_finish(UP_flags);
2784 }
2785 folio_batch_reinit(folios);
2786}
2787
2788/*
2789 * split_page takes a non-compound higher-order page, and splits it into
2790 * n (1<<order) sub-pages: page[0..n]
2791 * Each sub-page must be freed individually.
2792 *
2793 * Note: this is probably too low level an operation for use in drivers.
2794 * Please consult with lkml before using this in your driver.
2795 */
2796void split_page(struct page *page, unsigned int order)
2797{
2798 int i;
2799
2800 VM_BUG_ON_PAGE(PageCompound(page), page);
2801 VM_BUG_ON_PAGE(!page_count(page), page);
2802
2803 for (i = 1; i < (1 << order); i++)
2804 set_page_refcounted(page + i);
2805 split_page_owner(page, order, 0);
2806 pgalloc_tag_split(page_folio(page), order, 0);
2807 split_page_memcg(page, order, 0);
2808}
2809EXPORT_SYMBOL_GPL(split_page);
2810
2811int __isolate_free_page(struct page *page, unsigned int order)
2812{
2813 struct zone *zone = page_zone(page);
2814 int mt = get_pageblock_migratetype(page);
2815
2816 if (!is_migrate_isolate(mt)) {
2817 unsigned long watermark;
2818 /*
2819 * Obey watermarks as if the page was being allocated. We can
2820 * emulate a high-order watermark check with a raised order-0
2821 * watermark, because we already know our high-order page
2822 * exists.
2823 */
2824 watermark = zone->_watermark[WMARK_MIN] + (1UL << order);
2825 if (!zone_watermark_ok(zone, 0, watermark, 0, ALLOC_CMA))
2826 return 0;
2827 }
2828
2829 del_page_from_free_list(page, zone, order, mt);
2830
2831 /*
2832 * Set the pageblock if the isolated page is at least half of a
2833 * pageblock
2834 */
2835 if (order >= pageblock_order - 1) {
2836 struct page *endpage = page + (1 << order) - 1;
2837 for (; page < endpage; page += pageblock_nr_pages) {
2838 int mt = get_pageblock_migratetype(page);
2839 /*
2840 * Only change normal pageblocks (i.e., they can merge
2841 * with others)
2842 */
2843 if (migratetype_is_mergeable(mt))
2844 move_freepages_block(zone, page, mt,
2845 MIGRATE_MOVABLE);
2846 }
2847 }
2848
2849 return 1UL << order;
2850}
2851
2852/**
2853 * __putback_isolated_page - Return a now-isolated page back where we got it
2854 * @page: Page that was isolated
2855 * @order: Order of the isolated page
2856 * @mt: The page's pageblock's migratetype
2857 *
2858 * This function is meant to return a page pulled from the free lists via
2859 * __isolate_free_page back to the free lists they were pulled from.
2860 */
2861void __putback_isolated_page(struct page *page, unsigned int order, int mt)
2862{
2863 struct zone *zone = page_zone(page);
2864
2865 /* zone lock should be held when this function is called */
2866 lockdep_assert_held(&zone->lock);
2867
2868 /* Return isolated page to tail of freelist. */
2869 __free_one_page(page, page_to_pfn(page), zone, order, mt,
2870 FPI_SKIP_REPORT_NOTIFY | FPI_TO_TAIL);
2871}
2872
2873/*
2874 * Update NUMA hit/miss statistics
2875 */
2876static inline void zone_statistics(struct zone *preferred_zone, struct zone *z,
2877 long nr_account)
2878{
2879#ifdef CONFIG_NUMA
2880 enum numa_stat_item local_stat = NUMA_LOCAL;
2881
2882 /* skip numa counters update if numa stats is disabled */
2883 if (!static_branch_likely(&vm_numa_stat_key))
2884 return;
2885
2886 if (zone_to_nid(z) != numa_node_id())
2887 local_stat = NUMA_OTHER;
2888
2889 if (zone_to_nid(z) == zone_to_nid(preferred_zone))
2890 __count_numa_events(z, NUMA_HIT, nr_account);
2891 else {
2892 __count_numa_events(z, NUMA_MISS, nr_account);
2893 __count_numa_events(preferred_zone, NUMA_FOREIGN, nr_account);
2894 }
2895 __count_numa_events(z, local_stat, nr_account);
2896#endif
2897}
2898
2899static __always_inline
2900struct page *rmqueue_buddy(struct zone *preferred_zone, struct zone *zone,
2901 unsigned int order, unsigned int alloc_flags,
2902 int migratetype)
2903{
2904 struct page *page;
2905 unsigned long flags;
2906
2907 do {
2908 page = NULL;
2909 spin_lock_irqsave(&zone->lock, flags);
2910 if (alloc_flags & ALLOC_HIGHATOMIC)
2911 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2912 if (!page) {
2913 page = __rmqueue(zone, order, migratetype, alloc_flags);
2914
2915 /*
2916 * If the allocation fails, allow OOM handling and
2917 * order-0 (atomic) allocs access to HIGHATOMIC
2918 * reserves as failing now is worse than failing a
2919 * high-order atomic allocation in the future.
2920 */
2921 if (!page && (alloc_flags & (ALLOC_OOM|ALLOC_NON_BLOCK)))
2922 page = __rmqueue_smallest(zone, order, MIGRATE_HIGHATOMIC);
2923
2924 if (!page) {
2925 spin_unlock_irqrestore(&zone->lock, flags);
2926 return NULL;
2927 }
2928 }
2929 spin_unlock_irqrestore(&zone->lock, flags);
2930 } while (check_new_pages(page, order));
2931
2932 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
2933 zone_statistics(preferred_zone, zone, 1);
2934
2935 return page;
2936}
2937
2938static int nr_pcp_alloc(struct per_cpu_pages *pcp, struct zone *zone, int order)
2939{
2940 int high, base_batch, batch, max_nr_alloc;
2941 int high_max, high_min;
2942
2943 base_batch = READ_ONCE(pcp->batch);
2944 high_min = READ_ONCE(pcp->high_min);
2945 high_max = READ_ONCE(pcp->high_max);
2946 high = pcp->high = clamp(pcp->high, high_min, high_max);
2947
2948 /* Check for PCP disabled or boot pageset */
2949 if (unlikely(high < base_batch))
2950 return 1;
2951
2952 if (order)
2953 batch = base_batch;
2954 else
2955 batch = (base_batch << pcp->alloc_factor);
2956
2957 /*
2958 * If we had larger pcp->high, we could avoid to allocate from
2959 * zone.
2960 */
2961 if (high_min != high_max && !test_bit(ZONE_BELOW_HIGH, &zone->flags))
2962 high = pcp->high = min(high + batch, high_max);
2963
2964 if (!order) {
2965 max_nr_alloc = max(high - pcp->count - base_batch, base_batch);
2966 /*
2967 * Double the number of pages allocated each time there is
2968 * subsequent allocation of order-0 pages without any freeing.
2969 */
2970 if (batch <= max_nr_alloc &&
2971 pcp->alloc_factor < CONFIG_PCP_BATCH_SCALE_MAX)
2972 pcp->alloc_factor++;
2973 batch = min(batch, max_nr_alloc);
2974 }
2975
2976 /*
2977 * Scale batch relative to order if batch implies free pages
2978 * can be stored on the PCP. Batch can be 1 for small zones or
2979 * for boot pagesets which should never store free pages as
2980 * the pages may belong to arbitrary zones.
2981 */
2982 if (batch > 1)
2983 batch = max(batch >> order, 2);
2984
2985 return batch;
2986}
2987
2988/* Remove page from the per-cpu list, caller must protect the list */
2989static inline
2990struct page *__rmqueue_pcplist(struct zone *zone, unsigned int order,
2991 int migratetype,
2992 unsigned int alloc_flags,
2993 struct per_cpu_pages *pcp,
2994 struct list_head *list)
2995{
2996 struct page *page;
2997
2998 do {
2999 if (list_empty(list)) {
3000 int batch = nr_pcp_alloc(pcp, zone, order);
3001 int alloced;
3002
3003 alloced = rmqueue_bulk(zone, order,
3004 batch, list,
3005 migratetype, alloc_flags);
3006
3007 pcp->count += alloced << order;
3008 if (unlikely(list_empty(list)))
3009 return NULL;
3010 }
3011
3012 page = list_first_entry(list, struct page, pcp_list);
3013 list_del(&page->pcp_list);
3014 pcp->count -= 1 << order;
3015 } while (check_new_pages(page, order));
3016
3017 return page;
3018}
3019
3020/* Lock and remove page from the per-cpu list */
3021static struct page *rmqueue_pcplist(struct zone *preferred_zone,
3022 struct zone *zone, unsigned int order,
3023 int migratetype, unsigned int alloc_flags)
3024{
3025 struct per_cpu_pages *pcp;
3026 struct list_head *list;
3027 struct page *page;
3028 unsigned long __maybe_unused UP_flags;
3029
3030 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
3031 pcp_trylock_prepare(UP_flags);
3032 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
3033 if (!pcp) {
3034 pcp_trylock_finish(UP_flags);
3035 return NULL;
3036 }
3037
3038 /*
3039 * On allocation, reduce the number of pages that are batch freed.
3040 * See nr_pcp_free() where free_factor is increased for subsequent
3041 * frees.
3042 */
3043 pcp->free_count >>= 1;
3044 list = &pcp->lists[order_to_pindex(migratetype, order)];
3045 page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
3046 pcp_spin_unlock(pcp);
3047 pcp_trylock_finish(UP_flags);
3048 if (page) {
3049 __count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
3050 zone_statistics(preferred_zone, zone, 1);
3051 }
3052 return page;
3053}
3054
3055/*
3056 * Allocate a page from the given zone.
3057 * Use pcplists for THP or "cheap" high-order allocations.
3058 */
3059
3060/*
3061 * Do not instrument rmqueue() with KMSAN. This function may call
3062 * __msan_poison_alloca() through a call to set_pfnblock_flags_mask().
3063 * If __msan_poison_alloca() attempts to allocate pages for the stack depot, it
3064 * may call rmqueue() again, which will result in a deadlock.
3065 */
3066__no_sanitize_memory
3067static inline
3068struct page *rmqueue(struct zone *preferred_zone,
3069 struct zone *zone, unsigned int order,
3070 gfp_t gfp_flags, unsigned int alloc_flags,
3071 int migratetype)
3072{
3073 struct page *page;
3074
3075 if (likely(pcp_allowed_order(order))) {
3076 page = rmqueue_pcplist(preferred_zone, zone, order,
3077 migratetype, alloc_flags);
3078 if (likely(page))
3079 goto out;
3080 }
3081
3082 page = rmqueue_buddy(preferred_zone, zone, order, alloc_flags,
3083 migratetype);
3084
3085out:
3086 /* Separate test+clear to avoid unnecessary atomics */
3087 if ((alloc_flags & ALLOC_KSWAPD) &&
3088 unlikely(test_bit(ZONE_BOOSTED_WATERMARK, &zone->flags))) {
3089 clear_bit(ZONE_BOOSTED_WATERMARK, &zone->flags);
3090 wakeup_kswapd(zone, 0, 0, zone_idx(zone));
3091 }
3092
3093 VM_BUG_ON_PAGE(page && bad_range(zone, page), page);
3094 return page;
3095}
3096
3097static inline long __zone_watermark_unusable_free(struct zone *z,
3098 unsigned int order, unsigned int alloc_flags)
3099{
3100 long unusable_free = (1 << order) - 1;
3101
3102 /*
3103 * If the caller does not have rights to reserves below the min
3104 * watermark then subtract the free pages reserved for highatomic.
3105 */
3106 if (likely(!(alloc_flags & ALLOC_RESERVES)))
3107 unusable_free += READ_ONCE(z->nr_free_highatomic);
3108
3109#ifdef CONFIG_CMA
3110 /* If allocation can't use CMA areas don't use free CMA pages */
3111 if (!(alloc_flags & ALLOC_CMA))
3112 unusable_free += zone_page_state(z, NR_FREE_CMA_PAGES);
3113#endif
3114
3115 return unusable_free;
3116}
3117
3118/*
3119 * Return true if free base pages are above 'mark'. For high-order checks it
3120 * will return true of the order-0 watermark is reached and there is at least
3121 * one free page of a suitable size. Checking now avoids taking the zone lock
3122 * to check in the allocation paths if no pages are free.
3123 */
3124bool __zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3125 int highest_zoneidx, unsigned int alloc_flags,
3126 long free_pages)
3127{
3128 long min = mark;
3129 int o;
3130
3131 /* free_pages may go negative - that's OK */
3132 free_pages -= __zone_watermark_unusable_free(z, order, alloc_flags);
3133
3134 if (unlikely(alloc_flags & ALLOC_RESERVES)) {
3135 /*
3136 * __GFP_HIGH allows access to 50% of the min reserve as well
3137 * as OOM.
3138 */
3139 if (alloc_flags & ALLOC_MIN_RESERVE) {
3140 min -= min / 2;
3141
3142 /*
3143 * Non-blocking allocations (e.g. GFP_ATOMIC) can
3144 * access more reserves than just __GFP_HIGH. Other
3145 * non-blocking allocations requests such as GFP_NOWAIT
3146 * or (GFP_KERNEL & ~__GFP_DIRECT_RECLAIM) do not get
3147 * access to the min reserve.
3148 */
3149 if (alloc_flags & ALLOC_NON_BLOCK)
3150 min -= min / 4;
3151 }
3152
3153 /*
3154 * OOM victims can try even harder than the normal reserve
3155 * users on the grounds that it's definitely going to be in
3156 * the exit path shortly and free memory. Any allocation it
3157 * makes during the free path will be small and short-lived.
3158 */
3159 if (alloc_flags & ALLOC_OOM)
3160 min -= min / 2;
3161 }
3162
3163 /*
3164 * Check watermarks for an order-0 allocation request. If these
3165 * are not met, then a high-order request also cannot go ahead
3166 * even if a suitable page happened to be free.
3167 */
3168 if (free_pages <= min + z->lowmem_reserve[highest_zoneidx])
3169 return false;
3170
3171 /* If this is an order-0 request then the watermark is fine */
3172 if (!order)
3173 return true;
3174
3175 /* For a high-order request, check at least one suitable page is free */
3176 for (o = order; o < NR_PAGE_ORDERS; o++) {
3177 struct free_area *area = &z->free_area[o];
3178 int mt;
3179
3180 if (!area->nr_free)
3181 continue;
3182
3183 for (mt = 0; mt < MIGRATE_PCPTYPES; mt++) {
3184 if (!free_area_empty(area, mt))
3185 return true;
3186 }
3187
3188#ifdef CONFIG_CMA
3189 if ((alloc_flags & ALLOC_CMA) &&
3190 !free_area_empty(area, MIGRATE_CMA)) {
3191 return true;
3192 }
3193#endif
3194 if ((alloc_flags & (ALLOC_HIGHATOMIC|ALLOC_OOM)) &&
3195 !free_area_empty(area, MIGRATE_HIGHATOMIC)) {
3196 return true;
3197 }
3198 }
3199 return false;
3200}
3201
3202bool zone_watermark_ok(struct zone *z, unsigned int order, unsigned long mark,
3203 int highest_zoneidx, unsigned int alloc_flags)
3204{
3205 return __zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3206 zone_page_state(z, NR_FREE_PAGES));
3207}
3208
3209static inline bool zone_watermark_fast(struct zone *z, unsigned int order,
3210 unsigned long mark, int highest_zoneidx,
3211 unsigned int alloc_flags, gfp_t gfp_mask)
3212{
3213 long free_pages;
3214
3215 free_pages = zone_page_state(z, NR_FREE_PAGES);
3216
3217 /*
3218 * Fast check for order-0 only. If this fails then the reserves
3219 * need to be calculated.
3220 */
3221 if (!order) {
3222 long usable_free;
3223 long reserved;
3224
3225 usable_free = free_pages;
3226 reserved = __zone_watermark_unusable_free(z, 0, alloc_flags);
3227
3228 /* reserved may over estimate high-atomic reserves. */
3229 usable_free -= min(usable_free, reserved);
3230 if (usable_free > mark + z->lowmem_reserve[highest_zoneidx])
3231 return true;
3232 }
3233
3234 if (__zone_watermark_ok(z, order, mark, highest_zoneidx, alloc_flags,
3235 free_pages))
3236 return true;
3237
3238 /*
3239 * Ignore watermark boosting for __GFP_HIGH order-0 allocations
3240 * when checking the min watermark. The min watermark is the
3241 * point where boosting is ignored so that kswapd is woken up
3242 * when below the low watermark.
3243 */
3244 if (unlikely(!order && (alloc_flags & ALLOC_MIN_RESERVE) && z->watermark_boost
3245 && ((alloc_flags & ALLOC_WMARK_MASK) == WMARK_MIN))) {
3246 mark = z->_watermark[WMARK_MIN];
3247 return __zone_watermark_ok(z, order, mark, highest_zoneidx,
3248 alloc_flags, free_pages);
3249 }
3250
3251 return false;
3252}
3253
3254bool zone_watermark_ok_safe(struct zone *z, unsigned int order,
3255 unsigned long mark, int highest_zoneidx)
3256{
3257 long free_pages = zone_page_state(z, NR_FREE_PAGES);
3258
3259 if (z->percpu_drift_mark && free_pages < z->percpu_drift_mark)
3260 free_pages = zone_page_state_snapshot(z, NR_FREE_PAGES);
3261
3262 return __zone_watermark_ok(z, order, mark, highest_zoneidx, 0,
3263 free_pages);
3264}
3265
3266#ifdef CONFIG_NUMA
3267int __read_mostly node_reclaim_distance = RECLAIM_DISTANCE;
3268
3269static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3270{
3271 return node_distance(zone_to_nid(local_zone), zone_to_nid(zone)) <=
3272 node_reclaim_distance;
3273}
3274#else /* CONFIG_NUMA */
3275static bool zone_allows_reclaim(struct zone *local_zone, struct zone *zone)
3276{
3277 return true;
3278}
3279#endif /* CONFIG_NUMA */
3280
3281/*
3282 * The restriction on ZONE_DMA32 as being a suitable zone to use to avoid
3283 * fragmentation is subtle. If the preferred zone was HIGHMEM then
3284 * premature use of a lower zone may cause lowmem pressure problems that
3285 * are worse than fragmentation. If the next zone is ZONE_DMA then it is
3286 * probably too small. It only makes sense to spread allocations to avoid
3287 * fragmentation between the Normal and DMA32 zones.
3288 */
3289static inline unsigned int
3290alloc_flags_nofragment(struct zone *zone, gfp_t gfp_mask)
3291{
3292 unsigned int alloc_flags;
3293
3294 /*
3295 * __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
3296 * to save a branch.
3297 */
3298 alloc_flags = (__force int) (gfp_mask & __GFP_KSWAPD_RECLAIM);
3299
3300#ifdef CONFIG_ZONE_DMA32
3301 if (!zone)
3302 return alloc_flags;
3303
3304 if (zone_idx(zone) != ZONE_NORMAL)
3305 return alloc_flags;
3306
3307 /*
3308 * If ZONE_DMA32 exists, assume it is the one after ZONE_NORMAL and
3309 * the pointer is within zone->zone_pgdat->node_zones[]. Also assume
3310 * on UMA that if Normal is populated then so is DMA32.
