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