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