3311 */
3312 BUILD_BUG_ON(ZONE_NORMAL - ZONE_DMA32 != 1);
3313 if (nr_online_nodes > 1 && !populated_zone(--zone))
3314 return alloc_flags;
3315
3316 alloc_flags |= ALLOC_NOFRAGMENT;
3317#endif /* CONFIG_ZONE_DMA32 */
3318 return alloc_flags;
3319}
3320
3321/* Must be called after current_gfp_context() which can change gfp_mask */
3322static inline unsigned int gfp_to_alloc_flags_cma(gfp_t gfp_mask,
3323 unsigned int alloc_flags)
3324{
3325#ifdef CONFIG_CMA
3326 if (gfp_migratetype(gfp_mask) == MIGRATE_MOVABLE)
3327 alloc_flags |= ALLOC_CMA;
3328#endif
3329 return alloc_flags;
3330}
3331
3332/*
3333 * get_page_from_freelist goes through the zonelist trying to allocate
3334 * a page.
3335 */
3336static struct page *
3337get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
3338 const struct alloc_context *ac)
3339{
3340 struct zoneref *z;
3341 struct zone *zone;
3342 struct pglist_data *last_pgdat = NULL;
3343 bool last_pgdat_dirty_ok = false;
3344 bool no_fallback;
3345
3346retry:
3347 /*
3348 * Scan zonelist, looking for a zone with enough free.
3349 * See also cpuset_node_allowed() comment in kernel/cgroup/cpuset.c.
3350 */
3351 no_fallback = alloc_flags & ALLOC_NOFRAGMENT;
3352 z = ac->preferred_zoneref;
3353 for_next_zone_zonelist_nodemask(zone, z, ac->highest_zoneidx,
3354 ac->nodemask) {
3355 struct page *page;
3356 unsigned long mark;
3357
3358 if (cpusets_enabled() &&
3359 (alloc_flags & ALLOC_CPUSET) &&
3360 !__cpuset_zone_allowed(zone, gfp_mask))
3361 continue;
3362 /*
3363 * When allocating a page cache page for writing, we
3364 * want to get it from a node that is within its dirty
3365 * limit, such that no single node holds more than its
3366 * proportional share of globally allowed dirty pages.
3367 * The dirty limits take into account the node's
3368 * lowmem reserves and high watermark so that kswapd
3369 * should be able to balance it without having to
3370 * write pages from its LRU list.
3371 *
3372 * XXX: For now, allow allocations to potentially
3373 * exceed the per-node dirty limit in the slowpath
3374 * (spread_dirty_pages unset) before going into reclaim,
3375 * which is important when on a NUMA setup the allowed
3376 * nodes are together not big enough to reach the
3377 * global limit. The proper fix for these situations
3378 * will require awareness of nodes in the
3379 * dirty-throttling and the flusher threads.
3380 */
3381 if (ac->spread_dirty_pages) {
3382 if (last_pgdat != zone->zone_pgdat) {
3383 last_pgdat = zone->zone_pgdat;
3384 last_pgdat_dirty_ok = node_dirty_ok(zone->zone_pgdat);
3385 }
3386
3387 if (!last_pgdat_dirty_ok)
3388 continue;
3389 }
3390
3391 if (no_fallback && nr_online_nodes > 1 &&
3392 zone != zonelist_zone(ac->preferred_zoneref)) {
3393 int local_nid;
3394
3395 /*
3396 * If moving to a remote node, retry but allow
3397 * fragmenting fallbacks. Locality is more important
3398 * than fragmentation avoidance.
3399 */
3400 local_nid = zonelist_node_idx(ac->preferred_zoneref);
3401 if (zone_to_nid(zone) != local_nid) {
3402 alloc_flags &= ~ALLOC_NOFRAGMENT;
3403 goto retry;
3404 }
3405 }
3406
3407 cond_accept_memory(zone, order);
3408
3409 /*
3410 * Detect whether the number of free pages is below high
3411 * watermark. If so, we will decrease pcp->high and free
3412 * PCP pages in free path to reduce the possibility of
3413 * premature page reclaiming. Detection is done here to
3414 * avoid to do that in hotter free path.
3415 */
3416 if (test_bit(ZONE_BELOW_HIGH, &zone->flags))
3417 goto check_alloc_wmark;
3418
3419 mark = high_wmark_pages(zone);
3420 if (zone_watermark_fast(zone, order, mark,
3421 ac->highest_zoneidx, alloc_flags,
3422 gfp_mask))
3423 goto try_this_zone;
3424 else
3425 set_bit(ZONE_BELOW_HIGH, &zone->flags);
3426
3427check_alloc_wmark:
3428 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK);
3429 if (!zone_watermark_fast(zone, order, mark,
3430 ac->highest_zoneidx, alloc_flags,
3431 gfp_mask)) {
3432 int ret;
3433
3434 if (cond_accept_memory(zone, order))
3435 goto try_this_zone;
3436
3437 /*
3438 * Watermark failed for this zone, but see if we can
3439 * grow this zone if it contains deferred pages.
3440 */
3441 if (deferred_pages_enabled()) {
3442 if (_deferred_grow_zone(zone, order))
3443 goto try_this_zone;
3444 }
3445 /* Checked here to keep the fast path fast */
3446 BUILD_BUG_ON(ALLOC_NO_WATERMARKS < NR_WMARK);
3447 if (alloc_flags & ALLOC_NO_WATERMARKS)
3448 goto try_this_zone;
3449
3450 if (!node_reclaim_enabled() ||
3451 !zone_allows_reclaim(zonelist_zone(ac->preferred_zoneref), zone))
3452 continue;
3453
3454 ret = node_reclaim(zone->zone_pgdat, gfp_mask, order);
3455 switch (ret) {
3456 case NODE_RECLAIM_NOSCAN:
3457 /* did not scan */
3458 continue;
3459 case NODE_RECLAIM_FULL:
3460 /* scanned but unreclaimable */
3461 continue;
3462 default:
3463 /* did we reclaim enough */
3464 if (zone_watermark_ok(zone, order, mark,
3465 ac->highest_zoneidx, alloc_flags))
3466 goto try_this_zone;
3467
3468 continue;
3469 }
3470 }
3471
3472try_this_zone:
3473 page = rmqueue(zonelist_zone(ac->preferred_zoneref), zone, order,
3474 gfp_mask, alloc_flags, ac->migratetype);
3475 if (page) {
3476 prep_new_page(page, order, gfp_mask, alloc_flags);
3477
3478 /*
3479 * If this is a high-order atomic allocation then check
3480 * if the pageblock should be reserved for the future
3481 */
3482 if (unlikely(alloc_flags & ALLOC_HIGHATOMIC))
3483 reserve_highatomic_pageblock(page, order, zone);
3484
3485 return page;
3486 } else {
3487 if (cond_accept_memory(zone, order))
3488 goto try_this_zone;
3489
3490 /* Try again if zone has deferred pages */
3491 if (deferred_pages_enabled()) {
3492 if (_deferred_grow_zone(zone, order))
3493 goto try_this_zone;
3494 }
3495 }
3496 }
3497
3498 /*
3499 * It's possible on a UMA machine to get through all zones that are
3500 * fragmented. If avoiding fragmentation, reset and try again.
3501 */
3502 if (no_fallback) {
3503 alloc_flags &= ~ALLOC_NOFRAGMENT;
3504 goto retry;
3505 }
3506
3507 return NULL;
3508}
3509
3510static void warn_alloc_show_mem(gfp_t gfp_mask, nodemask_t *nodemask)
3511{
3512 unsigned int filter = SHOW_MEM_FILTER_NODES;
3513
3514 /*
3515 * This documents exceptions given to allocations in certain
3516 * contexts that are allowed to allocate outside current's set
3517 * of allowed nodes.
3518 */
3519 if (!(gfp_mask & __GFP_NOMEMALLOC))
3520 if (tsk_is_oom_victim(current) ||
3521 (current->flags & (PF_MEMALLOC | PF_EXITING)))
3522 filter &= ~SHOW_MEM_FILTER_NODES;
3523 if (!in_task() || !(gfp_mask & __GFP_DIRECT_RECLAIM))
3524 filter &= ~SHOW_MEM_FILTER_NODES;
3525
3526 __show_mem(filter, nodemask, gfp_zone(gfp_mask));
3527}
3528
3529void warn_alloc(gfp_t gfp_mask, nodemask_t *nodemask, const char *fmt, ...)
3530{
3531 struct va_format vaf;
3532 va_list args;
3533 static DEFINE_RATELIMIT_STATE(nopage_rs, 10*HZ, 1);
3534
3535 if ((gfp_mask & __GFP_NOWARN) ||
3536 !__ratelimit(&nopage_rs) ||
3537 ((gfp_mask & __GFP_DMA) && !has_managed_dma()))
3538 return;
3539
3540 va_start(args, fmt);
3541 vaf.fmt = fmt;
3542 vaf.va = &args;
3543 pr_warn("%s: %pV, mode:%#x(%pGg), nodemask=%*pbl",
3544 current->comm, &vaf, gfp_mask, &gfp_mask,
3545 nodemask_pr_args(nodemask));
3546 va_end(args);
3547
3548 cpuset_print_current_mems_allowed();
3549 pr_cont("\n");
3550 dump_stack();
3551 warn_alloc_show_mem(gfp_mask, nodemask);
3552}
3553
3554static inline struct page *
3555__alloc_pages_cpuset_fallback(gfp_t gfp_mask, unsigned int order,
3556 unsigned int alloc_flags,
3557 const struct alloc_context *ac)
3558{
3559 struct page *page;
3560
3561 page = get_page_from_freelist(gfp_mask, order,
3562 alloc_flags|ALLOC_CPUSET, ac);
3563 /*
3564 * fallback to ignore cpuset restriction if our nodes
3565 * are depleted
3566 */
3567 if (!page)
3568 page = get_page_from_freelist(gfp_mask, order,
3569 alloc_flags, ac);
3570
3571 return page;
3572}
3573
3574static inline struct page *
3575__alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
3576 const struct alloc_context *ac, unsigned long *did_some_progress)
3577{
3578 struct oom_control oc = {
3579 .zonelist = ac->zonelist,
3580 .nodemask = ac->nodemask,
3581 .memcg = NULL,
3582 .gfp_mask = gfp_mask,
3583 .order = order,
3584 };
3585 struct page *page;
3586
3587 *did_some_progress = 0;
3588
3589 /*
3590 * Acquire the oom lock. If that fails, somebody else is
3591 * making progress for us.
3592 */
3593 if (!mutex_trylock(&oom_lock)) {
3594 *did_some_progress = 1;
3595 schedule_timeout_uninterruptible(1);
3596 return NULL;
3597 }
3598
3599 /*
3600 * Go through the zonelist yet one more time, keep very high watermark
3601 * here, this is only to catch a parallel oom killing, we must fail if
3602 * we're still under heavy pressure. But make sure that this reclaim
3603 * attempt shall not depend on __GFP_DIRECT_RECLAIM && !__GFP_NORETRY
3604 * allocation which will never fail due to oom_lock already held.
3605 */
3606 page = get_page_from_freelist((gfp_mask | __GFP_HARDWALL) &
3607 ~__GFP_DIRECT_RECLAIM, order,
3608 ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
3609 if (page)
3610 goto out;
3611
3612 /* Coredumps can quickly deplete all memory reserves */
3613 if (current->flags & PF_DUMPCORE)
3614 goto out;
3615 /* The OOM killer will not help higher order allocs */
3616 if (order > PAGE_ALLOC_COSTLY_ORDER)
3617 goto out;
3618 /*
3619 * We have already exhausted all our reclaim opportunities without any
3620 * success so it is time to admit defeat. We will skip the OOM killer
3621 * because it is very likely that the caller has a more reasonable
3622 * fallback than shooting a random task.
3623 *
3624 * The OOM killer may not free memory on a specific node.
3625 */
3626 if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
3627 goto out;
3628 /* The OOM killer does not needlessly kill tasks for lowmem */
3629 if (ac->highest_zoneidx < ZONE_NORMAL)
3630 goto out;
3631 if (pm_suspended_storage())
3632 goto out;
3633 /*
3634 * XXX: GFP_NOFS allocations should rather fail than rely on
3635 * other request to make a forward progress.
3636 * We are in an unfortunate situation where out_of_memory cannot
3637 * do much for this context but let's try it to at least get
3638 * access to memory reserved if the current task is killed (see
3639 * out_of_memory). Once filesystems are ready to handle allocation
3640 * failures more gracefully we should just bail out here.
3641 */
3642
3643 /* Exhausted what can be done so it's blame time */
3644 if (out_of_memory(&oc) ||
3645 WARN_ON_ONCE_GFP(gfp_mask & __GFP_NOFAIL, gfp_mask)) {
3646 *did_some_progress = 1;
3647
3648 /*
3649 * Help non-failing allocations by giving them access to memory
3650 * reserves
3651 */
3652 if (gfp_mask & __GFP_NOFAIL)
3653 page = __alloc_pages_cpuset_fallback(gfp_mask, order,
3654 ALLOC_NO_WATERMARKS, ac);
3655 }
3656out:
3657 mutex_unlock(&oom_lock);
3658 return page;
3659}
3660
3661/*
3662 * Maximum number of compaction retries with a progress before OOM
3663 * killer is consider as the only way to move forward.
3664 */
3665#define MAX_COMPACT_RETRIES 16
3666
3667#ifdef CONFIG_COMPACTION
3668/* Try memory compaction for high-order allocations before reclaim */
3669static struct page *
3670__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3671 unsigned int alloc_flags, const struct alloc_context *ac,
3672 enum compact_priority prio, enum compact_result *compact_result)
3673{
3674 struct page *page = NULL;
3675 unsigned long pflags;
3676 unsigned int noreclaim_flag;
3677
3678 if (!order)
3679 return NULL;
3680
3681 psi_memstall_enter(&pflags);
3682 delayacct_compact_start();
3683 noreclaim_flag = memalloc_noreclaim_save();
3684
3685 *compact_result = try_to_compact_pages(gfp_mask, order, alloc_flags, ac,
3686 prio, &page);
3687
3688 memalloc_noreclaim_restore(noreclaim_flag);
3689 psi_memstall_leave(&pflags);
3690 delayacct_compact_end();
3691
3692 if (*compact_result == COMPACT_SKIPPED)
3693 return NULL;
3694 /*
3695 * At least in one zone compaction wasn't deferred or skipped, so let's
3696 * count a compaction stall
3697 */
3698 count_vm_event(COMPACTSTALL);
3699
3700 /* Prep a captured page if available */
3701 if (page)
3702 prep_new_page(page, order, gfp_mask, alloc_flags);
3703
3704 /* Try get a page from the freelist if available */
3705 if (!page)
3706 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3707
3708 if (page) {
3709 struct zone *zone = page_zone(page);
3710
3711 zone->compact_blockskip_flush = false;
3712 compaction_defer_reset(zone, order, true);
3713 count_vm_event(COMPACTSUCCESS);
3714 return page;
3715 }
3716
3717 /*
3718 * It's bad if compaction run occurs and fails. The most likely reason
3719 * is that pages exist, but not enough to satisfy watermarks.
3720 */
3721 count_vm_event(COMPACTFAIL);
3722
3723 cond_resched();
3724
3725 return NULL;
3726}
3727
3728static inline bool
3729should_compact_retry(struct alloc_context *ac, int order, int alloc_flags,
3730 enum compact_result compact_result,
3731 enum compact_priority *compact_priority,
3732 int *compaction_retries)
3733{
3734 int max_retries = MAX_COMPACT_RETRIES;
3735 int min_priority;
3736 bool ret = false;
3737 int retries = *compaction_retries;
3738 enum compact_priority priority = *compact_priority;
3739
3740 if (!order)
3741 return false;
3742
3743 if (fatal_signal_pending(current))
3744 return false;
3745
3746 /*
3747 * Compaction was skipped due to a lack of free order-0
3748 * migration targets. Continue if reclaim can help.
3749 */
3750 if (compact_result == COMPACT_SKIPPED) {
3751 ret = compaction_zonelist_suitable(ac, order, alloc_flags);
3752 goto out;
3753 }
3754
3755 /*
3756 * Compaction managed to coalesce some page blocks, but the
3757 * allocation failed presumably due to a race. Retry some.
3758 */
3759 if (compact_result == COMPACT_SUCCESS) {
3760 /*
3761 * !costly requests are much more important than
3762 * __GFP_RETRY_MAYFAIL costly ones because they are de
3763 * facto nofail and invoke OOM killer to move on while
3764 * costly can fail and users are ready to cope with
3765 * that. 1/4 retries is rather arbitrary but we would
3766 * need much more detailed feedback from compaction to
3767 * make a better decision.
3768 */
3769 if (order > PAGE_ALLOC_COSTLY_ORDER)
3770 max_retries /= 4;
3771
3772 if (++(*compaction_retries) <= max_retries) {
3773 ret = true;
3774 goto out;
3775 }
3776 }
3777
3778 /*
3779 * Compaction failed. Retry with increasing priority.
3780 */
3781 min_priority = (order > PAGE_ALLOC_COSTLY_ORDER) ?
3782 MIN_COMPACT_COSTLY_PRIORITY : MIN_COMPACT_PRIORITY;
3783
3784 if (*compact_priority > min_priority) {
3785 (*compact_priority)--;
3786 *compaction_retries = 0;
3787 ret = true;
3788 }
3789out:
3790 trace_compact_retry(order, priority, compact_result, retries, max_retries, ret);
3791 return ret;
3792}
3793#else
3794static inline struct page *
3795__alloc_pages_direct_compact(gfp_t gfp_mask, unsigned int order,
3796 unsigned int alloc_flags, const struct alloc_context *ac,
3797 enum compact_priority prio, enum compact_result *compact_result)
3798{
3799 *compact_result = COMPACT_SKIPPED;
3800 return NULL;
3801}
3802
3803static inline bool
3804should_compact_retry(struct alloc_context *ac, unsigned int order, int alloc_flags,
3805 enum compact_result compact_result,
3806 enum compact_priority *compact_priority,
3807 int *compaction_retries)
3808{
3809 struct zone *zone;
3810 struct zoneref *z;
3811
3812 if (!order || order > PAGE_ALLOC_COSTLY_ORDER)
3813 return false;
3814
3815 /*
3816 * There are setups with compaction disabled which would prefer to loop
3817 * inside the allocator rather than hit the oom killer prematurely.
3818 * Let's give them a good hope and keep retrying while the order-0
3819 * watermarks are OK.
3820 */
3821 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
3822 ac->highest_zoneidx, ac->nodemask) {
3823 if (zone_watermark_ok(zone, 0, min_wmark_pages(zone),
3824 ac->highest_zoneidx, alloc_flags))
3825 return true;
3826 }
3827 return false;
3828}
3829#endif /* CONFIG_COMPACTION */
3830
3831#ifdef CONFIG_LOCKDEP
3832static struct lockdep_map __fs_reclaim_map =
3833 STATIC_LOCKDEP_MAP_INIT("fs_reclaim", &__fs_reclaim_map);
3834
3835static bool __need_reclaim(gfp_t gfp_mask)
3836{
3837 /* no reclaim without waiting on it */
3838 if (!(gfp_mask & __GFP_DIRECT_RECLAIM))
3839 return false;
3840
3841 /* this guy won't enter reclaim */
3842 if (current->flags & PF_MEMALLOC)
3843 return false;
3844
3845 if (gfp_mask & __GFP_NOLOCKDEP)
3846 return false;
3847
3848 return true;
3849}
3850
3851void __fs_reclaim_acquire(unsigned long ip)
3852{
3853 lock_acquire_exclusive(&__fs_reclaim_map, 0, 0, NULL, ip);
3854}
3855
3856void __fs_reclaim_release(unsigned long ip)
3857{
3858 lock_release(&__fs_reclaim_map, ip);
3859}
3860
3861void fs_reclaim_acquire(gfp_t gfp_mask)
3862{
3863 gfp_mask = current_gfp_context(gfp_mask);
3864
3865 if (__need_reclaim(gfp_mask)) {
3866 if (gfp_mask & __GFP_FS)
3867 __fs_reclaim_acquire(_RET_IP_);
3868
3869#ifdef CONFIG_MMU_NOTIFIER
3870 lock_map_acquire(&__mmu_notifier_invalidate_range_start_map);
3871 lock_map_release(&__mmu_notifier_invalidate_range_start_map);
3872#endif
3873
3874 }
3875}
3876EXPORT_SYMBOL_GPL(fs_reclaim_acquire);
3877
3878void fs_reclaim_release(gfp_t gfp_mask)
3879{
3880 gfp_mask = current_gfp_context(gfp_mask);
3881
3882 if (__need_reclaim(gfp_mask)) {
3883 if (gfp_mask & __GFP_FS)
3884 __fs_reclaim_release(_RET_IP_);
3885 }
3886}
3887EXPORT_SYMBOL_GPL(fs_reclaim_release);
3888#endif
3889
3890/*
3891 * Zonelists may change due to hotplug during allocation. Detect when zonelists
3892 * have been rebuilt so allocation retries. Reader side does not lock and
3893 * retries the allocation if zonelist changes. Writer side is protected by the
3894 * embedded spin_lock.
