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