3895 */
3896static DEFINE_SEQLOCK(zonelist_update_seq);
3897
3898static unsigned int zonelist_iter_begin(void)
3899{
3900 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3901 return read_seqbegin(&zonelist_update_seq);
3902
3903 return 0;
3904}
3905
3906static unsigned int check_retry_zonelist(unsigned int seq)
3907{
3908 if (IS_ENABLED(CONFIG_MEMORY_HOTREMOVE))
3909 return read_seqretry(&zonelist_update_seq, seq);
3910
3911 return seq;
3912}
3913
3914/* Perform direct synchronous page reclaim */
3915static unsigned long
3916__perform_reclaim(gfp_t gfp_mask, unsigned int order,
3917 const struct alloc_context *ac)
3918{
3919 unsigned int noreclaim_flag;
3920 unsigned long progress;
3921
3922 cond_resched();
3923
3924 /* We now go into synchronous reclaim */
3925 cpuset_memory_pressure_bump();
3926 fs_reclaim_acquire(gfp_mask);
3927 noreclaim_flag = memalloc_noreclaim_save();
3928
3929 progress = try_to_free_pages(ac->zonelist, order, gfp_mask,
3930 ac->nodemask);
3931
3932 memalloc_noreclaim_restore(noreclaim_flag);
3933 fs_reclaim_release(gfp_mask);
3934
3935 cond_resched();
3936
3937 return progress;
3938}
3939
3940/* The really slow allocator path where we enter direct reclaim */
3941static inline struct page *
3942__alloc_pages_direct_reclaim(gfp_t gfp_mask, unsigned int order,
3943 unsigned int alloc_flags, const struct alloc_context *ac,
3944 unsigned long *did_some_progress)
3945{
3946 struct page *page = NULL;
3947 unsigned long pflags;
3948 bool drained = false;
3949
3950 psi_memstall_enter(&pflags);
3951 *did_some_progress = __perform_reclaim(gfp_mask, order, ac);
3952 if (unlikely(!(*did_some_progress)))
3953 goto out;
3954
3955retry:
3956 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
3957
3958 /*
3959 * If an allocation failed after direct reclaim, it could be because
3960 * pages are pinned on the per-cpu lists or in high alloc reserves.
3961 * Shrink them and try again
3962 */
3963 if (!page && !drained) {
3964 unreserve_highatomic_pageblock(ac, false);
3965 drain_all_pages(NULL);
3966 drained = true;
3967 goto retry;
3968 }
3969out:
3970 psi_memstall_leave(&pflags);
3971
3972 return page;
3973}
3974
3975static void wake_all_kswapds(unsigned int order, gfp_t gfp_mask,
3976 const struct alloc_context *ac)
3977{
3978 struct zoneref *z;
3979 struct zone *zone;
3980 pg_data_t *last_pgdat = NULL;
3981 enum zone_type highest_zoneidx = ac->highest_zoneidx;
3982
3983 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist, highest_zoneidx,
3984 ac->nodemask) {
3985 if (!managed_zone(zone))
3986 continue;
3987 if (last_pgdat != zone->zone_pgdat) {
3988 wakeup_kswapd(zone, gfp_mask, order, highest_zoneidx);
3989 last_pgdat = zone->zone_pgdat;
3990 }
3991 }
3992}
3993
3994static inline unsigned int
3995gfp_to_alloc_flags(gfp_t gfp_mask, unsigned int order)
3996{
3997 unsigned int alloc_flags = ALLOC_WMARK_MIN | ALLOC_CPUSET;
3998
3999 /*
4000 * __GFP_HIGH is assumed to be the same as ALLOC_MIN_RESERVE
4001 * and __GFP_KSWAPD_RECLAIM is assumed to be the same as ALLOC_KSWAPD
4002 * to save two branches.
4003 */
4004 BUILD_BUG_ON(__GFP_HIGH != (__force gfp_t) ALLOC_MIN_RESERVE);
4005 BUILD_BUG_ON(__GFP_KSWAPD_RECLAIM != (__force gfp_t) ALLOC_KSWAPD);
4006
4007 /*
4008 * The caller may dip into page reserves a bit more if the caller
4009 * cannot run direct reclaim, or if the caller has realtime scheduling
4010 * policy or is asking for __GFP_HIGH memory. GFP_ATOMIC requests will
4011 * set both ALLOC_NON_BLOCK and ALLOC_MIN_RESERVE(__GFP_HIGH).
4012 */
4013 alloc_flags |= (__force int)
4014 (gfp_mask & (__GFP_HIGH | __GFP_KSWAPD_RECLAIM));
4015
4016 if (!(gfp_mask & __GFP_DIRECT_RECLAIM)) {
4017 /*
4018 * Not worth trying to allocate harder for __GFP_NOMEMALLOC even
4019 * if it can't schedule.
4020 */
4021 if (!(gfp_mask & __GFP_NOMEMALLOC)) {
4022 alloc_flags |= ALLOC_NON_BLOCK;
4023
4024 if (order > 0)
4025 alloc_flags |= ALLOC_HIGHATOMIC;
4026 }
4027
4028 /*
4029 * Ignore cpuset mems for non-blocking __GFP_HIGH (probably
4030 * GFP_ATOMIC) rather than fail, see the comment for
4031 * cpuset_node_allowed().
4032 */
4033 if (alloc_flags & ALLOC_MIN_RESERVE)
4034 alloc_flags &= ~ALLOC_CPUSET;
4035 } else if (unlikely(rt_or_dl_task(current)) && in_task())
4036 alloc_flags |= ALLOC_MIN_RESERVE;
4037
4038 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, alloc_flags);
4039
4040 return alloc_flags;
4041}
4042
4043static bool oom_reserves_allowed(struct task_struct *tsk)
4044{
4045 if (!tsk_is_oom_victim(tsk))
4046 return false;
4047
4048 /*
4049 * !MMU doesn't have oom reaper so give access to memory reserves
4050 * only to the thread with TIF_MEMDIE set
4051 */
4052 if (!IS_ENABLED(CONFIG_MMU) && !test_thread_flag(TIF_MEMDIE))
4053 return false;
4054
4055 return true;
4056}
4057
4058/*
4059 * Distinguish requests which really need access to full memory
4060 * reserves from oom victims which can live with a portion of it
4061 */
4062static inline int __gfp_pfmemalloc_flags(gfp_t gfp_mask)
4063{
4064 if (unlikely(gfp_mask & __GFP_NOMEMALLOC))
4065 return 0;
4066 if (gfp_mask & __GFP_MEMALLOC)
4067 return ALLOC_NO_WATERMARKS;
4068 if (in_serving_softirq() && (current->flags & PF_MEMALLOC))
4069 return ALLOC_NO_WATERMARKS;
4070 if (!in_interrupt()) {
4071 if (current->flags & PF_MEMALLOC)
4072 return ALLOC_NO_WATERMARKS;
4073 else if (oom_reserves_allowed(current))
4074 return ALLOC_OOM;
4075 }
4076
4077 return 0;
4078}
4079
4080bool gfp_pfmemalloc_allowed(gfp_t gfp_mask)
4081{
4082 return !!__gfp_pfmemalloc_flags(gfp_mask);
4083}
4084
4085/*
4086 * Checks whether it makes sense to retry the reclaim to make a forward progress
4087 * for the given allocation request.
4088 *
4089 * We give up when we either have tried MAX_RECLAIM_RETRIES in a row
4090 * without success, or when we couldn't even meet the watermark if we
4091 * reclaimed all remaining pages on the LRU lists.
4092 *
4093 * Returns true if a retry is viable or false to enter the oom path.
4094 */
4095static inline bool
4096should_reclaim_retry(gfp_t gfp_mask, unsigned order,
4097 struct alloc_context *ac, int alloc_flags,
4098 bool did_some_progress, int *no_progress_loops)
4099{
4100 struct zone *zone;
4101 struct zoneref *z;
4102 bool ret = false;
4103
4104 /*
4105 * Costly allocations might have made a progress but this doesn't mean
4106 * their order will become available due to high fragmentation so
4107 * always increment the no progress counter for them
4108 */
4109 if (did_some_progress && order <= PAGE_ALLOC_COSTLY_ORDER)
4110 *no_progress_loops = 0;
4111 else
4112 (*no_progress_loops)++;
4113
4114 if (*no_progress_loops > MAX_RECLAIM_RETRIES)
4115 goto out;
4116
4117
4118 /*
4119 * Keep reclaiming pages while there is a chance this will lead
4120 * somewhere. If none of the target zones can satisfy our allocation
4121 * request even if all reclaimable pages are considered then we are
4122 * screwed and have to go OOM.
4123 */
4124 for_each_zone_zonelist_nodemask(zone, z, ac->zonelist,
4125 ac->highest_zoneidx, ac->nodemask) {
4126 unsigned long available;
4127 unsigned long reclaimable;
4128 unsigned long min_wmark = min_wmark_pages(zone);
4129 bool wmark;
4130
4131 if (cpusets_enabled() &&
4132 (alloc_flags & ALLOC_CPUSET) &&
4133 !__cpuset_zone_allowed(zone, gfp_mask))
4134 continue;
4135
4136 available = reclaimable = zone_reclaimable_pages(zone);
4137 available += zone_page_state_snapshot(zone, NR_FREE_PAGES);
4138
4139 /*
4140 * Would the allocation succeed if we reclaimed all
4141 * reclaimable pages?
4142 */
4143 wmark = __zone_watermark_ok(zone, order, min_wmark,
4144 ac->highest_zoneidx, alloc_flags, available);
4145 trace_reclaim_retry_zone(z, order, reclaimable,
4146 available, min_wmark, *no_progress_loops, wmark);
4147 if (wmark) {
4148 ret = true;
4149 break;
4150 }
4151 }
4152
4153 /*
4154 * Memory allocation/reclaim might be called from a WQ context and the
4155 * current implementation of the WQ concurrency control doesn't
4156 * recognize that a particular WQ is congested if the worker thread is
4157 * looping without ever sleeping. Therefore we have to do a short sleep
4158 * here rather than calling cond_resched().
4159 */
4160 if (current->flags & PF_WQ_WORKER)
4161 schedule_timeout_uninterruptible(1);
4162 else
4163 cond_resched();
4164out:
4165 /* Before OOM, exhaust highatomic_reserve */
4166 if (!ret)
4167 return unreserve_highatomic_pageblock(ac, true);
4168
4169 return ret;
4170}
4171
4172static inline bool
4173check_retry_cpuset(int cpuset_mems_cookie, struct alloc_context *ac)
4174{
4175 /*
4176 * It's possible that cpuset's mems_allowed and the nodemask from
4177 * mempolicy don't intersect. This should be normally dealt with by
4178 * policy_nodemask(), but it's possible to race with cpuset update in
4179 * such a way the check therein was true, and then it became false
4180 * before we got our cpuset_mems_cookie here.
4181 * This assumes that for all allocations, ac->nodemask can come only
4182 * from MPOL_BIND mempolicy (whose documented semantics is to be ignored
4183 * when it does not intersect with the cpuset restrictions) or the
4184 * caller can deal with a violated nodemask.
4185 */
4186 if (cpusets_enabled() && ac->nodemask &&
4187 !cpuset_nodemask_valid_mems_allowed(ac->nodemask)) {
4188 ac->nodemask = NULL;
4189 return true;
4190 }
4191
4192 /*
4193 * When updating a task's mems_allowed or mempolicy nodemask, it is
4194 * possible to race with parallel threads in such a way that our
4195 * allocation can fail while the mask is being updated. If we are about
4196 * to fail, check if the cpuset changed during allocation and if so,
4197 * retry.
4198 */
4199 if (read_mems_allowed_retry(cpuset_mems_cookie))
4200 return true;
4201
4202 return false;
4203}
4204
4205static inline struct page *
4206__alloc_pages_slowpath(gfp_t gfp_mask, unsigned int order,
4207 struct alloc_context *ac)
4208{
4209 bool can_direct_reclaim = gfp_mask & __GFP_DIRECT_RECLAIM;
4210 bool can_compact = gfp_compaction_allowed(gfp_mask);
4211 bool nofail = gfp_mask & __GFP_NOFAIL;
4212 const bool costly_order = order > PAGE_ALLOC_COSTLY_ORDER;
4213 struct page *page = NULL;
4214 unsigned int alloc_flags;
4215 unsigned long did_some_progress;
4216 enum compact_priority compact_priority;
4217 enum compact_result compact_result;
4218 int compaction_retries;
4219 int no_progress_loops;
4220 unsigned int cpuset_mems_cookie;
4221 unsigned int zonelist_iter_cookie;
4222 int reserve_flags;
4223
4224 if (unlikely(nofail)) {
4225 /*
4226 * We most definitely don't want callers attempting to
4227 * allocate greater than order-1 page units with __GFP_NOFAIL.
4228 */
4229 WARN_ON_ONCE(order > 1);
4230 /*
4231 * Also we don't support __GFP_NOFAIL without __GFP_DIRECT_RECLAIM,
4232 * otherwise, we may result in lockup.
4233 */
4234 WARN_ON_ONCE(!can_direct_reclaim);
4235 /*
4236 * PF_MEMALLOC request from this context is rather bizarre
4237 * because we cannot reclaim anything and only can loop waiting
4238 * for somebody to do a work for us.
4239 */
4240 WARN_ON_ONCE(current->flags & PF_MEMALLOC);
4241 }
4242
4243restart:
4244 compaction_retries = 0;
4245 no_progress_loops = 0;
4246 compact_result = COMPACT_SKIPPED;
4247 compact_priority = DEF_COMPACT_PRIORITY;
4248 cpuset_mems_cookie = read_mems_allowed_begin();
4249 zonelist_iter_cookie = zonelist_iter_begin();
4250
4251 /*
4252 * The fast path uses conservative alloc_flags to succeed only until
4253 * kswapd needs to be woken up, and to avoid the cost of setting up
4254 * alloc_flags precisely. So we do that now.
4255 */
4256 alloc_flags = gfp_to_alloc_flags(gfp_mask, order);
4257
4258 /*
4259 * We need to recalculate the starting point for the zonelist iterator
4260 * because we might have used different nodemask in the fast path, or
4261 * there was a cpuset modification and we are retrying - otherwise we
4262 * could end up iterating over non-eligible zones endlessly.
4263 */
4264 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4265 ac->highest_zoneidx, ac->nodemask);
4266 if (!zonelist_zone(ac->preferred_zoneref))
4267 goto nopage;
4268
4269 /*
4270 * Check for insane configurations where the cpuset doesn't contain
4271 * any suitable zone to satisfy the request - e.g. non-movable
4272 * GFP_HIGHUSER allocations from MOVABLE nodes only.
4273 */
4274 if (cpusets_insane_config() && (gfp_mask & __GFP_HARDWALL)) {
4275 struct zoneref *z = first_zones_zonelist(ac->zonelist,
4276 ac->highest_zoneidx,
4277 &cpuset_current_mems_allowed);
4278 if (!zonelist_zone(z))
4279 goto nopage;
4280 }
4281
4282 if (alloc_flags & ALLOC_KSWAPD)
4283 wake_all_kswapds(order, gfp_mask, ac);
4284
4285 /*
4286 * The adjusted alloc_flags might result in immediate success, so try
4287 * that first
4288 */
4289 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4290 if (page)
4291 goto got_pg;
4292
4293 /*
4294 * For costly allocations, try direct compaction first, as it's likely
4295 * that we have enough base pages and don't need to reclaim. For non-
4296 * movable high-order allocations, do that as well, as compaction will
4297 * try prevent permanent fragmentation by migrating from blocks of the
4298 * same migratetype.
4299 * Don't try this for allocations that are allowed to ignore
4300 * watermarks, as the ALLOC_NO_WATERMARKS attempt didn't yet happen.
4301 */
4302 if (can_direct_reclaim && can_compact &&
4303 (costly_order ||
4304 (order > 0 && ac->migratetype != MIGRATE_MOVABLE))
4305 && !gfp_pfmemalloc_allowed(gfp_mask)) {
4306 page = __alloc_pages_direct_compact(gfp_mask, order,
4307 alloc_flags, ac,
4308 INIT_COMPACT_PRIORITY,
4309 &compact_result);
4310 if (page)
4311 goto got_pg;
4312
4313 /*
4314 * Checks for costly allocations with __GFP_NORETRY, which
4315 * includes some THP page fault allocations
4316 */
4317 if (costly_order && (gfp_mask & __GFP_NORETRY)) {
4318 /*
4319 * If allocating entire pageblock(s) and compaction
4320 * failed because all zones are below low watermarks
4321 * or is prohibited because it recently failed at this
4322 * order, fail immediately unless the allocator has
4323 * requested compaction and reclaim retry.
4324 *
4325 * Reclaim is
4326 * - potentially very expensive because zones are far
4327 * below their low watermarks or this is part of very
4328 * bursty high order allocations,
4329 * - not guaranteed to help because isolate_freepages()
4330 * may not iterate over freed pages as part of its
4331 * linear scan, and
4332 * - unlikely to make entire pageblocks free on its
4333 * own.
4334 */
4335 if (compact_result == COMPACT_SKIPPED ||
4336 compact_result == COMPACT_DEFERRED)
4337 goto nopage;
4338
4339 /*
4340 * Looks like reclaim/compaction is worth trying, but
4341 * sync compaction could be very expensive, so keep
4342 * using async compaction.
4343 */
4344 compact_priority = INIT_COMPACT_PRIORITY;
4345 }
4346 }
4347
4348retry:
4349 /* Ensure kswapd doesn't accidentally go to sleep as long as we loop */
4350 if (alloc_flags & ALLOC_KSWAPD)
4351 wake_all_kswapds(order, gfp_mask, ac);
4352
4353 reserve_flags = __gfp_pfmemalloc_flags(gfp_mask);
4354 if (reserve_flags)
4355 alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, reserve_flags) |
4356 (alloc_flags & ALLOC_KSWAPD);
4357
4358 /*
4359 * Reset the nodemask and zonelist iterators if memory policies can be
4360 * ignored. These allocations are high priority and system rather than
4361 * user oriented.
4362 */
4363 if (!(alloc_flags & ALLOC_CPUSET) || reserve_flags) {
4364 ac->nodemask = NULL;
4365 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4366 ac->highest_zoneidx, ac->nodemask);
4367 }
4368
4369 /* Attempt with potentially adjusted zonelist and alloc_flags */
4370 page = get_page_from_freelist(gfp_mask, order, alloc_flags, ac);
4371 if (page)
4372 goto got_pg;
4373
4374 /* Caller is not willing to reclaim, we can't balance anything */
4375 if (!can_direct_reclaim)
4376 goto nopage;
4377
4378 /* Avoid recursion of direct reclaim */
4379 if (current->flags & PF_MEMALLOC)
4380 goto nopage;
4381
4382 /* Try direct reclaim and then allocating */
4383 page = __alloc_pages_direct_reclaim(gfp_mask, order, alloc_flags, ac,
4384 &did_some_progress);
4385 if (page)
4386 goto got_pg;
4387
4388 /* Try direct compaction and then allocating */
4389 page = __alloc_pages_direct_compact(gfp_mask, order, alloc_flags, ac,
4390 compact_priority, &compact_result);
4391 if (page)
4392 goto got_pg;
4393
4394 /* Do not loop if specifically requested */
4395 if (gfp_mask & __GFP_NORETRY)
4396 goto nopage;
4397
4398 /*
4399 * Do not retry costly high order allocations unless they are
4400 * __GFP_RETRY_MAYFAIL and we can compact
4401 */
4402 if (costly_order && (!can_compact ||
4403 !(gfp_mask & __GFP_RETRY_MAYFAIL)))
4404 goto nopage;
4405
4406 if (should_reclaim_retry(gfp_mask, order, ac, alloc_flags,
4407 did_some_progress > 0, &no_progress_loops))
4408 goto retry;
4409
4410 /*
4411 * It doesn't make any sense to retry for the compaction if the order-0
4412 * reclaim is not able to make any progress because the current
4413 * implementation of the compaction depends on the sufficient amount
4414 * of free memory (see __compaction_suitable)
4415 */
4416 if (did_some_progress > 0 && can_compact &&
4417 should_compact_retry(ac, order, alloc_flags,
4418 compact_result, &compact_priority,
4419 &compaction_retries))
4420 goto retry;
4421
4422
4423 /*
4424 * Deal with possible cpuset update races or zonelist updates to avoid
4425 * a unnecessary OOM kill.
4426 */
4427 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4428 check_retry_zonelist(zonelist_iter_cookie))
4429 goto restart;
4430
4431 /* Reclaim has failed us, start killing things */
4432 page = __alloc_pages_may_oom(gfp_mask, order, ac, &did_some_progress);
4433 if (page)
4434 goto got_pg;
4435
4436 /* Avoid allocations with no watermarks from looping endlessly */
4437 if (tsk_is_oom_victim(current) &&
4438 (alloc_flags & ALLOC_OOM ||
4439 (gfp_mask & __GFP_NOMEMALLOC)))
4440 goto nopage;
4441
4442 /* Retry as long as the OOM killer is making progress */
4443 if (did_some_progress) {
4444 no_progress_loops = 0;
4445 goto retry;
4446 }
4447
4448nopage:
4449 /*
4450 * Deal with possible cpuset update races or zonelist updates to avoid
4451 * a unnecessary OOM kill.
4452 */
4453 if (check_retry_cpuset(cpuset_mems_cookie, ac) ||
4454 check_retry_zonelist(zonelist_iter_cookie))
4455 goto restart;
4456
4457 /*
4458 * Make sure that __GFP_NOFAIL request doesn't leak out and make sure
4459 * we always retry
4460 */
4461 if (unlikely(nofail)) {
4462 /*
4463 * Lacking direct_reclaim we can't do anything to reclaim memory,
4464 * we disregard these unreasonable nofail requests and still
4465 * return NULL
4466 */
4467 if (!can_direct_reclaim)
4468 goto fail;
4469
4470 /*
4471 * Help non-failing allocations by giving some access to memory
4472 * reserves normally used for high priority non-blocking
4473 * allocations but do not use ALLOC_NO_WATERMARKS because this
4474 * could deplete whole memory reserves which would just make
4475 * the situation worse.
4476 */
4477 page = __alloc_pages_cpuset_fallback(gfp_mask, order, ALLOC_MIN_RESERVE, ac);
4478 if (page)
4479 goto got_pg;
4480
4481 cond_resched();
4482 goto retry;
4483 }
4484fail:
4485 warn_alloc(gfp_mask, ac->nodemask,
4486 "page allocation failure: order:%u", order);
4487got_pg:
4488 return page;
4489}
4490
4491static inline bool prepare_alloc_pages(gfp_t gfp_mask, unsigned int order,
4492 int preferred_nid, nodemask_t *nodemask,
4493 struct alloc_context *ac, gfp_t *alloc_gfp,
4494 unsigned int *alloc_flags)
4495{
4496 ac->highest_zoneidx = gfp_zone(gfp_mask);
4497 ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
4498 ac->nodemask = nodemask;
4499 ac->migratetype = gfp_migratetype(gfp_mask);
4500
4501 if (cpusets_enabled()) {
4502 *alloc_gfp |= __GFP_HARDWALL;
4503 /*
4504 * When we are in the interrupt context, it is irrelevant
4505 * to the current task context. It means that any node ok.
4506 */
4507 if (in_task() && !ac->nodemask)
4508 ac->nodemask = &cpuset_current_mems_allowed;
4509 else
4510 *alloc_flags |= ALLOC_CPUSET;
4511 }
4512
4513 might_alloc(gfp_mask);
4514
4515 if (should_fail_alloc_page(gfp_mask, order))
4516 return false;
4517
4518 *alloc_flags = gfp_to_alloc_flags_cma(gfp_mask, *alloc_flags);
4519
4520 /* Dirty zone balancing only done in the fast path */
4521 ac->spread_dirty_pages = (gfp_mask & __GFP_WRITE);
4522
4523 /*
4524 * The preferred zone is used for statistics but crucially it is
4525 * also used as the starting point for the zonelist iterator. It
4526 * may get reset for allocations that ignore memory policies.
4527 */
4528 ac->preferred_zoneref = first_zones_zonelist(ac->zonelist,
4529 ac->highest_zoneidx, ac->nodemask);
4530
4531 return true;
4532}
4533
4534/*
4535 * __alloc_pages_bulk - Allocate a number of order-0 pages to a list or array
4536 * @gfp: GFP flags for the allocation
4537 * @preferred_nid: The preferred NUMA node ID to allocate from
4538 * @nodemask: Set of nodes to allocate from, may be NULL
4539 * @nr_pages: The number of pages desired on the list or array
4540 * @page_list: Optional list to store the allocated pages
4541 * @page_array: Optional array to store the pages
4542 *
4543 * This is a batched version of the page allocator that attempts to
4544 * allocate nr_pages quickly. Pages are added to page_list if page_list
4545 * is not NULL, otherwise it is assumed that the page_array is valid.
4546 *
4547 * For lists, nr_pages is the number of pages that should be allocated.
4548 *
4549 * For arrays, only NULL elements are populated with pages and nr_pages
4550 * is the maximum number of pages that will be stored in the array.
4551 *
4552 * Returns the number of pages on the list or array.
4553 */
4554unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
4555 nodemask_t *nodemask, int nr_pages,
4556 struct list_head *page_list,
4557 struct page **page_array)
4558{
4559 struct page *page;
4560 unsigned long __maybe_unused UP_flags;
4561 struct zone *zone;
4562 struct zoneref *z;
4563 struct per_cpu_pages *pcp;
4564 struct list_head *pcp_list;
4565 struct alloc_context ac;
4566 gfp_t alloc_gfp;
4567 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4568 int nr_populated = 0, nr_account = 0;
4569
4570 /*
4571 * Skip populated array elements to determine if any pages need
4572 * to be allocated before disabling IRQs.
4573 */
4574 while (page_array && nr_populated < nr_pages && page_array[nr_populated])
4575 nr_populated++;
4576
4577 /* No pages requested? */
4578 if (unlikely(nr_pages <= 0))
4579 goto out;
4580
4581 /* Already populated array? */
4582 if (unlikely(page_array && nr_pages - nr_populated == 0))
4583 goto out;
4584
4585 /* Bulk allocator does not support memcg accounting. */
4586 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT))
4587 goto failed;
4588
4589 /* Use the single page allocator for one page. */
4590 if (nr_pages - nr_populated == 1)
4591 goto failed;
4592
4593#ifdef CONFIG_PAGE_OWNER
4594 /*
4595 * PAGE_OWNER may recurse into the allocator to allocate space to
4596 * save the stack with pagesets.lock held. Releasing/reacquiring
4597 * removes much of the performance benefit of bulk allocation so
4598 * force the caller to allocate one page at a time as it'll have
4599 * similar performance to added complexity to the bulk allocator.
4600 */
4601 if (static_branch_unlikely(&page_owner_inited))
4602 goto failed;
4603#endif
4604
4605 /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */
4606 gfp &= gfp_allowed_mask;
4607 alloc_gfp = gfp;
4608 if (!prepare_alloc_pages(gfp, 0, preferred_nid, nodemask, &ac, &alloc_gfp, &alloc_flags))
4609 goto out;
4610 gfp = alloc_gfp;
4611
4612 /* Find an allowed local zone that meets the low watermark. */
4613 z = ac.preferred_zoneref;
4614 for_next_zone_zonelist_nodemask(zone, z, ac.highest_zoneidx, ac.nodemask) {
4615 unsigned long mark;
4616
4617 if (cpusets_enabled() && (alloc_flags & ALLOC_CPUSET) &&
4618 !__cpuset_zone_allowed(zone, gfp)) {
4619 continue;
4620 }
4621
4622 if (nr_online_nodes > 1 && zone != zonelist_zone(ac.preferred_zoneref) &&
4623 zone_to_nid(zone) != zonelist_node_idx(ac.preferred_zoneref)) {
4624 goto failed;
4625 }
4626
4627 cond_accept_memory(zone, 0);
4628retry_this_zone:
4629 mark = wmark_pages(zone, alloc_flags & ALLOC_WMARK_MASK) + nr_pages;
4630 if (zone_watermark_fast(zone, 0, mark,
4631 zonelist_zone_idx(ac.preferred_zoneref),
4632 alloc_flags, gfp)) {
4633 break;
4634 }
4635
4636 if (cond_accept_memory(zone, 0))
4637 goto retry_this_zone;
4638
4639 /* Try again if zone has deferred pages */
4640 if (deferred_pages_enabled()) {
4641 if (_deferred_grow_zone(zone, 0))
4642 goto retry_this_zone;
4643 }
4644 }
4645
4646 /*
4647 * If there are no allowed local zones that meets the watermarks then
4648 * try to allocate a single page and reclaim if necessary.
4649 */
4650 if (unlikely(!zone))
4651 goto failed;
4652
4653 /* spin_trylock may fail due to a parallel drain or IRQ reentrancy. */
4654 pcp_trylock_prepare(UP_flags);
4655 pcp = pcp_spin_trylock(zone->per_cpu_pageset);
4656 if (!pcp)
4657 goto failed_irq;
4658
4659 /* Attempt the batch allocation */
4660 pcp_list = &pcp->lists[order_to_pindex(ac.migratetype, 0)];
4661 while (nr_populated < nr_pages) {
4662
4663 /* Skip existing pages */
4664 if (page_array && page_array[nr_populated]) {
4665 nr_populated++;
4666 continue;
4667 }
4668
4669 page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
4670 pcp, pcp_list);
4671 if (unlikely(!page)) {
4672 /* Try and allocate at least one page */
4673 if (!nr_account) {
4674 pcp_spin_unlock(pcp);
4675 goto failed_irq;
4676 }
4677 break;
4678 }
4679 nr_account++;
4680
4681 prep_new_page(page, 0, gfp, 0);
4682 if (page_list)
4683 list_add(&page->lru, page_list);
4684 else
4685 page_array[nr_populated] = page;
4686 nr_populated++;
4687 }
4688
4689 pcp_spin_unlock(pcp);
4690 pcp_trylock_finish(UP_flags);
4691
4692 __count_zid_vm_events(PGALLOC, zone_idx(zone), nr_account);
4693 zone_statistics(zonelist_zone(ac.preferred_zoneref), zone, nr_account);
4694
4695out:
4696 return nr_populated;
4697
4698failed_irq:
4699 pcp_trylock_finish(UP_flags);
4700
4701failed:
4702 page = __alloc_pages_noprof(gfp, 0, preferred_nid, nodemask);
4703 if (page) {
4704 if (page_list)
4705 list_add(&page->lru, page_list);
4706 else
4707 page_array[nr_populated] = page;
4708 nr_populated++;
4709 }
4710
4711 goto out;
4712}
4713EXPORT_SYMBOL_GPL(alloc_pages_bulk_noprof);
4714
4715/*
4716 * This is the 'heart' of the zoned buddy allocator.
4717 */
4718struct page *__alloc_pages_noprof(gfp_t gfp, unsigned int order,
4719 int preferred_nid, nodemask_t *nodemask)
4720{
4721 struct page *page;
4722 unsigned int alloc_flags = ALLOC_WMARK_LOW;
4723 gfp_t alloc_gfp; /* The gfp_t that was actually used for allocation */
4724 struct alloc_context ac = { };
4725
4726 /*
4727 * There are several places where we assume that the order value is sane
4728 * so bail out early if the request is out of bound.
4729 */
4730 if (WARN_ON_ONCE_GFP(order > MAX_PAGE_ORDER, gfp))
4731 return NULL;
4732
4733 gfp &= gfp_allowed_mask;
4734 /*
4735 * Apply scoped allocation constraints. This is mainly about GFP_NOFS
4736 * resp. GFP_NOIO which has to be inherited for all allocation requests
4737 * from a particular context which has been marked by
4738 * memalloc_no{fs,io}_{save,restore}. And PF_MEMALLOC_PIN which ensures
4739 * movable zones are not used during allocation.
4740 */
4741 gfp = current_gfp_context(gfp);
4742 alloc_gfp = gfp;
4743 if (!prepare_alloc_pages(gfp, order, preferred_nid, nodemask, &ac,
4744 &alloc_gfp, &alloc_flags))
4745 return NULL;
4746
4747 /*
4748 * Forbid the first pass from falling back to types that fragment
4749 * memory until all local zones are considered.
4750 */
4751 alloc_flags |= alloc_flags_nofragment(zonelist_zone(ac.preferred_zoneref), gfp);
4752
4753 /* First allocation attempt */
4754 page = get_page_from_freelist(alloc_gfp, order, alloc_flags, &ac);
4755 if (likely(page))
4756 goto out;
4757
4758 alloc_gfp = gfp;
4759 ac.spread_dirty_pages = false;
4760
4761 /*
4762 * Restore the original nodemask if it was potentially replaced with
4763 * &cpuset_current_mems_allowed to optimize the fast-path attempt.
4764 */
4765 ac.nodemask = nodemask;
4766
4767 page = __alloc_pages_slowpath(alloc_gfp, order, &ac);
4768
4769out:
4770 if (memcg_kmem_online() && (gfp & __GFP_ACCOUNT) && page &&
4771 unlikely(__memcg_kmem_charge_page(page, gfp, order) != 0)) {
4772 __free_pages(page, order);
4773 page = NULL;
4774 }
4775
4776 trace_mm_page_alloc(page, order, alloc_gfp, ac.migratetype);
4777 kmsan_alloc_page(page, order, alloc_gfp);
4778
4779 return page;
4780}
4781EXPORT_SYMBOL(__alloc_pages_noprof);
4782
4783struct folio *__folio_alloc_noprof(gfp_t gfp, unsigned int order, int preferred_nid,
4784 nodemask_t *nodemask)
4785{
4786 struct page *page = __alloc_pages_noprof(gfp | __GFP_COMP, order,
4787 preferred_nid, nodemask);
4788 return page_rmappable_folio(page);
4789}
4790EXPORT_SYMBOL(__folio_alloc_noprof);
4791
4792/*
4793 * Common helper functions. Never use with __GFP_HIGHMEM because the returned
4794 * address cannot represent highmem pages. Use alloc_pages and then kmap if
4795 * you need to access high mem.
4796 */
4797unsigned long get_free_pages_noprof(gfp_t gfp_mask, unsigned int order)
4798{
4799 struct page *page;
4800
4801 page = alloc_pages_noprof(gfp_mask & ~__GFP_HIGHMEM, order);
4802 if (!page)
4803 return 0;
4804 return (unsigned long) page_address(page);
4805}
4806EXPORT_SYMBOL(get_free_pages_noprof);
4807
4808unsigned long get_zeroed_page_noprof(gfp_t gfp_mask)
4809{
4810 return get_free_pages_noprof(gfp_mask | __GFP_ZERO, 0);
4811}
4812EXPORT_SYMBOL(get_zeroed_page_noprof);
4813
4814/**
4815 * __free_pages - Free pages allocated with alloc_pages().
4816 * @page: The page pointer returned from alloc_pages().
4817 * @order: The order of the allocation.
4818 *
4819 * This function can free multi-page allocations that are not compound
4820 * pages. It does not check that the @order passed in matches that of
4821 * the allocation, so it is easy to leak memory. Freeing more memory
4822 * than was allocated will probably emit a warning.
4823 *
4824 * If the last reference to this page is speculative, it will be released
4825 * by put_page() which only frees the first page of a non-compound
4826 * allocation. To prevent the remaining pages from being leaked, we free
4827 * the subsequent pages here. If you want to use the page's reference
4828 * count to decide when to free the allocation, you should allocate a
4829 * compound page, and use put_page() instead of __free_pages().
4830 *
4831 * Context: May be called in interrupt context or while holding a normal
4832 * spinlock, but not in NMI context or while holding a raw spinlock.
4833 */
4834void __free_pages(struct page *page, unsigned int order)
4835{
4836 /* get PageHead before we drop reference */
4837 int head = PageHead(page);
4838 struct alloc_tag *tag = pgalloc_tag_get(page);
4839
4840 if (put_page_testzero(page))
4841 free_unref_page(page, order);
4842 else if (!head) {
4843 pgalloc_tag_sub_pages(tag, (1 << order) - 1);
4844 while (order-- > 0)
4845 free_unref_page(page + (1 << order), order);
4846 }
4847}
4848EXPORT_SYMBOL(__free_pages);
4849
4850void free_pages(unsigned long addr, unsigned int order)
4851{
4852 if (addr != 0) {
4853 VM_BUG_ON(!virt_addr_valid((void *)addr));
4854 __free_pages(virt_to_page((void *)addr), order);
4855 }
4856}
4857
4858EXPORT_SYMBOL(free_pages);
4859
4860static void *make_alloc_exact(unsigned long addr, unsigned int order,
4861 size_t size)
4862{
4863 if (addr) {
4864 unsigned long nr = DIV_ROUND_UP(size, PAGE_SIZE);
4865 struct page *page = virt_to_page((void *)addr);
4866 struct page *last = page + nr;
4867
4868 split_page_owner(page, order, 0);
4869 pgalloc_tag_split(page_folio(page), order, 0);
4870 split_page_memcg(page, order, 0);
4871 while (page < --last)
4872 set_page_refcounted(last);
4873
4874 last = page + (1UL << order);
4875 for (page += nr; page < last; page++)
4876 __free_pages_ok(page, 0, FPI_TO_TAIL);
4877 }
4878 return (void *)addr;
4879}
4880
4881/**
4882 * alloc_pages_exact - allocate an exact number physically-contiguous pages.
4883 * @size: the number of bytes to allocate
4884 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4885 *
4886 * This function is similar to alloc_pages(), except that it allocates the
4887 * minimum number of pages to satisfy the request. alloc_pages() can only
4888 * allocate memory in power-of-two pages.
4889 *
4890 * This function is also limited by MAX_PAGE_ORDER.
4891 *
4892 * Memory allocated by this function must be released by free_pages_exact().
4893 *
4894 * Return: pointer to the allocated area or %NULL in case of error.
4895 */
4896void *alloc_pages_exact_noprof(size_t size, gfp_t gfp_mask)
4897{
4898 unsigned int order = get_order(size);
4899 unsigned long addr;
4900
4901 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4902 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4903
4904 addr = get_free_pages_noprof(gfp_mask, order);
4905 return make_alloc_exact(addr, order, size);
4906}
4907EXPORT_SYMBOL(alloc_pages_exact_noprof);
4908
4909/**
4910 * alloc_pages_exact_nid - allocate an exact number of physically-contiguous
4911 * pages on a node.
4912 * @nid: the preferred node ID where memory should be allocated
4913 * @size: the number of bytes to allocate
4914 * @gfp_mask: GFP flags for the allocation, must not contain __GFP_COMP
4915 *
4916 * Like alloc_pages_exact(), but try to allocate on node nid first before falling
4917 * back.
4918 *
4919 * Return: pointer to the allocated area or %NULL in case of error.
4920 */
4921void * __meminit alloc_pages_exact_nid_noprof(int nid, size_t size, gfp_t gfp_mask)
4922{
4923 unsigned int order = get_order(size);
4924 struct page *p;
4925
4926 if (WARN_ON_ONCE(gfp_mask & (__GFP_COMP | __GFP_HIGHMEM)))
4927 gfp_mask &= ~(__GFP_COMP | __GFP_HIGHMEM);
4928
4929 p = alloc_pages_node_noprof(nid, gfp_mask, order);
4930 if (!p)
4931 return NULL;
4932 return make_alloc_exact((unsigned long)page_address(p), order, size);
4933}
4934
4935/**
4936 * free_pages_exact - release memory allocated via alloc_pages_exact()
4937 * @virt: the value returned by alloc_pages_exact.
4938 * @size: size of allocation, same value as passed to alloc_pages_exact().
4939 *
4940 * Release the memory allocated by a previous call to alloc_pages_exact.
4941 */
4942void free_pages_exact(void *virt, size_t size)
4943{
4944 unsigned long addr = (unsigned long)virt;
4945 unsigned long end = addr + PAGE_ALIGN(size);
4946
4947 while (addr < end) {
4948 free_page(addr);
4949 addr += PAGE_SIZE;
4950 }
4951}
4952EXPORT_SYMBOL(free_pages_exact);
4953
4954/**
4955 * nr_free_zone_pages - count number of pages beyond high watermark
4956 * @offset: The zone index of the highest zone
4957 *
4958 * nr_free_zone_pages() counts the number of pages which are beyond the
4959 * high watermark within all zones at or below a given zone index. For each
4960 * zone, the number of pages is calculated as:
4961 *
4962 * nr_free_zone_pages = managed_pages - high_pages
4963 *
4964 * Return: number of pages beyond high watermark.
4965 */
4966static unsigned long nr_free_zone_pages(int offset)
4967{
4968 struct zoneref *z;
4969 struct zone *zone;
4970
4971 /* Just pick one node, since fallback list is circular */
4972 unsigned long sum = 0;
4973
4974 struct zonelist *zonelist = node_zonelist(numa_node_id(), GFP_KERNEL);
4975
4976 for_each_zone_zonelist(zone, z, zonelist, offset) {
4977 unsigned long size = zone_managed_pages(zone);
4978 unsigned long high = high_wmark_pages(zone);
4979 if (size > high)
4980 sum += size - high;
4981 }
4982
4983 return sum;
4984}
4985
4986/**
4987 * nr_free_buffer_pages - count number of pages beyond high watermark
4988 *
4989 * nr_free_buffer_pages() counts the number of pages which are beyond the high
4990 * watermark within ZONE_DMA and ZONE_NORMAL.
4991 *
4992 * Return: number of pages beyond high watermark within ZONE_DMA and
4993 * ZONE_NORMAL.
4994 */
4995unsigned long nr_free_buffer_pages(void)
4996{
4997 return nr_free_zone_pages(gfp_zone(GFP_USER));
4998}
4999EXPORT_SYMBOL_GPL(nr_free_buffer_pages);
5000
5001static void zoneref_set_zone(struct zone *zone, struct zoneref *zoneref)
5002{
5003 zoneref->zone = zone;
5004 zoneref->zone_idx = zone_idx(zone);
5005}
5006
5007/*
5008 * Builds allocation fallback zone lists.
5009 *
5010 * Add all populated zones of a node to the zonelist.
5011 */
5012static int build_zonerefs_node(pg_data_t *pgdat, struct zoneref *zonerefs)
5013{
5014 struct zone *zone;
5015 enum zone_type zone_type = MAX_NR_ZONES;
5016 int nr_zones = 0;
5017
5018 do {
5019 zone_type--;
5020 zone = pgdat->node_zones + zone_type;
5021 if (populated_zone(zone)) {
5022 zoneref_set_zone(zone, &zonerefs[nr_zones++]);
5023 check_highest_zone(zone_type);
5024 }
5025 } while (zone_type);
5026
5027 return nr_zones;
5028}
5029
5030#ifdef CONFIG_NUMA
5031
5032static int __parse_numa_zonelist_order(char *s)
5033{
5034 /*
5035 * We used to support different zonelists modes but they turned
5036 * out to be just not useful. Let's keep the warning in place
5037 * if somebody still use the cmd line parameter so that we do
5038 * not fail it silently
5039 */
5040 if (!(*s == 'd' || *s == 'D' || *s == 'n' || *s == 'N')) {
5041 pr_warn("Ignoring unsupported numa_zonelist_order value: %s\n", s);
5042 return -EINVAL;
5043 }
5044 return 0;
5045}
5046
5047static char numa_zonelist_order[] = "Node";
5048#define NUMA_ZONELIST_ORDER_LEN 16
5049/*
5050 * sysctl handler for numa_zonelist_order
5051 */
5052static int numa_zonelist_order_handler(const struct ctl_table *table, int write,
5053 void *buffer, size_t *length, loff_t *ppos)
5054{
5055 if (write)
5056 return __parse_numa_zonelist_order(buffer);
5057 return proc_dostring(table, write, buffer, length, ppos);
5058}
5059
5060static int node_load[MAX_NUMNODES];
5061
5062/**
5063 * find_next_best_node - find the next node that should appear in a given node's fallback list
5064 * @node: node whose fallback list we're appending
5065 * @used_node_mask: nodemask_t of already used nodes
5066 *
5067 * We use a number of factors to determine which is the next node that should
5068 * appear on a given node's fallback list. The node should not have appeared
5069 * already in @node's fallback list, and it should be the next closest node
5070 * according to the distance array (which contains arbitrary distance values
5071 * from each node to each node in the system), and should also prefer nodes
5072 * with no CPUs, since presumably they'll have very little allocation pressure
5073 * on them otherwise.
5074 *
5075 * Return: node id of the found node or %NUMA_NO_NODE if no node is found.
5076 */
5077int find_next_best_node(int node, nodemask_t *used_node_mask)
5078{
5079 int n, val;
5080 int min_val = INT_MAX;
5081 int best_node = NUMA_NO_NODE;
5082
5083 /*
5084 * Use the local node if we haven't already, but for memoryless local
5085 * node, we should skip it and fall back to other nodes.
5086 */
5087 if (!node_isset(node, *used_node_mask) && node_state(node, N_MEMORY)) {
5088 node_set(node, *used_node_mask);
5089 return node;
5090 }
5091
5092 for_each_node_state(n, N_MEMORY) {
5093
5094 /* Don't want a node to appear more than once */
5095 if (node_isset(n, *used_node_mask))
5096 continue;
5097
5098 /* Use the distance array to find the distance */
5099 val = node_distance(node, n);
5100
5101 /* Penalize nodes under us ("prefer the next node") */
5102 val += (n < node);
5103
5104 /* Give preference to headless and unused nodes */
5105 if (!cpumask_empty(cpumask_of_node(n)))
5106 val += PENALTY_FOR_NODE_WITH_CPUS;
5107
5108 /* Slight preference for less loaded node */
5109 val *= MAX_NUMNODES;
5110 val += node_load[n];
5111
5112 if (val < min_val) {
5113 min_val = val;
5114 best_node = n;
5115 }
5116 }
5117
5118 if (best_node >= 0)
5119 node_set(best_node, *used_node_mask);
5120
5121 return best_node;
5122}
5123
5124
5125/*
5126 * Build zonelists ordered by node and zones within node.
5127 * This results in maximum locality--normal zone overflows into local
5128 * DMA zone, if any--but risks exhausting DMA zone.
5129 */
5130static void build_zonelists_in_node_order(pg_data_t *pgdat, int *node_order,
5131 unsigned nr_nodes)
5132{
5133 struct zoneref *zonerefs;
5134 int i;
5135
5136 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5137
5138 for (i = 0; i < nr_nodes; i++) {
5139 int nr_zones;
5140
5141 pg_data_t *node = NODE_DATA(node_order[i]);
5142
5143 nr_zones = build_zonerefs_node(node, zonerefs);
5144 zonerefs += nr_zones;
5145 }
5146 zonerefs->zone = NULL;
5147 zonerefs->zone_idx = 0;
5148}
5149
5150/*
5151 * Build __GFP_THISNODE zonelists
5152 */
5153static void build_thisnode_zonelists(pg_data_t *pgdat)
5154{
5155 struct zoneref *zonerefs;
5156 int nr_zones;
5157
5158 zonerefs = pgdat->node_zonelists[ZONELIST_NOFALLBACK]._zonerefs;
5159 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5160 zonerefs += nr_zones;
5161 zonerefs->zone = NULL;
5162 zonerefs->zone_idx = 0;
5163}
5164
5165/*
5166 * Build zonelists ordered by zone and nodes within zones.
5167 * This results in conserving DMA zone[s] until all Normal memory is
5168 * exhausted, but results in overflowing to remote node while memory
5169 * may still exist in local DMA zone.
5170 */
5171
5172static void build_zonelists(pg_data_t *pgdat)
5173{
5174 static int node_order[MAX_NUMNODES];
5175 int node, nr_nodes = 0;
5176 nodemask_t used_mask = NODE_MASK_NONE;
5177 int local_node, prev_node;
5178
5179 /* NUMA-aware ordering of nodes */
5180 local_node = pgdat->node_id;
5181 prev_node = local_node;
5182
5183 memset(node_order, 0, sizeof(node_order));
5184 while ((node = find_next_best_node(local_node, &used_mask)) >= 0) {
5185 /*
5186 * We don't want to pressure a particular node.
5187 * So adding penalty to the first node in same
5188 * distance group to make it round-robin.
5189 */
5190 if (node_distance(local_node, node) !=
5191 node_distance(local_node, prev_node))
5192 node_load[node] += 1;
5193
5194 node_order[nr_nodes++] = node;
5195 prev_node = node;
5196 }
5197
5198 build_zonelists_in_node_order(pgdat, node_order, nr_nodes);
5199 build_thisnode_zonelists(pgdat);
5200 pr_info("Fallback order for Node %d: ", local_node);
5201 for (node = 0; node < nr_nodes; node++)
5202 pr_cont("%d ", node_order[node]);
5203 pr_cont("\n");
5204}
5205
5206#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5207/*
5208 * Return node id of node used for "local" allocations.
5209 * I.e., first node id of first zone in arg node's generic zonelist.
5210 * Used for initializing percpu 'numa_mem', which is used primarily
5211 * for kernel allocations, so use GFP_KERNEL flags to locate zonelist.
5212 */
5213int local_memory_node(int node)
5214{
5215 struct zoneref *z;
5216
5217 z = first_zones_zonelist(node_zonelist(node, GFP_KERNEL),
5218 gfp_zone(GFP_KERNEL),
5219 NULL);
5220 return zonelist_node_idx(z);
5221}
5222#endif
5223
5224static void setup_min_unmapped_ratio(void);
5225static void setup_min_slab_ratio(void);
5226#else /* CONFIG_NUMA */
5227
5228static void build_zonelists(pg_data_t *pgdat)
5229{
5230 struct zoneref *zonerefs;
5231 int nr_zones;
5232
5233 zonerefs = pgdat->node_zonelists[ZONELIST_FALLBACK]._zonerefs;
5234 nr_zones = build_zonerefs_node(pgdat, zonerefs);
5235 zonerefs += nr_zones;
5236
5237 zonerefs->zone = NULL;
5238 zonerefs->zone_idx = 0;
5239}
5240
5241#endif /* CONFIG_NUMA */
5242
5243/*
5244 * Boot pageset table. One per cpu which is going to be used for all
5245 * zones and all nodes. The parameters will be set in such a way
5246 * that an item put on a list will immediately be handed over to
5247 * the buddy list. This is safe since pageset manipulation is done
5248 * with interrupts disabled.
5249 *
5250 * The boot_pagesets must be kept even after bootup is complete for
5251 * unused processors and/or zones. They do play a role for bootstrapping
5252 * hotplugged processors.
5253 *
5254 * zoneinfo_show() and maybe other functions do
5255 * not check if the processor is online before following the pageset pointer.
5256 * Other parts of the kernel may not check if the zone is available.
5257 */
5258static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats);
5259/* These effectively disable the pcplists in the boot pageset completely */
5260#define BOOT_PAGESET_HIGH 0
5261#define BOOT_PAGESET_BATCH 1
5262static DEFINE_PER_CPU(struct per_cpu_pages, boot_pageset);
5263static DEFINE_PER_CPU(struct per_cpu_zonestat, boot_zonestats);
5264
5265static void __build_all_zonelists(void *data)
5266{
5267 int nid;
5268 int __maybe_unused cpu;
5269 pg_data_t *self = data;
5270 unsigned long flags;
5271
5272 /*
5273 * The zonelist_update_seq must be acquired with irqsave because the
5274 * reader can be invoked from IRQ with GFP_ATOMIC.
5275 */
5276 write_seqlock_irqsave(&zonelist_update_seq, flags);
5277 /*
5278 * Also disable synchronous printk() to prevent any printk() from
5279 * trying to hold port->lock, for
5280 * tty_insert_flip_string_and_push_buffer() on other CPU might be
5281 * calling kmalloc(GFP_ATOMIC | __GFP_NOWARN) with port->lock held.
5282 */
5283 printk_deferred_enter();
5284
5285#ifdef CONFIG_NUMA
5286 memset(node_load, 0, sizeof(node_load));
5287#endif
5288
5289 /*
5290 * This node is hotadded and no memory is yet present. So just
5291 * building zonelists is fine - no need to touch other nodes.
5292 */
5293 if (self && !node_online(self->node_id)) {
5294 build_zonelists(self);
5295 } else {
5296 /*
5297 * All possible nodes have pgdat preallocated
5298 * in free_area_init
5299 */
5300 for_each_node(nid) {
5301 pg_data_t *pgdat = NODE_DATA(nid);
5302
5303 build_zonelists(pgdat);
5304 }
5305
5306#ifdef CONFIG_HAVE_MEMORYLESS_NODES
5307 /*
5308 * We now know the "local memory node" for each node--
5309 * i.e., the node of the first zone in the generic zonelist.
5310 * Set up numa_mem percpu variable for on-line cpus. During
5311 * boot, only the boot cpu should be on-line; we'll init the
5312 * secondary cpus' numa_mem as they come on-line. During
5313 * node/memory hotplug, we'll fixup all on-line cpus.
5314 */
5315 for_each_online_cpu(cpu)
5316 set_cpu_numa_mem(cpu, local_memory_node(cpu_to_node(cpu)));
5317#endif
5318 }
5319
5320 printk_deferred_exit();
5321 write_sequnlock_irqrestore(&zonelist_update_seq, flags);
5322}
5323
5324static noinline void __init
5325build_all_zonelists_init(void)
5326{
5327 int cpu;
5328
5329 __build_all_zonelists(NULL);
5330
5331 /*
5332 * Initialize the boot_pagesets that are going to be used
5333 * for bootstrapping processors. The real pagesets for
5334 * each zone will be allocated later when the per cpu
5335 * allocator is available.
5336 *
5337 * boot_pagesets are used also for bootstrapping offline
5338 * cpus if the system is already booted because the pagesets
5339 * are needed to initialize allocators on a specific cpu too.
5340 * F.e. the percpu allocator needs the page allocator which
5341 * needs the percpu allocator in order to allocate its pagesets
5342 * (a chicken-egg dilemma).
5343 */
5344 for_each_possible_cpu(cpu)
5345 per_cpu_pages_init(&per_cpu(boot_pageset, cpu), &per_cpu(boot_zonestats, cpu));
5346
5347 mminit_verify_zonelist();
5348 cpuset_init_current_mems_allowed();
5349}
5350
5351/*
5352 * unless system_state == SYSTEM_BOOTING.
5353 *
5354 * __ref due to call of __init annotated helper build_all_zonelists_init
5355 * [protected by SYSTEM_BOOTING].
5356 */
5357void __ref build_all_zonelists(pg_data_t *pgdat)
5358{
5359 unsigned long vm_total_pages;
5360
5361 if (system_state == SYSTEM_BOOTING) {
5362 build_all_zonelists_init();
5363 } else {
5364 __build_all_zonelists(pgdat);
5365 /* cpuset refresh routine should be here */
5366 }
5367 /* Get the number of free pages beyond high watermark in all zones. */
5368 vm_total_pages = nr_free_zone_pages(gfp_zone(GFP_HIGHUSER_MOVABLE));
5369 /*
5370 * Disable grouping by mobility if the number of pages in the
5371 * system is too low to allow the mechanism to work. It would be
5372 * more accurate, but expensive to check per-zone. This check is
5373 * made on memory-hotadd so a system can start with mobility
5374 * disabled and enable it later
5375 */
5376 if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
5377 page_group_by_mobility_disabled = 1;
5378 else
5379 page_group_by_mobility_disabled = 0;
5380
5381 pr_info("Built %u zonelists, mobility grouping %s. Total pages: %ld\n",
5382 nr_online_nodes,
5383 str_off_on(page_group_by_mobility_disabled),
5384 vm_total_pages);
5385#ifdef CONFIG_NUMA
5386 pr_info("Policy zone: %s\n", zone_names[policy_zone]);
5387#endif
5388}
5389
5390static int zone_batchsize(struct zone *zone)
5391{
5392#ifdef CONFIG_MMU
5393 int batch;
5394
5395 /*
5396 * The number of pages to batch allocate is either ~0.1%
5397 * of the zone or 1MB, whichever is smaller. The batch
5398 * size is striking a balance between allocation latency
5399 * and zone lock contention.
5400 */
5401 batch = min(zone_managed_pages(zone) >> 10, SZ_1M / PAGE_SIZE);
5402 batch /= 4; /* We effectively *= 4 below */
5403 if (batch < 1)
5404 batch = 1;
5405
5406 /*
5407 * Clamp the batch to a 2^n - 1 value. Having a power
5408 * of 2 value was found to be more likely to have
5409 * suboptimal cache aliasing properties in some cases.
5410 *
5411 * For example if 2 tasks are alternately allocating
5412 * batches of pages, one task can end up with a lot
5413 * of pages of one half of the possible page colors
5414 * and the other with pages of the other colors.
5415 */
5416 batch = rounddown_pow_of_two(batch + batch/2) - 1;
5417
5418 return batch;
5419
5420#else
5421 /* The deferral and batching of frees should be suppressed under NOMMU
5422 * conditions.
5423 *
5424 * The problem is that NOMMU needs to be able to allocate large chunks
5425 * of contiguous memory as there's no hardware page translation to
5426 * assemble apparent contiguous memory from discontiguous pages.
5427 *
5428 * Queueing large contiguous runs of pages for batching, however,
5429 * causes the pages to actually be freed in smaller chunks. As there
5430 * can be a significant delay between the individual batches being
5431 * recycled, this leads to the once large chunks of space being
5432 * fragmented and becoming unavailable for high-order allocations.
5433 */
5434 return 0;
5435#endif
5436}
5437
5438static int percpu_pagelist_high_fraction;
5439static int zone_highsize(struct zone *zone, int batch, int cpu_online,
5440 int high_fraction)
5441{
5442#ifdef CONFIG_MMU
5443 int high;
5444 int nr_split_cpus;
5445 unsigned long total_pages;
5446
5447 if (!high_fraction) {
5448 /*
5449 * By default, the high value of the pcp is based on the zone
5450 * low watermark so that if they are full then background
5451 * reclaim will not be started prematurely.
5452 */
5453 total_pages = low_wmark_pages(zone);
5454 } else {
5455 /*
5456 * If percpu_pagelist_high_fraction is configured, the high
5457 * value is based on a fraction of the managed pages in the
5458 * zone.
5459 */
5460 total_pages = zone_managed_pages(zone) / high_fraction;
5461 }
5462
5463 /*
5464 * Split the high value across all online CPUs local to the zone. Note
5465 * that early in boot that CPUs may not be online yet and that during
5466 * CPU hotplug that the cpumask is not yet updated when a CPU is being
5467 * onlined. For memory nodes that have no CPUs, split the high value
5468 * across all online CPUs to mitigate the risk that reclaim is triggered
5469 * prematurely due to pages stored on pcp lists.
5470 */
5471 nr_split_cpus = cpumask_weight(cpumask_of_node(zone_to_nid(zone))) + cpu_online;
5472 if (!nr_split_cpus)
5473 nr_split_cpus = num_online_cpus();
5474 high = total_pages / nr_split_cpus;
5475
5476 /*
5477 * Ensure high is at least batch*4. The multiple is based on the
5478 * historical relationship between high and batch.
5479 */
5480 high = max(high, batch << 2);
5481
5482 return high;
5483#else
5484 return 0;
5485#endif
5486}
5487
5488/*
5489 * pcp->high and pcp->batch values are related and generally batch is lower
5490 * than high. They are also related to pcp->count such that count is lower
5491 * than high, and as soon as it reaches high, the pcplist is flushed.
5492 *
5493 * However, guaranteeing these relations at all times would require e.g. write
5494 * barriers here but also careful usage of read barriers at the read side, and
5495 * thus be prone to error and bad for performance. Thus the update only prevents
5496 * store tearing. Any new users of pcp->batch, pcp->high_min and pcp->high_max
5497 * should ensure they can cope with those fields changing asynchronously, and
5498 * fully trust only the pcp->count field on the local CPU with interrupts
5499 * disabled.
5500 *
5501 * mutex_is_locked(&pcp_batch_high_lock) required when calling this function
5502 * outside of boot time (or some other assurance that no concurrent updaters
5503 * exist).
5504 */
5505static void pageset_update(struct per_cpu_pages *pcp, unsigned long high_min,
5506 unsigned long high_max, unsigned long batch)
5507{
5508 WRITE_ONCE(pcp->batch, batch);
5509 WRITE_ONCE(pcp->high_min, high_min);
5510 WRITE_ONCE(pcp->high_max, high_max);
5511}
5512
5513static void per_cpu_pages_init(struct per_cpu_pages *pcp, struct per_cpu_zonestat *pzstats)
5514{
5515 int pindex;
5516
5517 memset(pcp, 0, sizeof(*pcp));
5518 memset(pzstats, 0, sizeof(*pzstats));
5519
5520 spin_lock_init(&pcp->lock);
5521 for (pindex = 0; pindex < NR_PCP_LISTS; pindex++)
5522 INIT_LIST_HEAD(&pcp->lists[pindex]);
5523
5524 /*
5525 * Set batch and high values safe for a boot pageset. A true percpu
5526 * pageset's initialization will update them subsequently. Here we don't
5527 * need to be as careful as pageset_update() as nobody can access the
5528 * pageset yet.
5529 */
5530 pcp->high_min = BOOT_PAGESET_HIGH;
5531 pcp->high_max = BOOT_PAGESET_HIGH;
5532 pcp->batch = BOOT_PAGESET_BATCH;
5533 pcp->free_count = 0;
5534}
5535
5536static void __zone_set_pageset_high_and_batch(struct zone *zone, unsigned long high_min,
5537 unsigned long high_max, unsigned long batch)
5538{
5539 struct per_cpu_pages *pcp;
5540 int cpu;
5541
5542 for_each_possible_cpu(cpu) {
5543 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5544 pageset_update(pcp, high_min, high_max, batch);
5545 }
5546}
5547
5548/*
5549 * Calculate and set new high and batch values for all per-cpu pagesets of a
5550 * zone based on the zone's size.
5551 */
5552static void zone_set_pageset_high_and_batch(struct zone *zone, int cpu_online)
5553{
5554 int new_high_min, new_high_max, new_batch;
5555
5556 new_batch = max(1, zone_batchsize(zone));
5557 if (percpu_pagelist_high_fraction) {
5558 new_high_min = zone_highsize(zone, new_batch, cpu_online,
5559 percpu_pagelist_high_fraction);
5560 /*
5561 * PCP high is tuned manually, disable auto-tuning via
5562 * setting high_min and high_max to the manual value.
5563 */
5564 new_high_max = new_high_min;
5565 } else {
5566 new_high_min = zone_highsize(zone, new_batch, cpu_online, 0);
5567 new_high_max = zone_highsize(zone, new_batch, cpu_online,
5568 MIN_PERCPU_PAGELIST_HIGH_FRACTION);
5569 }
5570
5571 if (zone->pageset_high_min == new_high_min &&
5572 zone->pageset_high_max == new_high_max &&
5573 zone->pageset_batch == new_batch)
5574 return;
5575
5576 zone->pageset_high_min = new_high_min;
5577 zone->pageset_high_max = new_high_max;
5578 zone->pageset_batch = new_batch;
5579
5580 __zone_set_pageset_high_and_batch(zone, new_high_min, new_high_max,
5581 new_batch);
5582}
5583
5584void __meminit setup_zone_pageset(struct zone *zone)
5585{
5586 int cpu;
5587
5588 /* Size may be 0 on !SMP && !NUMA */
5589 if (sizeof(struct per_cpu_zonestat) > 0)
5590 zone->per_cpu_zonestats = alloc_percpu(struct per_cpu_zonestat);
5591
5592 zone->per_cpu_pageset = alloc_percpu(struct per_cpu_pages);
5593 for_each_possible_cpu(cpu) {
5594 struct per_cpu_pages *pcp;
5595 struct per_cpu_zonestat *pzstats;
5596
5597 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5598 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
5599 per_cpu_pages_init(pcp, pzstats);
5600 }
5601
5602 zone_set_pageset_high_and_batch(zone, 0);
5603}
5604
5605/*
5606 * The zone indicated has a new number of managed_pages; batch sizes and percpu
5607 * page high values need to be recalculated.
5608 */
5609static void zone_pcp_update(struct zone *zone, int cpu_online)
5610{
5611 mutex_lock(&pcp_batch_high_lock);
5612 zone_set_pageset_high_and_batch(zone, cpu_online);
5613 mutex_unlock(&pcp_batch_high_lock);
5614}
5615
5616static void zone_pcp_update_cacheinfo(struct zone *zone, unsigned int cpu)
5617{
5618 struct per_cpu_pages *pcp;
5619 struct cpu_cacheinfo *cci;
5620
5621 pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
5622 cci = get_cpu_cacheinfo(cpu);
5623 /*
5624 * If data cache slice of CPU is large enough, "pcp->batch"
5625 * pages can be preserved in PCP before draining PCP for
5626 * consecutive high-order pages freeing without allocation.
5627 * This can reduce zone lock contention without hurting
5628 * cache-hot pages sharing.
5629 */
5630 spin_lock(&pcp->lock);
5631 if ((cci->per_cpu_data_slice_size >> PAGE_SHIFT) > 3 * pcp->batch)
5632 pcp->flags |= PCPF_FREE_HIGH_BATCH;
5633 else
5634 pcp->flags &= ~PCPF_FREE_HIGH_BATCH;
5635 spin_unlock(&pcp->lock);
5636}
5637
5638void setup_pcp_cacheinfo(unsigned int cpu)
5639{
5640 struct zone *zone;
5641
5642 for_each_populated_zone(zone)
5643 zone_pcp_update_cacheinfo(zone, cpu);
5644}
5645
5646/*
5647 * Allocate per cpu pagesets and initialize them.
5648 * Before this call only boot pagesets were available.
5649 */
5650void __init setup_per_cpu_pageset(void)
5651{
5652 struct pglist_data *pgdat;
5653 struct zone *zone;
5654 int __maybe_unused cpu;
5655
5656 for_each_populated_zone(zone)
5657 setup_zone_pageset(zone);
5658
5659#ifdef CONFIG_NUMA
5660 /*
5661 * Unpopulated zones continue using the boot pagesets.
5662 * The numa stats for these pagesets need to be reset.
5663 * Otherwise, they will end up skewing the stats of
5664 * the nodes these zones are associated with.
5665 */
5666 for_each_possible_cpu(cpu) {
5667 struct per_cpu_zonestat *pzstats = &per_cpu(boot_zonestats, cpu);
5668 memset(pzstats->vm_numa_event, 0,
5669 sizeof(pzstats->vm_numa_event));
5670 }
5671#endif
5672
5673 for_each_online_pgdat(pgdat)
5674 pgdat->per_cpu_nodestats =
5675 alloc_percpu(struct per_cpu_nodestat);
5676}
5677
5678__meminit void zone_pcp_init(struct zone *zone)
5679{
5680 /*
5681 * per cpu subsystem is not up at this point. The following code
5682 * relies on the ability of the linker to provide the
5683 * offset of a (static) per cpu variable into the per cpu area.
5684 */
5685 zone->per_cpu_pageset = &boot_pageset;
5686 zone->per_cpu_zonestats = &boot_zonestats;
5687 zone->pageset_high_min = BOOT_PAGESET_HIGH;
5688 zone->pageset_high_max = BOOT_PAGESET_HIGH;
5689 zone->pageset_batch = BOOT_PAGESET_BATCH;
5690
5691 if (populated_zone(zone))
5692 pr_debug(" %s zone: %lu pages, LIFO batch:%u\n", zone->name,
5693 zone->present_pages, zone_batchsize(zone));
5694}
5695
5696static void setup_per_zone_lowmem_reserve(void);
5697
5698void adjust_managed_page_count(struct page *page, long count)
5699{
5700 atomic_long_add(count, &page_zone(page)->managed_pages);
5701 totalram_pages_add(count);
5702 setup_per_zone_lowmem_reserve();
5703}
5704EXPORT_SYMBOL(adjust_managed_page_count);
5705
5706unsigned long free_reserved_area(void *start, void *end, int poison, const char *s)
5707{
5708 void *pos;
5709 unsigned long pages = 0;
5710
5711 start = (void *)PAGE_ALIGN((unsigned long)start);
5712 end = (void *)((unsigned long)end & PAGE_MASK);
5713 for (pos = start; pos < end; pos += PAGE_SIZE, pages++) {
5714 struct page *page = virt_to_page(pos);
5715 void *direct_map_addr;
5716
5717 /*
5718 * 'direct_map_addr' might be different from 'pos'
5719 * because some architectures' virt_to_page()
5720 * work with aliases. Getting the direct map
5721 * address ensures that we get a _writeable_
5722 * alias for the memset().
5723 */
5724 direct_map_addr = page_address(page);
5725 /*
5726 * Perform a kasan-unchecked memset() since this memory
5727 * has not been initialized.
5728 */
5729 direct_map_addr = kasan_reset_tag(direct_map_addr);
5730 if ((unsigned int)poison <= 0xFF)
5731 memset(direct_map_addr, poison, PAGE_SIZE);
5732
5733 free_reserved_page(page);
5734 }
5735
5736 if (pages && s)
5737 pr_info("Freeing %s memory: %ldK\n", s, K(pages));
5738
5739 return pages;
5740}
5741
5742void free_reserved_page(struct page *page)
5743{
5744 clear_page_tag_ref(page);
5745 ClearPageReserved(page);
5746 init_page_count(page);
5747 __free_page(page);
5748 adjust_managed_page_count(page, 1);
5749}
5750EXPORT_SYMBOL(free_reserved_page);
5751
5752static int page_alloc_cpu_dead(unsigned int cpu)
5753{
5754 struct zone *zone;
5755
5756 lru_add_drain_cpu(cpu);
5757 mlock_drain_remote(cpu);
5758 drain_pages(cpu);
5759
5760 /*
5761 * Spill the event counters of the dead processor
5762 * into the current processors event counters.
5763 * This artificially elevates the count of the current
5764 * processor.
5765 */
5766 vm_events_fold_cpu(cpu);
5767
5768 /*
5769 * Zero the differential counters of the dead processor
5770 * so that the vm statistics are consistent.
5771 *
5772 * This is only okay since the processor is dead and cannot
5773 * race with what we are doing.
5774 */
5775 cpu_vm_stats_fold(cpu);
5776
5777 for_each_populated_zone(zone)
5778 zone_pcp_update(zone, 0);
5779
5780 return 0;
5781}
5782
5783static int page_alloc_cpu_online(unsigned int cpu)
5784{
5785 struct zone *zone;
5786
5787 for_each_populated_zone(zone)
5788 zone_pcp_update(zone, 1);
5789 return 0;
5790}
5791
5792void __init page_alloc_init_cpuhp(void)
5793{
5794 int ret;
5795
5796 ret = cpuhp_setup_state_nocalls(CPUHP_PAGE_ALLOC,
5797 "mm/page_alloc:pcp",
5798 page_alloc_cpu_online,
5799 page_alloc_cpu_dead);
5800 WARN_ON(ret < 0);
5801}
5802
5803/*
5804 * calculate_totalreserve_pages - called when sysctl_lowmem_reserve_ratio
5805 * or min_free_kbytes changes.
5806 */
5807static void calculate_totalreserve_pages(void)
5808{
5809 struct pglist_data *pgdat;
5810 unsigned long reserve_pages = 0;
5811 enum zone_type i, j;
5812
5813 for_each_online_pgdat(pgdat) {
5814
5815 pgdat->totalreserve_pages = 0;
5816
5817 for (i = 0; i < MAX_NR_ZONES; i++) {
5818 struct zone *zone = pgdat->node_zones + i;
5819 long max = 0;
5820 unsigned long managed_pages = zone_managed_pages(zone);
5821
5822 /* Find valid and maximum lowmem_reserve in the zone */
5823 for (j = i; j < MAX_NR_ZONES; j++) {
5824 if (zone->lowmem_reserve[j] > max)
5825 max = zone->lowmem_reserve[j];
5826 }
5827
5828 /* we treat the high watermark as reserved pages. */
5829 max += high_wmark_pages(zone);
5830
5831 if (max > managed_pages)
5832 max = managed_pages;
5833
5834 pgdat->totalreserve_pages += max;
5835
5836 reserve_pages += max;
5837 }
5838 }
5839 totalreserve_pages = reserve_pages;
5840}
5841
5842/*
5843 * setup_per_zone_lowmem_reserve - called whenever
5844 * sysctl_lowmem_reserve_ratio changes. Ensures that each zone
5845 * has a correct pages reserved value, so an adequate number of
5846 * pages are left in the zone after a successful __alloc_pages().
5847 */
5848static void setup_per_zone_lowmem_reserve(void)
5849{
5850 struct pglist_data *pgdat;
5851 enum zone_type i, j;
5852
5853 for_each_online_pgdat(pgdat) {
5854 for (i = 0; i < MAX_NR_ZONES - 1; i++) {
5855 struct zone *zone = &pgdat->node_zones[i];
5856 int ratio = sysctl_lowmem_reserve_ratio[i];
5857 bool clear = !ratio || !zone_managed_pages(zone);
5858 unsigned long managed_pages = 0;
5859
5860 for (j = i + 1; j < MAX_NR_ZONES; j++) {
5861 struct zone *upper_zone = &pgdat->node_zones[j];
5862
5863 managed_pages += zone_managed_pages(upper_zone);
5864
5865 if (clear)
5866 zone->lowmem_reserve[j] = 0;
5867 else
5868 zone->lowmem_reserve[j] = managed_pages / ratio;
5869 }
5870 }
5871 }
5872
5873 /* update totalreserve_pages */
5874 calculate_totalreserve_pages();
5875}
5876
5877static void __setup_per_zone_wmarks(void)
5878{
5879 unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
5880 unsigned long lowmem_pages = 0;
5881 struct zone *zone;
5882 unsigned long flags;
5883
5884 /* Calculate total number of !ZONE_HIGHMEM and !ZONE_MOVABLE pages */
5885 for_each_zone(zone) {
5886 if (!is_highmem(zone) && zone_idx(zone) != ZONE_MOVABLE)
5887 lowmem_pages += zone_managed_pages(zone);
5888 }
5889
5890 for_each_zone(zone) {
5891 u64 tmp;
5892
5893 spin_lock_irqsave(&zone->lock, flags);
5894 tmp = (u64)pages_min * zone_managed_pages(zone);
5895 tmp = div64_ul(tmp, lowmem_pages);
5896 if (is_highmem(zone) || zone_idx(zone) == ZONE_MOVABLE) {
5897 /*
5898 * __GFP_HIGH and PF_MEMALLOC allocations usually don't
5899 * need highmem and movable zones pages, so cap pages_min
5900 * to a small value here.
5901 *
5902 * The WMARK_HIGH-WMARK_LOW and (WMARK_LOW-WMARK_MIN)
5903 * deltas control async page reclaim, and so should
5904 * not be capped for highmem and movable zones.
5905 */
5906 unsigned long min_pages;
5907
5908 min_pages = zone_managed_pages(zone) / 1024;
5909 min_pages = clamp(min_pages, SWAP_CLUSTER_MAX, 128UL);
5910 zone->_watermark[WMARK_MIN] = min_pages;
5911 } else {
5912 /*
5913 * If it's a lowmem zone, reserve a number of pages
5914 * proportionate to the zone's size.
5915 */
5916 zone->_watermark[WMARK_MIN] = tmp;
5917 }
5918
5919 /*
5920 * Set the kswapd watermarks distance according to the
5921 * scale factor in proportion to available memory, but
5922 * ensure a minimum size on small systems.
5923 */
5924 tmp = max_t(u64, tmp >> 2,
5925 mult_frac(zone_managed_pages(zone),
5926 watermark_scale_factor, 10000));
5927
5928 zone->watermark_boost = 0;
5929 zone->_watermark[WMARK_LOW] = min_wmark_pages(zone) + tmp;
5930 zone->_watermark[WMARK_HIGH] = low_wmark_pages(zone) + tmp;
5931 zone->_watermark[WMARK_PROMO] = high_wmark_pages(zone) + tmp;
5932
5933 spin_unlock_irqrestore(&zone->lock, flags);
5934 }
5935
5936 /* update totalreserve_pages */
5937 calculate_totalreserve_pages();
5938}
5939
5940/**
5941 * setup_per_zone_wmarks - called when min_free_kbytes changes
5942 * or when memory is hot-{added|removed}
5943 *
5944 * Ensures that the watermark[min,low,high] values for each zone are set
5945 * correctly with respect to min_free_kbytes.
5946 */
5947void setup_per_zone_wmarks(void)
5948{
5949 struct zone *zone;
5950 static DEFINE_SPINLOCK(lock);
5951
5952 spin_lock(&lock);
5953 __setup_per_zone_wmarks();
5954 spin_unlock(&lock);
5955
5956 /*
5957 * The watermark size have changed so update the pcpu batch
5958 * and high limits or the limits may be inappropriate.
5959 */
5960 for_each_zone(zone)
5961 zone_pcp_update(zone, 0);
5962}
5963
5964/*
5965 * Initialise min_free_kbytes.
5966 *
5967 * For small machines we want it small (128k min). For large machines
5968 * we want it large (256MB max). But it is not linear, because network
5969 * bandwidth does not increase linearly with machine size. We use
5970 *
5971 * min_free_kbytes = 4 * sqrt(lowmem_kbytes), for better accuracy:
5972 * min_free_kbytes = sqrt(lowmem_kbytes * 16)
5973 *
5974 * which yields
5975 *
5976 * 16MB: 512k
5977 * 32MB: 724k
5978 * 64MB: 1024k
5979 * 128MB: 1448k
5980 * 256MB: 2048k
5981 * 512MB: 2896k
5982 * 1024MB: 4096k
5983 * 2048MB: 5792k
5984 * 4096MB: 8192k
5985 * 8192MB: 11584k
5986 * 16384MB: 16384k
5987 */
5988void calculate_min_free_kbytes(void)
5989{
5990 unsigned long lowmem_kbytes;
5991 int new_min_free_kbytes;
5992
5993 lowmem_kbytes = nr_free_buffer_pages() * (PAGE_SIZE >> 10);
5994 new_min_free_kbytes = int_sqrt(lowmem_kbytes * 16);
5995
5996 if (new_min_free_kbytes > user_min_free_kbytes)
5997 min_free_kbytes = clamp(new_min_free_kbytes, 128, 262144);
5998 else
5999 pr_warn("min_free_kbytes is not updated to %d because user defined value %d is preferred\n",
6000 new_min_free_kbytes, user_min_free_kbytes);
6001
6002}
6003
6004int __meminit init_per_zone_wmark_min(void)
6005{
6006 calculate_min_free_kbytes();
6007 setup_per_zone_wmarks();
6008 refresh_zone_stat_thresholds();
6009 setup_per_zone_lowmem_reserve();
6010
6011#ifdef CONFIG_NUMA
6012 setup_min_unmapped_ratio();
6013 setup_min_slab_ratio();
6014#endif
6015
6016 khugepaged_min_free_kbytes_update();
6017
6018 return 0;
6019}
6020postcore_initcall(init_per_zone_wmark_min)
6021
6022/*
6023 * min_free_kbytes_sysctl_handler - just a wrapper around proc_dointvec() so
6024 * that we can call two helper functions whenever min_free_kbytes
6025 * changes.
6026 */
6027static int min_free_kbytes_sysctl_handler(const struct ctl_table *table, int write,
6028 void *buffer, size_t *length, loff_t *ppos)
6029{
6030 int rc;
6031
6032 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6033 if (rc)
6034 return rc;
6035
6036 if (write) {
6037 user_min_free_kbytes = min_free_kbytes;
6038 setup_per_zone_wmarks();
6039 }
6040 return 0;
6041}
6042
6043static int watermark_scale_factor_sysctl_handler(const struct ctl_table *table, int write,
6044 void *buffer, size_t *length, loff_t *ppos)
6045{
6046 int rc;
6047
6048 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6049 if (rc)
6050 return rc;
6051
6052 if (write)
6053 setup_per_zone_wmarks();
6054
6055 return 0;
6056}
6057
6058#ifdef CONFIG_NUMA
6059static void setup_min_unmapped_ratio(void)
6060{
6061 pg_data_t *pgdat;
6062 struct zone *zone;
6063
6064 for_each_online_pgdat(pgdat)
6065 pgdat->min_unmapped_pages = 0;
6066
6067 for_each_zone(zone)
6068 zone->zone_pgdat->min_unmapped_pages += (zone_managed_pages(zone) *
6069 sysctl_min_unmapped_ratio) / 100;
6070}
6071
6072
6073static int sysctl_min_unmapped_ratio_sysctl_handler(const struct ctl_table *table, int write,
6074 void *buffer, size_t *length, loff_t *ppos)
6075{
6076 int rc;
6077
6078 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6079 if (rc)
6080 return rc;
6081
6082 setup_min_unmapped_ratio();
6083
6084 return 0;
6085}
6086
6087static void setup_min_slab_ratio(void)
6088{
6089 pg_data_t *pgdat;
6090 struct zone *zone;
6091
6092 for_each_online_pgdat(pgdat)
6093 pgdat->min_slab_pages = 0;
6094
6095 for_each_zone(zone)
6096 zone->zone_pgdat->min_slab_pages += (zone_managed_pages(zone) *
6097 sysctl_min_slab_ratio) / 100;
6098}
6099
6100static int sysctl_min_slab_ratio_sysctl_handler(const struct ctl_table *table, int write,
6101 void *buffer, size_t *length, loff_t *ppos)
6102{
6103 int rc;
6104
6105 rc = proc_dointvec_minmax(table, write, buffer, length, ppos);
6106 if (rc)
6107 return rc;
6108
6109 setup_min_slab_ratio();
6110
6111 return 0;
6112}
6113#endif
6114
6115/*
6116 * lowmem_reserve_ratio_sysctl_handler - just a wrapper around
6117 * proc_dointvec() so that we can call setup_per_zone_lowmem_reserve()
6118 * whenever sysctl_lowmem_reserve_ratio changes.
6119 *
6120 * The reserve ratio obviously has absolutely no relation with the
6121 * minimum watermarks. The lowmem reserve ratio can only make sense
6122 * if in function of the boot time zone sizes.
6123 */
6124static int lowmem_reserve_ratio_sysctl_handler(const struct ctl_table *table,
6125 int write, void *buffer, size_t *length, loff_t *ppos)
6126{
6127 int i;
6128
6129 proc_dointvec_minmax(table, write, buffer, length, ppos);
6130
6131 for (i = 0; i < MAX_NR_ZONES; i++) {
6132 if (sysctl_lowmem_reserve_ratio[i] < 1)
6133 sysctl_lowmem_reserve_ratio[i] = 0;
6134 }
6135
6136 setup_per_zone_lowmem_reserve();
6137 return 0;
6138}
6139
6140/*
6141 * percpu_pagelist_high_fraction - changes the pcp->high for each zone on each
6142 * cpu. It is the fraction of total pages in each zone that a hot per cpu
6143 * pagelist can have before it gets flushed back to buddy allocator.
6144 */
6145static int percpu_pagelist_high_fraction_sysctl_handler(const struct ctl_table *table,
6146 int write, void *buffer, size_t *length, loff_t *ppos)
6147{
6148 struct zone *zone;
6149 int old_percpu_pagelist_high_fraction;
6150 int ret;
6151
6152 mutex_lock(&pcp_batch_high_lock);
6153 old_percpu_pagelist_high_fraction = percpu_pagelist_high_fraction;
6154
6155 ret = proc_dointvec_minmax(table, write, buffer, length, ppos);
6156 if (!write || ret < 0)
6157 goto out;
6158
6159 /* Sanity checking to avoid pcp imbalance */
6160 if (percpu_pagelist_high_fraction &&
6161 percpu_pagelist_high_fraction < MIN_PERCPU_PAGELIST_HIGH_FRACTION) {
6162 percpu_pagelist_high_fraction = old_percpu_pagelist_high_fraction;
6163 ret = -EINVAL;
6164 goto out;
6165 }
6166
6167 /* No change? */
6168 if (percpu_pagelist_high_fraction == old_percpu_pagelist_high_fraction)
6169 goto out;
6170
6171 for_each_populated_zone(zone)
6172 zone_set_pageset_high_and_batch(zone, 0);
6173out:
6174 mutex_unlock(&pcp_batch_high_lock);
6175 return ret;
6176}
6177
6178static struct ctl_table page_alloc_sysctl_table[] = {
6179 {
6180 .procname = "min_free_kbytes",
6181 .data = &min_free_kbytes,
6182 .maxlen = sizeof(min_free_kbytes),
6183 .mode = 0644,
6184 .proc_handler = min_free_kbytes_sysctl_handler,
6185 .extra1 = SYSCTL_ZERO,
6186 },
6187 {
6188 .procname = "watermark_boost_factor",
6189 .data = &watermark_boost_factor,
6190 .maxlen = sizeof(watermark_boost_factor),
6191 .mode = 0644,
6192 .proc_handler = proc_dointvec_minmax,
6193 .extra1 = SYSCTL_ZERO,
6194 },
6195 {
6196 .procname = "watermark_scale_factor",
6197 .data = &watermark_scale_factor,
6198 .maxlen = sizeof(watermark_scale_factor),
6199 .mode = 0644,
6200 .proc_handler = watermark_scale_factor_sysctl_handler,
6201 .extra1 = SYSCTL_ONE,
6202 .extra2 = SYSCTL_THREE_THOUSAND,
6203 },
6204 {
6205 .procname = "percpu_pagelist_high_fraction",
6206 .data = &percpu_pagelist_high_fraction,
6207 .maxlen = sizeof(percpu_pagelist_high_fraction),
6208 .mode = 0644,
6209 .proc_handler = percpu_pagelist_high_fraction_sysctl_handler,
6210 .extra1 = SYSCTL_ZERO,
6211 },
6212 {
6213 .procname = "lowmem_reserve_ratio",
6214 .data = &sysctl_lowmem_reserve_ratio,
6215 .maxlen = sizeof(sysctl_lowmem_reserve_ratio),
6216 .mode = 0644,
6217 .proc_handler = lowmem_reserve_ratio_sysctl_handler,
6218 },
6219#ifdef CONFIG_NUMA
6220 {
6221 .procname = "numa_zonelist_order",
6222 .data = &numa_zonelist_order,
6223 .maxlen = NUMA_ZONELIST_ORDER_LEN,
6224 .mode = 0644,
6225 .proc_handler = numa_zonelist_order_handler,
6226 },
6227 {
6228 .procname = "min_unmapped_ratio",
6229 .data = &sysctl_min_unmapped_ratio,
6230 .maxlen = sizeof(sysctl_min_unmapped_ratio),
6231 .mode = 0644,
6232 .proc_handler = sysctl_min_unmapped_ratio_sysctl_handler,
6233 .extra1 = SYSCTL_ZERO,
6234 .extra2 = SYSCTL_ONE_HUNDRED,
6235 },
6236 {
6237 .procname = "min_slab_ratio",
6238 .data = &sysctl_min_slab_ratio,
6239 .maxlen = sizeof(sysctl_min_slab_ratio),
6240 .mode = 0644,
6241 .proc_handler = sysctl_min_slab_ratio_sysctl_handler,
6242 .extra1 = SYSCTL_ZERO,
6243 .extra2 = SYSCTL_ONE_HUNDRED,
6244 },
6245#endif
6246};
6247
6248void __init page_alloc_sysctl_init(void)
6249{
6250 register_sysctl_init("vm", page_alloc_sysctl_table);
6251}
6252
6253#ifdef CONFIG_CONTIG_ALLOC
6254/* Usage: See admin-guide/dynamic-debug-howto.rst */
6255static void alloc_contig_dump_pages(struct list_head *page_list)
6256{
6257 DEFINE_DYNAMIC_DEBUG_METADATA(descriptor, "migrate failure");
6258
6259 if (DYNAMIC_DEBUG_BRANCH(descriptor)) {
6260 struct page *page;
6261
6262 dump_stack();
6263 list_for_each_entry(page, page_list, lru)
6264 dump_page(page, "migration failure");
6265 }
6266}
6267
6268/*
6269 * [start, end) must belong to a single zone.
6270 * @migratetype: using migratetype to filter the type of migration in
6271 * trace_mm_alloc_contig_migrate_range_info.
6272 */
6273int __alloc_contig_migrate_range(struct compact_control *cc,
6274 unsigned long start, unsigned long end,
6275 int migratetype)
6276{
6277 /* This function is based on compact_zone() from compaction.c. */
6278 unsigned int nr_reclaimed;
6279 unsigned long pfn = start;
6280 unsigned int tries = 0;
6281 int ret = 0;
6282 struct migration_target_control mtc = {
6283 .nid = zone_to_nid(cc->zone),
6284 .gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
6285 .reason = MR_CONTIG_RANGE,
6286 };
6287 struct page *page;
6288 unsigned long total_mapped = 0;
6289 unsigned long total_migrated = 0;
6290 unsigned long total_reclaimed = 0;
6291
6292 lru_cache_disable();
6293
6294 while (pfn < end || !list_empty(&cc->migratepages)) {
6295 if (fatal_signal_pending(current)) {
6296 ret = -EINTR;
6297 break;
6298 }
6299
6300 if (list_empty(&cc->migratepages)) {
6301 cc->nr_migratepages = 0;
6302 ret = isolate_migratepages_range(cc, pfn, end);
6303 if (ret && ret != -EAGAIN)
6304 break;
6305 pfn = cc->migrate_pfn;
6306 tries = 0;
6307 } else if (++tries == 5) {
6308 ret = -EBUSY;
6309 break;
6310 }
6311
6312 nr_reclaimed = reclaim_clean_pages_from_list(cc->zone,
6313 &cc->migratepages);
6314 cc->nr_migratepages -= nr_reclaimed;
6315
6316 if (trace_mm_alloc_contig_migrate_range_info_enabled()) {
6317 total_reclaimed += nr_reclaimed;
6318 list_for_each_entry(page, &cc->migratepages, lru) {
6319 struct folio *folio = page_folio(page);
6320
6321 total_mapped += folio_mapped(folio) *
6322 folio_nr_pages(folio);
6323 }
6324 }
6325
6326 ret = migrate_pages(&cc->migratepages, alloc_migration_target,
6327 NULL, (unsigned long)&mtc, cc->mode, MR_CONTIG_RANGE, NULL);
6328
6329 if (trace_mm_alloc_contig_migrate_range_info_enabled() && !ret)
6330 total_migrated += cc->nr_migratepages;
6331
6332 /*
6333 * On -ENOMEM, migrate_pages() bails out right away. It is pointless
6334 * to retry again over this error, so do the same here.
6335 */
6336 if (ret == -ENOMEM)
6337 break;
6338 }
6339
6340 lru_cache_enable();
6341 if (ret < 0) {
6342 if (!(cc->gfp_mask & __GFP_NOWARN) && ret == -EBUSY)
6343 alloc_contig_dump_pages(&cc->migratepages);
6344 putback_movable_pages(&cc->migratepages);
6345 }
6346
6347 trace_mm_alloc_contig_migrate_range_info(start, end, migratetype,
6348 total_migrated,
6349 total_reclaimed,
6350 total_mapped);
6351 return (ret < 0) ? ret : 0;
6352}
6353
6354static void split_free_pages(struct list_head *list)
6355{
6356 int order;
6357
6358 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6359 struct page *page, *next;
6360 int nr_pages = 1 << order;
6361
6362 list_for_each_entry_safe(page, next, &list[order], lru) {
6363 int i;
6364
6365 post_alloc_hook(page, order, __GFP_MOVABLE);
6366 if (!order)
6367 continue;
6368
6369 split_page(page, order);
6370
6371 /* Add all subpages to the order-0 head, in sequence. */
6372 list_del(&page->lru);
6373 for (i = 0; i < nr_pages; i++)
6374 list_add_tail(&page[i].lru, &list[0]);
6375 }
6376 }
6377}
6378
6379/**
6380 * alloc_contig_range() -- tries to allocate given range of pages
6381 * @start: start PFN to allocate
6382 * @end: one-past-the-last PFN to allocate
6383 * @migratetype: migratetype of the underlying pageblocks (either
6384 * #MIGRATE_MOVABLE or #MIGRATE_CMA). All pageblocks
6385 * in range must have the same migratetype and it must
6386 * be either of the two.
6387 * @gfp_mask: GFP mask to use during compaction
6388 *
6389 * The PFN range does not have to be pageblock aligned. The PFN range must
6390 * belong to a single zone.
6391 *
6392 * The first thing this routine does is attempt to MIGRATE_ISOLATE all
6393 * pageblocks in the range. Once isolated, the pageblocks should not
6394 * be modified by others.
6395 *
6396 * Return: zero on success or negative error code. On success all
6397 * pages which PFN is in [start, end) are allocated for the caller and
6398 * need to be freed with free_contig_range().
6399 */
6400int alloc_contig_range_noprof(unsigned long start, unsigned long end,
6401 unsigned migratetype, gfp_t gfp_mask)
6402{
6403 unsigned long outer_start, outer_end;
6404 int ret = 0;
6405
6406 struct compact_control cc = {
6407 .nr_migratepages = 0,
6408 .order = -1,
6409 .zone = page_zone(pfn_to_page(start)),
6410 .mode = MIGRATE_SYNC,
6411 .ignore_skip_hint = true,
6412 .no_set_skip_hint = true,
6413 .gfp_mask = current_gfp_context(gfp_mask),
6414 .alloc_contig = true,
6415 };
6416 INIT_LIST_HEAD(&cc.migratepages);
6417
6418 /*
6419 * What we do here is we mark all pageblocks in range as
6420 * MIGRATE_ISOLATE. Because pageblock and max order pages may
6421 * have different sizes, and due to the way page allocator
6422 * work, start_isolate_page_range() has special handlings for this.
6423 *
6424 * Once the pageblocks are marked as MIGRATE_ISOLATE, we
6425 * migrate the pages from an unaligned range (ie. pages that
6426 * we are interested in). This will put all the pages in
6427 * range back to page allocator as MIGRATE_ISOLATE.
6428 *
6429 * When this is done, we take the pages in range from page
6430 * allocator removing them from the buddy system. This way
6431 * page allocator will never consider using them.
6432 *
6433 * This lets us mark the pageblocks back as
6434 * MIGRATE_CMA/MIGRATE_MOVABLE so that free pages in the
6435 * aligned range but not in the unaligned, original range are
6436 * put back to page allocator so that buddy can use them.
6437 */
6438
6439 ret = start_isolate_page_range(start, end, migratetype, 0, gfp_mask);
6440 if (ret)
6441 goto done;
6442
6443 drain_all_pages(cc.zone);
6444
6445 /*
6446 * In case of -EBUSY, we'd like to know which page causes problem.
6447 * So, just fall through. test_pages_isolated() has a tracepoint
6448 * which will report the busy page.
6449 *
6450 * It is possible that busy pages could become available before
6451 * the call to test_pages_isolated, and the range will actually be
6452 * allocated. So, if we fall through be sure to clear ret so that
6453 * -EBUSY is not accidentally used or returned to caller.
6454 */
6455 ret = __alloc_contig_migrate_range(&cc, start, end, migratetype);
6456 if (ret && ret != -EBUSY)
6457 goto done;
6458 ret = 0;
6459
6460 /*
6461 * Pages from [start, end) are within a pageblock_nr_pages
6462 * aligned blocks that are marked as MIGRATE_ISOLATE. What's
6463 * more, all pages in [start, end) are free in page allocator.
6464 * What we are going to do is to allocate all pages from
6465 * [start, end) (that is remove them from page allocator).
6466 *
6467 * The only problem is that pages at the beginning and at the
6468 * end of interesting range may be not aligned with pages that
6469 * page allocator holds, ie. they can be part of higher order
6470 * pages. Because of this, we reserve the bigger range and
6471 * once this is done free the pages we are not interested in.
6472 *
6473 * We don't have to hold zone->lock here because the pages are
6474 * isolated thus they won't get removed from buddy.
6475 */
6476 outer_start = find_large_buddy(start);
6477
6478 /* Make sure the range is really isolated. */
6479 if (test_pages_isolated(outer_start, end, 0)) {
6480 ret = -EBUSY;
6481 goto done;
6482 }
6483
6484 /* Grab isolated pages from freelists. */
6485 outer_end = isolate_freepages_range(&cc, outer_start, end);
6486 if (!outer_end) {
6487 ret = -EBUSY;
6488 goto done;
6489 }
6490
6491 if (!(gfp_mask & __GFP_COMP)) {
6492 split_free_pages(cc.freepages);
6493
6494 /* Free head and tail (if any) */
6495 if (start != outer_start)
6496 free_contig_range(outer_start, start - outer_start);
6497 if (end != outer_end)
6498 free_contig_range(end, outer_end - end);
6499 } else if (start == outer_start && end == outer_end && is_power_of_2(end - start)) {
6500 struct page *head = pfn_to_page(start);
6501 int order = ilog2(end - start);
6502
6503 check_new_pages(head, order);
6504 prep_new_page(head, order, gfp_mask, 0);
6505 } else {
6506 ret = -EINVAL;
6507 WARN(true, "PFN range: requested [%lu, %lu), allocated [%lu, %lu)\n",
6508 start, end, outer_start, outer_end);
6509 }
6510done:
6511 undo_isolate_page_range(start, end, migratetype);
6512 return ret;
6513}
6514EXPORT_SYMBOL(alloc_contig_range_noprof);
6515
6516static int __alloc_contig_pages(unsigned long start_pfn,
6517 unsigned long nr_pages, gfp_t gfp_mask)
6518{
6519 unsigned long end_pfn = start_pfn + nr_pages;
6520
6521 return alloc_contig_range_noprof(start_pfn, end_pfn, MIGRATE_MOVABLE,
6522 gfp_mask);
6523}
6524
6525static bool pfn_range_valid_contig(struct zone *z, unsigned long start_pfn,
6526 unsigned long nr_pages)
6527{
6528 unsigned long i, end_pfn = start_pfn + nr_pages;
6529 struct page *page;
6530
6531 for (i = start_pfn; i < end_pfn; i++) {
6532 page = pfn_to_online_page(i);
6533 if (!page)
6534 return false;
6535
6536 if (page_zone(page) != z)
6537 return false;
6538
6539 if (PageReserved(page))
6540 return false;
6541
6542 if (PageHuge(page))
6543 return false;
6544 }
6545 return true;
6546}
6547
6548static bool zone_spans_last_pfn(const struct zone *zone,
6549 unsigned long start_pfn, unsigned long nr_pages)
6550{
6551 unsigned long last_pfn = start_pfn + nr_pages - 1;
6552
6553 return zone_spans_pfn(zone, last_pfn);
6554}
6555
6556/**
6557 * alloc_contig_pages() -- tries to find and allocate contiguous range of pages
6558 * @nr_pages: Number of contiguous pages to allocate
6559 * @gfp_mask: GFP mask to limit search and used during compaction
6560 * @nid: Target node
6561 * @nodemask: Mask for other possible nodes
6562 *
6563 * This routine is a wrapper around alloc_contig_range(). It scans over zones
6564 * on an applicable zonelist to find a contiguous pfn range which can then be
6565 * tried for allocation with alloc_contig_range(). This routine is intended
6566 * for allocation requests which can not be fulfilled with the buddy allocator.
6567 *
6568 * The allocated memory is always aligned to a page boundary. If nr_pages is a
6569 * power of two, then allocated range is also guaranteed to be aligned to same
6570 * nr_pages (e.g. 1GB request would be aligned to 1GB).
6571 *
6572 * Allocated pages can be freed with free_contig_range() or by manually calling
6573 * __free_page() on each allocated page.
6574 *
6575 * Return: pointer to contiguous pages on success, or NULL if not successful.
6576 */
6577struct page *alloc_contig_pages_noprof(unsigned long nr_pages, gfp_t gfp_mask,
6578 int nid, nodemask_t *nodemask)
6579{
6580 unsigned long ret, pfn, flags;
6581 struct zonelist *zonelist;
6582 struct zone *zone;
6583 struct zoneref *z;
6584
6585 zonelist = node_zonelist(nid, gfp_mask);
6586 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6587 gfp_zone(gfp_mask), nodemask) {
6588 spin_lock_irqsave(&zone->lock, flags);
6589
6590 pfn = ALIGN(zone->zone_start_pfn, nr_pages);
6591 while (zone_spans_last_pfn(zone, pfn, nr_pages)) {
6592 if (pfn_range_valid_contig(zone, pfn, nr_pages)) {
6593 /*
6594 * We release the zone lock here because
6595 * alloc_contig_range() will also lock the zone
6596 * at some point. If there's an allocation
6597 * spinning on this lock, it may win the race
6598 * and cause alloc_contig_range() to fail...
6599 */
6600 spin_unlock_irqrestore(&zone->lock, flags);
6601 ret = __alloc_contig_pages(pfn, nr_pages,
6602 gfp_mask);
6603 if (!ret)
6604 return pfn_to_page(pfn);
6605 spin_lock_irqsave(&zone->lock, flags);
6606 }
6607 pfn += nr_pages;
6608 }
6609 spin_unlock_irqrestore(&zone->lock, flags);
6610 }
6611 return NULL;
6612}
6613#endif /* CONFIG_CONTIG_ALLOC */
6614
6615void free_contig_range(unsigned long pfn, unsigned long nr_pages)
6616{
6617 unsigned long count = 0;
6618 struct folio *folio = pfn_folio(pfn);
6619
6620 if (folio_test_large(folio)) {
6621 int expected = folio_nr_pages(folio);
6622
6623 if (nr_pages == expected)
6624 folio_put(folio);
6625 else
6626 WARN(true, "PFN %lu: nr_pages %lu != expected %d\n",
6627 pfn, nr_pages, expected);
6628 return;
6629 }
6630
6631 for (; nr_pages--; pfn++) {
6632 struct page *page = pfn_to_page(pfn);
6633
6634 count += page_count(page) != 1;
6635 __free_page(page);
6636 }
6637 WARN(count != 0, "%lu pages are still in use!\n", count);
6638}
6639EXPORT_SYMBOL(free_contig_range);
6640
6641/*
6642 * Effectively disable pcplists for the zone by setting the high limit to 0
6643 * and draining all cpus. A concurrent page freeing on another CPU that's about
6644 * to put the page on pcplist will either finish before the drain and the page
6645 * will be drained, or observe the new high limit and skip the pcplist.
6646 *
6647 * Must be paired with a call to zone_pcp_enable().
6648 */
6649void zone_pcp_disable(struct zone *zone)
6650{
6651 mutex_lock(&pcp_batch_high_lock);
6652 __zone_set_pageset_high_and_batch(zone, 0, 0, 1);
6653 __drain_all_pages(zone, true);
6654}
6655
6656void zone_pcp_enable(struct zone *zone)
6657{
6658 __zone_set_pageset_high_and_batch(zone, zone->pageset_high_min,
6659 zone->pageset_high_max, zone->pageset_batch);
6660 mutex_unlock(&pcp_batch_high_lock);
6661}
6662
6663void zone_pcp_reset(struct zone *zone)
6664{
6665 int cpu;
6666 struct per_cpu_zonestat *pzstats;
6667
6668 if (zone->per_cpu_pageset != &boot_pageset) {
6669 for_each_online_cpu(cpu) {
6670 pzstats = per_cpu_ptr(zone->per_cpu_zonestats, cpu);
6671 drain_zonestat(zone, pzstats);
6672 }
6673 free_percpu(zone->per_cpu_pageset);
6674 zone->per_cpu_pageset = &boot_pageset;
6675 if (zone->per_cpu_zonestats != &boot_zonestats) {
6676 free_percpu(zone->per_cpu_zonestats);
6677 zone->per_cpu_zonestats = &boot_zonestats;
6678 }
6679 }
6680}
6681
6682#ifdef CONFIG_MEMORY_HOTREMOVE
6683/*
6684 * All pages in the range must be in a single zone, must not contain holes,
6685 * must span full sections, and must be isolated before calling this function.
6686 *
6687 * Returns the number of managed (non-PageOffline()) pages in the range: the
6688 * number of pages for which memory offlining code must adjust managed page
6689 * counters using adjust_managed_page_count().
6690 */
6691unsigned long __offline_isolated_pages(unsigned long start_pfn,
6692 unsigned long end_pfn)
6693{
6694 unsigned long already_offline = 0, flags;
6695 unsigned long pfn = start_pfn;
6696 struct page *page;
6697 struct zone *zone;
6698 unsigned int order;
6699
6700 offline_mem_sections(pfn, end_pfn);
6701 zone = page_zone(pfn_to_page(pfn));
6702 spin_lock_irqsave(&zone->lock, flags);
6703 while (pfn < end_pfn) {
6704 page = pfn_to_page(pfn);
6705 /*
6706 * The HWPoisoned page may be not in buddy system, and
6707 * page_count() is not 0.
6708 */
6709 if (unlikely(!PageBuddy(page) && PageHWPoison(page))) {
6710 pfn++;
6711 continue;
6712 }
6713 /*
6714 * At this point all remaining PageOffline() pages have a
6715 * reference count of 0 and can simply be skipped.
6716 */
6717 if (PageOffline(page)) {
6718 BUG_ON(page_count(page));
6719 BUG_ON(PageBuddy(page));
6720 already_offline++;
6721 pfn++;
6722 continue;
6723 }
6724
6725 BUG_ON(page_count(page));
6726 BUG_ON(!PageBuddy(page));
6727 VM_WARN_ON(get_pageblock_migratetype(page) != MIGRATE_ISOLATE);
6728 order = buddy_order(page);
6729 del_page_from_free_list(page, zone, order, MIGRATE_ISOLATE);
6730 pfn += (1 << order);
6731 }
6732 spin_unlock_irqrestore(&zone->lock, flags);
6733
6734 return end_pfn - start_pfn - already_offline;
6735}
6736#endif
6737
6738/*
6739 * This function returns a stable result only if called under zone lock.
6740 */
6741bool is_free_buddy_page(const struct page *page)
6742{
6743 unsigned long pfn = page_to_pfn(page);
6744 unsigned int order;
6745
6746 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6747 const struct page *head = page - (pfn & ((1 << order) - 1));
6748
6749 if (PageBuddy(head) &&
6750 buddy_order_unsafe(head) >= order)
6751 break;
6752 }
6753
6754 return order <= MAX_PAGE_ORDER;
6755}
6756EXPORT_SYMBOL(is_free_buddy_page);
6757
6758#ifdef CONFIG_MEMORY_FAILURE
6759static inline void add_to_free_list(struct page *page, struct zone *zone,
6760 unsigned int order, int migratetype,
6761 bool tail)
6762{
6763 __add_to_free_list(page, zone, order, migratetype, tail);
6764 account_freepages(zone, 1 << order, migratetype);
6765}
6766
6767/*
6768 * Break down a higher-order page in sub-pages, and keep our target out of
6769 * buddy allocator.
6770 */
6771static void break_down_buddy_pages(struct zone *zone, struct page *page,
6772 struct page *target, int low, int high,
6773 int migratetype)
6774{
6775 unsigned long size = 1 << high;
6776 struct page *current_buddy;
6777
6778 while (high > low) {
6779 high--;
6780 size >>= 1;
6781
6782 if (target >= &page[size]) {
6783 current_buddy = page;
6784 page = page + size;
6785 } else {
6786 current_buddy = page + size;
6787 }
6788
6789 if (set_page_guard(zone, current_buddy, high))
6790 continue;
6791
6792 add_to_free_list(current_buddy, zone, high, migratetype, false);
6793 set_buddy_order(current_buddy, high);
6794 }
6795}
6796
6797/*
6798 * Take a page that will be marked as poisoned off the buddy allocator.
6799 */
6800bool take_page_off_buddy(struct page *page)
6801{
6802 struct zone *zone = page_zone(page);
6803 unsigned long pfn = page_to_pfn(page);
6804 unsigned long flags;
6805 unsigned int order;
6806 bool ret = false;
6807
6808 spin_lock_irqsave(&zone->lock, flags);
6809 for (order = 0; order < NR_PAGE_ORDERS; order++) {
6810 struct page *page_head = page - (pfn & ((1 << order) - 1));
6811 int page_order = buddy_order(page_head);
6812
6813 if (PageBuddy(page_head) && page_order >= order) {
6814 unsigned long pfn_head = page_to_pfn(page_head);
6815 int migratetype = get_pfnblock_migratetype(page_head,
6816 pfn_head);
6817
6818 del_page_from_free_list(page_head, zone, page_order,
6819 migratetype);
6820 break_down_buddy_pages(zone, page_head, page, 0,
6821 page_order, migratetype);
6822 SetPageHWPoisonTakenOff(page);
6823 ret = true;
6824 break;
6825 }
6826 if (page_count(page_head) > 0)
6827 break;
6828 }
6829 spin_unlock_irqrestore(&zone->lock, flags);
6830 return ret;
6831}
6832
6833/*
6834 * Cancel takeoff done by take_page_off_buddy().
6835 */
6836bool put_page_back_buddy(struct page *page)
6837{
6838 struct zone *zone = page_zone(page);
6839 unsigned long flags;
6840 bool ret = false;
6841
6842 spin_lock_irqsave(&zone->lock, flags);
6843 if (put_page_testzero(page)) {
6844 unsigned long pfn = page_to_pfn(page);
6845 int migratetype = get_pfnblock_migratetype(page, pfn);
6846
6847 ClearPageHWPoisonTakenOff(page);
6848 __free_one_page(page, pfn, zone, 0, migratetype, FPI_NONE);
6849 if (TestClearPageHWPoison(page)) {
6850 ret = true;
6851 }
6852 }
6853 spin_unlock_irqrestore(&zone->lock, flags);
6854
6855 return ret;
6856}
6857#endif
6858
6859#ifdef CONFIG_ZONE_DMA
6860bool has_managed_dma(void)
6861{
6862 struct pglist_data *pgdat;
6863
6864 for_each_online_pgdat(pgdat) {
6865 struct zone *zone = &pgdat->node_zones[ZONE_DMA];
6866
6867 if (managed_zone(zone))
6868 return true;
6869 }
6870 return false;
6871}
6872#endif /* CONFIG_ZONE_DMA */
6873
6874#ifdef CONFIG_UNACCEPTED_MEMORY
6875
6876/* Counts number of zones with unaccepted pages. */
6877static DEFINE_STATIC_KEY_FALSE(zones_with_unaccepted_pages);
6878
6879static bool lazy_accept = true;
6880
6881static int __init accept_memory_parse(char *p)
6882{
6883 if (!strcmp(p, "lazy")) {
6884 lazy_accept = true;
6885 return 0;
6886 } else if (!strcmp(p, "eager")) {
6887 lazy_accept = false;
6888 return 0;
6889 } else {
6890 return -EINVAL;
6891 }
6892}
6893early_param("accept_memory", accept_memory_parse);
6894
6895static bool page_contains_unaccepted(struct page *page, unsigned int order)
6896{
6897 phys_addr_t start = page_to_phys(page);
6898
6899 return range_contains_unaccepted_memory(start, PAGE_SIZE << order);
6900}
6901
6902static void __accept_page(struct zone *zone, unsigned long *flags,
6903 struct page *page)
6904{
6905 bool last;
6906
6907 list_del(&page->lru);
6908 last = list_empty(&zone->unaccepted_pages);
6909
6910 account_freepages(zone, -MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
6911 __mod_zone_page_state(zone, NR_UNACCEPTED, -MAX_ORDER_NR_PAGES);
6912 __ClearPageUnaccepted(page);
6913 spin_unlock_irqrestore(&zone->lock, *flags);
6914
6915 accept_memory(page_to_phys(page), PAGE_SIZE << MAX_PAGE_ORDER);
6916
6917 __free_pages_ok(page, MAX_PAGE_ORDER, FPI_TO_TAIL);
6918
6919 if (last)
6920 static_branch_dec(&zones_with_unaccepted_pages);
6921}
6922
6923void accept_page(struct page *page)
6924{
6925 struct zone *zone = page_zone(page);
6926 unsigned long flags;
6927
6928 spin_lock_irqsave(&zone->lock, flags);
6929 if (!PageUnaccepted(page)) {
6930 spin_unlock_irqrestore(&zone->lock, flags);
6931 return;
6932 }
6933
6934 /* Unlocks zone->lock */
6935 __accept_page(zone, &flags, page);
6936}
6937
6938static bool try_to_accept_memory_one(struct zone *zone)
6939{
6940 unsigned long flags;
6941 struct page *page;
6942
6943 spin_lock_irqsave(&zone->lock, flags);
6944 page = list_first_entry_or_null(&zone->unaccepted_pages,
6945 struct page, lru);
6946 if (!page) {
6947 spin_unlock_irqrestore(&zone->lock, flags);
6948 return false;
6949 }
6950
6951 /* Unlocks zone->lock */
6952 __accept_page(zone, &flags, page);
6953
6954 return true;
6955}
6956
6957static inline bool has_unaccepted_memory(void)
6958{
6959 return static_branch_unlikely(&zones_with_unaccepted_pages);
6960}
6961
6962static bool cond_accept_memory(struct zone *zone, unsigned int order)
6963{
6964 long to_accept;
6965 bool ret = false;
6966
6967 if (!has_unaccepted_memory())
6968 return false;
6969
6970 if (list_empty(&zone->unaccepted_pages))
6971 return false;
6972
6973 /* How much to accept to get to promo watermark? */
6974 to_accept = promo_wmark_pages(zone) -
6975 (zone_page_state(zone, NR_FREE_PAGES) -
6976 __zone_watermark_unusable_free(zone, order, 0) -
6977 zone_page_state(zone, NR_UNACCEPTED));
6978
6979 while (to_accept > 0) {
6980 if (!try_to_accept_memory_one(zone))
6981 break;
6982 ret = true;
6983 to_accept -= MAX_ORDER_NR_PAGES;
6984 }
6985
6986 return ret;
6987}
6988
6989static bool __free_unaccepted(struct page *page)
6990{
6991 struct zone *zone = page_zone(page);
6992 unsigned long flags;
6993 bool first = false;
6994
6995 if (!lazy_accept)
6996 return false;
6997
6998 spin_lock_irqsave(&zone->lock, flags);
6999 first = list_empty(&zone->unaccepted_pages);
7000 list_add_tail(&page->lru, &zone->unaccepted_pages);
7001 account_freepages(zone, MAX_ORDER_NR_PAGES, MIGRATE_MOVABLE);
7002 __mod_zone_page_state(zone, NR_UNACCEPTED, MAX_ORDER_NR_PAGES);
7003 __SetPageUnaccepted(page);
7004 spin_unlock_irqrestore(&zone->lock, flags);
7005
7006 if (first)
7007 static_branch_inc(&zones_with_unaccepted_pages);
7008
7009 return true;
7010}
7011
7012#else
7013
7014static bool page_contains_unaccepted(struct page *page, unsigned int order)
7015{
7016 return false;
7017}
7018
7019static bool cond_accept_memory(struct zone *zone, unsigned int order)
7020{
7021 return false;
7022}
7023
7024static bool __free_unaccepted(struct page *page)
7025{
7026 BUILD_BUG();
7027 return false;
7028}
7029
7030#endif /* CONFIG_UNACCEPTED_MEMORY */